https://wiki.swarma.org/api.php?action=feedcontributions&user=Aceyuan&feedformat=atom集智百科 - 复杂系统|人工智能|复杂科学|复杂网络|自组织 - 用户贡献 [zh-cn]2024-03-28T10:54:22Z用户贡献MediaWiki 1.35.0https://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32540Judea Pearl2022-06-21T05:27:43Z<p>Aceyuan:奖项与成就</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究的严谨和深度提升到一个新水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[8]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段时间做出了非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将 Pearl 的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了计算机协会 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,他获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了计算机协会 ACM 的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015 年,他获得了美国国家科学院院士称号<br />
<br />
2015 年,他获得了计算机协会 ACM Fellow 资深会员称号<br />
<br />
2018 年,他获得了美国数学学会 AMS 乌尔夫·格林南德奖(Ulf Grenander Prize)<br />
<br />
2019 年,他获得美国统计协会 ASA Fellow 资深会员称号<br />
<br />
2020 年,他获得了皇家统计学会 RSS Honorary Fellow 名誉研究员称号<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
<br />
1984 年,《启发式:计算机问题解决的智能搜索策略》 Heuristics, Addison-Wesley, 1<nowiki/>984。<br />
<br />
* 发展了传统搜索算法和游戏算法;提出了自动推导启发式的新想法。<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
* Pearl 具有里程碑意义的著作,将他的哲学、人类认知理论和所有的技术材料整合为成一个整体,提出采用了概率方法(或简称为现代方法)来研究人工智能的革命性方法。<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
* 展示了因果关系如何从一个模糊的概念发展成为一种数学理论,Pearl 提出并统一了因果关系的概率、干预、反事实和结构方法,并设计了简单的数学工具来研究因果关系和统计关联之间的关系。<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
* 为统计初学者提供了一本全面介绍因果关系领域的理想书籍,书中对传统统计方法和因果方法做了比较。在每个章节末尾都提供了练习题以帮助学生学习。<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
* 全面介绍了由 Judea Pearl 和他的同事发起的因果革命——如何打破100年以来的对因果的讨论禁忌,在坚实的科学基础上建立了对因果关系的研究。展示了人类思想的本质和人工智能的关键,并为探索现在的世界和可能存在的世界提供坚实的思想和理论工具。<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [http://causality.cs.ucla.edu/blog/ 博客]<br />
* [https://twitter.com/yudapearl 推特]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32539Judea Pearl2022-06-21T05:08:08Z<p>Aceyuan:相关链接</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究的严谨和深度提升到一个新水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[8]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段时间做出了非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将 Pearl 的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,他获得了 ACM fellow 资深会员称号<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
<br />
1984 年,《启发式:计算机问题解决的智能搜索策略》 Heuristics, Addison-Wesley, 1<nowiki/>984。<br />
<br />
* 发展了传统搜索算法和游戏算法;提出了自动推导启发式的新想法。<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
* Pearl 具有里程碑意义的著作,将他的哲学、人类认知理论和所有的技术材料整合为成一个整体,提出采用了概率方法(或简称为现代方法)来研究人工智能的革命性方法。<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
* 展示了因果关系如何从一个模糊的概念发展成为一种数学理论,Pearl 提出并统一了因果关系的概率、干预、反事实和结构方法,并设计了简单的数学工具来研究因果关系和统计关联之间的关系。<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
* 为统计初学者提供了一本全面介绍因果关系领域的理想书籍,书中对传统统计方法和因果方法做了比较。在每个章节末尾都提供了练习题以帮助学生学习。<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
* 全面介绍了由 Judea Pearl 和他的同事发起的因果革命——如何打破100年以来的对因果的讨论禁忌,在坚实的科学基础上建立了对因果关系的研究。展示了人类思想的本质和人工智能的关键,并为探索现在的世界和可能存在的世界提供坚实的思想和理论工具。<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [http://causality.cs.ucla.edu/blog/ 博客]<br />
* [https://twitter.com/yudapearl 推特]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32538Judea Pearl2022-06-21T05:00:59Z<p>Aceyuan:撤销Aceyuan(讨论)的版本32535</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究的严谨和深度提升到一个新水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[8]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段时间做出了非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将 Pearl 的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,他获得了 ACM fellow 资深会员称号<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
<br />
1984 年,《启发式:计算机问题解决的智能搜索策略》 Heuristics, Addison-Wesley, 1<nowiki/>984。<br />
<br />
* 发展了传统搜索算法和游戏算法;提出了自动推导启发式的新想法。<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
* Pearl 具有里程碑意义的著作,将他的哲学、人类认知理论和所有的技术材料整合为成一个整体,提出采用了概率方法(或简称为现代方法)来研究人工智能的革命性方法。<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
* 展示了因果关系如何从一个模糊的概念发展成为一种数学理论,Pearl 提出并统一了因果关系的概率、干预、反事实和结构方法,并设计了简单的数学工具来研究因果关系和统计关联之间的关系。<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
* 为统计初学者提供了一本全面介绍因果关系领域的理想书籍,书中对传统统计方法和因果方法做了比较。在每个章节末尾都提供了练习题以帮助学生学习。<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
* 全面介绍了由 Judea Pearl 和他的同事发起的因果革命——如何打破100年以来的对因果的讨论禁忌,在坚实的科学基础上建立了对因果关系的研究。展示了人类思想的本质和人工智能的关键,并为探索现在的世界和可能存在的世界提供坚实的思想和理论工具。<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32537Judea Pearl2022-06-21T05:00:38Z<p>Aceyuan:撤销Aceyuan(讨论)的版本32536</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32536Judea Pearl2022-06-21T04:59:21Z<p>Aceyuan:</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32535Judea Pearl2022-06-21T04:59:04Z<p>Aceyuan:</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32534Judea Pearl2022-06-21T04:57:29Z<p>Aceyuan:</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究的严谨和深度提升到一个新水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[8]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段时间做出了非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将 Pearl 的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,他获得了 ACM fellow 资深会员称号<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
<br />
1984 年,《启发式:计算机问题解决的智能搜索策略》 Heuristics, Addison-Wesley, 1<nowiki/>984。<br />
<br />
* 发展了传统搜索算法和游戏算法;提出了自动推导启发式的新想法。<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
* Pearl 具有里程碑意义的著作,将他的哲学、人类认知理论和所有的技术材料整合为成一个整体,提出采用了概率方法(或简称为现代方法)来研究人工智能的革命性方法。<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
* 展示了因果关系如何从一个模糊的概念发展成为一种数学理论,Pearl 提出并统一了因果关系的概率、干预、反事实和结构方法,并设计了简单的数学工具来研究因果关系和统计关联之间的关系。<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
* 为统计初学者提供了一本全面介绍因果关系领域的理想书籍,书中对传统统计方法和因果方法做了比较。在每个章节末尾都提供了练习题以帮助学生学习。<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
* 全面介绍了由 Judea Pearl 和他的同事发起的因果革命——如何打破100年以来的对因果的讨论禁忌,在坚实的科学基础上建立了对因果关系的研究。展示了人类思想的本质和人工智能的关键,并为探索现在的世界和可能存在的世界提供坚实的思想和理论工具。<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32533Judea Pearl2022-06-21T04:56:35Z<p>Aceyuan:文章及著作</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究的严谨和深度提升到一个新水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[8]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段时间做出了非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将 Pearl 的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,他获得了 ACM fellow 资深会员称号<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发式:计算机问题解决的智能搜索策略》 Heuristics, Addison-Wesley, 1984。<br />
<br />
* 发展了传统搜索算法和游戏算法;提出了自动推导启发式的新想法。<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
* Pearl 具有里程碑意义的著作,将他的哲学、人类认知理论和所有的技术材料整合为成一个整体,提出采用了概率方法(或简称为现代方法)来研究人工智能的革命性方法。<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
* 展示了因果关系如何从一个模糊的概念发展成为一种数学理论,Pearl 提出并统一了因果关系的概率、干预、反事实和结构方法,并设计了简单的数学工具来研究因果关系和统计关联之间的关系。<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
* 为统计初学者提供了一本全面介绍因果关系领域的理想书籍,书中对传统统计方法和因果方法做了比较。在每个章节末尾都提供了练习题以帮助学生学习。<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
* 全面介绍了由 Judea Pearl 和他的同事发起的因果革命——如何打破100年以来的对因果的讨论禁忌,在坚实的科学基础上建立了对因果关系的研究。展示了人类思想的本质和人工智能的关键,并为探索现在的世界和可能存在的世界提供坚实的思想和理论工具。<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32524Judea Pearl2022-06-20T03:47:37Z<p>Aceyuan:</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了500余篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32523Judea Pearl2022-06-20T03:46:36Z<p>Aceyuan:</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|533x533px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32522Judea Pearl2022-06-20T03:35:55Z<p>Aceyuan:</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|508x508px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32521Judea Pearl2022-06-20T03:35:32Z<p>Aceyuan:</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|542x542px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32520Judea Pearl2022-06-20T03:35:11Z<p>Aceyuan:形式初稿</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|566x566px]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
do算子(do-calculus)<br />
珀尔因果层次模型(PCH: Pearl Causal Hierarchy)<br />
|-<br />
|主要研究方向<br />
|人工智能<br />
因果推理<br />
<br />
科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
节选自[https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
即使在开发贝叶斯概率网络的理论和技术时,Pearl 也怀疑需要一种不同的方法来解决他多年来一直关注的因果关系问题。在他2000年关于因果关系的著作《因果关系:模型、论证、推理》中,他描述了他早期的兴趣如下:<blockquote>在我高中三年级的时候,我第一次看到了因果关系的黑暗世界。我的科学老师 Feuchtwanger 博士通过讨论19世纪的发现,向我们介绍了逻辑研究,发现死于天花接种的人比死于天花本身的人多。一些人利用这些信息争辩说接种是有害的,而事实上,数据证明恰恰相反,接种通过根除天花来挽救生命。<br />
<br />
Feuchtwanger 博士总结道:“逻辑的用武之地就是保护我们免受此类因果谬误的影响。” 当时的我们都为逻辑的奇迹而折服,尽管 Feuchtwanger 博士从未真正向我们展示过逻辑如何保护我们免受这些谬误的影响。<br />
<br />
然而多年后我作为一名人工智能研究员意识到,事实并非如此。逻辑学和数学的任何分支都没有开发出足够的工具来管理类似天花疫苗这样涉及因果关系的问题。</blockquote>实际上,贝叶斯网络无法捕获因果信息,例如“吸烟”-->“肺癌”,它在数学上等同于网络“肺癌”-->“吸烟”。因果网络的关键特征是它能捕捉外生干预变量的潜在效果。在因果网络X-->Y中,人为设定Y的值对Y实施干预,不应该改变人对X的先验认知,即,对Y的干预切断了从X到Y的影响链;因此,因果网络“吸烟”-->“肺癌”能够反映我们关于真实世界如何运作的信念(迫使受试者吸烟确实能改变一个人的信念,使他相信这会让受试者会患上癌症),而“癌症”-->“吸烟”则不能反映我们对真实世界的理解(如果受试者因为人为诱导而患上癌症,则不会改变一个人对该受试者是否吸烟的信念)。这个 Pearl 称之为do-calculus的简单分析,导致了一套完整数学框架的出现,对因果模型做了形式化,并能通过分析数据确定因果关系。这项工作推翻了长期以来人们对统计学的看法,即因果关系只能通过受控随机试验来确定——多数情况下,在生物和社会科学等领域实施随机受控实验是不可能的。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
2015年,ACM fellow<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 相关链接 ==<br />
<br />
* [http://bayes.cs.ucla.edu/jp_home.html 个人主页]<br />
* [https://amturing.acm.org/award_winners/pearl_2658896.cfm AMC图灵奖人物主页]<br />
* [https://scholar.google.com/citations?hl=en&user=bAipNH8AAAAJ 谷歌学术个人主页]<br />
* [[wikipedia:Judea_Pearl|维基主页]]<br />
<br />
== 参考文献 ==<br />
<br />
<br />
<br />
# Pearl, J., “Asymptotic properties of minimax trees and game-searching procedures,” ''Artificial Intelligence'', 14, pp. 113–138, September 1980. ''One of the first papers to establish “phase transition” properties for a combinatorial problem; introduced new mathematical techniques into the AI literature''.<br />
# <span name="bib_2"></span>Pearl, J., “Knowledge versus search: A quantitative analysis using A*,” ''Artificial Intelligence'', Vol. 20, pp. 1–13, 1983. ''Proved the first results relating heuristic accuracy to search algorithm complexity.''<br />
# <span name="bib_3"></span>Pearl, J., “On the nature of pathology in game searching,” ''Artificial Intelligence'', Vol. 20, pp. 427–453, 1983. ''Proved that, under the standard model of game trees, deeper search does not necessarily improve play; and showed that this paradox is resolved by correct probabilistic updating of beliefs.''<br />
# <span name="bib_4"></span>Karp, R. and J. Pearl, “Searching for an optimal path in a tree with random costs," ''Artificial Intelligence'', Vol. 21, pp. 99–116, 1983. ''Identified a phase transition property for a very simple path-finding problem, with significant complexity implications.''<br />
# <span name="bib_5"></span>Pearl, J., “On the discovery and generation of certain heuristics,” ''AI Magazine'', Winter/Spring, pp. 23–33, 1983. ''The first paper on the systematic generation of admissible heuristics (lower bounds on optimal solution costs) by relaxing formally represented problem definitions; this idea led to dramatic advances in automated planning systems.''<br />
# <span name="bib_6"></span>Pearl, J., ''Heuristics'': ''Intelligent Search Strategies for Computer Problem Solving,'' Addison-Wesley, 1984. ''Synthesized essentially everything known up to that point about intelligent methods for search and game playing, much of it Pearl’s own work; also the first textbook to treat AI topics formally at a technically advanced level.''<br />
# <span name="bib_7"></span>Dechter, R. and J. Pearl, “Generalized best-first search strategies and the optimality of A*,” ''Journal of the Association for Computing Machinery'', Vol. 32, pp. 505–536, 1985. Available here.''Proved that A* is the most efficient member of a very broad class of problem-solving algorithms.''<br />
# <span name="bib_7"></span>Pearl, J., “Reverend Bayes on inference engines: A distributed hierarchical approach,” ''Proceedings, AAAI-82'', 1982. ''The paper that began the probabilistic revolution in AI by showing how several desirable properties of reasoning systems can be obtained through sound probabilistic inference. It introduced tree-structured networks as concise representations of complex probability models, identified conditional independence relationships as the key organizing principle for uncertain knowledge, and described an efficient, distributed, exact inference algorithm.''<br />
# Kim, J. and J. Pearl, “A computational model for combined causal and diagnostic reasoning in inference systems,” ''Proceedings, IJCAI-83'', 1983. ''Generalized the tree-structured network to allow for multiple parents, or causal influences, on any given variable.''<br />
# <span name="bib_10"></span>Pearl, J., “Learning hidden causes from empirical data,” ''Proceedings, IJCAI-85'', 1985. ''Initiated the study of methods for learning the structure of probabilistic causal models.''<br />
# <span name="bib_11"></span>Pearl, J., “On the logic of probabilistic dependencies,” ''Proceedings, AAAI-86'', 1986. ''One of several papers establishing the connection between graphical models and conditional independence relationships.''<br />
# <span name="bib_12"></span>Pearl, J., “Fusion, propagation and structuring in belief networks,” ''Artificial Intelligence'', Vol. 29, pp. 241–288, 1986. ''The key technical paper on representation and exact inference in general Bayesian networks; by 1991 this had become the most cited paper in the history of the Artificial Intelligence journal.''<br />
# <span name="bib_13"></span>Pearl J. and A. Paz, “Graphoids: A graph-based logic for reasoning about relevance relations,” In B. du Boulay et al. (Eds.), ''Advances in Artificial Intelligence II'', North-Holland, 1987. ''Establishes an axiomatic characterization of the properties that enable probabilities and other relational systems to be represented by graphs.''<br />
# <span name="bib_14"></span>Pearl, J., “Evidential reasoning using stochastic simulation of causal models,” ''Artificial Intelligence'', Vol. 32, pp. 245–258, 1987. ''Derived a general approximation algorithm for Bayesian network inference using Markov chain Monte Carlo (MCMC); this was the first significant use of MCMC in mainstream AI.''<br />
# <span name="bib_15"></span>Pearl, J., ''Probabilistic Reasoning in Intelligent Systems'', Morgan Kaufmann, 1988. ''Explained the philosophical, cognitive, and technical basis for a probabilistic view of knowledge, reasoning, and decision making. One of the most cited works in the history of computer science, this book initiated the modern era in AI and converted many researchers who had previously worked in the logical and neural-network communities.''<br />
# <span name="bib_15"></span> Pearl J. and T.S. Verma, “A theory of inferred causation,” ''Proceedings, KR-91'', 1991. ''Introduces minimal-model semantics as a basis for causal discovery, and shows that causal directionality can be inferred from patterns of correlations without resorting to temporal information.''<br />
# Pearl, J., “Graphical models, causality, and intervention,” ''Statistical Science'', Vol. 8, pp. 266–269, 1993. ''Introduces the back-door criterion for covariate selection, the first to guarantee bias-free estimation of causal effects.''<br />
# Pearl, J., “Causal diagrams for empirical research,” ''Biometrika'', Vol. 82, Num. 4, pp. 669–710, 1995. ''Introduces the theory of causal diagrams and its associated do-calculus; the first (and still the only) mathematical method to enable a systematic removal of confounding bias in observations.''<br />
# Pearl, J., “The Art and Science of Cause and Effect,” ''UCLA Cognitive Systems Laboratory, Technical Report R-248'', 1996. ''Transcript of lecture given Thursday, October 29, 1996, as part of the UCLA 81st Faculty Research Lecture Series.Used later as epilogue to the book Causality (2000). Provides a panoramic view of the historical development of causal thoughts from antiquity to modern days.''<br />
# Pearl, J., ''Causality'': ''Models, Reasoning, and Inference,'' Cambridge University Press, 2000. ''Building on theoretical results from 1987 to 2000, lays out a complete framework for causal discovery, interventional analysis and counterfactual reasoning, bringing mathematical rigor and conceptual clarity to an area previously considered off-limits for statistics. Winner of the 2001 Lakatos Prize for the most significant new work in the philosophy of science.''<br />
# Pearl, J., “The logic of counterfactuals in causal inference (Discussion of `Causal inference without counterfactuals' by A.P. Dawid),” ''Journal of American Statistical Association'', Vol. 95, pp. 428–435, 2000. ''Demonstrates how counterfactual reasoning underlines scientific thought and argues against its exclusion from statistical analysis.''<br />
# Tian, J. and J. Pearl, “Probabilities of causation: Bounds and identification,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 28, pp. 287–313, 2000. ''Derives tight bounds on the probability that one observed event was the cause of another, in the legal sense of "but for," thus providing a principled way of substantiating guilt and innocence from data.''<br />
# Pearl, J., “Robustness of causal claims,” ''Proceedings, UAI-04'', 2004. ''Offers a formal definition of robustness and develops a method for assessing the degree to which causal claims are robust to model misspecification.''<br />
# Pearl, J., “Direct and indirect effects,” ''Proceedings, UAI-01'', 2001. ''Establishes the theoretical basis of modern mediation analysis. Derives the "Mediation Formula" and provides graphical conditions for the identification of direct and indirect effect.''<br />
# Tian, J. and J. Pearl, “A general identification condition for causal effects,” ''Proceedings, AAAI-02'', 2002. ''Uses the do-calculus to derive a general graphical condition for identifying causal effects from a combination of data and assumptions.''<br />
# Halpern, J. and J. Pearl, “Causes and explanations: A structural-model approach—Parts I and II,” ''British Journal for the Philosophy of Science'', Vol. 56, pp. 843–887 and 889–911, 2005. ''Establishes counterfactual conditions for one event to be perceived as the “actual cause” of another and for one event to provide an “explanation” of another.''<br />
# Pearl, J., “Causal inference in statistics: An overview,” ''Statistics Surveys'', Vol. 3, pp. 96–146, 2009. ''Describes a unified methodology for causal inference based on a symbiosis between graphs and counterfactual logic.''<br />
# Pearl, J., “The algorithmization of counterfactuals,” ''Annals of Mathematics and Artificial Intelligence'', Vol. 61, pp. 29–39, 2011. ''Describes a computational model that explains how humans generate, evaluate and distinguish counterfactual statements so swiftly and consistently.''<br />
# Pearl J. and E. Bareinboim, “Transportability of causal and statistical relations: A formal approach,” ''Proceedings, AAAI-11'', 2011. ''Reduces the classical problem of external validity to mathematical transformations in the do-calculus, and establishes conditions under which experimental results can be generalized to new environments in which only passive observation can be conducted.''<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32518Judea Pearl2022-06-20T02:08:40Z<p>Aceyuan:研究领域</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|483x483像素]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称因果之梯 The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法(发音:A star algorithm),以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32517Judea Pearl2022-06-20T02:07:13Z<p>Aceyuan:研究领域</p>
<hr />
<div><br />
== 基本信息 ==<br />
[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔|边框|左|483x483像素]]<br />
{| class="wikitable"<br />
|+<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称因果之梯 The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
<br />
<br />
<br />
<br />
<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
=== 搜索和启发式 ===<br />
Pearl 在计算机科学领域的声誉最初不是建立在概率推理(这在当时是一个有争议的话题)上,而是建立在组合搜索上。从1980年开始发表一系列期刊论文,最终于1984年 Pearl 出版了《启发式:计算机问题解决的智能搜索策略》一书。这项工作包括许多关于传统搜索算法的新结果,例如A*算法,以及在游戏算法方面,将人工智能研究提升到一个新的严谨和深度水平。它还提出了关于如何从宽松的问题定义中自动推导出可接受的启发式的新想法,这种方法导致了规划系统的巨大进步。<br />
<br />
=== 贝叶斯网络 ===<br />
Pearl 认为,对问题进行合理的概率分析会给出直观正确的结果,即使在基于规则的系统行为不正确的情况下也是如此。一个这样的案例与因果推理(从原因到结果)和诊断推理的能力有关(从结果到原因)。“如果使用诊断规则,则无法进行预测,如果使用预测规则,则无法进行诊断推理,如果同时使用两者,则会遇到正反馈不稳定性,这是我们在概率论中从未遇到过的。” 另一个案例涉及“解释消失”现象,即当观察到给定结果时,对导致结果的任何原因的相信程度会增加,但当发现其他原因也能导致观察到的结果时,对之前原因的相信程度就会降低。基于规则的系统不能表现出“解释消失”现象,而它在概率分析中会自动发生。<br />
<br />
Pearl 意识到条件独立的概念将是构建具有多项式多参数的复杂概率模型和组织分布式概率计算的关键。论文“Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach”[2]介绍了由有向无环图定义的概率模型,并推导出了一种精确的、分布式的、异步的、线性时间的树推理算法——我们现在称之为信念传播的算法,turbocodes的基础。随后,Pearl 有一段非凡的创意产出,发表了 50 多篇论文,涵盖一般图的精确推理、使用马尔可夫链蒙特卡罗的近似推理算法、条件独立属性、学习算法等,直到 1988 年出版了《智能系统中的概率推理》。这部具有里程碑意义的著作将珀尔的哲学、他的人类认知理论和他所有的技术材料结合成一个有说服力的整体这引发了人工智能领域的一场革命。在短短几年内,来自人工智能内部逻辑阵营和神经网络阵营的主要研究人员采用了一种概率(通常简称为现代)方法来研究人工智能。<br />
<br />
Pearl 的贝叶斯网络为多元概率模型提供了句法和演算,就像乔治·布尔为逻辑模型提供句法和演算一样。与贝叶斯网络相关的理论和算法问题是机器学习和统计学现代研究议程的重要组成部分,它们的使用也渗透到其他领域,如自然语言处理、计算机视觉、机器人技术、计算生物学和认知科学。截至2012年,已经出现了大约 50,000 篇以贝叶斯网络为主要关注点的出版物。<br />
<br />
=== 因果关系 ===<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32513Judea Pearl2022-06-19T16:02:31Z<p>Aceyuan:更新成长经历</p>
<hr />
<div><br />
== 基本信息 ==<br />
{| class="wikitable"<br />
|+[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔]]<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称因果之梯 The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
从前的人工智能领域以布尔代数为基础,陈述要么是真要么是假。朱迪亚·珀尔创造了贝叶斯网络,将现实世界中的模糊性引入了该领域。贝叶斯网络利用图论(通常需要结合贝叶斯统计,但不必需),允许机器在面对不确定或零散信息时能做出合理的假设。他在著作《智能系统中的概率推理:合理推断网络》(1988)中描述了这项工作。 <br />
<br />
朱迪亚·珀尔还全面研究了因果关系,即原因和结果的关系,并对这些关系做了数学形式化描述。他的著作《因果关系:模型、论证、推理》(2000)在许多不同的学科中产生了影响,包括心理学、社会学、医学和科学哲学。<br />
<br />
朱迪亚·珀尔目前的研究兴趣包括:人工智能,概率和因果推理,科学哲学和科学史。<br />
<br />
=== 搜索和启发式 ===<br />
<br />
=== 贝叶斯网络 ===<br />
<br />
=== 因果关系 ===<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32488Judea Pearl2022-06-19T10:46:15Z<p>Aceyuan:更新简介及成长经历</p>
<hr />
<div><br />
== 基本信息 ==<br />
{| class="wikitable"<br />
|+[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔]]<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称因果之梯 The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他因发明了贝叶斯网络、定义复杂概率模型的数学形式以及这些模型中用于推理的主要算法而受到赞誉。这项工作不仅彻底改变了人工智能领域,而且成为许多其他工程和自然科学分支的重要工具。他后来创建了一个因果推理的数学框架,该框架对社会科学产生了重大影响。ACM授予Judea Pearl 2011年度图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。<br />
<br />
他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了因果之梯的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
=== 求学 ===<br />
Judea Pearl 于1960年在海法的以色列理工学院获得电气工程学士学位。他于1961年在纽瓦克工程学院(现为新泽西理工学院)获得电气工程硕士学位。1965年,他在新泽西州新不伦瑞克市的罗格斯大学获得物理学硕士学位,同年,在纽约布鲁克林理工学院(现纽约大学理工学院)获得电气工程博士学位。他的博士学位论文是“超导记忆的涡旋理论”,“Pearl 涡旋(Pearl Vortex)”就是用来描述他所研究的超导电流的类型,这个词在物理学家中很流行。<br />
<br />
=== 工作 ===<br />
Pearl 曾在新泽西州普林斯顿的 RCA 研究实验室从事超导参数放大器和存储器件方面的工作,并在加利福尼亚州霍桑市的 Electronic Memory, Inc. 从事高级存储系统方面的工作。尽管当时他的工作聚焦在物理器件方面,Pearl 说从那时起他就对智能系统潜在应用充满向往。<br />
<br />
=== 学术 ===<br />
当磁性和超导存储器的工业研究因大规模半导体存储器的出现而减少时,Pearl 决定进入学术界以追求他对逻辑和推理的长期兴趣。1969 年,他加入加州大学洛杉矶分校,最初在工程系统系任教,并于1970年在新成立的计算机科学系获得终身教职。1976年晋升为正教授。1978年,他创立了认知系统实验室——这个名称强调了他对理解人类认知的愿望。<br />
<br />
== 研究领域 ==<br />
从前的人工智能领域以布尔代数为基础,陈述要么是真要么是假。朱迪亚·珀尔创造了贝叶斯网络,将现实世界中的模糊性引入了该领域。贝叶斯网络利用图论(通常需要结合贝叶斯统计,但不必需),允许机器在面对不确定或零散信息时能做出合理的假设。他在著作《智能系统中的概率推理:合理推断网络》(1988)中描述了这项工作。 <br />
<br />
朱迪亚·珀尔还全面研究了因果关系,即原因和结果的关系,并对这些关系做了数学形式化描述。他的著作《因果关系:模型、论证、推理》(2000)在许多不同的学科中产生了影响,包括心理学、社会学、医学和科学哲学。<br />
<br />
朱迪亚·珀尔目前的研究兴趣包括:人工智能,概率和因果推理,科学哲学和科学史。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32487Judea Pearl2022-06-19T08:55:27Z<p>Aceyuan:/* 基本信息 */</p>
<hr />
<div><br />
== 基本信息 ==<br />
{| class="wikitable"<br />
|+[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔]]<br />
!类别<br />
!信息<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称<nowiki><font color="#ff8000">因果之梯</font></nowiki> The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他还因在结构模型的基础上发展出因果和反事实推论而受到广泛称赞。2011年,ACM授予Judea Pearl图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了<nowiki><font color="#ff8000"> 因果之梯</font></nowiki>的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
朱迪亚·珀尔于1960年在海法的以色列理工学院获得电气工程学士学位。<br />
<br />
他于1961年在纽瓦克工程学院(Newark College of Engineering,现为新泽西理工学院,New Jersey Institute of Technology,的一部分)获得电气工程硕士学位。<br />
<br />
然后,他在新泽西州新不伦瑞克市的罗格斯大学(Rutgers University)获得了物理学硕士学位。<br />
<br />
1965年他在纽约布鲁克林理工学院(Polytechnic Institute of Brooklyn,现纽约大学理工学院,Polytechnic Institute of New York University)获得电气工程博士学位。<br />
<br />
他曾在位于新泽西州普林斯顿的RCA公司(现为Sarnoff Corporation)的David Sarnoff实验室工作。<br />
<br />
之后还在位于加利福尼亚州霍桑市的制造商Electronic Memory, Inc. 公司(后来的Electronics Memory and Magnetics Corp.)研发计算机磁性存储器。<br />
<br />
1970年,他成为加州大学洛杉矶分校的计算机科学教授。<br />
<br />
== 研究领域 ==<br />
从前的人工智能领域以布尔代数为基础,陈述要么是真要么是假。朱迪亚·珀尔创造了贝叶斯网络,将现实世界中的模糊性引入了该领域。贝叶斯网络利用图论(通常需要结合贝叶斯统计,但不必需),允许机器在面对不确定或零散信息时能做出合理的假设。他在著作《智能系统中的概率推理:合理推断网络》(1988)中描述了这项工作。 <br />
<br />
朱迪亚·珀尔还全面研究了因果关系,即原因和结果的关系,并对这些关系做了数学形式化描述。他的著作《因果关系:模型、论证、推理》(2000)在许多不同的学科中产生了影响,包括心理学、社会学、医学和科学哲学。<br />
<br />
朱迪亚·珀尔目前的研究兴趣包括:人工智能,概率和因果推理,科学哲学和科学史。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32486Judea Pearl2022-06-19T08:45:28Z<p>Aceyuan:恢复简介</p>
<hr />
<div><br />
== 基本信息 ==<br />
{| class="wikitable"<br />
|+[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔]]<br />
!生日<br />
!1936年9月4日<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称<nowiki><font color="#ff8000">因果之梯</font></nowiki> The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
朱迪亚·珀尔(Judea Pearl)——以色列裔美籍计算机科学家、哲学家,以倡导人工智能的概率方法和贝叶斯网络而闻名。他还因在结构模型的基础上发展出因果和反事实推论而受到广泛称赞。2011年,ACM授予Judea Pearl图灵奖,以表彰他“通过发展概率和因果推理演算对人工智能做出的基础性贡献”。他早在40多年前便通过贝叶斯网络的设计,使机器实现概率推理而在人工智能领域声名大噪,并被誉为“贝叶斯网络之父”,但近年却公开声称自己其实是人工智能社区的一名“叛徒”:离开了主流追逐、并且也是由他奠定重要理论基础和方法论的概率推理,而去追求一项更具挑战性的任务——因果推理。Judea Pearl 认为当今深度学习所有令人印象深刻的成就,都只不过是为了适应“曲线拟合(Curve fitting)”。而今,这也导致深度学习的研究员们困在了<nowiki><font color="#ff8000"> 因果之梯</font></nowiki>的最底层——“关联”层次的问题窘境里。Judea Pearl 期望能掀起一场“因果革命”,采用因果推理模型,从因果而非单纯的数据关联角度去研究人工智能。<br />
<br />
== 成长经历 ==<br />
<br />
朱迪亚·珀尔于1960年在海法的以色列理工学院获得电气工程学士学位。<br />
<br />
他于1961年在纽瓦克工程学院(Newark College of Engineering,现为新泽西理工学院,New Jersey Institute of Technology,的一部分)获得电气工程硕士学位。<br />
<br />
然后,他在新泽西州新不伦瑞克市的罗格斯大学(Rutgers University)获得了物理学硕士学位。<br />
<br />
1965年他在纽约布鲁克林理工学院(Polytechnic Institute of Brooklyn,现纽约大学理工学院,Polytechnic Institute of New York University)获得电气工程博士学位。<br />
<br />
他曾在位于新泽西州普林斯顿的RCA公司(现为Sarnoff Corporation)的David Sarnoff实验室工作。<br />
<br />
之后还在位于加利福尼亚州霍桑市的制造商Electronic Memory, Inc. 公司(后来的Electronics Memory and Magnetics Corp.)研发计算机磁性存储器。<br />
<br />
1970年,他成为加州大学洛杉矶分校的计算机科学教授。<br />
<br />
== 研究领域 ==<br />
从前的人工智能领域以布尔代数为基础,陈述要么是真要么是假。朱迪亚·珀尔创造了贝叶斯网络,将现实世界中的模糊性引入了该领域。贝叶斯网络利用图论(通常需要结合贝叶斯统计,但不必需),允许机器在面对不确定或零散信息时能做出合理的假设。他在著作《智能系统中的概率推理:合理推断网络》(1988)中描述了这项工作。 <br />
<br />
朱迪亚·珀尔还全面研究了因果关系,即原因和结果的关系,并对这些关系做了数学形式化描述。他的著作《因果关系:模型、论证、推理》(2000)在许多不同的学科中产生了影响,包括心理学、社会学、医学和科学哲学。<br />
<br />
朱迪亚·珀尔目前的研究兴趣包括:人工智能,概率和因果推理,科学哲学和科学史。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)<br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32449Judea Pearl2022-06-18T15:25:29Z<p>Aceyuan:/* 奖项与成就 */</p>
<hr />
<div><br />
== 基本信息 ==<br />
{| class="wikitable"<br />
|+[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔]]<br />
!生日<br />
!1936年9月4日<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称因果之梯 The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
朱迪亚·珀尔 Judea Pearl,是一位著名的计算机科学家和哲学家,他在人工智能、因果关系和贝叶斯网络领域的工作赢得了国际声誉。朱迪亚·珀尔被公认为人工智能领域的巨人之一,这主要是因为他的工作彻底改变了统计学、心理学、医学和社会科学领域对因果关系的理解。朱迪亚·珀尔是美国国家工程院院士,也是美国人工智能协会的创始成员。他发明了贝叶斯网络,被称为“贝叶斯网络之父”。他对科学哲学、知识表示、非标准逻辑和机器学习感兴趣,提出了一个高级认知模型。在2011年他因创立因果推理演算法荣获美国计算机工程协会(ACM)的图灵奖, 这是计算机工程领域最高荣誉。截止2022年,他撰写了超过500篇科学论文和五本著作。他在人工智能领域的开创性贡献无人能及。<br />
<br />
== 成长经历 ==<br />
<br />
朱迪亚·珀尔于1960年在海法的以色列理工学院获得电气工程学士学位。<br />
<br />
他于1961年在纽瓦克工程学院(Newark College of Engineering,现为新泽西理工学院,New Jersey Institute of Technology,的一部分)获得电气工程硕士学位。<br />
<br />
然后,他在新泽西州新不伦瑞克市的罗格斯大学(Rutgers University)获得了物理学硕士学位。<br />
<br />
1965年他在纽约布鲁克林理工学院(Polytechnic Institute of Brooklyn,现纽约大学理工学院,Polytechnic Institute of New York University)获得电气工程博士学位。<br />
<br />
他曾在位于新泽西州普林斯顿的RCA公司(现为Sarnoff Corporation)的David Sarnoff实验室工作。<br />
<br />
之后还在位于加利福尼亚州霍桑市的制造商Electronic Memory, Inc. 公司(后来的Electronics Memory and Magnetics Corp.)研发计算机磁性存储器。<br />
<br />
1970年,他成为加州大学洛杉矶分校的计算机科学教授。<br />
<br />
== 研究领域 ==<br />
从前的人工智能领域以布尔代数为基础,陈述要么是真要么是假。朱迪亚·珀尔创造了贝叶斯网络,将现实世界中的模糊性引入了该领域。贝叶斯网络利用图论(通常需要结合贝叶斯统计,但不必需),允许机器在面对不确定或零散信息时能做出合理的假设。他在著作《智能系统中的概率推理:合理推断网络》(1988)中描述了这项工作。 <br />
<br />
朱迪亚·珀尔还全面研究了因果关系,即原因和结果的关系,并对这些关系做了数学形式化描述。他的著作《因果关系:模型、论证、推理》(2000)在许多不同的学科中产生了影响,包括心理学、社会学、医学和科学哲学。<br />
<br />
朱迪亚·珀尔目前的研究兴趣包括:人工智能,概率和因果推理,科学哲学和科学史。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>) <br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作:<br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=Judea_Pearl&diff=32448Judea Pearl2022-06-18T15:24:31Z<p>Aceyuan:Re-organize the contents</p>
<hr />
<div><br />
== 基本信息 ==<br />
{| class="wikitable"<br />
|+[[文件:Judea Pearl.png|替代=朱迪亚·珀尔|缩略图|朱迪亚·珀尔]]<br />
!生日<br />
!1936年9月4日<br />
|-<br />
|生日<br />
|1936年9月4日<br />
|-<br />
|出生地<br />
|以色列特拉维夫<br />
|-<br />
|国籍<br />
|美国<br />
|-<br />
|居住地<br />
|美国加利福尼亚州洛杉矶市<br />
|-<br />
|任职<br />
|加州大学洛杉矶分校,计算机科学教授<br />
|-<br />
|著名成就<br />
|贝叶斯网络<br />
珀尔因果层次模型 PCH(Pearl Causal Hierarchy),又称因果之梯 The Ladder of Causation<br />
|-<br />
|主要研究方向<br />
|人工智能、因果推理和科学哲学<br />
|-<br />
|教育院校<br />
|以色列理工学院<br />
纽瓦克工程学院(现新泽西理工学院)<br />
<br />
新泽西州新不伦瑞克市罗格斯大学<br />
<br />
纽约布鲁克林理工学院(现纽约大学理工学院)<br />
|}<br />
朱迪亚·珀尔 Judea Pearl,是一位著名的计算机科学家和哲学家,他在人工智能、因果关系和贝叶斯网络领域的工作赢得了国际声誉。朱迪亚·珀尔被公认为人工智能领域的巨人之一,这主要是因为他的工作彻底改变了统计学、心理学、医学和社会科学领域对因果关系的理解。朱迪亚·珀尔是美国国家工程院院士,也是美国人工智能协会的创始成员。他发明了贝叶斯网络,被称为“贝叶斯网络之父”。他对科学哲学、知识表示、非标准逻辑和机器学习感兴趣,提出了一个高级认知模型。在2011年他因创立因果推理演算法荣获美国计算机工程协会(ACM)的图灵奖, 这是计算机工程领域最高荣誉。截止2022年,他撰写了超过500篇科学论文和五本著作。他在人工智能领域的开创性贡献无人能及。<br />
<br />
== 成长经历 ==<br />
<br />
朱迪亚·珀尔于1960年在海法的以色列理工学院获得电气工程学士学位。<br />
<br />
他于1961年在纽瓦克工程学院(Newark College of Engineering,现为新泽西理工学院,New Jersey Institute of Technology,的一部分)获得电气工程硕士学位。<br />
<br />
然后,他在新泽西州新不伦瑞克市的罗格斯大学(Rutgers University)获得了物理学硕士学位。<br />
<br />
1965年他在纽约布鲁克林理工学院(Polytechnic Institute of Brooklyn,现纽约大学理工学院,Polytechnic Institute of New York University)获得电气工程博士学位。<br />
<br />
他曾在位于新泽西州普林斯顿的RCA公司(现为Sarnoff Corporation)的David Sarnoff实验室工作。<br />
<br />
之后还在位于加利福尼亚州霍桑市的制造商Electronic Memory, Inc. 公司(后来的Electronics Memory and Magnetics Corp.)研发计算机磁性存储器。<br />
<br />
1970年,他成为加州大学洛杉矶分校的计算机科学教授。<br />
<br />
== 研究领域 ==<br />
从前的人工智能领域以布尔代数为基础,陈述要么是真要么是假。朱迪亚·珀尔创造了贝叶斯网络,将现实世界中的模糊性引入了该领域。贝叶斯网络利用图论(通常需要结合贝叶斯统计,但不必需),允许机器在面对不确定或零散信息时能做出合理的假设。他在著作《智能系统中的概率推理:合理推断网络》(1988)中描述了这项工作。 <br />
<br />
朱迪亚·珀尔还全面研究了因果关系,即原因和结果的关系,并对这些关系做了数学形式化描述。他的著作《因果关系:模型、论证、推理》(2000)在许多不同的学科中产生了影响,包括心理学、社会学、医学和科学哲学。<br />
<br />
朱迪亚·珀尔目前的研究兴趣包括:人工智能,概率和因果推理,科学哲学和科学史。<br />
<br />
== 奖项与成就 ==<br />
多年来,他因在人工智能、人类推理和科学哲学领域做出重大贡献而享誉国际。获得近50项各类奖项,全部列表请参见:http://bayes.cs.ucla.edu/jp_home.html <br />
<br />
以下为重要奖项: <br />
<br />
2001 年,他因提出科学哲学方面的最佳著作而获得伦敦经济学院授予的拉科塔斯奖。 <br />
<br />
2003 年,他获得了 ACM 艾伦纽厄尔奖。<br />
<br />
2006 年,获得了 Civic Venture 的首届目的奖,该奖项旨在表彰 60 岁及以上在解决社区和国家问题方面表现出非凡远见的个人。<br />
<br />
2008 年,富兰克林研究所授予他本杰明富兰克林计算机和认知科学奖章。<br />
<br />
2011 年,他因对人类认知理论基础的贡献而获得大卫 E. Rumelhart 奖。他的母校授予他哈维科学技术奖。<br />
<br />
2011 年,他获得了ACM的图灵奖,这是计算机工程领域的最高荣誉,以表彰他“通过开发用于概率和因果推理的微积分对人工智能的根本贡献”。<br />
<br />
== 主要文章及著作 ==<br />
他就人工智能的各个主题发表了近500篇科学论文(http://bayes.cs.ucla.edu/jp_home.html<nowiki/>)。此外,他在上述感兴趣的领域共出版五本著作: <br />
<br />
1984 年,《启发法》 Heuristics, Addison-Wesley, 1984<br />
<br />
1988 年,《智能系统中的概率推理:合理推断网络》 Probabilistic Reasoning in Intelligent Systems, Morgan-Kaufmann, 1988<br />
<br />
2009 年,《因果关系:模型、论证、推理》 Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000; 2nd edition, 2009.<br />
<br />
2016 年,《统计因果推理入门》 Causal Inference in Statistics: A Primer, (with Madelyn Glymour and Nicholas P. Jewell) Wiley, 2016.<br />
<br />
2018 年,《为什么:关于因果关系的新科学》 The Book of Why: The New Science of Cause and Effect (with Dana Mackenzie), New York: Basic Books, May 2018<br />
<br />
== 个人生活 ==<br />
他和露丝结婚。这对夫妇生下了三个孩子,其中包括在巴基斯坦被基地组织武装分子绑架并杀害的记者丹尼尔·珀尔。这位以色列裔美国计算机科学家和哲学家的儿子、华尔街日报记者丹尼尔·珀尔(Daniel Pearl)于 2002 年在巴基斯坦被激进分子绑架并杀害。 <br />
<br />
在他的记者儿子去世后,他与家人一起创立了丹尼尔·珀尔基金会。该组织的主要目的是促进诚实的报道和东西方的理解。它还旨在达到犹太人和穆斯林之间的理解水平。该组织在 2002 年和 2003 年同时获得了两个奖项。<br />
<br />
== 研究领域 ==<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
<br />
(R-513): [pdf] <br />
<br />
S. Mueller and J. Pearl "Personalized Decision Making -- A Conceptual Introduction,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-513), April 2022.<br />
<br />
(R-511): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Bounds on Causal Effects and Application to High Dimensional Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-511), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-510): [pdf] [bib] <br />
<br />
A. Li and J. Pearl "Unit Selection with Causal Diagram,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-510), March 2022.<br />
<br />
In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22).<br />
<br />
(R-509): [pdf] [bib] <br />
<br />
A. Forney and S. Mueller "Causal Inference in AI Education: A Primer,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-509), June 2022.<br />
<br />
Forthcoming, Journal of Causal Inference.<br />
<br />
(R-508): [pdf] [bib] <br />
<br />
C. Cinelli "Transparent and Robust Causal Inferences in the Social and Health Sciences,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-508), July 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-507): [pdf] [bib] <br />
<br />
A. Li, "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-507), June 2021.<br />
<br />
Ph.D. Thesis<br />
<br />
(R-506): [pdf] [bib] <br />
<br />
S. Mueller, "Estimating Individualized Causes of Effects by Leveraging Population Data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-506), June 2021.<br />
<br />
Master's Thesis<br />
<br />
(R-505): [pdf] [bib]<br />
<br />
S. Mueller, A. Li, and J. Pearl "Causes of effects: Learning individual responses from population data,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-505), Revised May 2022.<br />
<br />
Forthcoming, Proceedings of IJCAI-2022. 5br><br />
<br />
(R-505-Supplemental): [Supplemental]<br />
<br />
(R-504): [pdf] [bib]<br />
<br />
C. Zhang, C. Cinelli, B. Chen, and J. Pearl "Exploiting equality constraints in causal inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-504), April 2021.<br />
<br />
Proceedings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. PMLR: Volume 130, 1630-1638, April 2021.<br />
<br />
(R-504-Supplemental): [Supplemental]<br />
<br />
(R-503): [pdf] [bib]<br />
<br />
J. Pearl "Causally Colored Reflections on Leo Breiman's `Statistical Modeling: The Two Cultures' (2001),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-503), March 2021.<br />
<br />
Observational Studies, Vol. 7.1:187-190, 2021.<br />
<br />
(R-502): [pdf] [bib]<br />
<br />
J. Pearl "Radical Empiricism and Machine Learning Research,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-502), May 2021.<br />
<br />
Journal of Causal Inference, 9:78–82, 2021.<br />
<br />
(R-501): [pdf] [bib]<br />
<br />
J. Pearl "Causal, Casual, and Curious (2013-2020): A collage in the art of causal reasoning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-501), October 2020.<br />
<br />
(R-493): [pdf] [bib]<br />
<br />
C. Cinelli, A. Forney, and J. Pearl "A Crash Course in Good and Bad Controls,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-493), Revised, March 2022.<br />
<br />
Forthcoming, Journal Sociological Methods and Research.<br />
<br />
(R-492): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl "Generalizing experimental results by leveraging knowledge of mechanisms,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-492), September 2020.<br />
<br />
European Journal of Epidemiology, 36:149--164, 2021. URL <nowiki>https://doi.org/10.1007/s10654-020-00687-4</nowiki>.<br />
<br />
(R-491-L): [pdf] [bib]<br />
<br />
C. Zhang, B. Chen, and J. Pearl "A Simultaneous Discover-Identify Approach to Causal Inference in Linear Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-491-L), February 2020.<br />
<br />
Extended version of paper in Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-2020), 34(6): 10318--10325, Palo Alto, CA: AAAI Press, 2020.<br />
<br />
(R-489): [pdf] [bib]<br />
<br />
J. Pearl "The Limitations of Opaque Learning Machines,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-489), May 2019.<br />
<br />
Chapter 2 in John Brockman (Ed.), Possible Minds: 25 Ways of Looking at AI, Penguin Press, 2019.<br />
<br />
(R-488): [pdf] [bib]<br />
<br />
A. Li and J. Pearl "Unit Selection Based on Counterfactual Logic,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-488), June 2019.<br />
<br />
In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19), 1793-1799, 2019.<br />
<br />
(R-488-Supplemental): [Supplemental]<br />
<br />
(R-487): [pdf] [bib]<br />
<br />
J. Pearl and Co-authored by D. Mackenzie, "Telling and re-telling history: The case for a whiggish account of the history of causation,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-487), March 2019.<br />
<br />
(R-486): [pdf] [bib]<br />
<br />
J. Pearl, "On the interpretation of do(x),"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-486), February 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 7(1), online, March 2019.<br />
<br />
(R-485): [pdf] [bib]<br />
<br />
J. Pearl, "Causal and counterfactual inference,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-485), December 2021.<br />
<br />
In Markus Knauff and Wolfgang Spohn (Eds.), The Handbook of Rationality, Section 7.1, pp. 427-438, The MIT Press, 2021.<br />
<br />
(R-484): [pdf] [bib]<br />
<br />
J. Pearl, "Sufficient Causes: On Oxygen, Matches, and Fires,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-484), September 2019.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, AOP, <nowiki>https://doi.org/10.1515/jci-2019-0026</nowiki>, September 2019.<br />
<br />
(R-483): [pdf] [bib]<br />
<br />
J. Pearl, "Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-483), August 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(2), online, September 2018.<br />
<br />
(R-482): [pdf] [bib]<br />
<br />
C. Cinelli, D. Kumor, B. Chen, J. Pearl, and E. Bareinboim<br />
<br />
"Sensitivity Analysis of Linear Structural Causal Models,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-482), June 2019.<br />
<br />
Proceedings of the 36th International Conference on Machine Learning, PMLR 97, 1252-1261, 2019.<br />
<br />
(R-481): [pdf] [bib]<br />
<br />
J. Pearl, "The Seven Tools of Causal Inference with Reflections on Machine Learning,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-481), July 2018.<br />
<br />
Communications of ACM, 62(3): 54-60, March 2019<br />
<br />
(R-480): [pdf] [bib]<br />
<br />
K. Mohan, F. Thoemmes, J. Pearl, "Estimation with Incomplete Data: The Linear Case,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-480), May 2018.<br />
<br />
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), 5082-5088, 2018.<br />
<br />
(R-479): [pdf] [bib]<br />
<br />
C. Cinelli and J. Pearl, "RE: A Practical Example Demonstrating the Utility of Single-world Intervention Graphs,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-479), April 2018.<br />
<br />
Journal of Epidemiology, 29(6): e50--e51, November 2018.<br />
<br />
(R-478): [pdf] [bib]<br />
<br />
J. Pearl and E. Bareinboim, "A note on `Generalizability of Study Results',"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-478), April 2018.<br />
<br />
Epidemiology, 30(2):186--188, March 2019.<br />
<br />
(R-477): [pdf] [bib]<br />
<br />
J. Pearl, "Challenging the Hegemony of Randomized Controlled Trials: Comments on Deaton and Cartwright,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-477), April 2018.<br />
<br />
Social Science and Medicine, published online, April 2018.<br />
<br />
(R-476): [pdf] [bib]<br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-475): [pdf] [bib]<br />
<br />
J. Pearl, "Theoretical Impediments to Machine Learning with Seven Sparks from the Causal Revolution"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-475), July 2018.<br />
<br />
Paper supporting Keynote Talk WSDM'18, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, DOI: <nowiki>http://dx.doi.org/10.1145/3159652.3160601</nowiki>, February 2018.<br />
<br />
(R-474): [pdf] [bib]<br />
<br />
J. Pearl, "Comments on `The Tale Wagged by the DAG'"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-474), January 2018.<br />
<br />
International Journal of Epidemiology, 47(3):1002-1004, 2018.<br />
<br />
(R-473): [pdf] [bib]<br />
<br />
K. Mohan and J. Pearl, "Graphical Models for Processing Missing Data"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-473-L), June 2019.<br />
<br />
Journal of American Statistical Association (JASA). Online March 2021 (<nowiki>https://doi.org/10.1080/01621459.2021.1874961</nowiki>).<br />
<br />
(R-472): [pdf] [bib]<br />
<br />
J. Pearl, "What is Gained from Past Learning"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-472), March 2018.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 6(1), Article 20180005, March 2018. <nowiki>https://doi.org/10.1515/jci-2018-0005</nowiki><br />
<br />
(R-471): [pdf] [bib]<br />
<br />
A. Forney, J. Pearl, and E. Bareinboim, "Counterfactual Data-Fusion for Online Reinforcement Learners"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-471), June 2017.<br />
<br />
Presented at the Transfer in Reinforcement Learning workshop at AAMAS-2017.<br />
<br />
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1156-1164, 2017.<br />
<br />
(R-470): [pdf] [bib]<br />
<br />
J. Pearl, "The Eight Pillars of Causal Wisdom"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-470), April 2017.<br />
<br />
(R-469): <br />
<br />
J. Pearl, "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
(R-466): [pdf] [bib]<br />
<br />
J. Pearl "The Sure-Thing Principle"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-466), February 2016.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, 4(1):81-86, March 2016.<br />
<br />
(R-461): [pdf] [bib]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461), July 2016.<br />
<br />
In S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-461-L): [pdf]<br />
<br />
B. Chen, J. Pearl, and E. Bareinboim, "Incorporating Knowledge into Structural Equation Models using Auxiliary Variables"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-461-L), April 2016.<br />
<br />
Extended version of paper in S. Kambhampati (Ed.), Proceedings of the 25 International Joint Conference on Artificial Intelligence (IJCAI), Palo Alto: AAAI Press, 3577-3583, 2016.<br />
<br />
(R-460): [pdf] [bib]<br />
<br />
E. Bareinboim, Andrew Forney, and J. Pearl, "Bandits with Unobserved Confounders: A Causal Approach"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-460), November 2015.<br />
<br />
In C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, and R. Garnett (Eds.), Neural Information Processing Systems (NIPS) Conference, Advances in Neural Information Processing Systems 28, Curran Associates, Inc., pp. 1342-1350, 2015.<br />
<br />
(R-459): [pdf] [bib]<br />
<br />
J. Pearl, "A Linear `Microscope' for Interventions and Counterfactuals"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-459), March 2017.<br />
<br />
Journal of Causal Inference, Causal, Casual, and Curious Section, published online 5(1):1-15, March 2017.<br />
<br />
== 参考文献 ==<br />
'''[1]J. Pearl''', "A Personal Journey into Bayesian Networks,"<br />
<br />
UCLA Cognitive Systems Laboratory, Technical Report (R-476), May 2018.<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部推文推荐 ===<br />
[https://swarma.org/?p=26472 统计学权威盘点过去50年最重要的统计学思想,因果推理、bootstrap等上榜,Judea Pearl点赞 | 集智俱乐部]<br />
<br />
[https://swarma.org/?p=34329 福利 | 因果推断会是下一个AI热潮吗?Judea Pearl《因果论》重磅上市!]<br />
<br />
[https://swarma.org/?p=32656 Stephen Wolfram专访Judea Pearl:从贝叶斯网络到元胞自动机 | 集智俱乐部]<br />
<br />
== 说明 ==<br />
J. Pearl 发表很多论文,是困难的 去编写问题 从Pearl 的论文 使用自己的语言。因此,我采用多轮次去编写。每个轮次编写1~2个问题。更多问题将编写。<br />
<br />
* 如何编写问题? 请参考“[https://www.bigphysics.org/index.php/%E5%88%86%E7%B1%BB:%E7%A0%94%E7%A9%B6%E6%8A%A5%E5%91%8A%E5%86%99%E4%BD%9C V形图]”<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnbL31C8uMNOd2M0LrnZcGhg 如何编写学者词条?]<br />
<br />
* [https://faixz6wsb8.feishu.cn/docs/doccnz7ACovB7pTqhbS3nQMn5Pb 如何多人一起 编写学者词条?]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E6%96%87%E4%BB%B6:Judea_Pearl.png&diff=32440文件:Judea Pearl.png2022-06-18T15:10:22Z<p>Aceyuan:</p>
<hr />
<div>Title image</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E6%BD%9C%E5%9C%A8%E7%BB%93%E6%9E%9C&diff=32001潜在结果2022-06-11T12:22:47Z<p>Aceyuan:审校完成</p>
<hr />
<div>此词条由因果科学读书会词条梳理志愿者我是猫(74989)翻译审校,未经专家审核,带来阅读不便,请见谅<br />
<br />
== 概念来源 ==<br />
潜在结果是指给定一个单元,和一系列动作,我们把一个“动作-单元”确定为一个潜在结果。“潜在(potential)”这个词表达的意思是我们并不总是能在现实中观察到这个结果(outcome),但原则上它们可能发生。<br />
<br />
潜在结果最初的提出是在Neyman的论文中,但是这篇文章只在随机对照试验中使用了潜在结果的概念,且直到1990年翻译成英文后才为人所知。Rubin在他1974年的论文中也提出了潜在结果的概念,并将这个概念推广到了观察性数据中,真正开启了统计学界对因果推断的广泛研究。<br />
<br />
何为潜在结果?又如何基于潜在结果定义因果?假设我们关心某个变量A(例如,在某个时间点是否服用阿莫西林,A=1是服用,A=0是没有服用)对Y(服用后三小时的是否还感冒,Y=1表示感冒,Y=0表示没有感冒)的因果关系。那么我们观察到的某个个体就存在两个“潜在”的状态:一个是如果他服药,他三小时后是否感冒,不妨记作Y(1);另一个如果他没有服药,他三小时后是否感冒,不妨记作Y(0)。这里Y(1)和Y(0)就是潜在结果。(注意,在实际中,Y(1)和Y(0)这二者中只有一个可以被观察到。另外,严格地说,此处实际上做了“个体处理稳定性”即SUTVA的假设)那么对这个人,就可能有以下四种情况:<br />
<br />
a) Y(0)=0, Y(1)=0。即不论吃不吃药,这个人在三小时后均不会感冒。<br />
<br />
b) Y(0)=1, Y(1)=1。即不论吃不吃药,这个人在三小时后均会感冒。<br />
<br />
c) Y(0)=1, Y(1)=0。即此人如果不吃药,三小时后会感冒,但是如果吃药,三小时后不会感冒。<br />
<br />
d) Y(0)=0, Y(1)=1。即此人如果不吃药,三小时后不会感冒,但是如果吃药,三小时后会感冒。<br />
<br />
在a和b两种情况下,Y(1)=Y(0),即吃不吃药不会影响三小时后是否感冒的状态,这种情况下我们说吃药对三小时后是否感冒没有因果作用,相反,在c和d两种情况下,Y(1)≠Y(0),这种情况下我们说吃药对三小时后是否感冒有因果作用。使用潜在结果,我们便可以方便地定义感兴趣的因果作用,例如平均因果效应E[Y(1)-Y(0)],这个量代表了在一个群体中,如果每一个人都采取某种处理和都不接受处理相比,这两种情况下平均意义上的结果差值。<br />
<br />
使用潜在结果我们或许可以理解为什么人们不会认为“太阳升起是因为鸡打鸣”,因为根据我们的常识,如果某天鸡不打鸣(或许是因为生病或劳累),太阳仍然会照常升起。<br />
<br />
'''因此从分析潜在结果出发,诞生了<font color="#ff8000">潜在结果框架,有时也称为</font>鲁宾因果模型 Rubin Causal Model (RCM)''' ,'''Neyman-Rubin 因果模型'''<ref name="sekhon">{{cite book |last=Sekhon |first=Jasjeet |chapter=The Neyman–Rubin Model of Causal Inference and Estimation via Matching Methods |title=The Oxford Handbook of Political Methodology |year=2007 |chapter-url=http://sekhon.berkeley.edu/papers/SekhonOxfordHandbook.pdf }}</ref>。它是一种基于潜在结果框架的因果统计分析方法,以Donald Rubin的名字命名。“鲁宾因果模型”这个名字最早是由 Paul W. Holland 创造的。 <ref name="holland:causal86">{{cite journal |last=Holland |first=Paul W. |title=Statistics and Causal Inference |journal=Journal of the American Statistical Association |volume=81 |issue=396 |year=1986 |pages=945–960 |jstor=2289064 |doi=10.1080/01621459.1986.10478354}}</ref> '''<font color="#ff8000"> 潜在结果框架 Potential Outcomes Framework</font>'''最初是由 Jerzy Neyman 在他 1923 年的硕士论文中提出的,<ref name="neyman:masters">Neyman, Jerzy. ''Sur les applications de la theorie des probabilites aux experiences agricoles: Essai des principes.'' Master's Thesis (1923). Excerpts reprinted in English, Statistical Science, Vol. 5, pp.&nbsp;463–472. (Dorota Dabrowska, and T. P. Speed, Translators.)</ref>尽管他只在完全随机实验的背景下讨论了它。 <ref name="Jasa1">{{cite journal |last=Rubin |first=Donald |year=2005 |title=Causal Inference Using Potential Outcomes |journal=Journal of the American Statistical Association|volume=100 |issue=469 |pages=322–331 |doi=10.1198/016214504000001880 }}</ref>鲁宾将其扩展为在观察性和实验性研究中思考因果关系的一般框架。<ref name="sekhon" /><br />
<br />
== 思想介绍 ==<br />
鲁宾因果模型是基于潜在结果的想法。例如,如果一个人上过大学,他在 40 岁时会有特定的收入,而如果他没有上过大学,他在 40 岁时会有不同的收入。为了衡量这个人上大学的因果效应,我们需要比较同一个人在两种不同的未来中的结果。由于不可能同时看到两种潜在结果,因此总是缺少其中一种潜在结果。这种困境就是“因果推理的基本问题”。<br />
<br />
由于因果推理的根本问题,无法直接观察到单元级别的因果效应。然而,随机实验允许估计人口水平的因果效应。<ref name=":01">{{cite journal |last=Rubin |first=Donald |title=Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies |journal=Journal of Educational Psychology|volume=66 |issue=5 |year=1974 |pages=688–701 [p. 689] |doi=10.1037/h0037350 }}</ref>随机实验将人们随机分配到对照组:大学或非大学。由于这种随机分配,各组(平均)相等,40 岁时的收入差异可归因于大学分配,因为这是各组之间的唯一差异。然后可以通过计算处理(上大学)和对照(非上大学)样本之间的平均值差异来获得平均因果效应(也称为平均处理效应)的估计值。<br />
<br />
然而,在许多情况下,由于伦理或实际问题,随机实验是不可能的。在这种情况下,存在非随机分配机制。上大学的例子就是这种情况:人们不是随机分配上大学的。相反,人们可能会根据他们的经济状况、父母的教育等来选择上大学。已经开发了许多用于因果推断的统计方法,例如倾向得分匹配。这些方法试图通过寻找类似于处理单元的控制单元来纠正分配机制。<br />
== 样例介绍 ==<br />
假设乔正在参与 FDA 对一种新的高血压药物的测试。如果我们是无所不知的,我们就会知道乔在治疗组和控制组下的结果。我们想要探究的因果效应,或者说治疗效果,就是指这两种潜在结果之间的差异。<br />
{| class="wikitable"<br />
!subject<br />
!Yt(u)<br />
!Yc (u)<br />
!Yt (u)-Yc (u)<br />
|-<br />
!乔<br />
|130<br />
|135<br />
|−5<br />
|}<br />
Y<sub>t</sub>(u) 表示如果乔服用了这种新药物之后对应的血压。一般来说,这个符号表示在个体 u 上的实施治疗 t 的潜在结果。类似地,Yc (u)是在个体 u 上的不做治疗(控制 )c 的潜在结果,即Yc (u)表示乔不吃这种新药物时对应的血压。则在这种情况下,Yt (u)-Yc (u)也就是服用这种新药物对乔的血压的因果效应。<br />
<br />
从这个表格中我们只知道对乔的因果效应。研究中的其他人如果服用新药,血压可能会升高。然而,不管其他受试者的因果效应如何,我们可以得出结论,对于乔来说,相比于他没有服用新药的情况,服用该药,他的血压会降低。<br />
<br />
考虑更多的病患样本<br />
{| class="wikitable"<br />
!subject<br />
!Yt(u)<br />
!Yc (u)<br />
!Yt (u)-Yc (u)<br />
|-<br />
!乔<br />
|130<br />
|135<br />
|−5<br />
|-<br />
!玛丽<br />
|140<br />
|150<br />
|−10<br />
|-<br />
!莎莉<br />
|135<br />
|125<br />
|10<br />
|-<br />
!鲍勃<br />
|135<br />
|150<br />
|−15<br />
|}<br />
每个实验对象的因果效应是不同的。从该表中可知乔,玛丽和鲍勃的因果效应为负值,说明药物仅对乔,玛丽和鲍勃起作用。他们服用这种药物后的血压比没有服用这种药物时的血压要低。另一方面,对于莎莉 来说,这种药物会导致血压升高。<br />
<br />
为了让一个潜在的结果有意义,它必须是可测试的,至少是先验的。例如,如果乔在任何情况下都没有办法获得新药,那么他就不可能获得效应。这永远不可能发生在乔身上。如果不能观察到效应,即使在理论上,那么治疗对乔的血压的因果效应也不能确定。<br />
<br />
== 没有干预就没有因果关系 ==<br />
新药的因果效应是明确定义的,因为它是两种可能发生的潜在结果的简单差异。在这种情况下,我们(或其他事物)可以干预世界,至少在概念上是这样,因此可能会发生不同的事。<br />
<br />
如果永远不可能发生其中一种潜在结果,那么这种因果效应的定义就会变得更加棘手。例如,乔的身高对他的体重有什么因果关系?这似乎与我们的其他示例相似。我们只需要比较两个潜在的结果:乔 在处理下的体重(处理被定义为增高3英寸)和 乔 在控制下的体重(控制被定义为他当前的身高)。<br />
<br />
问题在于:我们无法增加乔的身高。没有办法观察如果乔更高,他的体重会是多少,因为我们没有办法干预乔的身高从而让他变得更高,这就让研究乔的身高和体重的因果关系变得没有意义。因此有一个口号:没有干预就没有因果关系。<br />
<br />
== 个体处理稳定性假设 (SUTVA) ==<br />
我们要求“对一个个体 [潜在结果] 的观察不应受到其他个体的特定处理分配的影响”(Cox 1958,第 2.4 节)。这被称为个体处理稳定性假设(SUTVA),它超越了独立性的概念。<br />
<br />
在我们的例子中,乔 的血压不应该取决于 玛丽 是否接受了药物。但如果真的发生了呢?假设乔和玛丽住在同一所房子里,玛丽总是做饭。这种药物会导致玛丽渴望咸的食物,所以如果她服用这种药物,她会用比其他情况下更多的盐来烹饪。高盐饮食会增加乔的血压。因此,乔的血压结果将同时取决于他接受的处理和玛丽接受的处理。<br />
<br />
在不满足SUTVA的情况下,因果推断会更加困难。我们可以通过考虑更多的处理来解释相关的观察结果。我们通过考虑 玛丽 是否接受处理来创建 4 个处理。<br />
{| class="wikitable" align="center"<br />
!主题||乔 = c,玛丽 = t||乔 = t,玛丽 = t||乔 = c,玛丽 = c||乔 = t,玛丽 = c<br />
|-<br />
!乔||140||130||125||120<br />
|}<br />
回想一下,因果效应被定义为两个潜在结果之间的差异。在这种情况下,存在多种因果效应,因为存在两个以上的潜在结果。一是玛丽接受处理时药物对乔的因果效应{130,140}。另一个是当玛丽没有接受处理时对乔的因果效应{120,125}。第三是在乔没有得到处理的情况下,玛丽的处理对乔的因果效应{125,140}。玛丽 接受的处理对 乔 的因果影响比 乔 接受的处理对 乔 的影响更大,而且是相反的方向。<br />
<br />
通过以这种方式考虑更多潜在结果,我们可以使SUTVA成立。但是,如果 乔 以外的任何个体都依赖于 玛丽,那么我们必须考虑进一步的潜在结果。依赖个体的数量越多,我们必须考虑的潜在结果就越多,计算也变得越复杂(考虑对不同的20个人进行的实验,每个人的处理状态都会影响其他人的结果)。为了(轻松)估计单一处理相对于对照的因果效应,SUTVA 应该成立。<br />
<br />
== 分配机制 ==<br />
分配机制,即给个体分配处理(或者治疗)的方法,影响平均因果效应的计算。换句话说,当把一个接受处理的组和一个没有接受处理的组进行比较时,我们需要知道(或者做出一个假设)为什么某些人被分配到处理组,而其他人没有。<br />
<br />
一种分配机制是随机化。对于每个受试者,我们可以抛硬币来确定她是否接受处理。在最简单的情况下,这种分配是随机的(就像在临床试验中一样) ,而且不会混淆,因为分配并不依赖于潜在的结果。<br />
<br />
如果我们希望五个受试者接受处理,我们可以将处理分配给我们从组里里挑选出来的前五个名字。当我们随机分配处理时,我们可能会得到不同的答案。<br />
<br />
另一种分配机制是非随机化的,如果所有接受处理的个体都是因为他们最有可能受益而接受处理,那么处理结果和对照组之间的直接比较不能代表处理的因果效应。<br />
<br />
从以下这个例子来理解分配机制,假设我们同时知道每个个体接受处理和未接受处理的真实潜在结果是什么:<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)||Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||115||15<br />
|-<br />
!玛丽||120||125||−5<br />
|-<br />
!莎莉||100||125||−25<br />
|-<br />
!鲍勃||110||130||−20<br />
|-<br />
!詹姆士||115||120||−5<br />
|-<br />
!平均||115||123||−8<br />
|}可以计算得出真实的平均因果效应是-8。但实际上每个人的因果效应各异,都不会等于这个平均值。而且在现实生活中个体因果效应通常是未知的,因为没办法让一个个体即接受处理又不接受处理,而能同时得到两种情况下的潜在结果。但是在随机分配处理后,我们可以对多个人中接受处理的和未接受处理的潜在结果分别求平均,得到人群的平均因果效应:<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)||Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||?||?<br />
|-<br />
!玛丽||120||?||?<br />
|-<br />
!莎莉||?||125||?<br />
|-<br />
!鲍勃||?||130||?<br />
|-<br />
!詹姆士||115||?||?<br />
|-<br />
!平均||121.66||127.5||−5.83<br />
|}<br />
<br />
当然,不同的随机分配接受处理个体,产生的平均因果效应的估计值也不同:<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)<br />
!Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||?||?<br />
|-<br />
!玛丽||120||?||?<br />
|-<br />
!莎莉||100||?||?<br />
|-<br />
!鲍勃||?||130||?<br />
|-<br />
!詹姆士||?||120||?<br />
|-<br />
!平均||116.67||125||−8.33<br />
|}平均因果效应会有所不同,因为我们的样本很小并且反馈效应的方差很大。如果样本较大且方差较小,则无论随机分配给处理的特定单位如何,平均因果效应将更接近真实的平均因果效应。<br />
<br />
以上是随机分配机制的结果。再看一种可能不太合理的分配机制,假设仅将处理分配给所有男性。<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)<br />
!Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||?||?<br />
|-<br />
!鲍勃||110||?||?<br />
|-<br />
!詹姆士||105||?||?<br />
|-<br />
!玛丽||?||130||?<br />
|-<br />
!莎莉||?||125||?<br />
|-<br />
!苏茜||?||135||?<br />
|-<br />
!平均||115||130||−15<br />
|}虽然在不区分性别的情况下可以得到一个平均因果效应值,但是在这种分配机制下,因为没有接受处理的女性,所以无法单独对女性受试者确定她们的平均因果效应。此时对女性受试者而言,她们接受处理的概率为0,而为了能对受试者做出因果效应的任何推断,受试者接受处理的概率必须大于 0 且小于 1(Positivity 假设)。<br />
<br />
== 另见 ==<br />
<br />
* [[鲁宾因果框架]]<br />
<br />
== 编者推荐 ==<br />
'''集智俱乐部文章'''<br />
<br />
[https://mp.weixin.qq.com/s/vFDu-g2qy-sUfIl8EhrMJg Donald Rubin的因果推断学术贡献:超出统计学范畴的划时代影响]<br />
<br />
[https://swarma.org/?p=29559 因果推断的潜在结果框架在实验性研究的应用 | 周日直播·因果科学读书会 | 集智俱乐部 (swarma.org)]<br />
<br />
[https://swarma.org/?p=24656 两套因果框架深度剖析:潜在结果模型与结构因果模型 | 因果科学读书会 | 集智俱乐部 (swarma.org)]<br />
<br />
'''集智课程'''<br />
<br />
[https://campus.swarma.org/course/3527 因果科学读书会第三季:因果+X (swarma.org)]<br />
<br />
“因果”并不是一个新概念,而是一个已经在多个学科中使用了数十年的分析技术。通过前两季的分享,我们主要梳理了因果科学在计算机领域的前沿进展。如要融会贯通,我们需要回顾数十年来在社会学、经济学、医学、生物学等多个领域中,都是使用了什么样的因果模型、以什么样的范式、解决了什么样的问题。我们还要尝试进行对比和创新,看能否以现在的眼光,用其他的模型,为这些研究提供新的解决思路。<br />
<br />
“因果+X”就是要让因果真正地应用于我们的科学研究中,不管你是来自计算机、数理统计领域,还是社会学、经济学、管理学领域,还是医学、生物学领域,我们希望共同探究出因果研究的范式,真正解决因果的多学科应用问题,乃至解决工业界的问题。<br />
<br />
== 参考文献 ==<br />
<references /></div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E6%BD%9C%E5%9C%A8%E7%BB%93%E6%9E%9C&diff=31643潜在结果2022-06-07T17:29:00Z<p>Aceyuan:审校</p>
<hr />
<div>此词条由因果科学读书会词条梳理志愿者我是猫(74989)翻译审校,未经专家审核,带来阅读不便,请见谅<br />
<br />
== 概念来源 ==<br />
潜在结果最初的提出是在Neyman的论文[1]中,但是这篇文章只在随机对照试验中使用了潜在结果的概念,且直到1990年翻译成英文后才为人所知。Rubin在他1974年的论文中也提出了潜在结果的概念,并将这个概念推广到了观察性数据中[2],真正开启了统计学界对因果推断的广泛研究。<br />
<br />
何为潜在结果?又如何基于潜在结果定义因果?假设我们关心某个变量A(例如,在某个时间点是否服用阿莫西林,A=1是服用,A=0是没有服用)对Y(服用后三小时的是否还感冒,Y=1表示感冒,Y=0表示没有感冒)的因果关系。那么我们观察到的某个个体就存在两个“潜在”的状态:一个是如果他服药,他三小时后是否感冒,不妨记作Y(1);另一个如果他没有服药,他三小时后是否感冒,不妨记作Y(0)。这里Y(1)和Y(0)就是潜在结果。(注意,在实际中,Y(1)和Y(0)这二者中只有一个可以被观察到。另外,严格地说,此处实际上做了“个体处理性稳定性”即SUTVA的假设)那么对这个人,就可能有以下四种情况:<br />
<br />
a) Y(0)=0, Y(1)=0。即不论吃不吃药,这个人在三小时后均不会感冒。<br />
<br />
b) Y(0)=1, Y(1)=1。即不论吃不吃药,这个人在三小时后均会感冒。<br />
<br />
c) Y(0)=1, Y(1)=0。即此人如果不吃药,三小时后会感冒,但是如果吃药,三小时后不会感冒。<br />
<br />
d) Y(0)=0, Y(1)=1。即此人如果不吃药,三小时后不会感冒,但是如果吃药,三小时后会感冒。<br />
<br />
在a和b两种情况下,Y(1)=Y(0),即吃不吃药不会影响三小时后是否感冒的状态,这种情况下我们说吃药对三小时后是否感冒没有因果作用,相反,在c和d两种情况下,Y(1)≠Y(0),这种情况下我们说吃药对三小时后是否感冒有因果作用。使用潜在结果,我们便可以方便地定义感兴趣的因果作用,例如平均因果效应E[Y(1)-Y(0)],这个量代表了在一个群体中,如果每一个人都采取某种处理和都不接受处理相比,这两种情况下平均意义上的结果差值。<br />
<br />
使用潜在结果我们或许可以理解为什么人们不会认为“太阳升起是因为鸡打鸣”,因为根据我们的常识,如果某天鸡不打鸣(或许是因为生病或劳累),太阳仍然会照常升起。<br />
<br />
'''与此同时,也诞生了<font color="#ff8000">潜在结果框架,有时也称为</font>鲁宾因果模型 Rubin Causal Model (RCM)''' ,'''Neyman-Rubin 因果模型'''<ref name="sekhon">{{cite book |last=Sekhon |first=Jasjeet |chapter=The Neyman–Rubin Model of Causal Inference and Estimation via Matching Methods |title=The Oxford Handbook of Political Methodology |year=2007 |chapter-url=http://sekhon.berkeley.edu/papers/SekhonOxfordHandbook.pdf }}</ref>。它是一种基于潜在结果框架的因果统计分析方法,以Donald Rubin的名字命名。“鲁宾因果模型”这个名字最早是由 Paul W. Holland 创造的。 <ref name="holland:causal86">{{cite journal |last=Holland |first=Paul W. |title=Statistics and Causal Inference |journal=Journal of the American Statistical Association |volume=81 |issue=396 |year=1986 |pages=945–960 |jstor=2289064 |doi=10.1080/01621459.1986.10478354}}</ref> '''<font color="#ff8000"> 潜在结果框架 Potential Outcomes Framework</font>'''最初是由 Jerzy Neyman 在他 1923 年的硕士论文中提出的,<ref name="neyman:masters">Neyman, Jerzy. ''Sur les applications de la theorie des probabilites aux experiences agricoles: Essai des principes.'' Master's Thesis (1923). Excerpts reprinted in English, Statistical Science, Vol. 5, pp.&nbsp;463–472. (Dorota Dabrowska, and T. P. Speed, Translators.)</ref>尽管他只在完全随机实验的背景下讨论了它。 <ref name="Jasa1">{{cite journal |last=Rubin |first=Donald |year=2005 |title=Causal Inference Using Potential Outcomes |journal=Journal of the American Statistical Association|volume=100 |issue=469 |pages=322–331 |doi=10.1198/016214504000001880 }}</ref>鲁宾将其扩展为在观察性和实验性研究中思考因果关系的一般框架。<ref name="sekhon" /><br />
<br />
== 思想介绍 ==<br />
鲁宾因果模型是基于潜在结果的想法。例如,如果一个人上过大学,他在 40 岁时会有特定的收入,而如果他没有上过大学,他在 40 岁时会有不同的收入。为了衡量这个人上大学的因果效应,我们需要比较同一个人在两种不同的未来中的结果。由于不可能同时看到两种潜在结果,因此总是缺少其中一种潜在结果。这种困境就是“因果推理的基本问题”。<br />
<br />
由于因果推理的根本问题,无法直接观察到单元级别的因果效应。然而,随机实验允许估计人口水平的因果效应。<ref name=":01">{{cite journal |last=Rubin |first=Donald |title=Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies |journal=Journal of Educational Psychology|volume=66 |issue=5 |year=1974 |pages=688–701 [p. 689] |doi=10.1037/h0037350 }}</ref>随机实验将人们随机分配到对照组:大学或非大学。由于这种随机分配,各组(平均)相等,40 岁时的收入差异可归因于大学分配,因为这是各组之间的唯一差异。然后可以通过计算处理(上大学)和对照(非上大学)样本之间的平均值差异来获得平均因果效应(也称为平均处理效应)的估计值。<br />
<br />
然而,在许多情况下,由于伦理或实际问题,随机实验是不可能的。在这种情况下,存在非随机分配机制。上大学的例子就是这种情况:人们不是随机分配上大学的。相反,人们可能会根据他们的经济状况、父母的教育等来选择上大学。已经开发了许多用于因果推断的统计方法,例如倾向得分匹配。这些方法试图通过寻找类似于处理单元的控制单元来纠正分配机制。<br />
== 概念定义 ==<br />
潜在结果:给定一个单元,和一系列动作,我们把一个“动作-单元”确定为一个潜在结果。“潜在(potential)”这个词表达的意思是我们并不总是能在现实中观察到这个结果(outcome),但原则上它们可能发生。<br />
<br />
== 样例介绍 ==<br />
假设乔正在参与 FDA 对一种新的高血压药物的测试。如果我们是无所不知的,我们就会知道乔在治疗组和控制组下的结果。我们想要探究的因果效应,或者说治疗效果,就是指这两种潜在结果之间的差异。<br />
{| class="wikitable"<br />
!subject<br />
!Yt(u)<br />
!Yc (u)<br />
!Yt (u)-Yc (u)<br />
|-<br />
!Joe<br />
|130<br />
|135<br />
|−5<br />
|}<br />
Yt(u) 表示如果Joe服用了这种新药物之后对应的血压。一般来说,这个符号表示在个体 u 上的实施治疗 t 的潜在结果。类似地,Yc (u)是在个体 u 上的不做治疗(控制 )c 的潜在结果,即Yc (u)表示Joe不吃这种新药物时对应的血压。则在这种情况下,Yt (u)-Yc (u)也就是服用这种新药物对Joe的血压的因果效应。<br />
<br />
从这个表格中我们只知道对Joe的因果效应。研究中的其他人如果服用新药,血压可能会升高。然而,不管其他受试者的因果效应如何,我们可以得出结论,对于Joe来说,相比于他没有服用新药的情况,服用该药,他的血压会降低。<br />
<br />
考虑更多的病患样本<br />
{| class="wikitable"<br />
!subject<br />
!Yt(u)<br />
!Yc (u)<br />
!Yt (u)-Yc (u)<br />
|-<br />
!Joe<br />
|130<br />
|135<br />
|−5<br />
|-<br />
!Mary<br />
|140<br />
|150<br />
|−10<br />
|-<br />
!Sally<br />
|135<br />
|125<br />
|10<br />
|-<br />
!Bob<br />
|135<br />
|150<br />
|−15<br />
|}<br />
每个实验对象的因果效应是不同的。从该表中可知Joe,Mary和Bob的因果效应为负值,说明药物仅对Joe,Mary和Bob起作用。他们服用这种药物后的血压比没有服用这种药物时的血压要低。另一方面,对于Sally 来说,这种药物会导致血压升高。<br />
<br />
为了让一个潜在的结果有意义,它必须是可测试的,至少是先验的。例如,如果Joe在任何情况下都没有办法获得新药,那么他就不可能获得效应。这永远不可能发生在Joe身上。如果不能观察到效应,即使在理论上,那么治疗对Joe的血压的因果效应也不能确定。<br />
<br />
== 没有干预就没有因果关系 ==<br />
新药的因果效应是明确定义的,因为它是两种可能发生的潜在结果的简单差异。在这种情况下,我们(或其他事物)可以干预世界,至少在概念上是这样,因此可能会发生不同的事。<br />
<br />
如果永远不可能发生其中一种潜在结果,那么这种因果效应的定义就会变得更加棘手。例如,Joe的身高对他的体重有什么因果关系?这似乎与我们的其他示例相似。我们只需要比较两个潜在的结果:Joe 在处理下的体重(处理被定义为增高3英寸)和 Joe 在控制下的体重(控制被定义为他当前的身高)。<br />
<br />
问题在于:我们无法增加Joe的身高。没有办法观察如果Joe更高,他的体重会是多少,因为我们没有办法干预Joe的身高从而让他变得更高,这就让研究Joe的身高和体重的因果关系变得没有意义。因此有一个口号:没有干预就没有因果关系。<br />
<br />
== 个体处理稳定性假设 (SUTVA) ==<br />
我们要求“对一个个体的 [潜在结果] 观察不应受到其他个体的特定处理分配的影响”(Cox 1958,第 2.4 节)。这被称为个体处理稳定性假设(SUTVA),它超越了独立性的概念。<br />
<br />
在我们的例子中,Joe 的血压不应该取决于 Mary 是否接受了药物。但如果真的发生了呢?假设Joe和Mary住在同一所房子里,Mary总是做饭。这种药物会导致Mary渴望咸的食物,所以如果她服用这种药物,她会用比其他情况下更多的盐来烹饪。高盐饮食会增加Joe的血压。因此,Joe的血压结果将同时取决于他接受的处理和Mary接受的处理。<br />
<br />
在不满足SUTVA的情况下,因果推断会更加困难。我们可以通过考虑更多的处理来解释相关的观察结果。我们通过考虑 Mary 是否接受处理来创建 4 个处理。<br />
{| class="wikitable" align="center"<br />
!主题||乔 = c,玛丽 = t||乔 = t,玛丽 = t||乔 = c,玛丽 = c||乔 = t,玛丽 = c<br />
|-<br />
!乔||140||130||125||120<br />
|}<br />
回想一下,因果效应被定义为两个潜在结果之间的差异。在这种情况下,存在多种因果效应,因为存在两个以上的潜在结果。一是Mary接受处理时药物对Joe的因果效应【130-140】。另一个是当Mary没有接受处理时对Joe的因果效应【120-125】。第三是在Joe没有得到处理的情况下,Mary的处理对Joe的因果效应【125-140】。Mary 接受的处理对 Joe 的因果影响比 Joe 接受的处理对 Joe 的影响更大,而且是相反的方向。<br />
<br />
通过以这种方式考虑更多潜在结果,我们可以使SUTVA成立。但是,如果 Joe 以外的任何个体都依赖于 Mary,那么我们必须考虑进一步的潜在结果。依赖个体的数量越多,我们必须考虑的潜在结果就越多,计算也变得越复杂(考虑对不同的20个人进行的实验,每个人的处理状态都会影响其他人的结果)。为了(轻松)估计单一处理相对于对照的因果效应,SUTVA 应该成立。<br />
<br />
== 分配机制 ==<br />
分配机制,即分配单位处理的方法,影响平均因果效应的计算。换句话说,当把一个接受治疗的组和一个没有接受治疗的组进行比较时,我们需要知道(或者做出一个假设)为什么某些人被分配到治疗组,而其他人没有。<br />
<br />
一种分配机制是随机化。对于每个受试者,我们可以抛硬币来确定她是否接受处理。在最简单的情况下,这种分配是随机的(就像在临床试验中一样) ,而且不会混淆,因为分配并不依赖于潜在的结果。<br />
<br />
如果我们希望五个受试者接受处理,我们可以将处理分配给我们从组里里挑选出来的前五个名字。当我们随机分配处理时,我们可能会得到不同的答案。<br />
<br />
另一种分配机制是非随机化的,如果所有接受治疗的个体都是因为他们最有可能受益而接受治疗,那么治疗结果和对照组之间的直接比较不能代表治疗的因果效应。<br />
<br />
从以下这个例子来理解分配机制,假设这个数据是真实的:<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)||Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||115||15<br />
|-<br />
!玛丽||120||125||−5<br />
|-<br />
!莎莉||100||125||−25<br />
|-<br />
!鲍勃||110||130||−20<br />
|-<br />
!詹姆士||115||120||−5<br />
|-<br />
!平均||115||123||−8<br />
|}真正的平均因果效应是 -8。但是对这些人的因果效应永远不会等于这个平均值。因果效应各不相同,因为它通常(总是未知(?))在现实生活中也是如此。在随机分配处理后,我们可以估计因果效应为:<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)||Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||?||?<br />
|-<br />
!玛丽||120||?||?<br />
|-<br />
!莎莉||?||125||?<br />
|-<br />
!鲍勃||?||130||?<br />
|-<br />
!詹姆士||115||?||?<br />
|-<br />
!平均||121.66||127.5||−5.83<br />
|}<br />
<br />
<br />
处理的不同随机分配产生对平均因果效应的不同估计。<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)<br />
!Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||?||?<br />
|-<br />
!玛丽||120||?||?<br />
|-<br />
!莎莉||100||?||?<br />
|-<br />
!鲍勃||?||130||?<br />
|-<br />
!詹姆士||?||120||?<br />
|-<br />
!平均||116.67||125||−8.33<br />
|}平均因果效应会有所不同,因为我们的样本很小并且反馈效应的方差很大。如果样本较大且方差较小,则无论随机分配给处理的特定单位如何,平均因果效应将更接近真实的平均因果效应。<br />
<br />
或者,假设该机制将处理分配给所有男性且仅分配给他们。<br />
{| class="wikitable" align="center"<br />
!主题||Y_{t}(u)||Y_{c}(u)<br />
!Y_{t}(u)-Y_{c}(u)<br />
|-<br />
!乔||130||?||?<br />
|-<br />
!鲍勃||110||?||?<br />
|-<br />
!詹姆士||105||?||?<br />
|-<br />
!玛丽||?||130||?<br />
|-<br />
!莎莉||?||125||?<br />
|-<br />
!苏茜||?||135||?<br />
|-<br />
!平均||115||130||−15<br />
|}在这种分配机制下,女性不可能接受处理,因此无法确定对女性受试者的平均因果效应。为了对受试者做出因果效应的任何推断,受试者接受治疗的概率必须大于 0 且小于 1。<br />
<br />
== 另见 ==<br />
<br />
* [[鲁宾因果框架]]<br />
<br />
== 编者推荐 ==<br />
'''集智俱乐部文章'''<br />
<br />
[https://mp.weixin.qq.com/s/vFDu-g2qy-sUfIl8EhrMJg Donald Rubin的因果推断学术贡献:超出统计学范畴的划时代影响]<br />
<br />
[https://swarma.org/?p=29559 因果推断的潜在结果框架在实验性研究的应用 | 周日直播·因果科学读书会 | 集智俱乐部 (swarma.org)]<br />
<br />
[https://swarma.org/?p=24656 两套因果框架深度剖析:潜在结果模型与结构因果模型 | 因果科学读书会 | 集智俱乐部 (swarma.org)]<br />
<br />
'''集智课程'''<br />
<br />
[https://campus.swarma.org/course/3527 因果科学读书会第三季:因果+X (swarma.org)]<br />
<br />
“因果”并不是一个新概念,而是一个已经在多个学科中使用了数十年的分析技术。通过前两季的分享,我们主要梳理了因果科学在计算机领域的前沿进展。如要融会贯通,我们需要回顾数十年来在社会学、经济学、医学、生物学等多个领域中,都是使用了什么样的因果模型、以什么样的范式、解决了什么样的问题。我们还要尝试进行对比和创新,看能否以现在的眼光,用其他的模型,为这些研究提供新的解决思路。<br />
<br />
“因果+X”就是要让因果真正地应用于我们的科学研究中,不管你是来自计算机、数理统计领域,还是社会学、经济学、管理学领域,还是医学、生物学领域,我们希望共同探究出因果研究的范式,真正解决因果的多学科应用问题,乃至解决工业界的问题。<br />
<br />
== 参考文献 ==<br />
<references /></div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%9B%A0%E6%9E%9C%E4%B9%8B%E6%A2%AF&diff=31642因果之梯2022-06-07T16:53:27Z<p>Aceyuan:审校补充</p>
<hr />
<div>此词条由因果科学读书会词条梳理志愿者我是猫(74989)编撰<br />
<br />
未经专家审核,带来阅读不便,请见谅。<br />
<br />
== 起源 ==<br />
因果关系对于人类感知和理解世界,采取行动以及理解自己起着核心作用。大约二十年前,计算机科学家 Judea Pearl 通过发现和系统地研究「因果阶梯」(Ladder of Causation),在理解因果关系方面取得了突破,该框架着重说明了观察、做事和想象的独特作用。为了纪念这一具有里程碑意义的发现,人们将其命名为「Pearl 因果层次结构」(Pearl Causal Hierarchy,PCH)。<br />
<br />
去年 7 月,来自哥伦比亚大学和斯坦福大学的四位研究者撰写了一篇关于 PCH 和因果推理的技术报告,从逻辑概率和推理图两个方面,对 PCH 进行了新颖全面的解读。<br />
<br />
首先介绍了 PCH 是如何从规范的因果机制集合(结构因果模型,SCM)中有机出现的。然后文章转向逻辑层面。该报告的第一个结果是因果层次定理(CHT),该定理表明 PCH 的三个层级从测度论的角度上来看几乎总是分离的。粗略地讲,CHT 表明一层的数据实际上不足以确定较高层的信息。由于在大多数实际情况下,科学家无法获得潜在因果机制的精确形式(只能访问他们生成的关于某些 PCH 层的数据),这促使研究者从图的角度来研究 PCH 内部的推理<br />
<br />
具体来说,研究者探索了一组因果推理方法,这些方法可以在给定 SCM 部分规格的情况下,将 PCH 的各层进行桥接,以进行推理。例如,当只有被动观察结果(第一层数据)可用时,你会推断将发生的情况会在环境(第二层语句)中遭到干预。研究者提出了一系列图模型,这些模型让科学家能够以认知上有意义且简约的方式来表示 SCM 的部分规格。<br />
<br />
最后,研究者探究了被称为“do算子“(do calculus)的推理系统,展示了在必要情况下,它足以实现 PCH 各层之间的推理。研究者表示:与 PCH 所描绘的人类经验的基本层面相联系是迈向创建下一代 AI 系统的关键一步,该系统将是安全、强大、与人类兼容并符合社会利益的。<br />
<br />
== 定义 ==<br />
[[文件:因果之梯.png]]<br />
<br />
图灵曾经提出了图灵测试来进行一个二元分类——人类和非人类。但Pearl是提出了一个三元分类:<br />
<br />
第一层级的梯子上站着的是机器人和动物,能够做的就是基于被动观察来做出预测。Pearl认为,目前为止我们的机器学习进展都还是在这一层级的,无论大家认为它有多么强大。<br />
<br />
第二层级的梯子上站着的是原始人类和婴儿,它们学会了有意图地去使用工具,对周遭环境进行干预。<br />
<br />
第三个层级上的底子上站着的是有较高智慧的人类,拥有反思的能力,能够在大脑中将真实的世界与虚构的世界进行对比。 <br />
<br />
在这三个层级上,能够提出和解决的问题是不同的:<br />
<br />
在第一个层级上,问题都是基于相关性的,比如:“我的肺部有很多焦油沉积,我未来患肺癌的概率是多少?”<br />
<br />
而在第二个层级上,就涉及到了对现实世界的干预,并预测干预结果,比如:“我现在已经吸烟三年了,如果我现在戒烟,我还会患肺癌吗?”<br />
<br />
第三个层级上,就是要构建一个虚拟世界,并将虚拟世界与现在进行对比,问题的答案就是对比的结果,比如“如果过去的三年我都没有吸烟,现在我还会患肺癌吗?” Pearl在数学上证明了,这三个层级之间是有着根本的区别的。<br />
<br />
因果关系之梯是Judea Pearl提出的一种分类法(也可以称为框架),该分类法回答了“因果推理主体可以做什么”这一问题。该问题的另一种表述是——“相较于不具备因果模型的生物,拥有因果模型的生物能推算出什么前者推算不出的东西”。这种分类法的好处在于,它绕过了关于因果论究竟为何物的漫长而徒劳的讨论,聚焦于具体的可回答的问题。<br />
<br />
其中因果关系之梯包括三个层级:关联(association)、干预(intervention)和反事实(counterfactual),分别对应逐级复杂的因果问题。下图为因果关系之梯的示意图。[[文件:因果机制集合及三层架构.png|无框|749x749像素]]<br />
<br />
== 第一层级:关联 ==<br />
在该层级中,主体通过观察寻找规律。这种观察是被动的,不对世界做出干涉,而是通过我们观察到的世界对问题做出回答。关联的例子有:<br />
<br />
* 猫头鹰通过观察老鼠的活动判断老鼠下一刻可能出现的位置<br />
* 计算机围棋程序通过对数百万棋谱的研究发现胜率高的走法<br />
<br />
典型问题:<br />
<br />
在关联层级,我们会问的典型问题是“如果我观察到......会怎样?”。<br />
<br />
例如,超市经理会问“购买牙膏的顾客同时购买牙线的可能性有多大?”。<br />
<br />
回答方法:<br />
<br />
统计学可以用于回答这类问题,我们可以利用收集到的历史数据计算P(牙线|牙膏)这一条件概率测算购买牙线和牙膏的关联程度。但是,统计学无法告诉我们事物之间的因果关系。<br />
<br />
== 第二层级:干预 ==<br />
干预涉及到主体对现状的主动改变,根据可以被观察到的世界回答问题。<br />
<br />
典型问题:<br />
<br />
* “如果我们实施......行动,将会怎样?” 例如:“如果我们把牙膏的价格翻倍,牙线的销售额将会怎样?”<br />
* “怎么做?” 例如:“我们应当如何定价以卖掉仓库里积压的牙膏?”<br />
<br />
回答方法:<br />
<br />
* 进行严格控制条件下的实验,例如网站通过AB test判断用户对页面颜色的偏好<br />
* 建立因果模型(DAG),结合数据进行预测<br />
<br />
== 第三层级:反事实 ==<br />
数据顾名思义就是事实,数据无法告诉我们在反事实或虚构的世界里会发生什么。在反事实世界里,观察到的事实被直截了当地否定了。然而,人类的思维却能可靠地、重复地进行这种寻求背后解释的推断。...... 这种能力彻底地区分了人类智能与动物智能,以及人类与模型盲版本的人工智能和机器学习。<br />
<br />
我们通过一个情景来理解反事实:<br />
<br />
假设乔在服用了药物D一个月后死亡,那么我们要关注的问题就是这种药物是否导致了他的死亡。为了回答这个问题,我们需要想象这样一种情况:假如乔在即将服药时改变了主意,他现在会活着吗?<br />
<br />
在反事实层级,我们需要回答与一个无法被观察的世界相关的问题。<br />
<br />
典型问题:<br />
<br />
* “假如我当时做了......会怎样?” 例如:“假如我们把牙膏的价格提高一倍,则之前买了牙膏的顾客仍然选择购买的概率是多少?”<br />
* “为什么?”<br />
<br />
回答方法:<br />
<br />
第二层级中的干预实验无法回答反事实问题,因为我们无法对过去的事实进行改变。<br />
<br />
通过结构因果模型(SCM),我们可以回答反事实问题。<br />
<br />
== 样例理解 ==<br />
为了更好的理解因果之梯得三层架构,我们可以通过Judea Pearl著作中所提的一个例子来理解三个层级概念。<br />
<br />
假设要将一个犯人进行枪决,需要经过下述流程:<br />
[[文件:Sample.png]]<br />
<br />
# 首先,需要法院发布处决犯人的命令<br />
# 行刑队队长收到法院命令后,对士兵A和士兵B发布处决指令<br />
# 士兵A或士兵B接到命令开枪<br />
<br />
我们假设士兵A和B只听队长的命令开枪,不会擅自开枪。此外,只要任一枪手开枪,犯人都会死亡。<br />
<br />
请考虑以下问题:<br />
<br />
# 不考虑自然死亡的情况,如果犯人死了,那么这是否意味着法院已下令处决犯人?<br />
# 假设我们发现士兵A射击了,它告诉了我们关于B的什么信息?<br />
# 如果士兵A决定按自己的意愿射击,而不等待队长的命令,情况会怎样?犯人会不会死?<br />
# 假设犯人现在已倒地身亡,从这一点我们可以得出结论:A射击了,B射击了,行刑队队长发出了指令,法院下了判决。但是,假如A决定不开枪,犯人是否还活着?<br />
<br />
上述问题中,1和2为关联层级的问题(一个事实告诉我们有关另一事实的什么信息),我们可以通过观察到的情况“犯人已死”、“士兵A射击”做出推理回答问题,得到法院下令和士兵B也射击了的回答。<br />
<br />
3为干预层级的问题,我们需要对现实世界做出调整(现实:士兵A只听队长命令,调整:士兵A按照自己的意愿射击),从而根据可以被观察到的世界回答问题。<br />
<br />
4为反事实层级的问题。在现实世界中,A已经开枪了,但我们需要了解另一个与现实相矛盾的世界(A没有开枪,其它情况不变)中犯人的情况。对于该问题,我们无法根据观察到的世界回答,也无法对现在的世界进行干涉从而回答该问题。但是,我们可以通过构建因果模型从而推理出问题的回答。<br />
<br />
== 推荐资料 ==<br />
参考链接:<br />
<br />
凉某人的笔记本 (cnblogs.com),<br />
<br />
Judea Pearl提出的“因果阶梯”到底是什么?哥大、斯坦福研究者60页文章详解该问题 (thepaper.cn)<br />
<br />
集智俱乐部文章<br />
<br />
[https://swarma.org/?p=28073 600+学者共攀因果之梯,因果科学风暴再升级! | 集智俱乐部 (swarma.org)]<br />
<br />
[https://swarma.org/?p=26056 周日直播丨攀登因果之梯第三阶:反事实推理及其应用分享 | 集智俱乐部 (swarma.org)]<br />
<br />
[https://swarma.org/?p=28945 构建因果引擎,创新科研范式——因果科学的学习路线图 | 集智俱乐部 (swarma.org)]<br />
<br />
集智课程<br />
<br />
[https://campus.swarma.org/course/1798 因果科学与 Causal AI 系列读书会 (swarma.org)]<br />
<br />
[https://campus.swarma.org/course/2460 因果科学与Causal AI读书会第二季 (swarma.org)]<br />
<br />
[https://campus.swarma.org/course/3527 因果科学读书会第三季:因果+X (swarma.org)]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%9B%A0%E6%9E%9C%E6%8E%A8%E6%96%AD%E5%BC%95%E6%93%8E&diff=31641因果推断引擎2022-06-07T16:37:19Z<p>Aceyuan:消除4.中可能的表达歧义</p>
<hr />
<div>此词条由因果科学读书会词条梳理志愿者我是猫(74989)编撰<br />
<br />
未经专家审核,带来阅读不便,请见谅。<br />
<br />
== 来源 ==<br />
科学研究是如何开展的?在大多数情况下,科研人员会花费大量时间学习前置知识和目标研究领域知识体系,这可能会耗费从小学开始到本科毕业的15年以上的时间;之后,科研人员会根据研究目标,调研相关方向的经典文献和前沿文献,寻找其中的可创新点;在提出猜想与假设后,根据现有知识和能力设计实验;通过实验收集到相关数据后,进行数据分析,以确定当前数据会支持还是驳回原有猜想;最后通过深入的论证,详述研究对本领域产生了什么样的贡献,给出了什么样的新知识。<br />
<br />
如果未来让“因果科学”的相关技术辅助科研人员的研究工作,甚至是指导科研工作方向或者改变科研的范式,会是什么样的呢?朱迪亚·珀尔教授在《为什么》一书中给出了一个“因果推断引擎”的蓝图,在这里又补充上了“因果发现引擎”在体系中应处的位置,构成了一个完整的“因果引擎”。因果发现引擎是一种从观测数据中发现出可能的因果模型的机器,它接受两种输入——假设和数据,产生的输出为一个或一簇因果模型。下图展示了“因果引擎”的概貌。[[文件:因果引擎.png|无|有框|因果推断引擎]]<br />
<br />
== 流程介绍 ==<br />
因果推断引擎是一种问题处理机器,它接收三种不同的输入——假设、问题和数据,并能够产生三种输出。<br />
<br />
第一种输出是“是/否”判断,用于判定在现有的因果模型下,假设我们拥有完美的、无限的数据,那么给定的问题在理论上是否有解。如果答案为“是”,则接下来推断引擎会生成一个被估量。这是一个数学公式,可以被理解为一种能从任何假设数据中生成答案的方法,只要这些数据是可获取的。最后,在推断引擎接收到数据输入后,它将用上述方法生成一个问题答案的实际估计值,并给出对该估计值的不确定性大小的统计估计。这种不确定性反映了样本数据集的代表性以及可能存在的测量误差或数据缺失。<br />
<br />
== '''因果引擎的十要素''' ==<br />
'''1. 知识'''<br />
<br />
知识指的是推理主体(Reasoning Agent)过去的经验,包括过去的观察、以往的行为、接受过的教育和文化习俗等所有被认为与目标问题有关的内容。“知识”周围的虚线框表示它仍隐藏在推理主体的思想中,尚未在模型中得到正式表达。<br />
<br />
例如,我们如何去判断和界定一个病人的存活期、药物对于病人生理机能的可能的影响、其他的何种因素极有可能影响到药物疗效等等信息,都属于这个部分。<br />
<br />
'''2. 假设'''<br />
<br />
科学研究总是要求我们给出简化的假设,这些假设也就是研究者在现有知识的基础上认为有必要明确表述出来的陈述。研究者所拥有的大部分知识都隐藏于他的大脑,只有假设能将其公之于世,也只有假设才能被嵌入模型。<br />
<br />
例如,我们可以根据已有知识判断出药物的作用机制与病人的性别是无关的,与病人的年龄是有关的,与病人的视力水平是无关的,与病人的身高是无关的,但与病人的体重是有关的等等。<br />
<br />
'''3. 因果模型'''<br />
<br />
因果模型有多种表现形式,包括因果图、结构方程、逻辑语句等。从构建因果图的角度来看,“因果关系”的定义就非常简单了:如果变量Y“听从于”变量X,并根据所“听到”的内容决定自己的值,那么变量X就是变量Y的一个因。这里的因果模型可以直接从假设进行构建,也可以从因果发现引擎中获得。<br />
<br />
例如,如果我们怀疑一位病人的存活期L“听从于”该病人是否服用了药物D,那么我们便可以称D为L的因,并在因果图里绘制一个从D到L的箭头。当然,关于D和L之间的关系问题的答案很可能还取决于其他变量,因而我们也必须将这些变量及其因果关系在因果图中表示出来。在这里,我们统一用Z来表示其他变量。<br />
<br />
'''4. 可验证的蕴涵'''<br />
<br />
以因果模型的路径来表示的变量之间的听从模式通常会导向数据中某种显而易见的模式或相关关系。这些模式可被用于测试模型,因此也被称为“可验证的蕴涵”(Testable Implications)。<br />
<br />
例如,将“D和L之间没有连接路径”翻译成统计学语言,就是“D和L相互独立”,也就是说,发现D的存在不会改变L发生的可能性。而如果实际数据与这一推断相抵触,那么我们就需要修改模型。<br />
<br />
对于因果发现引擎输出的模型,绝大多数情况下并不会发生这种抵触;但对于直接从假设构建的模型,这种抵触是可能发生的,这就需要在原有的模型里补充额外的模块,输入可验证的蕴涵和数据后,计算数据与模型假设的匹配程度。<br />
<br />
'''5. 问题'''<br />
<br />
向推理引擎提交的问题就是我们希望获得解答的科学问题,这一问题必须用因果词汇来表述。因果革命的主要成就之一就是确保了这一语言在科学上容易理解,同时在数学上精确严谨。<br />
<br />
例如,我们现在感兴趣的问题是“药物D对病人生存期L的影响是什么”,可以表述为“P(L|do(D))是什么”。<br />
<br />
'''6. 被估量'''<br />
<br />
被估量“Estimand”来自拉丁语,意思是“需要估计的东西”。它是我们从数据中估算出来的统计量。一旦这个量被估算出来,我们便可以用它来合理地表示问题的答案。虽然被估量的表现形式是一个概率公式,如P(L|D,Z)×P(Z),但实际上它是一种方法,可以让我们根据我们所掌握的数据类型回答因果问题(前提是推断引擎证实了这种数据类型就是我们需要的)。<br />
<br />
重要的是要认识到,与传统的统计学所提供的估计方法不同,在当前的因果模型下,无论我们收集到多少数据,有些问题可能仍然无法得到解答。这个被称作“可识别性”问题。例如,如果我们的模型显示D和L都依赖于第三变量Z(比如疾病的发展阶段),并且,如果我们没有任何方法可以测量Z的值,那么问题P(L|do(D))就无法得到解答。在这种情况下,收集数据完全就是浪费时间。相反,我们需要做的是回过头完善模型,具体方式则是输入新的科学知识,使我们可以估计Z的值,或者简化假设(注意,此处存在犯错的风险),例如假设Z对D的影响是可以忽略不计的。<br />
<br />
'''7. 数据'''<br />
<br />
数据可以被视作填充被估量的原料。这里我们一定要认识到,数据本身不具备表述因果关系的能力。数据告诉我们的只是数量信息,如P(L|D)或P(L|D,Z)的值。而被估量则能够告诉我们如何将这些统计量转化为一个表达式。基于模型假设,该表达式在逻辑上等价于我们所要回答的因果问题,比说P(L|do(D))。<br />
<br />
在传统的统计方法中,被估量就等同于有待解决的问题。例如,如果我们对存活期为L的人群中服用过药物D的病人的比例感兴趣,那么我们可以将这个问题简记为P(D|L)。该表达式的值也就是我们的被估量。这一表达式已经确切地说明了数据中的哪个概率有待被估计,而并不涉及任何因果知识。<br />
<br />
'''8. 统计估计'''<br />
<br />
直接将现有数据代入被估量就可以得到粗糙的估计值了。但需要注意的是,数据永远是从理论上无限的总体中抽取的有限样本。比如在我们所讨论的这个例子中,数据样本由我们筛选出来进行研究的病人组成,即使这种筛选是随机的,我们也无法避免根据样本测量的概率无法代表整个总体的相应概率的可能性。幸运的是,依靠机器学习领域所提供的先进技术,统计学科为我们提供了很多方法来应对这种不确定性,这些方法包括最大似然估计、倾向评分、置信区间、显著性检验等,使我们从已有数据可以得到更为“精准”的估计值。<br />
<br />
'''9. 估计值'''<br />
<br />
如果我们的模型是正确的且数据是充分的,那么我们就获得了这个待解决的因果问题的答案,比如“药物D使糖尿病患者Z的生存期L增加了30%,误差±20%"。这一答案可以被添加到我们的【知识】中。而如果这一答案与我们的预期不符,则很可能说明我们需要对因果模型做一些改进。<br />
<br />
'''10. 因果发现引擎'''<br />
<br />
因果发现引擎是促进整体因果引擎自动化执行的重要推手。因果发现引擎可以根据现有的大量数据得到一个因果图,以辅助因果推断引擎的执行。除此之外,对于部分因果发现的算法,可以输入已有的知识来改进解决。<br />
<br />
例如,通过药物研究组其他项目收集到的数据,我们可以对人的其他生理指标或背景因素之间的关系进行建模,例如家庭富裕程度、学历、婚姻状态、年龄、身高体重、血压等等多个变量之间的因果关系模型。通过这个模型,也可以指导我们如何去探究一个新药的有效性,确定要去测量哪些变量。<br />
<br />
这个流程图乍看起来很复杂,因而你可能会怀疑它是否确有必要。事实上,在日常生活中,我们总能用某种方法做出一些因果判断,而与此同时并没有意识到自己经历了如此复杂的推断过程,当然也不会诉诸计算概率和比例的数学工具。我们的因果直觉通常足以让我们应付日常生活乃至职业生活中的不确定性。但是如果我们想教一个笨拙的机器人借助因果思维来思考问题,或者如果我们正试图推动无法依靠直觉来指引的前沿科学的发展,那么这一经过精心设计的推断流程就很有必要了。<br />
<br />
朱迪亚·珀尔教授认为,对于因果关系方面的知识来说,数据没有任何发言权。例如,有关行动或干预结果的信息根本无法从原始数据中获得,这些信息只能从对照试验操作中收集。相比之下,如果拥有一个因果模型,我们就可以在大部分情况下从未经干预处理的数据中预测干预的结果了。<br />
<br />
因果模型所具备而数据挖掘和深度学习所缺乏的另一个优势就是适应性。被估量是在我们真正检查数据的特性之前仅仅根据因果模型计算出来的,这就使得因果推断引擎适应性极强,因为无论变量之间的数值关系如何,被估量都能适用于与定性模型适配的数据。<br />
<br />
在未来,也许同领域的科研工作者们可以共享一个大规模的因果图,代表了共有的知识和假设。要去探究一个新的变量与原有知识体系的关系时,可以直接截取与这个变量有关系的子图。之后通过因果引擎计算出被估量,依照被估量中出现的变量来设计实验。实验采集到数据后可以计算出估计值并得到结果。相关的实验数据还可以融入到已有的数据之后,供因果引擎来检查是否可以更新现有的领域大模型,如果出现了不一致性,就自动展现出来,成为新的研究点。<br />
<br />
这种因果引擎或许在未来可以帮助建立负责任、可解释的人工智能,还可以完全改变现有的科学范式,让领域知识变得明确和可共享,方便科研人员设计实验和得出更加准确的结果,引发科学界的“因果浪潮”。<br />
<br />
== 编者推荐 ==<br />
<br />
=== 集智俱乐部文章 ===<br />
[https://mp.weixin.qq.com/s/sfu_UneK8buxy2XO8Q3_sg 构建因果引擎,创新科研范式——因果科学的学习路线图]<br />
===集智课程===<br />
====[https://campus.swarma.org/course/3527 因果科学读书会第三季:因果+X]====<br />
“因果”并不是一个新概念,而是一个已经在多个学科中使用了数十年的分析技术。通过前两季的分享,我们主要梳理了因果科学在计算机领域的前沿进展。如要融会贯通,我们需要回顾数十年来在社会学、经济学、医学、生物学等多个领域中,都是使用了什么样的因果模型、以什么样的范式、解决了什么样的问题。我们还要尝试进行对比和创新,看能否以现在的眼光,用其他的模型,为这些研究提供新的解决思路。<br />
<br />
<br />
“因果+X”就是要让因果真正地应用于我们的科学研究中,不管你是来自计算机、数理统计领域,还是社会学、经济学、管理学领域,还是医学、生物学领域,我们希望共同探究出因果研究的范式,真正解决因果的多学科应用问题,乃至解决工业界的问题。</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%90%88%E6%88%90%E5%AF%B9%E7%85%A7&diff=29412合成对照2022-03-22T03:39:04Z<p>Aceyuan:翻译稿-1</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共492,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
The '''synthetic control method''' is a statistical method used to evaluate the effect of an intervention in [[comparative case study|comparative case studies]]. It involves the construction of a weighted combination of groups used as controls, to which the [[treatment group]] is compared.<ref>{{Cite journal|last=Abadie|first=Alberto|date=2021|title=Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects|url=https://www.aeaweb.org/articles?id=10.1257/jel.20191450|journal=Journal of Economic Literature|language=en|volume=59|issue=2|pages=391–425|doi=10.1257/jel.20191450|issn=0022-0515|doi-access=free}}</ref> This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike [[difference in differences]] approaches, this method can account for the effects of [[confounder]]s changing over time, by weighting the control group to better match the treatment group before the intervention.<ref name=he>{{cite journal|last1=Kreif|first1=Noémi|last2=Grieve|first2=Richard|last3=Hangartner|first3=Dominik|last4=Turner|first4=Alex James|last5=Nikolova|first5=Silviya|last6=Sutton|first6=Matt|title=Examination of the Synthetic Control Method for Evaluating Health Policies with Multiple Treated Units|journal=Health Economics|date=December 2016|volume=25|issue=12|pages=1514–1528|doi=10.1002/hec.3258|pmid=26443693|pmc=5111584}}</ref> Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups.<ref name=ajps>{{cite journal|last1=Abadie|first1=Alberto|authorlink1=Alberto Abadie|last2=Diamond|first2=Alexis|last3=Hainmueller|first3=Jens|title=Comparative Politics and the Synthetic Control Method|journal=American Journal of Political Science|date=February 2015|volume=59|issue=2|pages=495–510|doi=10.1111/ajps.12116}}</ref> It has been applied to the fields of [[political science]],<ref name=ajps/> [[health policy]],<ref name=he/> [[criminology]],<ref>{{cite journal|last1=Saunders|first1=Jessica|last2=Lundberg|first2=Russell|last3=Braga|first3=Anthony A.|last4=Ridgeway|first4=Greg|last5=Miles|first5=Jeremy|title=A Synthetic Control Approach to Evaluating Place-Based Crime Interventions|journal=Journal of Quantitative Criminology|date=3 June 2014|volume=31|issue=3|pages=413–434|doi=10.1007/s10940-014-9226-5}}</ref> and [[economics]].<ref>{{cite journal|last1=Billmeier|first1=Andreas|last2=Nannicini|first2=Tommaso|title=Assessing Economic Liberalization Episodes: A Synthetic Control Approach|journal=Review of Economics and Statistics|date=July 2013|volume=95|issue=3|pages=983–1001|doi=10.1162/REST_a_00324}}</ref><br />
<br />
The synthetic control method is a statistical method used to evaluate the effect of an intervention in comparative case studies. It involves the construction of a weighted combination of groups used as controls, to which the treatment group is compared. This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike difference in differences approaches, this method can account for the effects of confounders changing over time, by weighting the control group to better match the treatment group before the intervention. Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups. It has been applied to the fields of political science, health policy, criminology, and economics.<br />
<br />
【翻译】合成对照是一种统计方法,在比较案例研究中用于评估干预措施的效果。它使用多组数据的加权组合来构建对照组,使之与治疗组进行比较。基于这种比较,可用来估计在干预之后的时间里,假如没有对治疗组进行干预的情况下治疗组将如何发展。与双重差分方法(Difference in difference)不同,这种方法可以考虑混杂因素随时间变化的影响,通过调整对照组的加权组合,可以对干预之前的治疗组数据做更好的匹配。合成对照还有个优点是,它允许研究人员在多组候选数据中做系统性选择。它已应用于政治学、卫生政策、犯罪学和经济学等多个领域。<br />
<br />
The synthetic control method combines elements from [[Matching (statistics)|matching]] and [[difference-in-differences]] techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the [[Minimum wage in the United States|minimum wage]] in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in [[Philadelphia]] that were unaffected by a minimum wage raise,<ref name="CardKrueger">{{cite journal |last=Card |first=D. |authorlink=David Card |first2=A. |last2=Krueger |authorlink2=Alan Krueger |year=1994 |title=Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania |journal=[[American Economic Review]] |volume=84 |issue=4 |pages=772–793 |jstor=2118030 }}</ref> and studies that look at [[crime rates]] in southern cities to evaluate the impact of the [[Mariel boat lift]] on crime.<ref>{{cite journal |last=Card |first=D. |year=1990 |title=The Impact of the Mariel Boatlift on the Miami Labor Market |journal=[[Industrial and Labor Relations Review]] |volume=43 |issue=2 |pages=245–257 |doi=10.1177/001979399004300205 |url=http://arks.princeton.edu/ark:/88435/dsp016h440s46f }}</ref> The control group in this specific scenario can be interpreted as a [[Weighted arithmetic mean|weighted average]], where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
The synthetic control method combines elements from matching and difference-in-differences techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the minimum wage in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in Philadelphia that were unaffected by a minimum wage raise, and studies that look at crime rates in southern cities to evaluate the impact of the Mariel boat lift on crime. The control group in this specific scenario can be interpreted as a weighted average, where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
综合控制方法结合了匹配技术和差中差技术的要素。差异中的差异法是一种常用的政策评估工具,用于在总体水平上评估干预措施的效果(例如:。州、国家、年龄组别等)平均超过一组未受影响的单位。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,比较对象是紧邻州边境的费城,那边的快餐店没有受到提高最低工资的影响,以及研究南部城市的犯罪率来评估马里埃尔船只提升对犯罪率的影响。在这个特定的场景中,控制组可以被解释为一个加权平均数,其中一些单位实际上得到了零重量,而其他单位得到了相等的,非零重量。<br />
<br />
【翻译】合成对照方法结合了匹配方法和双重差分方法的技术要素。双重差分法也是一种常用的政策评估工具,通过比较被干预单元和未被干预单元在总体水平上(例如:州、国家、年龄组别等)的均值差异来评估政策干预效果。著名的案例包括新泽西州快餐店提高最低工资政策对就业影响的研究,比较对象是在新泽西州边界另一侧,费城那边那些没受到该政策影响的快餐店;还有通过研究南部城市的犯罪率来评估马里埃尔移民潮如何影响犯罪的案例。在双重差分场景中,合成对照的控制组可被理解为一个加权平均,其中的一些单元相当于得到了零权重,而另外的一些单元则得到了非零权重(每个单元内的数据共享同一权重值)。<br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have ''J'' observations over ''T'' time periods where the relevant treatment occurs at time <math>T_{0}</math> where <math>T_{0}<T.</math> Let <br />
<br />
:<math>\alpha_{it}=Y_{it}-Y^N_{it},</math><br />
be the treatment effect for unit <math>i</math> at time <math>t</math>, where <math>Y^N_{it}</math> is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only <math>Y^N_{1t}</math>is not observed for <math>t>T_{0}</math>. We aim to estimate <math>(\alpha_{1T_{0}+1}......\alpha_{1T})</math>. <br />
<br />
:<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have J observations over T time periods where the relevant treatment occurs at time T_{0} where T_{0}<T. Let <br />
<br />
:\alpha_{it}=Y_{it}-Y^N_{it},<br />
be the treatment effect for unit i at time t, where Y^N_{it} is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only Y^N_{1t}is not observed for t>T_{0}. We aim to estimate (\alpha_{1T_{0}+1}......\alpha_{1T}). <br />
<br />
: <br />
<br />
综合控制方法试图为控制组的权重分配提供一种更加系统的方法。它通常使用一个相对较长的时间序列的结果之前的干预和估计权重的方式,控制组镜像治疗组尽可能接近。特别地,假设我们在 t 时间段有 j 观测值,在 t {0} < t 时相应的处理发生在 t {0} < t 时。让<br />
<br />
alpha _ { it } = y _ { it }-y ^ n _ { it } ,为时间 t 的单位 i 的治疗效果,其中 y ^ n _ { it }是未经治疗的结果。不失一般性,如果单位1接受相应的治疗,只有 y ^ n {1 t }没有观察到 t > t {0}。我们的目标是估计(alpha _ {1T _ {0} + 1} ... ... alpha _ {1T })。<br />
<br />
<br />
【翻译】合成对照方法试图用一种更加系统的方法为控制组分配权重。它通常用干预之前比较长一段时间内的多个时间序列作为输入数据,估计一组权重值使得这些输入数据加权的结果尽可能地拟合治疗组的时间序列数据,并将结果用作控制组时间序列数据。特别地,假设我们在T个时间段里共有J个观测量(单元),其中一个单元在T_{0}时间接受了治疗,T_{0}<T。让<br />
<br />
: \alpha_{it}=Y_{it}-Y^N_{it},<br />
<br />
为单元 i 的在时间 t 的治疗效果,其中 Y^N_{it} 是未经治疗的结果。不失一般性,如果指定单元1接受治疗,则只有单元1的数据 Y^N_{1t}在 t > T_{0} 时段是无法观测的。而我们的目标是要估计(\alpha_{1T_{0}+1} ... ... \alpha_{1T})估。<br />
<br />
<br />
Imposing some structure<br />
<br />
:<math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math><br />
and assuming there exist some optimal weights <math>w_2, \ldots, w_J</math> such that<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math><br />
<br />
for <math>t\leqslant T_{0}</math>, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
: <math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math><br />
for <math>t>T_{0}</math>. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<ref name=":0">{{cite journal |last=Abadie |first=A. |authorlink=Alberto Abadie |first2=A. |last2=Diamond |first3= J. |last3=Hainmüller |year=2010 |title=Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California's Tobacco Control Program |journal=[[Journal of the American Statistical Association]] |volume=105 |issue=490 |pages=493–505 |doi=10.1198/jasa.2009.ap08746 }}</ref><br />
:<br />
<br />
:Imposing some structure Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}<br />
<br />
and assuming there exist some optimal weights w_2, \ldots, w_J such that <br />
<br />
:Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}<br />
for t\leqslant T_{0}, the synthetic controls approach suggests using these weights to estimate the counterfactual<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}<br />
for t>T_{0}. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<br />
<br />
:<br />
强加一些结构<br />
: y ^ n { it } = delta { t } + theta { t } z { i } + lambda { t } mu { i } + varepsilon { it }<br />
假设存在一些最优权重 w _ 2,ldots,w _ j 使得<br />
:y {1 t } = Sigma ^ j { j = 2} w { j } y { jt }<br />
对于 t _ {0} ,综合控制方法建议使用这些权重来估计反事实<br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}<br />
<br />
for t>T_{0}.因此,在一定的正则性条件下,这种权重可以作为利息处理效果的估计量。本质上,该方法采用了匹配的思想,并利用训练数据进行干预前的权重设置,从而得到干预后的相应控制。<br />
:<br />
<br />
<br />
<br />
【翻译】<br />
<br />
强加一些结构<br />
<br />
: <math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math><br />
对于<math>t\leqslant T_{0}</math>,假设存在一些最优权重<math>w_2, \ldots, w_J</math>,使得<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math><br />
而对于<math>t>T_{0}</math>,合成对照方法建议使用这些权重来做出反事实估计<br />
:<math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math> <br />
因此,在一定的正则性条件下,此类权重可以作为我们所关心的治疗效果的估计量。本质上,该方法基于匹配的思想,利用干预前的数据训练得到加权组合的控制组,进而可以对干预后的控制组数据进行推断。<ref name=":0" /> <br />
<br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth,<ref>{{cite journal |last=Cavallo |first=E. |first2=S. |last2=Galliani |first3=I. |last3=Noy |first4=J. |last4=Pantano |year=2013 |title=Catastrophic Natural Disasters and Economic Growth |journal=[[Review of Economics and Statistics]] |volume=95 |issue=5 |pages=1549–1561 |doi=10.1162/REST_a_00413 |url=http://www.economics.hawaii.edu/research/workingpapers/WP_10-6.pdf }}</ref> and studies linking political murders to house prices.<ref>{{cite journal |last=Gautier |first=P. A. |first2=A. |last2=Siegmann |first3=A. |last3=Van Vuuren |year=2009 |title=Terrorism and Attitudes towards Minorities: The effect of the Theo van Gogh murder on house prices in Amsterdam |journal=[[Journal of Urban Economics]] |volume=65 |issue=2 |pages=113–126 |doi=10.1016/j.jue.2008.10.004 }}</ref> <br />
<!-- THE CITATION AT THE END OF THIS SENTENCE IS FOR A PAPER ABOUT "synthetic cohort models" (a.k.a. "pseudo-panel approach," using repeated cross-sections), WHICH IS NOT THE SAME AS "synthetic control": Yet, despite its intuitive appeal, it may be the case that synthetic controls could suffer from significant finite sample biases.<ref>{{cite journal |last=Devereux |first=P. J. |year=2007 |title=Small-sample bias in synthetic cohort models of labor supply |journal=[[Journal of Applied Econometrics]] |volume=22 |issue=4 |pages=839–848 |doi=10.1002/jae.938 }}</ref> --><br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth, and studies linking political murders to house prices. <br />
<br />
<br />
综合控制已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
【翻译】<br />
<br />
合成对照已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
==References==<br />
{{Reflist|30em}}<br />
<br />
[[Category:Design of experiments]]<br />
[[Category:Statistical methods]]<br />
[[Category:Observational study]]<br />
[[Category:Econometric modeling]]<br />
<br />
Category:Design of experiments<br />
Category:Statistical methods<br />
Category:Observational study<br />
Category:Econometric modeling<br />
<br />
类别: 实验设计类别: 统计方法类别: 观察性研究类别: 计量经济模型<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Synthetic control method]]. Its edit history can be viewed at [[合成对照/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%90%88%E6%88%90%E5%AF%B9%E7%85%A7&diff=29211合成对照2022-03-18T02:22:06Z<p>Aceyuan:Draft 1</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共492,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
The '''synthetic control method''' is a statistical method used to evaluate the effect of an intervention in [[comparative case study|comparative case studies]]. It involves the construction of a weighted combination of groups used as controls, to which the [[treatment group]] is compared.<ref>{{Cite journal|last=Abadie|first=Alberto|date=2021|title=Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects|url=https://www.aeaweb.org/articles?id=10.1257/jel.20191450|journal=Journal of Economic Literature|language=en|volume=59|issue=2|pages=391–425|doi=10.1257/jel.20191450|issn=0022-0515|doi-access=free}}</ref> This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike [[difference in differences]] approaches, this method can account for the effects of [[confounder]]s changing over time, by weighting the control group to better match the treatment group before the intervention.<ref name=he>{{cite journal|last1=Kreif|first1=Noémi|last2=Grieve|first2=Richard|last3=Hangartner|first3=Dominik|last4=Turner|first4=Alex James|last5=Nikolova|first5=Silviya|last6=Sutton|first6=Matt|title=Examination of the Synthetic Control Method for Evaluating Health Policies with Multiple Treated Units|journal=Health Economics|date=December 2016|volume=25|issue=12|pages=1514–1528|doi=10.1002/hec.3258|pmid=26443693|pmc=5111584}}</ref> Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups.<ref name=ajps>{{cite journal|last1=Abadie|first1=Alberto|authorlink1=Alberto Abadie|last2=Diamond|first2=Alexis|last3=Hainmueller|first3=Jens|title=Comparative Politics and the Synthetic Control Method|journal=American Journal of Political Science|date=February 2015|volume=59|issue=2|pages=495–510|doi=10.1111/ajps.12116}}</ref> It has been applied to the fields of [[political science]],<ref name=ajps/> [[health policy]],<ref name=he/> [[criminology]],<ref>{{cite journal|last1=Saunders|first1=Jessica|last2=Lundberg|first2=Russell|last3=Braga|first3=Anthony A.|last4=Ridgeway|first4=Greg|last5=Miles|first5=Jeremy|title=A Synthetic Control Approach to Evaluating Place-Based Crime Interventions|journal=Journal of Quantitative Criminology|date=3 June 2014|volume=31|issue=3|pages=413–434|doi=10.1007/s10940-014-9226-5}}</ref> and [[economics]].<ref>{{cite journal|last1=Billmeier|first1=Andreas|last2=Nannicini|first2=Tommaso|title=Assessing Economic Liberalization Episodes: A Synthetic Control Approach|journal=Review of Economics and Statistics|date=July 2013|volume=95|issue=3|pages=983–1001|doi=10.1162/REST_a_00324}}</ref><br />
<br />
The synthetic control method is a statistical method used to evaluate the effect of an intervention in comparative case studies. It involves the construction of a weighted combination of groups used as controls, to which the treatment group is compared. This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike difference in differences approaches, this method can account for the effects of confounders changing over time, by weighting the control group to better match the treatment group before the intervention. Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups. It has been applied to the fields of political science, health policy, criminology, and economics.<br />
<br />
【翻译】合成对照方法是一种统计方法,用于评估比较案例研究中的干预措施的效果。它使用多组数据加权组合构建对照组,并与治疗组进行比较。这种比较被用来估计如果治疗组没有接受治疗会发生什么。与双重差分(Difference in difference)方法不同,这种方法可以纳入随时间变化的混杂因素的影响,通过调整对照组的加权系数,能更好地对干预前的治疗组数据进行匹配。合成对照的另一个优点是,它允许研究人员在多组候选数据中做系统性选择。它已应用于政治学、卫生政策、犯罪学和经济学等领域。<br />
<br />
The synthetic control method combines elements from [[Matching (statistics)|matching]] and [[difference-in-differences]] techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the [[Minimum wage in the United States|minimum wage]] in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in [[Philadelphia]] that were unaffected by a minimum wage raise,<ref name="CardKrueger">{{cite journal |last=Card |first=D. |authorlink=David Card |first2=A. |last2=Krueger |authorlink2=Alan Krueger |year=1994 |title=Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania |journal=[[American Economic Review]] |volume=84 |issue=4 |pages=772–793 |jstor=2118030 }}</ref> and studies that look at [[crime rates]] in southern cities to evaluate the impact of the [[Mariel boat lift]] on crime.<ref>{{cite journal |last=Card |first=D. |year=1990 |title=The Impact of the Mariel Boatlift on the Miami Labor Market |journal=[[Industrial and Labor Relations Review]] |volume=43 |issue=2 |pages=245–257 |doi=10.1177/001979399004300205 |url=http://arks.princeton.edu/ark:/88435/dsp016h440s46f }}</ref> The control group in this specific scenario can be interpreted as a [[Weighted arithmetic mean|weighted average]], where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
The synthetic control method combines elements from matching and difference-in-differences techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the minimum wage in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in Philadelphia that were unaffected by a minimum wage raise, and studies that look at crime rates in southern cities to evaluate the impact of the Mariel boat lift on crime. The control group in this specific scenario can be interpreted as a weighted average, where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
综合控制方法结合了匹配技术和差中差技术的要素。差异中的差异法是一种常用的政策评估工具,用于在总体水平上评估干预措施的效果(例如:。州、国家、年龄组别等)平均超过一组未受影响的单位。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,比较对象是紧邻州边境的费城,那边的快餐店没有受到提高最低工资的影响,以及研究南部城市的犯罪率来评估马里埃尔移民潮对犯罪率的影响。在这个特定的场景中,控制组可以被解释为一个加权平均数,其中一些单位实际上得到了零重量,而其他单位得到了相等的,非零重量。<br />
<br />
【翻译】合成对照方法结合了匹配技术和双重差分技术的要素。双重差分法是一种常用的政策评估工具,通过对未被影响的单元求平均,在总体水平上(例如:州、国家、年龄组别等)评估在被干预单元上的政策干预效果。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,通过比较它们与费城边境那边没有受到最低工资提高影响的快餐店,以及研究南部城市的犯罪率来评估马里埃尔船只提升对犯罪的影响。在双重差分场景中,合成对照的控制组可被理解为一个加权平均,其中的一些单元相当于得到了零权值,而另外的单元则得到了非零且相等的权值。<br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have ''J'' observations over ''T'' time periods where the relevant treatment occurs at time <math>T_{0}</math> where <math>T_{0}<T.</math> Let <br />
<br />
:<math>\alpha_{it}=Y_{it}-Y^N_{it},</math><br />
be the treatment effect for unit <math>i</math> at time <math>t</math>, where <math>Y^N_{it}</math> is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only <math>Y^N_{1t}</math>is not observed for <math>t>T_{0}</math>. We aim to estimate <math>(\alpha_{1T_{0}+1}......\alpha_{1T})</math>. <br />
<br />
:<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have J observations over T time periods where the relevant treatment occurs at time T_{0} where T_{0}<T. Let <br />
<br />
:\alpha_{it}=Y_{it}-Y^N_{it},<br />
be the treatment effect for unit i at time t, where Y^N_{it} is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only Y^N_{1t}is not observed for t>T_{0}. We aim to estimate (\alpha_{1T_{0}+1}......\alpha_{1T}). <br />
<br />
: <br />
<br />
综合控制方法试图为控制组的权重分配提供一种更加系统的方法。它通常使用一个相对较长的时间序列的结果之前的干预和估计权重的方式,控制组镜像治疗组尽可能接近。特别地,假设我们在 t 时间段有 j 观测值,在 t {0} < t 时相应的处理发生在 t {0} < t 时。让<br />
<br />
alpha _ { it } = y _ { it }-y ^ n _ { it } ,为时间 t 的单位 i 的治疗效果,其中 y ^ n _ { it }是未经治疗的结果。不失一般性,如果单位1接受相应的治疗,只有 y ^ n {1 t }没有观察到 t > t {0}。我们的目标是估计(alpha _ {1T _ {0} + 1} ... ... alpha _ {1T })。<br />
<br />
<br />
【翻译】合成对照方法试图用一种更加系统的方法为控制组分配权重。它通常采用干预之前相对比较长时间段内的多个时间序列,估计一组权值使得对这些时间序列的结果进行加权后得到的控制组时间序列尽可能去拟合治疗组的时间序列。特别地,假设我们在总共T个时间段里有J个观测量(单位),其中一个单位接受了治疗,相应的治疗发生在T_{0}时间段,且T_{0}<T。让<br />
<br />
: \alpha_{it}=Y_{it}-Y^N_{it},<br />
<br />
为单位 i 的在时间 t 的治疗效果,其中 Y^N_{it} 是未经治疗的结果。不失一般性,如果单位1接受了相应的治疗,只有在 t > T_{0} 时段的 Y^N_{1t} 没有被观察到。我们的目标是估计(\alpha_{1T_{0}+1} ... ... \alpha_{1T})。<br />
<br />
<br />
Imposing some structure<br />
<br />
:<math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math><br />
and assuming there exist some optimal weights <math>w_2, \ldots, w_J</math> such that<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math><br />
<br />
for <math>t\leqslant T_{0}</math>, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
: <math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math><br />
for <math>t>T_{0}</math>. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<ref name=":0">{{cite journal |last=Abadie |first=A. |authorlink=Alberto Abadie |first2=A. |last2=Diamond |first3= J. |last3=Hainmüller |year=2010 |title=Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California's Tobacco Control Program |journal=[[Journal of the American Statistical Association]] |volume=105 |issue=490 |pages=493–505 |doi=10.1198/jasa.2009.ap08746 }}</ref><br />
:<br />
<br />
:Imposing some structure Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}<br />
<br />
and assuming there exist some optimal weights w_2, \ldots, w_J such that <br />
<br />
:Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}<br />
for t\leqslant T_{0}, the synthetic controls approach suggests using these weights to estimate the counterfactual<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}<br />
for t>T_{0}. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<br />
<br />
:<br />
强加一些结构<br />
: y ^ n { it } = delta { t } + theta { t } z { i } + lambda { t } mu { i } + varepsilon { it }<br />
假设存在一些最优权重 w _ 2,ldots,w _ j 使得<br />
:y {1 t } = Sigma ^ j { j = 2} w { j } y { jt }<br />
对于 t _ {0} ,综合控制方法建议使用这些权重来估计反事实<br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}<br />
<br />
for t>T_{0}.因此,在一定的正则性条件下,这种权重可以作为利息处理效果的估计量。本质上,该方法采用了匹配的思想,并利用训练数据进行干预前的权重设置,从而得到干预后的相应控制。<br />
:<br />
<br />
<br />
【翻译】<br />
<br />
强加一些结构<br />
<br />
: <math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math><br />
对于<math>t\leqslant T_{0}</math>,假设存在一些最优权重<math>w_2, \ldots, w_J</math>,使得<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math><br />
而对于<math>t>T_{0}</math>,合成对照方法建议使用这些权重来估计反事实<br />
:<math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math> <br />
因此,在一定的正则性条件下,此类权重可以作为我们所关心的治疗效果的估计量。 本质上,该方法使用匹配的思想,并利用干预前的训练数据计算权重,进而能够计算干预后的相关控制组数据。<ref name=":0" /> <br />
<br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth,<ref>{{cite journal |last=Cavallo |first=E. |first2=S. |last2=Galliani |first3=I. |last3=Noy |first4=J. |last4=Pantano |year=2013 |title=Catastrophic Natural Disasters and Economic Growth |journal=[[Review of Economics and Statistics]] |volume=95 |issue=5 |pages=1549–1561 |doi=10.1162/REST_a_00413 |url=http://www.economics.hawaii.edu/research/workingpapers/WP_10-6.pdf }}</ref> and studies linking political murders to house prices.<ref>{{cite journal |last=Gautier |first=P. A. |first2=A. |last2=Siegmann |first3=A. |last3=Van Vuuren |year=2009 |title=Terrorism and Attitudes towards Minorities: The effect of the Theo van Gogh murder on house prices in Amsterdam |journal=[[Journal of Urban Economics]] |volume=65 |issue=2 |pages=113–126 |doi=10.1016/j.jue.2008.10.004 }}</ref> <br />
<!-- THE CITATION AT THE END OF THIS SENTENCE IS FOR A PAPER ABOUT "synthetic cohort models" (a.k.a. "pseudo-panel approach," using repeated cross-sections), WHICH IS NOT THE SAME AS "synthetic control": Yet, despite its intuitive appeal, it may be the case that synthetic controls could suffer from significant finite sample biases.<ref>{{cite journal |last=Devereux |first=P. J. |year=2007 |title=Small-sample bias in synthetic cohort models of labor supply |journal=[[Journal of Applied Econometrics]] |volume=22 |issue=4 |pages=839–848 |doi=10.1002/jae.938 }}</ref> --><br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth, and studies linking political murders to house prices. <br />
<br />
<br />
综合控制已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
【翻译】<br />
<br />
合成对照已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
==References==<br />
{{Reflist|30em}}<br />
<br />
[[Category:Design of experiments]]<br />
[[Category:Statistical methods]]<br />
[[Category:Observational study]]<br />
[[Category:Econometric modeling]]<br />
<br />
Category:Design of experiments<br />
Category:Statistical methods<br />
Category:Observational study<br />
Category:Econometric modeling<br />
<br />
类别: 实验设计类别: 统计方法类别: 观察性研究类别: 计量经济模型<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Synthetic control method]]. Its edit history can be viewed at [[合成对照/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%90%88%E6%88%90%E5%AF%B9%E7%85%A7&diff=29209合成对照2022-03-17T02:21:25Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共492,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
The '''synthetic control method''' is a statistical method used to evaluate the effect of an intervention in [[comparative case study|comparative case studies]]. It involves the construction of a weighted combination of groups used as controls, to which the [[treatment group]] is compared.<ref>{{Cite journal|last=Abadie|first=Alberto|date=2021|title=Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects|url=https://www.aeaweb.org/articles?id=10.1257/jel.20191450|journal=Journal of Economic Literature|language=en|volume=59|issue=2|pages=391–425|doi=10.1257/jel.20191450|issn=0022-0515|doi-access=free}}</ref> This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike [[difference in differences]] approaches, this method can account for the effects of [[confounder]]s changing over time, by weighting the control group to better match the treatment group before the intervention.<ref name=he>{{cite journal|last1=Kreif|first1=Noémi|last2=Grieve|first2=Richard|last3=Hangartner|first3=Dominik|last4=Turner|first4=Alex James|last5=Nikolova|first5=Silviya|last6=Sutton|first6=Matt|title=Examination of the Synthetic Control Method for Evaluating Health Policies with Multiple Treated Units|journal=Health Economics|date=December 2016|volume=25|issue=12|pages=1514–1528|doi=10.1002/hec.3258|pmid=26443693|pmc=5111584}}</ref> Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups.<ref name=ajps>{{cite journal|last1=Abadie|first1=Alberto|authorlink1=Alberto Abadie|last2=Diamond|first2=Alexis|last3=Hainmueller|first3=Jens|title=Comparative Politics and the Synthetic Control Method|journal=American Journal of Political Science|date=February 2015|volume=59|issue=2|pages=495–510|doi=10.1111/ajps.12116}}</ref> It has been applied to the fields of [[political science]],<ref name=ajps/> [[health policy]],<ref name=he/> [[criminology]],<ref>{{cite journal|last1=Saunders|first1=Jessica|last2=Lundberg|first2=Russell|last3=Braga|first3=Anthony A.|last4=Ridgeway|first4=Greg|last5=Miles|first5=Jeremy|title=A Synthetic Control Approach to Evaluating Place-Based Crime Interventions|journal=Journal of Quantitative Criminology|date=3 June 2014|volume=31|issue=3|pages=413–434|doi=10.1007/s10940-014-9226-5}}</ref> and [[economics]].<ref>{{cite journal|last1=Billmeier|first1=Andreas|last2=Nannicini|first2=Tommaso|title=Assessing Economic Liberalization Episodes: A Synthetic Control Approach|journal=Review of Economics and Statistics|date=July 2013|volume=95|issue=3|pages=983–1001|doi=10.1162/REST_a_00324}}</ref><br />
<br />
The synthetic control method is a statistical method used to evaluate the effect of an intervention in comparative case studies. It involves the construction of a weighted combination of groups used as controls, to which the treatment group is compared. This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike difference in differences approaches, this method can account for the effects of confounders changing over time, by weighting the control group to better match the treatment group before the intervention. Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups. It has been applied to the fields of political science, health policy, criminology, and economics.<br />
<br />
【翻译】合成对照方法是一种统计方法,用于评估比较案例研究中的干预措施的效果。它使用多组数据加权组合构建对照组,并与治疗组进行比较。这种比较被用来估计如果治疗组没有接受治疗会发生什么。与双重差分(Difference in difference)方法不同,这种方法可以纳入随时间变化的混杂因素的影响,通过调整对照组的加权系数,能更好地对干预前的治疗组数据进行匹配。合成对照的另一个优点是,它允许研究人员在多组候选数据中做系统性选择。它已应用于政治学、卫生政策、犯罪学和经济学等领域。<br />
<br />
The synthetic control method combines elements from [[Matching (statistics)|matching]] and [[difference-in-differences]] techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the [[Minimum wage in the United States|minimum wage]] in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in [[Philadelphia]] that were unaffected by a minimum wage raise,<ref name="CardKrueger">{{cite journal |last=Card |first=D. |authorlink=David Card |first2=A. |last2=Krueger |authorlink2=Alan Krueger |year=1994 |title=Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania |journal=[[American Economic Review]] |volume=84 |issue=4 |pages=772–793 |jstor=2118030 }}</ref> and studies that look at [[crime rates]] in southern cities to evaluate the impact of the [[Mariel boat lift]] on crime.<ref>{{cite journal |last=Card |first=D. |year=1990 |title=The Impact of the Mariel Boatlift on the Miami Labor Market |journal=[[Industrial and Labor Relations Review]] |volume=43 |issue=2 |pages=245–257 |doi=10.1177/001979399004300205 |url=http://arks.princeton.edu/ark:/88435/dsp016h440s46f }}</ref> The control group in this specific scenario can be interpreted as a [[Weighted arithmetic mean|weighted average]], where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
The synthetic control method combines elements from matching and difference-in-differences techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the minimum wage in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in Philadelphia that were unaffected by a minimum wage raise, and studies that look at crime rates in southern cities to evaluate the impact of the Mariel boat lift on crime. The control group in this specific scenario can be interpreted as a weighted average, where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
综合控制方法结合了匹配技术和差中差技术的要素。差异中的差异法是一种常用的政策评估工具,用于在总体水平上评估干预措施的效果(例如:。州、国家、年龄组别等)平均超过一组未受影响的单位。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,比较对象是紧邻州边境的费城,那边的快餐店没有受到提高最低工资的影响,以及研究南部城市的犯罪率来评估马里埃尔移民潮对犯罪率的影响。在这个特定的场景中,控制组可以被解释为一个加权平均数,其中一些单位实际上得到了零重量,而其他单位得到了相等的,非零重量。<br />
<br />
【翻译】合成对照方法结合了匹配技术和双重差分技术的要素。双重差分法是一种常用的政策评估工具,通过对未被影响的单元求平均,在总体水平上(例如:州、国家、年龄组别等)评估在被干预单元上的政策干预效果。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,通过比较它们与费城边境那边没有受到最低工资提高影响的快餐店,以及研究南部城市的犯罪率来评估马里埃尔船只提升对犯罪的影响。在双重差分场景中,合成对照的控制组可被理解为一个加权平均,其中的一些单元相当于得到了零权值,而另外的单元则得到了非零且相等的权值。<br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have ''J'' observations over ''T'' time periods where the relevant treatment occurs at time <math>T_{0}</math> where <math>T_{0}<T.</math> Let <br />
<br />
:<math>\alpha_{it}=Y_{it}-Y^N_{it},</math><br />
be the treatment effect for unit <math>i</math> at time <math>t</math>, where <math>Y^N_{it}</math> is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only <math>Y^N_{1t}</math>is not observed for <math>t>T_{0}</math>. We aim to estimate <math>(\alpha_{1T_{0}+1}......\alpha_{1T})</math>. <br />
<br />
:<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have J observations over T time periods where the relevant treatment occurs at time T_{0} where T_{0}<T. Let <br />
<br />
:\alpha_{it}=Y_{it}-Y^N_{it},<br />
be the treatment effect for unit i at time t, where Y^N_{it} is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only Y^N_{1t}is not observed for t>T_{0}. We aim to estimate (\alpha_{1T_{0}+1}......\alpha_{1T}). <br />
<br />
: <br />
<br />
综合控制方法试图为控制组的权重分配提供一种更加系统的方法。它通常使用一个相对较长的时间序列的结果之前的干预和估计权重的方式,控制组镜像治疗组尽可能接近。特别地,假设我们在 t 时间段有 j 观测值,在 t {0} < t 时相应的处理发生在 t {0} < t 时。让<br />
<br />
alpha _ { it } = y _ { it }-y ^ n _ { it } ,为时间 t 的单位 i 的治疗效果,其中 y ^ n _ { it }是未经治疗的结果。不失一般性,如果单位1接受相应的治疗,只有 y ^ n {1 t }没有观察到 t > t {0}。我们的目标是估计(alpha _ {1T _ {0} + 1} ... ... alpha _ {1T })。<br />
<br />
<br />
【翻译】合成对照方法试图用一种更加系统的方法为控制组分配权重。它通常采用干预之前相对比较长时间段内的多个时间序列,估计一组权值使得对这些时间序列的输出进行加权后得到的控制组时间序列尽可能去拟合治疗组的时间序列。特别地,假设我们在总共T个时间段里有J个观测量(单位),其中一个单位接受了治疗,相应的治疗发生在T_{0}时间段,且T_{0}<T。让<br />
<br />
: \alpha_{it}=Y_{it}-Y^N_{it},<br />
<br />
为单位 i 的在时间 t 的治疗效果,其中 Y^N_{it} 是未经治疗的结果。不失一般性,如果单位1接受了相应的治疗,只有在 t > T_{0} 时段的 Y^N_{1t} 没有被观察到。我们的目标是估计(\alpha_{1T_{0}+1} ... ... \alpha_{1T})。<br />
<br />
<br />
Imposing some structure<br />
<br />
Imposing some structure<br />
<br />
强加一些结构<br />
<br />
<br />
【翻译】<br />
<br />
强加一些结构<br />
<br />
:<math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math> <br />
<br />
:Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it} <br />
<br />
: y ^ n { it } = delta { t } + theta { t } z { i } + lambda { t } mu { i } + varepsilon { it }<br />
<br />
and assuming there exist some optimal weights <math>w_2, \ldots, w_J</math> such that <br />
<br />
and assuming there exist some optimal weights w_2, \ldots, w_J such that <br />
<br />
假设存在一些最优权重 w _ 2,ldots,w _ j 使得<br />
<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math> <br />
<br />
:Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt} <br />
<br />
: y {1 t } = Sigma ^ j { j = 2} w { j } y { jt }<br />
<br />
for <math>t\leqslant T_{0}</math>, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
for t\leqslant T_{0}, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
对于 t _ {0} ,综合控制方法建议使用这些权重来估计反事实<br />
<br />
:<math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math> <br />
for <math>t>T_{0}</math>. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<ref>{{cite journal |last=Abadie |first=A. |authorlink=Alberto Abadie |first2=A. |last2=Diamond |first3= J. |last3=Hainmüller |year=2010 |title=Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California's Tobacco Control Program |journal=[[Journal of the American Statistical Association]] |volume=105 |issue=490 |pages=493–505 |doi=10.1198/jasa.2009.ap08746 }}</ref><br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}.因此,在一定的正则性条件下,这种权重可以作为利息处理效果的估计量。本质上,该方法采用了匹配的思想,并利用训练数据进行干预前的权重设置,从而得到干预后的相应控制。<br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth,<ref>{{cite journal |last=Cavallo |first=E. |first2=S. |last2=Galliani |first3=I. |last3=Noy |first4=J. |last4=Pantano |year=2013 |title=Catastrophic Natural Disasters and Economic Growth |journal=[[Review of Economics and Statistics]] |volume=95 |issue=5 |pages=1549–1561 |doi=10.1162/REST_a_00413 |url=http://www.economics.hawaii.edu/research/workingpapers/WP_10-6.pdf }}</ref> and studies linking political murders to house prices.<ref>{{cite journal |last=Gautier |first=P. A. |first2=A. |last2=Siegmann |first3=A. |last3=Van Vuuren |year=2009 |title=Terrorism and Attitudes towards Minorities: The effect of the Theo van Gogh murder on house prices in Amsterdam |journal=[[Journal of Urban Economics]] |volume=65 |issue=2 |pages=113–126 |doi=10.1016/j.jue.2008.10.004 }}</ref> <br />
<!-- THE CITATION AT THE END OF THIS SENTENCE IS FOR A PAPER ABOUT "synthetic cohort models" (a.k.a. "pseudo-panel approach," using repeated cross-sections), WHICH IS NOT THE SAME AS "synthetic control": Yet, despite its intuitive appeal, it may be the case that synthetic controls could suffer from significant finite sample biases.<ref>{{cite journal |last=Devereux |first=P. J. |year=2007 |title=Small-sample bias in synthetic cohort models of labor supply |journal=[[Journal of Applied Econometrics]] |volume=22 |issue=4 |pages=839–848 |doi=10.1002/jae.938 }}</ref> --><br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth, and studies linking political murders to house prices. <br />
<br />
<br />
综合控制已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
==References==<br />
{{Reflist|30em}}<br />
<br />
[[Category:Design of experiments]]<br />
[[Category:Statistical methods]]<br />
[[Category:Observational study]]<br />
[[Category:Econometric modeling]]<br />
<br />
Category:Design of experiments<br />
Category:Statistical methods<br />
Category:Observational study<br />
Category:Econometric modeling<br />
<br />
类别: 实验设计类别: 统计方法类别: 观察性研究类别: 计量经济模型<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Synthetic control method]]. Its edit history can be viewed at [[合成对照/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%90%88%E6%88%90%E5%AF%B9%E7%85%A7&diff=29175合成对照2022-03-16T01:04:37Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共492,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
The '''synthetic control method''' is a statistical method used to evaluate the effect of an intervention in [[comparative case study|comparative case studies]]. It involves the construction of a weighted combination of groups used as controls, to which the [[treatment group]] is compared.<ref>{{Cite journal|last=Abadie|first=Alberto|date=2021|title=Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects|url=https://www.aeaweb.org/articles?id=10.1257/jel.20191450|journal=Journal of Economic Literature|language=en|volume=59|issue=2|pages=391–425|doi=10.1257/jel.20191450|issn=0022-0515|doi-access=free}}</ref> This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike [[difference in differences]] approaches, this method can account for the effects of [[confounder]]s changing over time, by weighting the control group to better match the treatment group before the intervention.<ref name=he>{{cite journal|last1=Kreif|first1=Noémi|last2=Grieve|first2=Richard|last3=Hangartner|first3=Dominik|last4=Turner|first4=Alex James|last5=Nikolova|first5=Silviya|last6=Sutton|first6=Matt|title=Examination of the Synthetic Control Method for Evaluating Health Policies with Multiple Treated Units|journal=Health Economics|date=December 2016|volume=25|issue=12|pages=1514–1528|doi=10.1002/hec.3258|pmid=26443693|pmc=5111584}}</ref> Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups.<ref name=ajps>{{cite journal|last1=Abadie|first1=Alberto|authorlink1=Alberto Abadie|last2=Diamond|first2=Alexis|last3=Hainmueller|first3=Jens|title=Comparative Politics and the Synthetic Control Method|journal=American Journal of Political Science|date=February 2015|volume=59|issue=2|pages=495–510|doi=10.1111/ajps.12116}}</ref> It has been applied to the fields of [[political science]],<ref name=ajps/> [[health policy]],<ref name=he/> [[criminology]],<ref>{{cite journal|last1=Saunders|first1=Jessica|last2=Lundberg|first2=Russell|last3=Braga|first3=Anthony A.|last4=Ridgeway|first4=Greg|last5=Miles|first5=Jeremy|title=A Synthetic Control Approach to Evaluating Place-Based Crime Interventions|journal=Journal of Quantitative Criminology|date=3 June 2014|volume=31|issue=3|pages=413–434|doi=10.1007/s10940-014-9226-5}}</ref> and [[economics]].<ref>{{cite journal|last1=Billmeier|first1=Andreas|last2=Nannicini|first2=Tommaso|title=Assessing Economic Liberalization Episodes: A Synthetic Control Approach|journal=Review of Economics and Statistics|date=July 2013|volume=95|issue=3|pages=983–1001|doi=10.1162/REST_a_00324}}</ref><br />
<br />
The synthetic control method is a statistical method used to evaluate the effect of an intervention in comparative case studies. It involves the construction of a weighted combination of groups used as controls, to which the treatment group is compared. This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike difference in differences approaches, this method can account for the effects of confounders changing over time, by weighting the control group to better match the treatment group before the intervention. Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups. It has been applied to the fields of political science, health policy, criminology, and economics.<br />
<br />
【翻译】合成对照方法是一种统计方法,用于评估比较案例研究中的干预措施的效果。它使用多组数据加权组合构建对照组,并与治疗组进行比较。这种比较被用来估计如果治疗组没有接受治疗会发生什么。与双重差分(Difference in difference)方法不同,这种方法可以纳入随时间变化的混杂因素的影响,通过调整对照组的加权系数,能更好地对干预前的治疗组数据进行匹配。合成对照的另一个优点是,它允许研究人员在多组候选数据中做系统性选择。它已应用于政治学、卫生政策、犯罪学和经济学等领域。<br />
<br />
The synthetic control method combines elements from [[Matching (statistics)|matching]] and [[difference-in-differences]] techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the [[Minimum wage in the United States|minimum wage]] in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in [[Philadelphia]] that were unaffected by a minimum wage raise,<ref name="CardKrueger">{{cite journal |last=Card |first=D. |authorlink=David Card |first2=A. |last2=Krueger |authorlink2=Alan Krueger |year=1994 |title=Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania |journal=[[American Economic Review]] |volume=84 |issue=4 |pages=772–793 |jstor=2118030 }}</ref> and studies that look at [[crime rates]] in southern cities to evaluate the impact of the [[Mariel boat lift]] on crime.<ref>{{cite journal |last=Card |first=D. |year=1990 |title=The Impact of the Mariel Boatlift on the Miami Labor Market |journal=[[Industrial and Labor Relations Review]] |volume=43 |issue=2 |pages=245–257 |doi=10.1177/001979399004300205 |url=http://arks.princeton.edu/ark:/88435/dsp016h440s46f }}</ref> The control group in this specific scenario can be interpreted as a [[Weighted arithmetic mean|weighted average]], where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
The synthetic control method combines elements from matching and difference-in-differences techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the minimum wage in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in Philadelphia that were unaffected by a minimum wage raise, and studies that look at crime rates in southern cities to evaluate the impact of the Mariel boat lift on crime. The control group in this specific scenario can be interpreted as a weighted average, where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
综合控制方法结合了匹配技术和差中差技术的要素。差异中的差异法是一种常用的政策评估工具,用于在总体水平上评估干预措施的效果(例如:。州、国家、年龄组别等)平均超过一组未受影响的单位。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,比较对象是紧邻州边境的费城,那边的快餐店没有受到提高最低工资的影响,以及研究南部城市的犯罪率来评估马里埃尔移民潮对犯罪率的影响。在这个特定的场景中,控制组可以被解释为一个加权平均数,其中一些单位实际上得到了零重量,而其他单位得到了相等的,非零重量。<br />
<br />
【翻译】合成对照方法结合了匹配技术和双重差分技术的要素。双重差分法是一种常用的政策评估工具,通过对未被影响的单元求平均,在总体水平上(例如:州、国家、年龄组别等)评估在被干预单元上的政策干预效果。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,通过比较它们与费城边境那边没有受到最低工资提高影响的快餐店,以及研究南部城市的犯罪率来评估马里埃尔船只提升对犯罪的影响。在双重差分场景中,合成对照的控制组可被理解为一个加权平均,其中的一些单元相当于得到了零权值,而另外的单元则得到了非零且相等的权值。<br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have ''J'' observations over ''T'' time periods where the relevant treatment occurs at time <math>T_{0}</math> where <math>T_{0}<T.</math> Let <br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have J observations over T time periods where the relevant treatment occurs at time T_{0} where T_{0}<T. Let <br />
<br />
综合控制方法试图为控制组的权重分配提供一种更加系统的方法。它通常使用一个相对较长的时间序列的结果之前的干预和估计权重的方式,控制组镜像治疗组尽可能接近。特别地,假设我们在 t 时间段有 j 观测值,在 t {0} < t 时相应的处理发生在 t {0} < t 时。让<br />
<br />
【翻译】合成对照方法试图用一种更加系统的方法为控制组分配权重。它通常采用干预之前相对比较长时间段内的多个时间序列,估计一组权值使得对这些时间序列的输出进行加权后得到的控制组时间序列尽可能去拟合治疗组的时间序列。特别地,假设我们在时间段T里总共有J个观测值,其中T_{0},T_{0}<T代表治疗组发生的时间段。让<br />
<br />
:<math>\alpha_{it}=Y_{it}-Y^N_{it},</math> <br />
be the treatment effect for unit <math>i</math> at time <math>t</math>, where <math>Y^N_{it}</math> is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only <math>Y^N_{1t}</math>is not observed for <math>t>T_{0}</math>. We aim to estimate <math>(\alpha_{1T_{0}+1}......\alpha_{1T})</math>. <br />
<br />
:\alpha_{it}=Y_{it}-Y^N_{it}, <br />
be the treatment effect for unit i at time t, where Y^N_{it} is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only Y^N_{1t}is not observed for t>T_{0}. We aim to estimate (\alpha_{1T_{0}+1}......\alpha_{1T}). <br />
<br />
: alpha _ { it } = y _ { it }-y ^ n _ { it } ,为时间 t 的单位 i 的治疗效果,其中 y ^ n _ { it }是未经治疗的结果。不失一般性,如果单位1接受相应的治疗,只有 y ^ n {1 t }没有观察到 t > t {0}。我们的目标是估计(alpha _ {1T _ {0} + 1} ... ... alpha _ {1T })。<br />
<br />
Imposing some structure<br />
<br />
Imposing some structure<br />
<br />
强加一些结构<br />
<br />
:<math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math> <br />
<br />
:Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it} <br />
<br />
: y ^ n { it } = delta { t } + theta { t } z { i } + lambda { t } mu { i } + varepsilon { it }<br />
<br />
and assuming there exist some optimal weights <math>w_2, \ldots, w_J</math> such that <br />
<br />
and assuming there exist some optimal weights w_2, \ldots, w_J such that <br />
<br />
假设存在一些最优权重 w _ 2,ldots,w _ j 使得<br />
<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math> <br />
<br />
:Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt} <br />
<br />
: y {1 t } = Sigma ^ j { j = 2} w { j } y { jt }<br />
<br />
for <math>t\leqslant T_{0}</math>, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
for t\leqslant T_{0}, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
对于 t _ {0} ,综合控制方法建议使用这些权重来估计反事实<br />
<br />
:<math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math> <br />
for <math>t>T_{0}</math>. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<ref>{{cite journal |last=Abadie |first=A. |authorlink=Alberto Abadie |first2=A. |last2=Diamond |first3= J. |last3=Hainmüller |year=2010 |title=Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California's Tobacco Control Program |journal=[[Journal of the American Statistical Association]] |volume=105 |issue=490 |pages=493–505 |doi=10.1198/jasa.2009.ap08746 }}</ref><br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}.因此,在一定的正则性条件下,这种权重可以作为利息处理效果的估计量。本质上,该方法采用了匹配的思想,并利用训练数据进行干预前的权重设置,从而得到干预后的相应控制。<br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth,<ref>{{cite journal |last=Cavallo |first=E. |first2=S. |last2=Galliani |first3=I. |last3=Noy |first4=J. |last4=Pantano |year=2013 |title=Catastrophic Natural Disasters and Economic Growth |journal=[[Review of Economics and Statistics]] |volume=95 |issue=5 |pages=1549–1561 |doi=10.1162/REST_a_00413 |url=http://www.economics.hawaii.edu/research/workingpapers/WP_10-6.pdf }}</ref> and studies linking political murders to house prices.<ref>{{cite journal |last=Gautier |first=P. A. |first2=A. |last2=Siegmann |first3=A. |last3=Van Vuuren |year=2009 |title=Terrorism and Attitudes towards Minorities: The effect of the Theo van Gogh murder on house prices in Amsterdam |journal=[[Journal of Urban Economics]] |volume=65 |issue=2 |pages=113–126 |doi=10.1016/j.jue.2008.10.004 }}</ref> <br />
<!-- THE CITATION AT THE END OF THIS SENTENCE IS FOR A PAPER ABOUT "synthetic cohort models" (a.k.a. "pseudo-panel approach," using repeated cross-sections), WHICH IS NOT THE SAME AS "synthetic control": Yet, despite its intuitive appeal, it may be the case that synthetic controls could suffer from significant finite sample biases.<ref>{{cite journal |last=Devereux |first=P. J. |year=2007 |title=Small-sample bias in synthetic cohort models of labor supply |journal=[[Journal of Applied Econometrics]] |volume=22 |issue=4 |pages=839–848 |doi=10.1002/jae.938 }}</ref> --><br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth, and studies linking political murders to house prices. <br />
<br />
<br />
综合控制已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
==References==<br />
{{Reflist|30em}}<br />
<br />
[[Category:Design of experiments]]<br />
[[Category:Statistical methods]]<br />
[[Category:Observational study]]<br />
[[Category:Econometric modeling]]<br />
<br />
Category:Design of experiments<br />
Category:Statistical methods<br />
Category:Observational study<br />
Category:Econometric modeling<br />
<br />
类别: 实验设计类别: 统计方法类别: 观察性研究类别: 计量经济模型<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Synthetic control method]]. Its edit history can be viewed at [[合成对照/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%90%88%E6%88%90%E5%AF%B9%E7%85%A7&diff=29126合成对照2022-03-14T09:10:15Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共492,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
The '''synthetic control method''' is a statistical method used to evaluate the effect of an intervention in [[comparative case study|comparative case studies]]. It involves the construction of a weighted combination of groups used as controls, to which the [[treatment group]] is compared.<ref>{{Cite journal|last=Abadie|first=Alberto|date=2021|title=Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects|url=https://www.aeaweb.org/articles?id=10.1257/jel.20191450|journal=Journal of Economic Literature|language=en|volume=59|issue=2|pages=391–425|doi=10.1257/jel.20191450|issn=0022-0515|doi-access=free}}</ref> This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike [[difference in differences]] approaches, this method can account for the effects of [[confounder]]s changing over time, by weighting the control group to better match the treatment group before the intervention.<ref name=he>{{cite journal|last1=Kreif|first1=Noémi|last2=Grieve|first2=Richard|last3=Hangartner|first3=Dominik|last4=Turner|first4=Alex James|last5=Nikolova|first5=Silviya|last6=Sutton|first6=Matt|title=Examination of the Synthetic Control Method for Evaluating Health Policies with Multiple Treated Units|journal=Health Economics|date=December 2016|volume=25|issue=12|pages=1514–1528|doi=10.1002/hec.3258|pmid=26443693|pmc=5111584}}</ref> Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups.<ref name=ajps>{{cite journal|last1=Abadie|first1=Alberto|authorlink1=Alberto Abadie|last2=Diamond|first2=Alexis|last3=Hainmueller|first3=Jens|title=Comparative Politics and the Synthetic Control Method|journal=American Journal of Political Science|date=February 2015|volume=59|issue=2|pages=495–510|doi=10.1111/ajps.12116}}</ref> It has been applied to the fields of [[political science]],<ref name=ajps/> [[health policy]],<ref name=he/> [[criminology]],<ref>{{cite journal|last1=Saunders|first1=Jessica|last2=Lundberg|first2=Russell|last3=Braga|first3=Anthony A.|last4=Ridgeway|first4=Greg|last5=Miles|first5=Jeremy|title=A Synthetic Control Approach to Evaluating Place-Based Crime Interventions|journal=Journal of Quantitative Criminology|date=3 June 2014|volume=31|issue=3|pages=413–434|doi=10.1007/s10940-014-9226-5}}</ref> and [[economics]].<ref>{{cite journal|last1=Billmeier|first1=Andreas|last2=Nannicini|first2=Tommaso|title=Assessing Economic Liberalization Episodes: A Synthetic Control Approach|journal=Review of Economics and Statistics|date=July 2013|volume=95|issue=3|pages=983–1001|doi=10.1162/REST_a_00324}}</ref><br />
<br />
The synthetic control method is a statistical method used to evaluate the effect of an intervention in comparative case studies. It involves the construction of a weighted combination of groups used as controls, to which the treatment group is compared. This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike difference in differences approaches, this method can account for the effects of confounders changing over time, by weighting the control group to better match the treatment group before the intervention. Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups. It has been applied to the fields of political science, health policy, criminology, and economics.<br />
<br />
【翻译】合成对照方法是一种统计方法,用于评估比较案例研究中的干预措施的效果。它使用多组数据加权组合构建对照组,并与治疗组进行比较。这种比较被用来估计如果治疗组没有接受治疗会发生什么。与双重差分(Difference in difference)方法不同,这种方法可以纳入随时间变化的混杂因素的影响,通过调整对照组的加权系数,能更好地对干预前的治疗组数据进行匹配。合成对照的另一个优点是,它允许研究人员在多组候选数据中做系统性选择。它已应用于政治学、卫生政策、犯罪学和经济学等领域。<br />
<br />
The synthetic control method combines elements from [[Matching (statistics)|matching]] and [[difference-in-differences]] techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the [[Minimum wage in the United States|minimum wage]] in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in [[Philadelphia]] that were unaffected by a minimum wage raise,<ref name="CardKrueger">{{cite journal |last=Card |first=D. |authorlink=David Card |first2=A. |last2=Krueger |authorlink2=Alan Krueger |year=1994 |title=Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania |journal=[[American Economic Review]] |volume=84 |issue=4 |pages=772–793 |jstor=2118030 }}</ref> and studies that look at [[crime rates]] in southern cities to evaluate the impact of the [[Mariel boat lift]] on crime.<ref>{{cite journal |last=Card |first=D. |year=1990 |title=The Impact of the Mariel Boatlift on the Miami Labor Market |journal=[[Industrial and Labor Relations Review]] |volume=43 |issue=2 |pages=245–257 |doi=10.1177/001979399004300205 |url=http://arks.princeton.edu/ark:/88435/dsp016h440s46f }}</ref> The control group in this specific scenario can be interpreted as a [[Weighted arithmetic mean|weighted average]], where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
The synthetic control method combines elements from matching and difference-in-differences techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the minimum wage in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in Philadelphia that were unaffected by a minimum wage raise, and studies that look at crime rates in southern cities to evaluate the impact of the Mariel boat lift on crime. The control group in this specific scenario can be interpreted as a weighted average, where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
综合控制方法结合了匹配技术和差中差技术的要素。差异中的差异法是一种常用的政策评估工具,用于在总体水平上评估干预措施的效果(例如:。州、国家、年龄组别等)平均超过一组未受影响的单位。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,比较对象是紧邻州边境的费城,那边的快餐店没有受到提高最低工资的影响,以及研究南部城市的犯罪率来评估马里埃尔移民潮对犯罪率的影响。在这个特定的场景中,控制组可以被解释为一个加权平均数,其中一些单位实际上得到了零重量,而其他单位得到了相等的,非零重量。<br />
<br />
【翻译】合成对照方法结合了匹配技术和双重差分技术的要素。双重差分法是一种常用的政策评估工具,通过对未被影响的单元求平均,在总体水平上(例如:州、国家、年龄组别等)评估在被干预单元上的政策干预效果。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,通过比较它们与费城边境那边没有受到最低工资提高影响的快餐店,以及研究南部城市的犯罪率来评估马里埃尔船只提升对犯罪的影响。在双重差分场景中,合成对照的控制组可被理解为一个加权平均,其中的一些单元相当于得到了零权值,而另外的单元则得到了非零且相等的权值。<br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have ''J'' observations over ''T'' time periods where the relevant treatment occurs at time <math>T_{0}</math> where <math>T_{0}<T.</math> Let <br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have J observations over T time periods where the relevant treatment occurs at time T_{0} where T_{0}<T. Let <br />
<br />
综合控制方法试图为控制组的权重分配提供一种更加系统的方法。它通常使用一个相对较长的时间序列的结果之前的干预和估计权重的方式,控制组镜像治疗组尽可能接近。特别地,假设我们在 t 时间段有 j 观测值,在 t {0} < t 时相应的处理发生在 t {0} < t 时。让<br />
<br />
:<math>\alpha_{it}=Y_{it}-Y^N_{it},</math> <br />
be the treatment effect for unit <math>i</math> at time <math>t</math>, where <math>Y^N_{it}</math> is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only <math>Y^N_{1t}</math>is not observed for <math>t>T_{0}</math>. We aim to estimate <math>(\alpha_{1T_{0}+1}......\alpha_{1T})</math>. <br />
<br />
:\alpha_{it}=Y_{it}-Y^N_{it}, <br />
be the treatment effect for unit i at time t, where Y^N_{it} is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only Y^N_{1t}is not observed for t>T_{0}. We aim to estimate (\alpha_{1T_{0}+1}......\alpha_{1T}). <br />
<br />
: alpha _ { it } = y _ { it }-y ^ n _ { it } ,为时间 t 的单位 i 的治疗效果,其中 y ^ n _ { it }是未经治疗的结果。不失一般性,如果单位1接受相应的治疗,只有 y ^ n {1 t }没有观察到 t > t {0}。我们的目标是估计(alpha _ {1T _ {0} + 1} ... ... alpha _ {1T })。<br />
<br />
Imposing some structure<br />
<br />
Imposing some structure<br />
<br />
强加一些结构<br />
<br />
:<math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math> <br />
<br />
:Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it} <br />
<br />
: y ^ n { it } = delta { t } + theta { t } z { i } + lambda { t } mu { i } + varepsilon { it }<br />
<br />
and assuming there exist some optimal weights <math>w_2, \ldots, w_J</math> such that <br />
<br />
and assuming there exist some optimal weights w_2, \ldots, w_J such that <br />
<br />
假设存在一些最优权重 w _ 2,ldots,w _ j 使得<br />
<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math> <br />
<br />
:Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt} <br />
<br />
: y {1 t } = Sigma ^ j { j = 2} w { j } y { jt }<br />
<br />
for <math>t\leqslant T_{0}</math>, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
for t\leqslant T_{0}, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
对于 t _ {0} ,综合控制方法建议使用这些权重来估计反事实<br />
<br />
:<math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math> <br />
for <math>t>T_{0}</math>. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<ref>{{cite journal |last=Abadie |first=A. |authorlink=Alberto Abadie |first2=A. |last2=Diamond |first3= J. |last3=Hainmüller |year=2010 |title=Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California's Tobacco Control Program |journal=[[Journal of the American Statistical Association]] |volume=105 |issue=490 |pages=493–505 |doi=10.1198/jasa.2009.ap08746 }}</ref><br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}.因此,在一定的正则性条件下,这种权重可以作为利息处理效果的估计量。本质上,该方法采用了匹配的思想,并利用训练数据进行干预前的权重设置,从而得到干预后的相应控制。<br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth,<ref>{{cite journal |last=Cavallo |first=E. |first2=S. |last2=Galliani |first3=I. |last3=Noy |first4=J. |last4=Pantano |year=2013 |title=Catastrophic Natural Disasters and Economic Growth |journal=[[Review of Economics and Statistics]] |volume=95 |issue=5 |pages=1549–1561 |doi=10.1162/REST_a_00413 |url=http://www.economics.hawaii.edu/research/workingpapers/WP_10-6.pdf }}</ref> and studies linking political murders to house prices.<ref>{{cite journal |last=Gautier |first=P. A. |first2=A. |last2=Siegmann |first3=A. |last3=Van Vuuren |year=2009 |title=Terrorism and Attitudes towards Minorities: The effect of the Theo van Gogh murder on house prices in Amsterdam |journal=[[Journal of Urban Economics]] |volume=65 |issue=2 |pages=113–126 |doi=10.1016/j.jue.2008.10.004 }}</ref> <br />
<!-- THE CITATION AT THE END OF THIS SENTENCE IS FOR A PAPER ABOUT "synthetic cohort models" (a.k.a. "pseudo-panel approach," using repeated cross-sections), WHICH IS NOT THE SAME AS "synthetic control": Yet, despite its intuitive appeal, it may be the case that synthetic controls could suffer from significant finite sample biases.<ref>{{cite journal |last=Devereux |first=P. J. |year=2007 |title=Small-sample bias in synthetic cohort models of labor supply |journal=[[Journal of Applied Econometrics]] |volume=22 |issue=4 |pages=839–848 |doi=10.1002/jae.938 }}</ref> --><br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth, and studies linking political murders to house prices. <br />
<br />
<br />
综合控制已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
==References==<br />
{{Reflist|30em}}<br />
<br />
[[Category:Design of experiments]]<br />
[[Category:Statistical methods]]<br />
[[Category:Observational study]]<br />
[[Category:Econometric modeling]]<br />
<br />
Category:Design of experiments<br />
Category:Statistical methods<br />
Category:Observational study<br />
Category:Econometric modeling<br />
<br />
类别: 实验设计类别: 统计方法类别: 观察性研究类别: 计量经济模型<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Synthetic control method]]. Its edit history can be viewed at [[合成对照/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%90%88%E6%88%90%E5%AF%B9%E7%85%A7&diff=29118合成对照2022-03-14T03:39:26Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共492,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
The '''synthetic control method''' is a statistical method used to evaluate the effect of an intervention in [[comparative case study|comparative case studies]]. It involves the construction of a weighted combination of groups used as controls, to which the [[treatment group]] is compared.<ref>{{Cite journal|last=Abadie|first=Alberto|date=2021|title=Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects|url=https://www.aeaweb.org/articles?id=10.1257/jel.20191450|journal=Journal of Economic Literature|language=en|volume=59|issue=2|pages=391–425|doi=10.1257/jel.20191450|issn=0022-0515|doi-access=free}}</ref> This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike [[difference in differences]] approaches, this method can account for the effects of [[confounder]]s changing over time, by weighting the control group to better match the treatment group before the intervention.<ref name=he>{{cite journal|last1=Kreif|first1=Noémi|last2=Grieve|first2=Richard|last3=Hangartner|first3=Dominik|last4=Turner|first4=Alex James|last5=Nikolova|first5=Silviya|last6=Sutton|first6=Matt|title=Examination of the Synthetic Control Method for Evaluating Health Policies with Multiple Treated Units|journal=Health Economics|date=December 2016|volume=25|issue=12|pages=1514–1528|doi=10.1002/hec.3258|pmid=26443693|pmc=5111584}}</ref> Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups.<ref name=ajps>{{cite journal|last1=Abadie|first1=Alberto|authorlink1=Alberto Abadie|last2=Diamond|first2=Alexis|last3=Hainmueller|first3=Jens|title=Comparative Politics and the Synthetic Control Method|journal=American Journal of Political Science|date=February 2015|volume=59|issue=2|pages=495–510|doi=10.1111/ajps.12116}}</ref> It has been applied to the fields of [[political science]],<ref name=ajps/> [[health policy]],<ref name=he/> [[criminology]],<ref>{{cite journal|last1=Saunders|first1=Jessica|last2=Lundberg|first2=Russell|last3=Braga|first3=Anthony A.|last4=Ridgeway|first4=Greg|last5=Miles|first5=Jeremy|title=A Synthetic Control Approach to Evaluating Place-Based Crime Interventions|journal=Journal of Quantitative Criminology|date=3 June 2014|volume=31|issue=3|pages=413–434|doi=10.1007/s10940-014-9226-5}}</ref> and [[economics]].<ref>{{cite journal|last1=Billmeier|first1=Andreas|last2=Nannicini|first2=Tommaso|title=Assessing Economic Liberalization Episodes: A Synthetic Control Approach|journal=Review of Economics and Statistics|date=July 2013|volume=95|issue=3|pages=983–1001|doi=10.1162/REST_a_00324}}</ref><br />
<br />
The synthetic control method is a statistical method used to evaluate the effect of an intervention in comparative case studies. It involves the construction of a weighted combination of groups used as controls, to which the treatment group is compared. This comparison is used to estimate what would have happened to the treatment group if it had not received the treatment.<br />
Unlike difference in differences approaches, this method can account for the effects of confounders changing over time, by weighting the control group to better match the treatment group before the intervention. Another advantage of the synthetic control method is that it allows researchers to systematically select comparison groups. It has been applied to the fields of political science, health policy, criminology, and economics.<br />
<br />
'''合成对照'''方法是一种统计方法,用于评估比较案例研究中的干预措施的效果。它使用多组数据加权组合构建对照组,并与治疗组进行比较。这种比较被用来估计如果治疗组没有接受治疗会发生什么。与双重差分(Difference in difference)方法不同,这种方法可以纳入随时间变化的混杂因素的影响,通过调整对照组的加权系数,能更好地对干预前的治疗组数据进行匹配。合成对照的另一个优点是,它允许研究人员在多组候选数据中做系统性选择。它已应用于政治学、卫生政策、犯罪学和经济学等领域。<br />
<br />
The synthetic control method combines elements from [[Matching (statistics)|matching]] and [[difference-in-differences]] techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the [[Minimum wage in the United States|minimum wage]] in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in [[Philadelphia]] that were unaffected by a minimum wage raise,<ref name="CardKrueger">{{cite journal |last=Card |first=D. |authorlink=David Card |first2=A. |last2=Krueger |authorlink2=Alan Krueger |year=1994 |title=Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania |journal=[[American Economic Review]] |volume=84 |issue=4 |pages=772–793 |jstor=2118030 }}</ref> and studies that look at [[crime rates]] in southern cities to evaluate the impact of the [[Mariel boat lift]] on crime.<ref>{{cite journal |last=Card |first=D. |year=1990 |title=The Impact of the Mariel Boatlift on the Miami Labor Market |journal=[[Industrial and Labor Relations Review]] |volume=43 |issue=2 |pages=245–257 |doi=10.1177/001979399004300205 |url=http://arks.princeton.edu/ark:/88435/dsp016h440s46f }}</ref> The control group in this specific scenario can be interpreted as a [[Weighted arithmetic mean|weighted average]], where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
The synthetic control method combines elements from matching and difference-in-differences techniques. Difference-in-differences methods are often-used policy evaluation tools that estimate the effect of an intervention at an aggregate level (e.g. state, country, age group etc.) by averaging over a set of unaffected units. Famous examples include studies of the employment effects of a raise in the minimum wage in New Jersey fast food restaurants by comparing them to fast food restaurants just across the border in Philadelphia that were unaffected by a minimum wage raise, and studies that look at crime rates in southern cities to evaluate the impact of the Mariel boat lift on crime. The control group in this specific scenario can be interpreted as a weighted average, where some units effectively receive zero weight while others get an equal, non-zero weight.<br />
<br />
综合控制方法结合了匹配技术和差中差技术的要素。差异中的差异法是一种常用的政策评估工具,用于在总体水平上评估干预措施的效果(例如:。州、国家、年龄组别等)平均超过一组未受影响的单位。著名的例子包括新泽西州快餐店提高最低工资对就业影响的研究,通过比较它们与费城边境那边没有受到最低工资提高影响的快餐店,以及研究南部城市的犯罪率来评估马里埃尔船只提升对犯罪的影响。在这个特定的场景中,控制组可以被解释为一个加权平均数,其中一些单位实际上得到了零重量,而其他单位得到了相等的,非零重量。<br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have ''J'' observations over ''T'' time periods where the relevant treatment occurs at time <math>T_{0}</math> where <math>T_{0}<T.</math> Let <br />
<br />
The synthetic control method tries to offer a more systematic way to assign weights to the control group. It typically uses a relatively long time series of the outcome prior to the intervention and estimates weights in such a way that the control group mirrors the treatment group as closely as possible. In particular, assume we have J observations over T time periods where the relevant treatment occurs at time T_{0} where T_{0}<T. Let <br />
<br />
综合控制方法试图为控制组的权重分配提供一种更加系统的方法。它通常使用一个相对较长的时间序列的结果之前的干预和估计权重的方式,控制组镜像治疗组尽可能接近。特别地,假设我们在 t 时间段有 j 观测值,在 t {0} < t 时相应的处理发生在 t {0} < t 时。让<br />
<br />
:<math>\alpha_{it}=Y_{it}-Y^N_{it},</math> <br />
be the treatment effect for unit <math>i</math> at time <math>t</math>, where <math>Y^N_{it}</math> is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only <math>Y^N_{1t}</math>is not observed for <math>t>T_{0}</math>. We aim to estimate <math>(\alpha_{1T_{0}+1}......\alpha_{1T})</math>. <br />
<br />
:\alpha_{it}=Y_{it}-Y^N_{it}, <br />
be the treatment effect for unit i at time t, where Y^N_{it} is the outcome in absence of the treatment. Without loss of generality, if unit 1 receives the relevant treatment, only Y^N_{1t}is not observed for t>T_{0}. We aim to estimate (\alpha_{1T_{0}+1}......\alpha_{1T}). <br />
<br />
: alpha _ { it } = y _ { it }-y ^ n _ { it } ,为时间 t 的单位 i 的治疗效果,其中 y ^ n _ { it }是未经治疗的结果。不失一般性,如果单位1接受相应的治疗,只有 y ^ n {1 t }没有观察到 t > t {0}。我们的目标是估计(alpha _ {1T _ {0} + 1} ... ... alpha _ {1T })。<br />
<br />
Imposing some structure<br />
<br />
Imposing some structure<br />
<br />
强加一些结构<br />
<br />
:<math>Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it}</math> <br />
<br />
:Y^N_{it}=\delta_{t}+\theta_{t}Z_{i}+\lambda_{t}\mu_{i}+\varepsilon_{it} <br />
<br />
: y ^ n { it } = delta { t } + theta { t } z { i } + lambda { t } mu { i } + varepsilon { it }<br />
<br />
and assuming there exist some optimal weights <math>w_2, \ldots, w_J</math> such that <br />
<br />
and assuming there exist some optimal weights w_2, \ldots, w_J such that <br />
<br />
假设存在一些最优权重 w _ 2,ldots,w _ j 使得<br />
<br />
:<math>Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt}</math> <br />
<br />
:Y_{1t} = \Sigma^J_{j=2} w_{j}Y_{jt} <br />
<br />
: y {1 t } = Sigma ^ j { j = 2} w { j } y { jt }<br />
<br />
for <math>t\leqslant T_{0}</math>, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
for t\leqslant T_{0}, the synthetic controls approach suggests using these weights to estimate the counterfactual <br />
<br />
对于 t _ {0} ,综合控制方法建议使用这些权重来估计反事实<br />
<br />
:<math>Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt}</math> <br />
for <math>t>T_{0}</math>. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<ref>{{cite journal |last=Abadie |first=A. |authorlink=Alberto Abadie |first2=A. |last2=Diamond |first3= J. |last3=Hainmüller |year=2010 |title=Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California's Tobacco Control Program |journal=[[Journal of the American Statistical Association]] |volume=105 |issue=490 |pages=493–505 |doi=10.1198/jasa.2009.ap08746 }}</ref><br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}. So under some regularity conditions, such weights would provide estimators for the treatment effects of interest. In essence, the method uses the idea of matching and using the training data pre-intervention to set up the weights and hence a relevant control post-intervention.<br />
<br />
:Y^N_{1t}=\Sigma^J_{j=2}w_{j}Y_{jt} <br />
for t>T_{0}.因此,在一定的正则性条件下,这种权重可以作为利息处理效果的估计量。本质上,该方法采用了匹配的思想,并利用训练数据进行干预前的权重设置,从而得到干预后的相应控制。<br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth,<ref>{{cite journal |last=Cavallo |first=E. |first2=S. |last2=Galliani |first3=I. |last3=Noy |first4=J. |last4=Pantano |year=2013 |title=Catastrophic Natural Disasters and Economic Growth |journal=[[Review of Economics and Statistics]] |volume=95 |issue=5 |pages=1549–1561 |doi=10.1162/REST_a_00413 |url=http://www.economics.hawaii.edu/research/workingpapers/WP_10-6.pdf }}</ref> and studies linking political murders to house prices.<ref>{{cite journal |last=Gautier |first=P. A. |first2=A. |last2=Siegmann |first3=A. |last3=Van Vuuren |year=2009 |title=Terrorism and Attitudes towards Minorities: The effect of the Theo van Gogh murder on house prices in Amsterdam |journal=[[Journal of Urban Economics]] |volume=65 |issue=2 |pages=113–126 |doi=10.1016/j.jue.2008.10.004 }}</ref> <br />
<!-- THE CITATION AT THE END OF THIS SENTENCE IS FOR A PAPER ABOUT "synthetic cohort models" (a.k.a. "pseudo-panel approach," using repeated cross-sections), WHICH IS NOT THE SAME AS "synthetic control": Yet, despite its intuitive appeal, it may be the case that synthetic controls could suffer from significant finite sample biases.<ref>{{cite journal |last=Devereux |first=P. J. |year=2007 |title=Small-sample bias in synthetic cohort models of labor supply |journal=[[Journal of Applied Econometrics]] |volume=22 |issue=4 |pages=839–848 |doi=10.1002/jae.938 }}</ref> --><br />
<br />
Synthetic controls have been used in a number of empirical applications, ranging from studies examining natural catastrophes and growth, and studies linking political murders to house prices. <br />
<br />
<br />
综合控制已经被应用于许多实证研究中,从研究自然灾害和经济增长,到研究政治谋杀与房价之间的联系。<br />
<br />
==References==<br />
{{Reflist|30em}}<br />
<br />
[[Category:Design of experiments]]<br />
[[Category:Statistical methods]]<br />
[[Category:Observational study]]<br />
[[Category:Econometric modeling]]<br />
<br />
Category:Design of experiments<br />
Category:Statistical methods<br />
Category:Observational study<br />
Category:Econometric modeling<br />
<br />
类别: 实验设计类别: 统计方法类别: 观察性研究类别: 计量经济模型<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Synthetic control method]]. Its edit history can be viewed at [[合成对照/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%88%86%E5%B1%82%E9%9A%8F%E6%9C%BA%E8%AF%95%E9%AA%8C&diff=23924分层随机试验2021-06-25T02:29:55Z<p>Aceyuan:修正目录层次</p>
<hr />
<div><br />
[[File:Graphic_breakdown_of_stratified_random_sampling.jpeg|thumb|220x220px|分层随机抽样的图形分解 Graphic breakdown of stratified random sampling]]<br />
<br />
在统计学中,<font color="#ff8000"> '''分层随机试验 Stratified randomization''' </font>是一种抽样方法,首先将整个研究<font color="#ff8000"> '''总体 Population''' </font>层为具有相同属性或特征的子群,称为<font color="#ff8000"> '''分层 Attributes''' </font>,然后从分层组中进行简单随机抽样,在抽样过程的任何阶段,随机、完全偶然地无偏抽取同一子群中的元素。<ref name=":3" /><ref>{{Citation|title=Simple random sample|date=2020-03-18|url=https://en.wikipedia.org/w/index.php?title=Simple_random_sample&oldid=946144051|work=Wikipedia|language=en|access-date=2020-04-07}}</ref>分层随机试验被认为是<font color="#ff8000"> '''分层抽样 Stratified sampling''' </font>的一个细分。当共享属性部分存在,并且在被调查总体的不同亚群之间有很大差异时,应该采用分层随机试验。因此,在取样过程中需要特别考虑或明确区分。<ref>{{Citation|title=Stratified sampling|date=2020-02-09|url=https://en.wikipedia.org/w/index.php?title=Stratified_sampling&oldid=939938944|work=Wikipedia|language=en|access-date=2020-04-07}}</ref>这种抽样方法应区别于<font color="#ff8000"> '''整群抽样方法 Cluster sampling''' </font>,整群抽样方法是在整个群体中选择一个简单的随机抽样来代表整个总体,或分层系统抽样方法,在分层过程之后进行<font color="#ff8000"> '''系统抽样 Systematic sampling''' </font>。分层随机抽样有时也称为<font color="#ff8000"> '''定额随机抽样 Quota random sampling''' </font>。<ref name=":3">{{Cite web|url=https://www.investopedia.com/ask/answers/032615/what-are-some-examples-stratified-random-sampling.asp|title=How Stratified Random Sampling Works|last=Nickolas|first=Steven|date=July 14, 2019|website=Investopedia|language=en|access-date=2020-04-07}}</ref><br />
<br />
== 分层随机试验的步骤 Steps for stratified randomization==<br />
<br />
分层随机试验在目标总体异<font color="#ff8000"> '''质性 Heterogeneous'''</font>的情况下非常有用, 它能有效地显示研究中的趋势或特征在不同阶层之间的差异。<ref name=":3" />当进行分层随机试验时,应采取以下8个步骤:<ref name=":4">{{Cite web|url=https://www.statisticshowto.com/stratified-random-sample/|title=Stratified Random Sample: Definition, Examples|last=Stephanie|date=Dec 11, 2013|website=Statistics How To|language=en-US|access-date=2020-04-07}}</ref><ref name=":5">{{Cite web|url=https://www.questionpro.com/blog/stratified-random-sampling/|title=Stratified Random Sampling: Definition, Method and Examples|date=2018-03-13|website=QuestionPro|language=en|access-date=2020-04-07}}</ref><br />
<br />
#定义目标总体<br />
#定义分层<font color="#ff8000"> '''变量 Variables''' </font>并决定要创建的分层数量。确定分层变量的标准,包括年龄、社会经济地位、国籍、种族、教育程度等,并应与研究目标相一致。理想情况下,应该使用4-6个阶层,因为任何分层变量的增加将提高其中一些变量抵消其他变量的影响的概率。<ref name=":5" /><br />
#使用<font color="#ff8000"> '''抽样框架 Sampling frame''' </font>评估目标总体中的所有元素。之后根据<font color="#ff8000"> '''覆盖率 Coverage''' </font> 和分组进行更改。<br />
#列出所有的元素并考虑抽样结果。每个阶层应该相互排斥 Mutually exclusive,加起来涵盖总体的所有成员,而总体的每一个成员应该属于唯一的阶层,和其他差异最小的成员一起。<ref name=":4" /><br />
#决定随机抽样的选择标准。这可以手动完成,也可以用设计好的计算机程序完成。<br />
#为所有元素分配一个随机且唯一的编号,然后根据分配的编号对这些元素进行排序。<br />
#回顾每一层的大小(Size)和每一层中所有元素的<font color="#ff8000"> '''数值分布 Numerical distribution''' </font>。确定抽样类型,按比例或不按比例分层抽样。<br />
#按照第5步中的规定进行所选的随机抽样。至少,必须从每个阶层中选择一种元素,以便最终样品包括每个阶层的代表。如果从每个阶层中选择两个或两个以上的元素,则可以计算所收集数据的<font color="#ff8000"> '''误差范围 Error margins''' </font>。<ref name=":5" /><br />
<br />
== 技术 Techniques ==<br />
[[File:Simple_random_sampling_after_stratification_step.png|thumb|分层后简单随机抽样]]<br />
<br />
分层随机试验决定一个或多个预后因素(prognostic factors),在平均意义下,这些预后因素使每个亚组具有相似的进入特征。通过检查先前研究的结果,可以准确地确定患者因素。<ref>{{Cite journal|last=Sylvester|first=Richard|date=December 1982|title=Fundamentals of clinical trials|journal=Controlled Clinical Trials|volume=3|issue=4|pages=385–386|doi=10.1016/0197-2456(82)90029-0|issn=0197-2456}}</ref><br />
<br />
子群的数量可以通过乘以每个因素的层数来计算。在随机化前或随机化时测量因素,并根据测量结果将实验对象分为若干亚组或层。<br />
<br />
在每一层中,可以应用几种随机试验策略,包括<font color="#ff8000"> '''简单随机试验 Simple randomization''' </font>、<font color="#ff8000"> '''分块随机试验 Blocked randomization''' </font>和<font color="#ff8000"> '''最小化试验 Minimization''' </font>。<br />
=== 分层内简单随机抽样 Simple randomization within strata ===<br />
<br />
简单随机试验被认为是在每个阶层中分配受试者的最简单方法。对于每个任务,受试者被完全随机地分配到每个组中。尽管简单的随机化很容易进行,但由于取样量小,分配不均,因此在含有100多个样本的地层中,通常采用简单的随机化方法。尽管很容易进行,但简单随机试验通常应用于包含 100 个以上样本的层,因为小样本量会使分配不均等。<ref name=":0" /><br />
<br />
===分层内的区块随机试验 Block randomization within strata===<br />
<br />
'''<font color="#ff8000"> 区块随机试验 Block randomization </font>''',有时称为置换区块随机试验,应用区块将来自同一阶层的受试者平均分配到研究中的每个组。 在区块随机试验中,指定了分配比率(一个特定组与其他组的数量之比)和组大小。 块大小必须是处理次数的倍数,以便每个层中的样本可以按预期比例分配到处理组。<ref name=":0">{{Cite book|last=Pocock, Stuart J.|title=Clinical trials : a practical approach|publisher=John Wiley & Sons Ltd|date=Jul 1, 2013|isbn=978-1-118-79391-6|location=Chichester|oclc=894581169}}</ref>例如,在一项关于乳腺癌的临床试验中,应该有 4 或 8 个层次,其中年龄和淋巴结状态是两个预后因素(prognostic factors),每个因素分为两个水平。 可以通过多种方式将不同的区块分配给样本,包括随机列表(random list)和计算机编程。<ref>{{Cite web|url=https://www.sealedenvelope.com/help/redpill/latest/block/|title=Sealed Envelope {{!}} Random permuted blocks|date=Feb 25, 2020|website=www.sealedenvelope.com|access-date=2020-04-07}}</ref><ref>{{Citation|last1=Friedman|first1=Lawrence M.|title=Introduction to Clinical Trials|date=2010|work=Fundamentals of Clinical Trials|pages=1–18|publisher=Springer New York|isbn=978-1-4419-1585-6|last2=Furberg|first2=Curt D.|last3=DeMets|first3=David L.|doi=10.1007/978-1-4419-1586-3_1}}</ref><br />
<br />
区块随机试验通常用于样本量较大的实验,以避免具有重要特征的样本分配不平衡。 在某些对随机试验有严格要求的领域,如临床试验,当没有对导体(conductors)进行盲法处理且区块块大小有限时,分配是可预测的。 分层中的块置换随机试验可能会随着分层数量的增加和样本量的限制而导致分层之间的样本不平衡,例如,有可能找不到符合某些分层特征的样本<ref>{{Cite book|title=Fundamentals of clinical trials|others=Friedman, Lawrence M., 1942-, Furberg, Curt,, DeMets, David L., 1944-, Reboussin, David,, Granger, Christopher B.|date=27 August 2015|isbn=978-3-319-18539-2|edition=Fifth|location=New York|oclc=919463985}}</ref>。<br />
<br />
===最小化方法 Minimization method===<br />
<br />
为了保证每个处理组的相似性,尝试了“最小化”方法,这比分层内的随机排列块更直接。在最小化方法中,根据每个处理组中的样本总和将每个层中的样本分配到处理组中,这使得受试者数量在组间保持平衡。<ref name=":0" /> 如果多个治疗组的总和相同,则将进行简单的随机化以分配治疗。在实践中,最小化方法需要根据预后因素(prognostic factors)跟踪治疗分配的每日记录,这可以通过使用一组索引卡进行记录来有效完成。最小化方法有效地避免了组间不平衡,但比块随机化涉及的随机过程更少,因为随机过程仅在治疗的总人数相同时进行。一个可行的解决方案是应用额外的随机列表,这使得具有较小边际总数的总和的治疗组具有更高的机会(例如 ¾),而其他治疗具有较低的机会(例如 ¼)。<ref name=":1">{{Cite journal|last=Pocock|first=S. J.|date=March 1979|title=Allocation of Patients to Treatment in Clinical Trials|journal=Biometrics|volume=35|issue=1|pages=183–197|doi=10.2307/2529944|jstor=2529944|pmid=497334|issn=0006-341X}}</ref><br />
<br />
==应用 Application==<br />
<br />
[[File:Confounding_factors_are_important_to_consider_in_clinical_trials.png|thumb|219x219px|混杂因素在临床试验中很重要]]<br />
<br />
分层随机试验在需要对特定层进行不同权重的情况下非常有用且富有成效。 通过这种方式,研究人员可以操纵每个层次的选择机制,以放大或最小化调查结果中所需的特征。<ref>{{Cite web|url=https://www.thoughtco.com/stratified-sampling-3026731|title=Understanding Stratified Samples and How to Make Them|last=Crossman|first=Ashley|date=Jan 27, 2020|website=ThoughtCo|language=en|access-date=2020-04-07}}</ref><br />
<br />
当研究人员打算寻找两个或多个层次之间的关联时,分层随机化很有帮助,因为简单的随机抽样会导致更大的可能出现目标群体的不平等代表性。当研究人员希望消除观察性研究中的'''<font color="#ff8000"> 混杂因素 Confounder </font>'''时,它也很有用,因为分层随机试验允许调整'''<font color="#ff8000"> 协方差 Covariances </font>'''和 '''<font color="#ff8000"> p 值 p-values </font>'''以获得更准确的结果。 <ref>{{Cite book|last=Hennekens, Charles H.|title=Epidemiology in medicine|date=1987|publisher=Little, Brown|others=Buring, Julie E., Mayrent, Sherry L.|isbn=0-316-35636-0|edition=1st|location=Boston, Massachusetts|oclc=16890223}}</ref><br />
<br />
与简单随机抽样相比,分层随机抽样的统计准确度也更高,因为选择代表总体的元素具有高度相关性。<ref name=":5" />与分层之间的差异相比,分层内的差异要小得多。因此,随着样本间差异的最小化,'''<font color="#ff8000"> 标准差 Standard deviation </font>'''也会随之收紧,从而导致最终结果的准确性更高,误差更小。当研究资金紧张时,这有效地减少了所需的样本量并提高了抽样的'''<font color="#ff8000"> 成本效益 Cost-effectiveness </font>'''。<br />
<br />
在现实生活中,分层随机试验可应用于选举投票结果、社会群体收入差距调查或各国教育机会的衡量。 <ref name=":3" /><br />
<br />
==临床试验中的分层随机试验 Stratified randomization in clinical trials==<br />
<br />
在'''<font color="#ff8000"> 临床试验 Clinical trials </font>'''中,根据患者的社会和个人背景或与研究相关的任何因素对患者进行分层,以匹配整个患者群体中的每个组。 这样做的目的是建立临床/预后因素(prognostic factor)的平衡,因为如果研究设计不平衡,试验将不会产生有效的结果。<ref>{{Cite book|last1=Polit|first1=DF|title=Nursing Research: Generating and Assessing Evidence for Nursing Practice, 9th ed.|last2=Beck|first2=CT|publisher=Lippincott Williams & Wilkins.|year=2012|location=Philadelphia, USA: Wolters Klower Health}}</ref> 分层随机化的步骤非常重要,因为它试图确保没有偏见、有意或无意地影响所研究患者样本的代表性。 <ref>{{Cite web|url=https://www.omixon.com/patient-stratification-in-clinical-trials/|title=Patient Stratification in Clinical Trials|date=2014-12-01|website=Omixon {{!}} NGS for HLA|language=en-US|access-date=2020-04-26}}</ref> 它增加了研究能力,尤其是在小型临床试验中(n<400),因为这些已知的临床特征分层被认为会影响干预的结果。<ref>{{Cite web|url=https://www.statisticshowto.com/stratified-randomization/|title=Stratified Randomization in Clinical Trials|last=Stephanie|date=2016-05-20|website=Statistics How To|language=en-US|access-date=2020-04-26}}</ref>它有助于防止在临床研究中受到高度重视的 '''<font color="#ff8000"> I 型错误 Type I error </font>'''的发生。 <ref name=":6">{{Cite journal|last=Kernan|first=W|date=Jan 1999|title=Stratified Randomization for Clinical Trials|journal=Journal of Clinical Epidemiology|volume=52|issue=1|pages=19–26|doi=10.1016/S0895-4356(98)00138-3|pmid=9973070}}</ref>它还对主动对照等效试验的样本量产生重要影响,并且在理论上有助于'''<font color="#ff8000"> 亚组分析 Subgroup analysis </font>'''和'''<font color="#ff8000"> 中期分析 Interim analysis </font>'''。 <ref name=":6" /><br />
<br />
==优势 Advantage==<br />
分层随机试验的优点包括:<br />
<br />
#分层随机试验可以准确反映一般人群的结果,因为应用影响因素对整个样本进行分层并平衡样本在治疗组之间的重要特征。例如,采用分层随机化从人群中抽取 100 名样本可以保证每个治疗组的男女平衡,而使用简单随机化可能会导致一组只有 20 名男性,而另一组有 80 名男性。<ref name=":0" /><br />
#分层随机试验比其他抽样方法(例如'''<font color="#ff8000"> 整群抽样 Cluster sampling </font>'''、简单随机抽样 和'''<font color="#ff8000"> 系统抽样 Systematic sampling </font>'''或'''<font color="#ff8000"> 非概率方法 Non-probability methods </font>''')的误差更小,因为可以使分层内的测量具有较低的标准差。在某些情况下,将分割的分层随机试验比简单地随机试验一般样本更易于管理且成本更低。<ref name=":1" /><br />
#由于分层随机试验本质的精确性,团队更容易接受分层样本的训练。<br />
#由于这种方法的统计准确性,研究人员可以通过分析小样本得到非常有用的结果。<br />
#这种抽样技术涵盖了广泛的总体,因为已经对分层划分进行了完整的控制。<br />
# 有时需要分层随机试验来估计总体中各组的总体参数。<ref name=":1" /><br />
<br />
==缺点 Disadvantage ==<br />
<br />
分层随机试验的限制包括:<br />
<br />
#分层随机试验首先参考预后因素将样本分成若干层,但有可能无法划分样本。在应用中,在某些情况下,预后因素的重要性缺乏严格的认可,这可能进一步导致偏差。这就是为什么在将因素纳入分层之前应该检查因素产生影响的潜力的原因。在某些因素对结果的影响无法得到批准(approved)的情况下,建议进行无分层随机试验。 <ref>{{Cite web|url=https://www.investopedia.com/ask/answers/041615/what-are-advantages-and-disadvantages-stratified-random-sampling.asp|title=Pros and Cons of Stratified Random Sampling|last=Murphy|first=Chris B.|date=Apr 13, 2019|website=Investopedia|language=en|access-date=2020-04-07}}</ref><br />
#如果可用数据不能代表整个亚组总体,则认为亚组大小具有相同的重要性。在某些应用中,子组大小是根据可用数据量来决定的,而不是将样本大小缩放到子组大小,这会在因子效应中引入偏差。在某些需要对数据进行方差分层的情况下,子组方差差异显着,使得每个子组的抽样规模无法保证与整个子组总体成正比。<ref name=":2">{{Citation|last1=Glass|first1=Aenne|title=Potential Advantages and Disadvantages of Stratification in Methods of Randomization|date=2014|work=Springer Proceedings in Mathematics & Statistics|pages=239–246|publisher=Springer New York|isbn=978-1-4939-2103-4|last2=Kundt|first2=Guenther|doi=10.1007/978-1-4939-2104-1_23}}</ref><br />
#如果人口不能完全分配到层中,则不能应用分层抽样,这将导致样本大小与可用样本成正比,而不是与总体子组人口成正比。<ref name=":0" /><br />
#如果受试者符合多层次的纳入标准,则将样本分配到亚组的过程可能涉及重叠,这可能导致总体的错误陈述。<ref name=":2">{{Citation|last1=Glass|first1=Aenne|title=Potential Advantages and Disadvantages of Stratification in Methods of Randomization|date=2014|work=Springer Proceedings in Mathematics & Statistics|pages=239–246|publisher=Springer New York|isbn=978-1-4939-2103-4|last2=Kundt|first2=Guenther|doi=10.1007/978-1-4939-2104-1_23}}</ref><br />
<br />
==参考文献==<br />
{{Citation|last1=Glass|first1=Aenne|title=Potential Advantages and Disadvantages of Stratification in Methods of Randomization|date=2014|work=Springer Proceedings in Mathematics & Statistics|pages=239–246|publisher=Springer New York|isbn=978-1-4939-2103-4|last2=Kundt|first2=Guenther|doi=10.1007/978-1-4939-2104-1_23}}</ref></div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=23923倾向得分匹配2021-06-24T16:47:01Z<p>Aceyuan:补充遗漏内容</p>
<hr />
<div>{{#seo:<br />
|keywords=统计分析,评估<br />
|description=是一种用于估计治疗、政策或其他干预的效果统计匹配技术<br />
}}<br />
<br />
在观察数据的统计分析中,'''倾向性评分匹配 Propensity Score Matching (PSM)'''是一种用于估计治疗、政策或其他干预的效果统计匹配技术,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆 Paul R. Rosenbaum和唐纳德·鲁宾 Donald Rubin在1983年介绍了这项技术。<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化分配机制意味着每个协变量将在处理组和对照组中呈现类似的分布。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单地通过对比评估吸烟者和不吸烟者来估计平均处理效果将产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
<br />
==综述==<br />
PSM适用于非实验环境中[[因果推断]]和简单选择偏差的情况,其中: (i)对照组与处理组中的类似单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的协变量特征进行比较。<br />
<br />
<br />
在常规的匹配机制中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有显著的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
==一般步骤==<br />
<br />
1. 做Logistic回归:<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. 依照倾向性评分的估计量进行分层,检查协变量的倾向性评分的估计量在每层处理组和对照组是否均衡<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. 根据倾向性评分的估计量,将每个处理组个体与一个或多个对照组个体进行匹配,使用以下方法之一:<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
*[[Stratified sampling|分层匹配]]<br />
*双重差分匹配(核和局部线性加权)<br />
*精确匹配<br />
<br />
<br />
4. 对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡<br />
<br />
<br />
5. 基于新样本的多变量分析<br />
* 如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
==形式定义 ==<br />
===基本设置===<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55" />是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]个体。每个个体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示个体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个个体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。个体编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些个体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===强可忽略处理分配===<br />
设某个物体有协变量''X''(即:条件非混杂变量)向量,以及对应着控制和处理两种情况的'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp \!\!\!\! \perp Z \mid X </math><br />
<br />
这里<math>\perp \!\!\!\! \perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55" /><br />
<br />
<br />
=== 平衡得分===<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp \!\!\!\! \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
<br />
===倾向性评分 ===<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。倾向性评分匹配将使得处理组和对照组的协变量分布趋同,从而减少选择偏差。<br />
<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后这些倾向性评分即可作为用于逆概率加权方法的权重估计量。<br />
<br />
<br />
===主要定理===<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55" /><br />
<br />
*倾向性评分<math>e(x)</math>是平衡得分。<br />
*任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成只有一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数(保留全部维度)。<br />
*如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
:::<math> (r_0, r_1) \perp \!\!\!\! \perp Z \mid e(X).</math><br />
:*对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
*利用平衡得分的样本估计可产生在''X''上均衡的样本<br />
<br />
<br />
===与充分性的关系 ===<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小充分统计量。<br />
<br />
<br />
===混杂变量的图检测方法===<br />
<br />
朱迪亚·珀尔 Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
<br />
== 缺点 ==<br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。<ref>{{Cite journal|last=King|first=Gary|last2=Nielsen|first2=Richard|date=2019-05-07|title=Why Propensity Scores Should Not Be Used for Matching|journal=Political Analysis|volume=27|issue=4|pages=435–454|doi=10.1017/pan.2019.11|issn=1047-1987|doi-access=free}} | [https://gking.harvard.edu/files/gking/files/psnot.pdf link to the full article] (from the author's homepage)</ref>匹配方法背后的直观仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个协变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维数问题“,即每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。<ref>{{cite journal |vauthors=Garrido MM, etal |year=2014 |title=Methods for Constructing and Assessing Propensity Scores |journal= Health Services Research |doi= 10.1111/1475-6773.12182 |pmid= 24779867 |pmc=4213057 |volume=49 |issue=5 |pages=1701–20}}</ref>由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。<ref>{{cite book |last=Shadish |first=W. R. |last2=Cook |first2=T. D. |last3=Campbell |first3=D. T. |year=2002 |title=Experimental and Quasi-experimental Designs for Generalized Causal Inference |location=Boston |publisher=Houghton Mifflin |isbn=978-0-395-61556-0 }}</ref>另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
Judea Pearl也提出了关于匹配方法的普遍担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。<ref name=pearl:ch11-3-5>{{cite book |last=Pearl |first=J. |chapter=Understanding propensity scores |title=Causality: Models, Reasoning, and Inference |location=New York |publisher=Cambridge University Press |edition=Second |year=2009 |isbn=978-0-521-89560-6 }}</ref>当试验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它也很容易地手动实现。<ref name="pearl"/><br />
<br />
<br />
==统计包中的实现==<br />
* [[R (programming language)|R]]: 倾向得分匹配作为 <code>MatchIt</code> 包的一部分提供。<ref>{{cite journal |first=Daniel |last=Ho |first2=Kosuke |last2=Imai |first3=Gary |last3=King |author3-link=Gary King (political scientist) |first4=Elizabeth |last4=Stuart |year=2007 |title=Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference |journal=[[Political Analysis (journal)|Political Analysis]] |volume=15|issue=3 |pages=199–236 |doi=10.1093/pan/mpl013 |doi-access=free }}</ref><ref>{{cite web |title=MatchIt: Nonparametric Preprocessing for Parametric Causal Inference |work=R Project |url=https://cran.r-project.org/package=MatchIt }}</ref> 它也可以很容易地手工实现。<ref>{{cite book |first=Andrew |last=Gelman |first2=Jennifer |last2=Hill |title=Data Analysis Using Regression and Multilevel/Hierarchical Models |location=New York |publisher=Cambridge University Press |year=2007 |isbn=978-0-521-68689-1 |pages=206–212 |url=https://books.google.com/books?id=lV3DIdV0F9AC&pg=PA206 }}</ref><br />
* [[SAS_(software)|SAS]]: PSMatch过程,以及宏 <code>OneToManyMTCH</code>可根据倾向得分对观察数据进行匹配。<ref>{{cite web<br />
| first =Lori<br />
| last =Parsons<br />
| title =Performing a 1:N Case-Control Match on Propensity Score<br />
| publisher =SAS Institute<br />
| location =SUGI 29<br />
| url =http://www2.sas.com/proceedings/sugi29/165-29.pdf<br />
| access-date =June 10, 2016}}</ref><br />
* [[Stata]]: 有几个命令实现了倾向得分匹配,<ref>[http://fmwww.bc.edu/RePEc/usug2001/psmatch.pdf Implementing Propensity Score Matching Estimators with STATA]. Lecture notes 2001</ref> 包括用户编写的<code>psmatch2</code>。<ref>{{cite paper |first=E. |last=Leuven|author-link2=Barbara Sianesi|first2=B. |last2=Sianesi |date=2003 |title= PSMATCH2: Stata module to perform full Mahalanobis and propensity score matching, common support graphing, and covariate imbalance testing |url=http://ideas.repec.org/c/boc/bocode/s432001.html }}</ref> Stata 13 及更高版本还提供了内置命令 <code>teffects psmatch</code>。<ref>{{cite web |title=teffects psmatch — Propensity-score matching |work=Stata Manual |url=https://www.stata.com/manuals15/teteffectspsmatch.pdf }}</ref><br />
* [[SPSS]]: IBM SPSS Statistics菜单(数据/倾向评分匹配)中提供了一个倾向评分匹配对话框,允许用户设置匹配容差、抽取样本时随机化案例顺序、确定精确匹配的优先级、样本有或无替换、设置一个随机种子,并通过提高处理速度和最小化内存使用来最大化性能。 FUZZY Python过程也可以通过扩展对话框轻松添加为软件的扩展。此过程基于一组指定的关键变量,通过从控制中随机抽取来匹配案例和控制。FUZZY命令支持精确匹配和模糊匹配。<br />
<br />
==其他词条==<br />
*[[Rubin causal model|鲁宾因果框架]]<br />
*[[Ignorability|可忽略性]]<br />
*[[Heckman correction|赫克曼校正]]<br />
*[[Matching (statistics)|匹配 ]]<br />
<br />
<br />
<br />
==参考文献==<br />
<references /><br />
<br />
<br />
==编者推荐==<br />
===书籍推荐===<br />
[[File:统计因果推理入门.jpg|200px|thumb|right|《统计因果推理入门》封面|链接=https://wiki.swarma.org/index.php%3Ftitle=%E6%96%87%E4%BB%B6:%E7%BB%9F%E8%AE%A1%E5%9B%A0%E6%9E%9C%E6%8E%A8%E7%90%86%E5%85%A5%E9%97%A8.jpg]]<br />
*[https://wiki.swarma.org/index.php%3Ftitle=%E7%BB%9F%E8%AE%A1%E5%9B%A0%E6%9E%9C%E6%8E%A8%E7%90%86%E5%85%A5%E9%97%A8 统计因果推理入门] 对应英文[https://wiki.swarma.org/index.php%3Ftitle=Causal_Inference_in_Statistics:_A_Primer Causal Inference in Statistics: A Primer]<br />
关于因果的讨论很多,但是许多入门的教材只是为没有统计学基础的读者介绍如何使用统计学技术处理因果性问题,而没有讨论因果模型和因果参数,本书希望协助具有基础统计学知识的教师和学生应对几乎在所有自然科学和社会科学非试验研究中存在的因果性问题。本书聚焦于用简单和自然的方法定义因果参数,并且说明在观察研究中,哪些假设对于估计参数是必要的。我们也证明这些假设可以用显而易见的数学形式描述出来,也可以用简单的数学工具将这些假设转化为量化的因果关系,如治疗效果和政策干预,以确定其可检测的内在关系。<br />
*[https://wiki.swarma.org/index.php%3Ftitle=Counterfactuals_and_Causal_Inference:_Methods_and_Principles_for_Social_Research Counterfactuals and Causal Inference: Methods and Principles for Social Research]<br />
===课程推荐===<br />
*[https://campus.swarma.org/course/2526 两套因果框架深度剖析:潜在结果模型与结构因果模型]<br />
::这个视频内容来自[https://wiki.swarma.org/index.php%3Ftitle=%E9%9B%86%E6%99%BA%E4%BF%B1%E4%B9%90%E9%83%A8%E8%AF%BB%E4%B9%A6%E4%BC%9A 集智俱乐部读书会]-因果科学与Causal AI读书会第二季内容的分享,由英国剑桥大学及其学习组博士陆超超详细的阐述了潜在结果模型和结果因果模型,并介绍了两个框架的相互转化规律。<br />
::1. 讲述因果推断的两大框架:潜在结果模型和结构因果模型,讨论他们各自的优缺点以及他们之间的联系,详细介绍他们之间的转化规律。<br />
*[https://www.bilibili.com/video/BV1NJ411w7ms?from=search&seid=15960075946481426104 Average Effect of Treatment on the Treated (ATT) 实验组的平均干预效应/匹配方法]<br />
::B站搬运的杜克大学社会科学研究中心的分享视频,介绍了在使用匹配方法时会涉及到的ATT、CATE、ATE的方法。<br />
*[https://www.bilibili.com/video/BV19741137L2?from=search&seid=13934883753123755445 倾向性匹配得分]B站Up主分享的倾向性匹配得分的基本概念和R语言实现过程。<br />
*B站up主PSM系列视频: ([https://www.bilibili.com/video/BV1gV41117Md 一)基础知识][https://www.bilibili.com/video/BV1CK4y1E7sf (二)匹配估计量][https://www.bilibili.com/video/BV1Hf4y1q7Zz (三)倾向性得分匹配] [https://www.bilibili.com/video/BV1Az4y1C7UB/?spm_id_from=333.788.recommend_more_video.-1 原理和实践: PSM倾向性得分匹配最详细的讲解(四)]<br />
*[https://campus.swarma.org/course/2030 潜结果框架下的因果效应]<br />
什么是因果呢?“因”其实就是引起某种现象发生的原因,而“果”就是某种现象发生后产生的结果。因果问题在我们日常生活中十分常见,但是不管是传统的统计学还是当下很火的大数据、机器学习,更多的是解决相关性的问题。因果问题存在于很多领域,如医疗健康、经济、政治科学、数字营销等。该课程是由浙江大学助理教授况琨讲授的,主要回答以下一些重要的问题:因果性与相关性的区别是什么?相关性有哪几种来源?如何评估因果效应?有哪些常用且前沿的方法?<br />
===文章总结===<br />
*[https://mp.weixin.qq.com/s/f-rI5W6tc6qOzthbzK4oAw 崔鹏:稳定学习——挖掘因果推理和机器学习的共同基础]<br />
<br />
*知乎上RandomWalk总结的关于因果推断之Potential Outcome Framework的内容,其中提到因果退镀and额目标就是从观测数据中估计treatment effect。<br />
<br />
*Mesonychid在自己的个人主页上分享的关于[https://hanyuz1996.github.io/2017/08/30/Donald-Rubin/ Donald-Rubin潜在结果模型]的解释。<br />
<br />
*Yishi Lin在自己的个人主页上分享的关于因果推断的一些介绍[https://dango.rocks/blog/2019/01/08/Causal-Inference-Introduction1/ 因果推断漫谈(一):掀开 “因果推断” 的面纱]<br />
*[https://swarma.org/?p=22045 《因果科学周刊》第2期:如何解决混淆偏差?]本文围绕因果科学领域的“混淆偏差”问题展开介绍,并进行了相关论文的推荐。<br />
*[https://zhuanlan.zhihu.com/p/237723948 倾向得分匹配(PSM)的原理与步骤]这篇知乎文章里,详细介绍了PSM在stata的实现过程。<br />
*[https://zhuanlan.zhihu.com/p/46502579 用R实现倾向性评分代码]这篇知乎文章中介绍了如何使用R实现倾向性评分。<br />
===相关路径===<br />
*[https://pattern.swarma.org/path?id=99 因果科学与Casual AI读书会必读参考文献列表],这个是根据读书会中解读的论文,做的一个分类和筛选,方便大家梳理整个框架和内容。<br />
*[https://pattern.swarma.org/path?id=9 因果推断方法概述],这个路径对因果在哲学方面的探讨,以及因果在机器学习方面应用的分析。<br />
*[https://pattern.swarma.org/path?id=90 因果科学和 Causal AI入门路径],这条路径解释了因果科学是什么以及它的发展脉络。此路径将分为三个部分进行展开,第一部分是因果科学的基本定义及其哲学基础,第二部分是统计领域中的因果推断,第三个部分是机器学习中的因果(Causal AI)。<br />
<br />
<br />
----<br />
本中文词条由[[用户:Aceyuan|Aceyuan]]翻译、[[用户:李昊轩|李昊轩]]审校,[[用户:薄荷|薄荷]]编辑,欢迎在讨论页面留言。<br />
<br />
'''本词条内容源自wikipedia及公开资料,遵守 CC3.0协议。'''<br />
<br />
<br />
[[Category:统计分析]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E7%BB%93%E6%9E%84%E5%9B%A0%E6%9E%9C%E6%A8%A1%E5%9E%8B&diff=23922结构因果模型2021-06-24T16:44:50Z<p>Aceyuan:修正标题层级</p>
<hr />
<div>{{#seo:<br />
|keywords=结构因果模型,因果科学,因果之梯<br />
|description=结构因果模型,因果科学,因果之梯<br />
}}<br />
<br />
<br />
[[File:Diagram of Dynamic Causal Modelling - Causal Modelling and Brain Connectivity in Functional Magnetic Resonance Imaging by Karl Friston.png|thumb|300px|比较两个竞争的因果模型(DCM,GCM)用于解释[[fMRI 图像]]<ref>{{cite journal | doi=10.1371/journal.pbio.1000033 | pmid=19226186 | pmc=2642881 | author=Karl Friston | title=Causal Modelling and Brain Connectivity in Functional Magnetic Resonance Imaging | journal=PLOS Biology| volume=7 | number=2 | pages=e1000033 | date=Feb 2009}}</ref>]]<br />
<br />
<br />
在科学哲学中,'''<font color="#ff8000"> 因果模型 Causal Model</font>'''或'''<font color="#ff8000"> 结构因果模型 Structural Causal Model</font>''',是描述系统因果机制的概念模型。因果模型可以通过提供清晰的规则来决定需要考虑/控制哪些自变量,从而改进研究设计。<br />
<br />
<br />
因果模型可以从现有的观察数据中回答一些问题,而无需进行随机对照实验等干预性研究。一些干预性研究由于伦理或实践的原因是不合适的,这意味着如果没有一个因果模型,一些假设就无法被检验。<br />
<br />
<br />
因果模型可以帮助解决'''<font color="#ff8000"> 外部有效性 External Validity</font>'''问题(一项研究的结果是否适用于未研究的总体)。在某些情况下,因果模型可以允许多项研究的数据合并起来回答任何单个数据集都无法回答的问题。<br />
<br />
<br />
因果模型是可证伪的,因为如果一些因果模型与数据不匹配,这些因果模型就必须作为无效模型而被拒绝接受。因果模型还必须使得研究相关现象的科学家们信服。<ref>{{Cite journal|last1=Barlas|first1=Yaman|last2=Carpenter|first2=Stanley|date=1990|title=Philosophical roots of model validation: Two paradigms|journal=System Dynamics Review|language=en|volume=6|issue=2|pages=148–166|doi=10.1002/sdr.4260060203}}</ref><br />
<br />
<br />
因果模型在信号处理、流行病学和机器学习中都有应用。<ref name=":0">[http://bayes.cs.ucla.edu/jp_home.html Pearl 2009]</ref><br />
<br />
<br />
== 定义 ==<br />
<br />
<blockquote>因果模型是表示单个系统或群体内因果关系的数学模型。它有助于从统计数据中推断因果关系,可以教会我们关于因果关系的认识论,并展现因果关系和概率之间的关系。它还被应用于哲学家感兴趣的主题,例如反事实逻辑、决策理论和实际因果关系分析。<ref>{{Citation|last=Hitchcock|first=Christopher|title=Causal Models|date=2018|url=https://plato.stanford.edu/archives/fall2018/entries/causal-models/|encyclopedia=The Stanford Encyclopedia of Philosophy|editor-last=Zalta|editor-first=Edward N.|edition=Fall 2018|publisher=Metaphysics Research Lab, Stanford University|access-date=2018-09-08}}</ref></blockquote><br />
<br />
Judea Pearl将因果模型定义为一个有序的三元组<math>\langle U, V, E\rangle</math>,其中<math> U </math>是一组外生变量,其值由模型外部的因素决定;<math>V </math>是一组内生变量,其值由模型内部的因素决定;<math>E </math>是一组结构方程,把每个内生变量的值表示为<math> U</math> 和<math> V </math>中其他变量值的函数。<ref name=":0" /><br />
<br />
== 历史 ==<br />
<br />
亚里士多德定义了因果关系的分类法,包括质料因、形式因、动力因、目的因。休谟更偏爱反事实,他拒绝了亚里士多德的分类法。有段时间,他否认物体本身具有使得一个物体成为原因而另一个物体成为结果的“力量”。<ref name=":1"> [https://book.douban.com/subject/33438811/ Pearl, Judea; Mackenzie, Dana (2018-05-15). The Book of Why: The New Science of Cause and Effect. Basic Books. ISBN 9780465097616.]</ref>后来,他接受了“如果第一物体还没存在,第二个根本不存在”的观点(“but-for”因果关系)。<ref name=":1" /><br />
<br />
<br />
19世纪末,统计学学科开始形成。经过多年努力确定诸如生物遗传等领域的因果规则后,高尔顿引入了'''<font color="#ff8000"> 均值回归 Mean Regression </font>'''的概念(以二年生症候群为缩影),后来这将他引向了非因果的相关性概念。<ref name=":1" /><br />
<br />
<br />
作为一个实证主义者,皮尔逊将因果的概念从许多科学中去除,他认为因果关系是一种无法证明的特殊的关联,并引入相关系数作为关联强度的度量方法。他写道: “作为运动原因的力,与作为成长原因的树神完全一样”,而因果关系只是“现代科学高深奥秘中的迷信”。皮尔逊在伦敦大学学院创立了期刊“Biometrika”和生物识别实验室,该实验室成为了统计领域的全球领军者。<ref name=":1" /><br />
<br />
<br />
1908年,Hardy和Weinberg通过重拾孟德尔遗传律,解决了导致高尔顿放弃因果关系的性状稳定问题。<ref name=":1" /><br />
<br />
<br />
1921年,Wright的路径分析成为因果模型和因果图的理论雏形。<ref>{{Cite book|url={{google books |plainurl=y |id=yWWEIvNgUQ4C|page=707}} |title=The Oxford Handbook of Causation |volume=1 |editor-last=Beebee |editor-first=Helen|editor-last2=Hitchcock|editor-first2=Christopher|editor-last3=Menzies|editor-first3=Peter|date=2012-01-12|publisher=OUP Oxford|isbn=9780191629464|language=en|first=Samir |last=Okasha |chapter=Causation in Biology|chapter-url=http://www.oxfordhandbooks.com/view/10.1093/oxfordhb/9780199279739.001.0001/oxfordhb-9780199279739-e-0036}}</ref>他开发了这种路径分析方法,试图同时阐明遗传、发育和环境对豚鼠皮毛模式的相对影响。他通过一个分析过程如何解释豚鼠出生体重、子宫内时间和产仔数之间的关系来支持他旁门左道的观点。杰出的统计学家对这些想法的反对使因果关系在接下来的40年中被家畜育种学家之外的科学家所忽略。取而代之的是,科学家们依赖于相关性,一定程度上是在批评Wright的领军统计学家Fisher的授意下。<ref name=":1" />唯一的例外是一名叫Burks的学生,在1926年首先应用路径图来表示中介影响,并断言保持中介变量恒定会引起误差。她可能独立地发明了路径图。<ref name=":1" /><br />
<br />
<br />
1923年,Neyman提出了潜在结果(potential outcome)的概念,但是直到1990年他的论文才被从波兰语翻译成英语。<ref name=":1" /><br />
<br />
<br />
1958年,Cox警告说,仅当Z高概率不被自变量影响的时,控制变量Z才有效。<ref name=":1" /><br />
<br />
<br />
20世纪60年代,Duncan、Blalock、Goldberger等人重新发现了路径分析。Duncan在阅读Blalock关于路径图的著作时,想起了二十年前Ogburn的一次演讲,其中提到了Wright的论文,而后又提到了Burks。<ref name=":1" /><br />
<br />
<br />
社会学家最初将因果模型称为'''<font color="#ff8000"> 结构方程模型 Structural Equation Modeling </font>''',但一旦它成为教条式方法就失去了效用,导致一些从业者拒绝与因果关系的任何联系。经济学家采用了路径分析的代数部分,称其为'''<font color="#ff8000"> 联立方程建模 Simultaneous Equation Modeling </font>'''。但是,经济学家仍然避免将因果含义赋予他们的方程式。<ref name=":1" /><br />
<br />
<br />
Wright在发表第一篇论文60年后,根据Karlin等人的批评,发表了一篇概述该论文的文章,该论文反对仅处理线性关系,而鲁棒的、非模型的数据表示方式则更具揭示性。<ref name=":1" /><br />
<br />
<br />
1973年,Lewis提倡用but-for因果关系(反事实)代替相关性。他提到了人类具有想象某个原因是否发生和结果仅在原因后发生的不同可选世界的能力。<ref name=":1" />1974年Rubin引入了“潜在结果 potential outcome”的概念,作为询问因果问题的语言。<ref name=":1" /><br />
<br />
<br />
1983年,Cartwright提出与一个结果因果相关的任何因子都是有条件的,不再以简单的概率作为唯一指导。<ref name=":1" /><br />
<br />
<br />
1986年,Baron和Kenny引入了检测和评估线性方程系统中的中介的原理。截至2014年,他们的论文是有史以来被引用最多的第33篇。<ref name=":1" />那年,Greenland和Robins通过考虑反事实,引入了“可交换性”方法,来处理混杂问题。他们提出评估如果治疗组没有接受治疗会给治疗组带来什么后果,并将其结果与对照组进行比较。如果结果一致,说明没有混杂因子。<ref name=":1" /><br />
<br />
<br />
哥伦比亚大学设有因果人工智能实验室,该实验室正试图将因果建模理论与人工神经网络联系起来。<ref>{{Cite web|url=https://www.technologyreview.com/s/615189/what-ai-still-cant-do/|title=What AI still can't do|last=Bergstein|first=Brian|website=MIT Technology Review|language=en-US|access-date=2020-02-20}}</ref><br />
<br />
<br />
== 因果关系之梯==<br />
<br />
Judea Pearl的因果元模型涉及三个层次的抽象,他称之为因果之梯。最低层的“关联”(看到/观察)需要感知输入数据中的规律性或模式,用相关性表示。中间层的“干预 ”(do)可以预测有意识行动的后果,用因果关系表示。最高层的“反事实”(想象)涉及构建部分世界的理论,该理论解释为什么特定行为会产生特定后果,以及在没有此行为的情况下会发生什么。<ref name=":1" /><br />
<br />
<br />
=== 关联 ===<br />
<br />
如果观察一个对象改变了观察另一个对象的可能性,则这个对象与另一个对象相关联。例子:购买牙膏的购物者也更有可能购买牙线。数学上用<br />
:<math>P (买牙线 | 买牙膏) </math><br />
表示已知一个人购买牙膏时的其购买牙线的可能性。关联也可以通过计算两个事件的相关性来衡量。关联并不意味着因果。一个事件可能导致另一个事件,反过来也可能,或者两个事件都可能由某个第三事件引起(牙医对口腔健康的宣传使得购物者同时购买牙线和牙膏)。<ref name=":1" /><br />
<br />
<br />
=== 干预 ===<br />
<br />
该层涉及事件之间的特定因果关系。因果是通过实验性地执行影响事件的一些动作来评估。例如:如果我们将牙膏的价格提高一倍,那么人们购买牙线的概率将是多少?因果无法通过检验历史信息来确定,因为可能存在其他因素同时影响这两个变量,比如存在牙膏价格变化的其他原因,而且这种原因会影响牙线的价格(例如两种商品的关税增加)。数学上用<br />
<br />
:<math>P (牙线价格 | do(牙膏价格)) </math><br />
<br />
表示这种概率。其中do是一个算子,表示对谁做实验性干预(如价格翻倍)。<ref name=":1" />这个算子指示了要在创造所需效果的世界中进行最小的变化,即在现实模型上进行尽可能小的改变的“小手术”。<ref>{{cite journal |last1=Pearl |first1=Judea |title=Causal and Counterfactual Inference |date=29 Oct 2019 |url=https://ftp.cs.ucla.edu/pub/stat_ser/r485.pdf |access-date=14 December 2020}}</ref><br />
<br />
<br />
=== 反事实 ===<br />
<br />
最高层的反事实涉及对过去事件的其他可能版本的考虑,或者考虑同一实验个体中在不同情况下会发生的情况。例如,如果当初那家商店的牙线价格翻了一番,那么当时那些购买牙膏的购物者仍然会购买牙线的可能性是多少?<br />
<br />
:<math>P (买牙线 | 买牙膏, 当初牙线价格翻倍) </math><br />
<br />
反事实可以表明存在因果关系。回答反事实的模型允许进行精确的干预,这些干预的后果可被预测。在极端情况下,这样的模型被人们认为是物理定律(如惯性:若不将力施加到静止物体上物体将不会移动)。<ref name=":1" /><br />
<br />
<br />
== 因果 ==<br />
=== 因果和相关 ===<br />
<br />
统计学涉及分析多个变量之间的关系。传统上,这些关系被描述为相关性,即没有任何隐含因果关系的关联。因果模型试图通过添加因果关系的概念来扩展此框架,在因果关系中,一个变量的变化导致其他变量的变化。<ref name=":0" /><br />
<br />
<br />
20世纪因果的定义完全依赖于概率或关联。如果一个事件<math> X </math>增加了另一个事件<math> Y </math>的可能性,则认为它会导致另一个事件。在数学上,这表示为:<br />
:<math>P (Y | X) > P(Y)</math><br />
这样的定义是不充分的,因为可能有其他关系(例如,<math> X </math>和<math> Y </math>的共同原因)可以满足该条件。因果与因果之梯的第二层有关。关联处于第一层,仅向第二层提供证据。<ref name=":1" /><br />
<br />
<br />
之后的定义试图通过以背景因素为条件来解决这种歧义。数学上表示为:<br />
<br />
<br />
:<math>P (Y | X, K = k) > P(Y| K = k)</math><br />
<br />
<br />
其中<math> K </math>是背景变量的集合,<math> k </math>表示特定语境中背景变量的值。但是,只要概率是唯一准则,那么所需的背景变量集是难以确定的。<ref name=":1" /><br />
<br />
<br />
定义因果的其他尝试包括'''<font color="#ff8000"> 格兰杰因果 Granger Causality </font>''',这是一种统计假设检验,在经济学中,可以通过衡量用一个时间序列的过去值预测另一个时间序列的未来值的能力,来评估序列间的因果。<ref name=":1" /><br />
<br />
<br />
=== 类型===<br />
<br />
原因可以是必要的、充分的、部分的以及它们的组合。<ref>{{Cite book|url={{google books |plainurl=y |id=skIZAQAAIAAJ|page=25}} |title=Discrete Mathematics with Applications|last=Epp|first=Susanna S.|date=2004|publisher=Thomson-Brooks/Cole|isbn=9780534359454|language=en|pages= 25–26}}</ref><br />
<br />
<br />
==== 必要因 ====<br />
对于<math> y </math>的必要因<math> x </math>,<math>y </math>的存在意味着<math> x </math>在此前发生了。但是<math> x </math>的存在不意味着y会发生。必要因也被称为“若非(but-for)”因,即<math>y</math>不会发生若非<math> x </math>发生。<ref name=":1" /><br />
<br />
<br />
==== 充分因 ====<br />
对于<math>y</math>的充分因<math>x</math>,<math>x</math>的存在意味着<math>y</math>接下来会发生。然而另一个原因<math>z</math>也可能独立地造成<math>y</math>的发生。即<math>y</math>的发生不要求<math>x</math>的发生。 <ref name="CR" >[http://www.istarassessment.org/srdims/causal-reasoning-2/ "Causal Reasoning"]. www.istarassessment.org. Retrieved 2 March 2016.</ref><br />
<br />
<br />
==== 促成因====<br />
对于<math>y</math>的促成因<math>x</math>,<math>x</math>的存在会增加<math>y</math>的似然。如果似然是100%,那么<math>x</math>就是充分的。促成因也是必要的。<ref name="Riegelman">{{Cite journal|last1=Riegelman|first1=R.|year=1979|title=Contributory cause: Unnecessary and insufficient|journal=Postgraduate Medicine|volume=66|issue=2|pages=177–179|doi=10.1080/00325481.1979.11715231|pmid=450828}}</ref><br />
<br />
<br />
== 模型 ==<br />
=== 因果图 ===<br />
<br />
因果图是一个有向图,它显示了因果模型中变量间的因果关系。因果图包括一组变量(或节点),每个节点通过箭头连接到一个或多个对其具有因果效应的其他节点。箭头描绘了因果的方向,例如,将变量<math> A </math>和变量<math>B</math> 以指向 <math>B</math> 的箭头相连表示A的变化以某种概率导致<math>B</math>的变化。一条路径是两个节点间沿着因果箭头的图的遍历。<ref name=":1" /><br />
<br />
<br />
因果图包括'''<font color="#ff8000"> 因果环图 Causal Loop Diagrams </font>''','''<font color="#ff8000"> 有向无环图 Directed Acyclic Graphs </font>'''和'''<font color="#ff8000"> 鱼骨图 Ishikawa diagrams</font>'''。<ref name=":1" /><br />
<br />
<br />
因果图和它们的定量概率无关,对这些概率的更改不需要修改因果图。<ref name=":1" /><br />
<br />
<br />
=== 模型元素 ===<br />
因果模型具有形式结构,其元素具有特定的属性。<ref name=":1" /><br />
<br />
<br />
==== 连接方式 ====<br />
三个节点的连接类型有三种,分别是线型的链,分支型的叉和合并型的对撞。<ref name=":1" /><br />
<br />
<br />
===== 链 =====<br />
链(结构)是直线连接,箭头从原因指向结果。在这个模型中,<math>B</math>是中介变量,因为它调节了<math> A</math> 对<math> C</math> 的影响。<ref name=":1" /><br />
:<math> A \rightarrow B \rightarrow C</math><br />
<br />
<br />
===== 叉 =====<br />
在叉(结构)中,一个原因有多种结果,这两种结果有一个共同的原因。 <math>A </math>和<math> C</math> 之间存在非因果的虚假相关性,可以通过把<math> B</math> 作为条件(选取<math>B</math>的特定值)来消除虚假相关性。<ref name=":1" /><br />
<br />
:<math> A \leftarrow B \rightarrow C</math><br />
<br />
“把B作为条件”是指“给定B”(即B取某个值)。某些情况下叉(结构)是混杂因子:<br />
<br />
:<math> A \leftarrow B \rightarrow C \rightarrow A</math><br />
在这样的模型中,<math> B </math>是<math> A </math>和<math> C </math>的共同原因( <math>C</math> 也是<math> A </math>的原因),这使<math>B</math>成为'''<font color="#ff8000"> 混杂因子 Confounder </font>'''。<ref name=":1" /><br />
<br />
<br />
===== 对撞 =====<br />
在对撞(结构)中,多种原因会影响一种结果。以 <math>B</math> 为条件( <math>B</math> 取特定值)通常会揭示 <math>A</math> 与<math> C</math> 之间的非因果的负相关。这种负相关被称为对撞偏差和“辩解”效应,即 <math>B</math> 解释了<math> A</math> 与 <math>C</math> 之间的相关性。<ref name=":1" /> <math>A</math> 和<math> C</math> 两者都是影响 <math>B</math> 的必要因时,该相关性是正的。<ref name=":1" /><br />
:<math> A \rightarrow B \leftarrow C</math><br />
<br />
<br />
==== 节点类型 ====<br />
===== 中介变量 =====<br />
中介变量节点修改了其他原因对结果的影响(这与原因简单地影响结果不同)。<ref name=":1" />例如,在上面的链结构中,<math>B</math>是中介变量,因为它修改了 <math>C</math> 的间接原因<math> A</math> 对结果变量 <math>C</math> 的影响。<br />
<br />
<br />
===== 混杂因子 =====<br />
混杂因子节点影响多个结果,从而在它们之间产生正相关。<ref name=":1" /><br />
<br />
<br />
===== 工具变量 =====<br />
满足如下条件的是工具变量:<ref name=":1" /><br />
<br />
<br />
#有通往结果变量的路径<br />
#没有通往其他原因变量(解释变量)的路径<br />
#对结果没有直接影响<br />
<br />
<br />
回归系数可以用作工具变量对结果的因果影响的估计,只要该影响不被混杂即可。通过这种方式,工具变量允许对因果因子进行量化,而无需有关混杂因子的数据。<ref name=":1" />例如,给定模型:<br />
<br />
:<math> Z \rightarrow X \rightarrow Y \leftarrow U \rightarrow X</math><br />
<math>Z</math>是一种工具变量,因为它有一条通往结果<math>Y</math>的路径,并且不受<math>U</math>的混杂。<br />
<br />
<br />
在上面的例子中,如果Z和X是二进制值,那么Z=0,X=1不出现的假设称为单调性。<br />
<br />
<br />
对该技术的改进包括通过调节其他变量来创建工具变量,以阻断工具变量和混杂因子之间的路径,并组合多个变量以形成单个工具变量。<ref name=":1" /><br />
<br />
<br />
===== 孟德尔随机化 =====<br />
<br />
定义:孟德尔随机化使用已知功能的基因,来观察研究中可改变的部分对疾病的因果关系。<ref name="Katan1986">{{cite journal|author=Katan MB|date=March 1986|title=Apolipoprotein E isoforms, serum cholesterol, and cancer|journal=Lancet|volume=1|issue=8479|pages=507–8|doi=10.1016/s0140-6736(86)92972-7}}</ref><ref>{{Cite book|url=https://www.ncbi.nlm.nih.gov/books/NBK62433/|title=Mendelian Randomization: Genetic Variants as Instruments for Strengthening Causal Inference in Observational Studies|last1=Smith|first1=George Davey|last2=Ebrahim|first2=Shah|date=2008|publisher=National Academies Press }}</ref><br />
<br />
<br />
由于基因在人群中随机变化,基因的存在通常可以视为工具变量。这意味着在许多情况下,可以使用观察性研究中的回归来量化因果关系。<ref name=":1" /><br />
<br />
<br />
== 关联 ==<br />
=== 独立性条件 ===<br />
<br />
独立性条件是用于确定两个变量是否彼此独立的规则。如果一个变量的值不直接影响另一个变量的值,则两个变量是独立的。多个因果模型可以共享独立条件。例如,模型<br />
:<math> A \rightarrow B \rightarrow C</math><br />
和<br />
:<math> A \leftarrow B \rightarrow C</math><br />
具有相同的独立条件,因为<math> B </math>作为条件时<math> A </math>和<math> C </math>独立。但是,这两个模型的含义不同,还可能与数据不符(也就是说,如果观测数据显示在<math> B </math>作为条件后显示了<math> A </math>和<math> C </math>之间的关联,那么这两个模型都是不正确的)。相反,数据无法显示这两个模型中的哪个是正确的,因为它们具有相同的独立性条件。<br />
<br />
<br />
将变量作为条件是进行假设实验的一种机制。将变量作为条件即在条件变量的给定值下分析其他变量的值。在第一个示例中,<math> B </math>作为条件意味着给定<math> B</math> 的取值的观察,此时不应显示出<math> A </math>和<math> C</math> 之间的依赖关系。如果存在这种依赖关系,则该模型是不正确的。非因果模型无法进行这种区分的,因为它们不会做出因果断言。<ref name=":1" /><br />
<br />
<br />
=== 混杂/去混杂 ===<br />
<br />
设计相关性研究的基本要素是确定对所研究变量的潜在混杂影响。控制这些变量是为了消除这些影响。但是,这些混杂变量无法被先验地正确确定。因此,一项研究可能会控制不相关的变量,甚至(间接地)控制了所研究的变量。<ref name=":1" /><br />
<br />
<br />
因果模型为识别恰当的混杂变量提供了一种鲁棒的技术。形式上,如果“ <math>Y</math> 通过不经过<math> X</math> 的路径与 <math>Z</math> 关联”,则<math> Z</math> 是混杂因素。这些混杂变量通常可以使用其他研究所收集的数据来确定。数学上,如果<br />
:<math> P(Y|X) \neq P(Y|do(X))</math><br />
那么X是Y的混杂因子。<ref name=":1" /><br />
<br />
<br />
在此之前,混杂因子的不正确的定义包括:<ref name=":1" /><br />
<br />
#“与<math>X</math>和<math>Y</math>都相关的任何变量。”<br />
#<math>Y</math>和未观测变量<math>Z</math>有关联<br />
#不相容性:“原始相对风险和潜在混杂因素调整后产生的相对风险”之间的差异<br />
#流行病学:在大范围总体中与 <math>X</math> 相关的变量,而在未接触<math> X</math> 的人群中与<math> Y </math>相关的变量。<br />
<br />
<br />
在如下模型中,上述定义是有缺陷的:<br />
<br />
:<math> X \rightarrow Z \rightarrow Y</math><br />
<math>Z </math>符合定义,但 <math>Z</math> 是中介变量,而不是混杂因子,并且是控制结果的一个例子。<br />
<br />
<br />
在模型中<br />
:<math> X \leftarrow A \rightarrow B \leftarrow C \rightarrow Y</math><br />
传统上,<math> B</math> 被认为是混杂因子,因为它与<math> X </math>和<math> Y</math> 关联,但 <math>B</math> 既不在因果路径上,也不是因果路径上任何节点的后代。控制 <math>B</math> 将使<math> B</math> 成为混杂因子。这被称为M偏差。<ref name=":1" /><br />
<br />
<br />
==== 后门调整 ====<br />
为了分析因果模型中<math>X</math>对<math>Y</math>的因果效应,我们需要针对所有混杂变量进行调整(去混杂)。<ref name=":1" />为了确定混杂变量的集合,我们需要<br />
<br />
#通过该集合阻塞<math>X</math>和<math>Y</math>之间的每个非因果路径<br />
#不破坏任何原有的因果路径<br />
#不创建任何虚假路径<br />
<br />
<br />
定义:从<math>X</math>到<math>Y</math>的后门路径是指,从从<math> X</math> 到<math> Y</math> 的任何以指向<math> X</math> 的箭头为开始的路径。<ref name=":1" /><br />
<br />
<br />
定义:给定模型中的一对有序变量<math>(X,Y)</math>,如果<br />
<br />
#混杂变量集<math>Z</math>中没有<math>X</math>的后代,<br />
<br />
#<math>X</math>和<math>Y</math>之间的所有后门路径都被<math>Z</math>中的混杂变量阻断,<br />
<br />
则称混杂变量集<math>Z</math>满足后门准则。<br />
<br />
<br />
如果<math>(X,Y)</math>满足后门准则,则在控制混杂变量集<math> Z</math> 时<math> X</math> 和<math> Y</math> 是无混杂的。除了混杂变量外,没有必要控制其他任何变量。<ref name=":1" />后门准则是找到混杂变量<math> Z </math>的集合的充分条件,但不是分析<math> X </math>对<math> Y </math>的因果效应必要条件。<br />
<br />
<br />
<br />
当因果模型是现实的合理表示并且满足后门准则时,则对于线性关系可以将'''<font color="#ff8000"> 偏回归系数 Partial Regression Coefficients </font>'''作为'''<font color="#ff8000"> (因果)路径系数 (Causal) Path Coefficients </font>'''。<ref name=":1"/> <ref>[http://bayes.cs.ucla.edu/BOOK-2K/ch3-3.pdf chapter 3-3 Controlling Confounding Bias]</ref><br />
:<math> P(Y|do(X))=\sum_z{P(Y|X,Z=z)P(Z=z)}</math><br />
<br />
<br />
==== 前门调整 Frontdoor Adjustment ====<br />
<br />
如果阻塞路径的所有元素都不可观测,则后门路径不可计算,但是如果所有从<math> X </math>到<math> Y </math>的路径都有元素<math> z</math> ,并且<math> z</math> 到<math> Y </math>没有开放的路径,那么我们可以使用<math> z </math>的集合<math> Z </math>来测量<math> P(Y|do(X))</math>。实际上<math> Z </math>作为<math> X </math>的代理时有一些条件。<br />
<br />
<br />
定义<ref>{{Cite book|title=Causal Inference in Statistics: A Primer|ISBN=978-1-119-18684-7|last1=Pearl|first1=Judea|last2=Glymour|first2=Madelyn|first3=Nicholas P|last3=Jewell}}</ref>:前门路径是这样的直接因果路径<ref name=":1" /><br />
<br />
<br />
#<math>Z</math>阻断了所有<math>X</math>到<math>Y</math>的有向路径<br />
#<math>X</math>到<math>Y</math>没有后门路径<br />
#所有<math>Z</math>到<math>Y</math>的后门路径都被<math>X</math>阻断<br />
<br />
<br />
以下式子通过将前门路径上的变量集<math>Z</math>作条件,将含有do的表达式转化成不含do的表达式:<ref name=":1" /><br />
:<math> P(Y|do(X))=\sum_z{[P(Z=z|X)\sum_x{P(Y|X=x,Z=z)P(X=x)}]}</math><br />
<br />
<br />
假定上述概率涉及到的观察数据可用,则无需进行实验即可计算出最终概率,而不管是否存在其他混杂路径且无需进行后门调整。<ref name=":1" /><br />
<br />
<br />
== 干预 ==<br />
<br />
=== 查询 ===<br />
<br />
查询是根据特定模型提出的问题。通常通过进行干预实验来回答这些问题。“干预”会设定模型中一个变量的值并观察结果。从数学上讲,此类查询采用以下形式(例子):<ref name=":1" /><br />
<br />
:<math> P(牙线价格|do(牙膏价格))</math><br />
<br />
其中do算子表示该实验明确修改牙膏的价格。图模型上看,这可以阻止任何可能影响该变量的因果变量。这消除了所有指向实验变量(牙膏价格)的因果箭头。<ref name=":1" /><br />
<br />
<br />
do算子也可以应用于多个变量(使它们取值固定)进行更复杂的查询。<br />
<br />
<br />
=== Do演算===<br />
Do演算是一组可用于将一个表达式转换为另一个表达式的一系列操作,其总体目标是将包含do算子的表达式转换为不包含do算子的表达式。不含do算子的表达式可以仅从观察数据中估计出来,而无需进行实验干预;而实验干预可能是代价大,耗时长甚至是不道德的(例如,要求受试者吸烟)。<ref name=":1" />Do演算的规则集是完备的,可用于推导出该系统中的每个真命题。有一种算法可以确定对于给定模型,是否可以在多项式时间内求解。<ref name=":1" /><br />
<br />
<br />
====do演算规则集====<br />
该运算包括了三条涉及do算子的条件概率变换规则。其中规则1和3都是显然的,但规则2有些微妙。下面给出表达do演算规则集的三种版本。<br />
=====版本1=====<br />
该版本是维基百科上do演算的表达方式。<br />
<br />
规则1用来增删观测:<ref name=":1" /><br />
:<math> P(Y|do(X),Z,W)=P(Y|do(X),Z)</math><br />
在删除所有指向<math>X</math>的箭头的图中,<math>Z</math>阻塞了所有从<math>W</math>到<math>Y</math>的路径。<ref name=":1" /><br />
<br />
<br />
规则2用来互换干预和观测:<ref name=":1" /><br />
:<math> P(Y|do(X),Z)=P(Y|X,Z)</math><br />
在原图中<math>Z</math>满足后门准则。<ref name=":1" /><br />
<br />
<br />
规则3用来增删干预:<ref name=":1" /><br />
:<math> P(Y|do(X))=P(Y)</math><br />
在原图中<math>X</math>和<math>Y</math>间没有因果路径。<ref name=":1" /><br />
<br />
<br />
=====版本2=====<br />
该版本是Judea Pearl的《Causality: Models, Reasoning and Inference (2nd Edition)》中的表达方式。<br />
<br />
规则1用于增删观测:在 <math> G_{\overline{X}} </math> 中,当给定<math>X</math>和<math>W</math>,有<math>Y</math>和<math>Z</math>条件独立时,则<br />
:<math> P(Y|do(X),Z,W)=P(Y|do(X),Z)</math><br />
<br />
<br />
规则2用于互换干预和观察:在 <math> G_{\overline{X}\underline{Z}} </math> 中,当给定<math>X</math>和<math>W</math>,有<math>Y</math>和<math>Z</math>条件独立时,则<br />
:<math> P(Y|do(X),do(Z),W)=P(Y|do(X),Z,W)</math><br />
<br />
<br />
规则3用于增删干预:在 <math> G_{\overline{X}\underline{Z(W)}} </math> 中,当给定<math>X</math>和<math>W</math>,有<math>Y</math>和<math>Z</math>条件独立时,则<br />
:<math> P(Y|do(X),do(Z),W)=P(Y|do(X),W)</math><br />
<br />
<br />
其中 <math> Z(W) </math> 表示 <math> Z - An(W)_{ G_{ \overline{X} } } </math> , <math> An(W)_{G} </math> 表示<math>W</math>在图<math>G</math>中的祖先集(<math>W</math>及其祖先节点构成的点集), <math> G_{\overline{X}} </math> 表示删除<math>G</math>中所有指向<math>X</math>节点的边后得到的子图, <math> G_{\overline{X}\underline{Z}} </math> 表示删除<math>G</math>中所有指向<math>X</math>节点的边和从<math>Z</math>指向其他节点的边后得到的子图。<br />
<br />
<br />
=====版本3=====<br />
该版本是Daphne Koller和Nir Friedman的《概率图模型:原理与技术》中的表达方式。<br />
<br />
<br />
规则1用于增删观测:在 <math> G_{\overline{Z}}^{+} </math> 中,当给定<math>Z</math>和<math>X</math>,有<math>W</math>和<math>Y</math>'''<font color="#ff8000"> 有向分离 d-seperated </font>'''时,则<br />
:<math> P(Y|do(Z),X,W)=P(Y|do(Z),X) </math><br />
<br />
<br />
规则2用于互换干预和观察:在 <math> G_{\overline{Z}}^{+} </math> 中,当给定<math>X</math>、<math>Z</math>、<math>W</math>,有<math>Y</math>和 <math> \hat{X} </math> 有向分离时,则<br />
:<math> P(Y|do(Z),do(X),W)=P(Y|do(Z),X,W)</math><br />
<br />
<br />
规则3用于增删干预:在 <math> G_{\overline{Z}}^{+} </math> 中,当给定<math>Z</math>和<math>W</math>,有<math>Y</math>和 <math> \hat{X} </math> 有向分离时,则<br />
:<math> P(Y|do(Z),do(X),W)=P(Y|do(Z),W)</math><br />
<br />
<br />
其中 <math> G_{\overline{Z}}^{+} </math> 表示删除<math>G</math>中所有指向<math>Z</math>节点的边,添加独立决策变量<math> \hat{Z} </math>唯一指向<math>Z</math>,从而得到的G子图的拓展图。<br />
<br />
<br />
=====扩展=====<br />
这些规则并不意味着任何查询都能移除do算子。有些情况下,将一个不能进行的操作换成另一个可以进行的操作也是有意义的。例如:<br />
:<math> P(心脏病|do(血胆固醇))=P(心脏病|do(饮食))</math><br />
<br />
<br />
==反事实 ==<br />
<br />
反事实考虑那些无法从数据中得到的概率,如一个不吸烟的人在过去重度吸烟的话,他现在会不会得癌症。<br />
<br />
<br />
===潜在结果 Potential Outcome ===<br />
<br />
定义:Y的潜在结果是“如果<math>X</math>被赋值为<math>x</math>,对于个体<math>u</math>来说<math>Y</math>会怎么样”。数学上可以表达为<ref name=":1" /><br />
<br />
<br />
:<math> Y_X=Y_x(u)</math><br />
<br />
<br />
潜在结果是在个体<math>u</math>的层次定义的。<ref name=":1" />''<br />
<br />
<br />
传统的潜在结果是数据驱动的,而非模型驱动的,这限制了它辨析因果关系的能力。它将因果问题当作数据缺失问题,甚至在标准场景下都会给出错误的回答。<ref name=":1" /><br />
<br />
<br />
===因果推断 ===<br />
<br />
在因果模型的语境中,潜在结果是被从因果角度解释的,而非从统计角度解释。<br />
<br />
<br />
因果推断的第一定律意味着潜在结果<br />
<br />
:<math> Y_x(u)</math><br />
<br />
<br />
可以被这样计算:将因果模型<math>M</math>中指向<math>X</math>的箭头删除,计算特定的<math>x</math>的结果。形式上,<ref name=":1" /><br />
<br />
<br />
:<math> Y_x(u)=Y_{M_x}(u)</math><br />
<br />
<br />
=== 计算反事实 Conducting a counterfactual ===<br />
用一个因果模型计算反事实包括三步。这种方法不管模型是线性还是非线性都有效。当因果关系确定时,可以计算出一个点估计。在其他情况下(如仅能计算概率时),可以计算出一个概率区间,如原本不吸烟的人如果吸烟会增加10-20%的癌症概率。<br />
给定下列模型,<br />
:<math> Y \leftarrow X \rightarrow M \rightarrow Y \leftarrow U</math><br />
====归因 Abduct====<br />
应用归纳推理(使用观察来找到最简单/最可能的解释的逻辑推理)来估计<math>u</math>,它是支持反事实的特定观察上未观察到的变量的代理。根据命题证据计算<math>u</math>的概率。<br />
====行动 Act====<br />
对于特定观察,使用do算子建立反事实(如令<math>m = 0</math>),从而相应地修改方程式。<br />
====预测 Predict ====<br />
使用修改后的公式计算输出(<math>y</math>)的值。<br />
<br />
===中介Mediation ===<br />
直接原因和间接原因(中介)可以通过执行反事实区分。理解中介需要在干预直接原因时保持中介恒定。在模型<br />
:<math> Y \leftarrow M \leftarrow X \rightarrow Y</math><br />
中,<math>M</math>是<math>X</math>对<math>Y</math>影响的中介,<math>X</math>对<math>Y</math>也有非中介影响。这样保持<math>M</math>恒定,就可以计算<math>do(X)</math>。<br />
对于线性模型,可以通过取中介路径上所有路径系数的乘积来计算间接效应。总间接效应是通过各个间接效应的和计算得出的。对于线性模型,当拟合的不包括中介的方程式的系数与包含中介的方程式的系数显着不同时,这就意味着中介发生了。<br />
<br />
====直接效应 Direct effect====<br />
<br />
在这样模型的实验中,受控直接效应(CDE)通过将<math>M</math>强行赋值(<math>do(M=0)</math>)和随机化(<math>do(X=0),do(X=1),...</math>),然后观察<math>Y</math>的结果值获得。<br />
:<math> CDE(0)=P(Y=1|do(X=1),do(M=0))-P(Y=1|do(X=0),do(M=0))</math><br />
每个中介因子有一个相应的受控直接效应(CDE)。<br />
然而,更好的实验是计算自然直接效应(NDE)。这是通过保持<math>X</math>和<math>M</math>之间的关系不变,同时干预X和Y之间的关系而确定的效应。<br />
:<math> NDE(0)=P(Y_{M=M0}=1|do(X=1))-P(Y_{M=M0}=1|do(X=0))</math><br />
例如,考虑每年或几年去看牙科医生的次数(<math>X</math>)的直接效应,去看牙科医生会使牙科医生鼓励人们使用牙线(<math>M</math>)。牙龈(<math>Y</math>)因此变得更健康,这归因于牙科医生(直接)或牙线(中介/间接)。需要进行的实验是继续使用牙线,但不去看牙科医生。<br />
<br />
====间接效应 Indirect effect ====<br />
<br />
<math>X</math>对<math>Y</math>的间接效应是当<math>X</math>保持不变,<math>M</math>增加到<math>X</math>增加一个单位时<math>M</math>会达到的值。<br />
间接效应不能被控制,因为不能通过保持另一个变量恒定来禁用直接路径。自然间接效应(NIE)是使用牙线(<math>M</math>)对牙龈健康(<math>Y</math>)的影响。自然间接效应NIE的计算方式为(使用无牙线和无牙线的情况)给定牙医和没有牙医的情况下使用牙线的概率微分的和,或<br />
:<math> NIE=\sum _{m}[P(M=m|X=1)-P(M=m|X=0)]P(Y=1|X=0,M=m)</math><br />
自然直接效应NDE计算包括了反事实步骤(:<math>Y_{M=M0}</math>)。对于非线性模型,下列看上去显然的等式<br />
:<math> Total effect=Direct effect + Indirect effect</math><br />
是不成立的,因为阈值效应和二进制值等异常。然而,<br />
:<math>Total effect(X=0\rightarrow X=1)=NDE(X=0\rightarrow X=1)-NIE(X=1\rightarrow X=0)</math><br />
对于所有线性和非线性模型都是可以生效的。它允许NDE直接从观测的数据计算出了,不需要干预或使用反事实下标。<br />
<br />
==可移植性 Transportability==<br />
<br />
即使因果模型及对应的相关数据不同,因果模型也提供了一种工具来集成跨数据集的数据,称为移植。例如,调查数据可以与随机对照实验数据合并。<ref name=":1" />移植提供了一个外部有效性问题的解决方案,即一项研究是否可以在不同的背景下应用。<br />
<br />
一,如果两个模型在所有相关变量上都匹配,并且已知来自其中一个模型的数据是无偏的,则可以使用一个总体的数据得出关于另一个总体的结论。<br />
<br />
二,已知数据存在偏差,则重加权可以允许模型在数据集间移植。<br />
<br />
三,可以从不完整的数据集中得出结论。<br />
<br />
四,可以组合(移植)来自多个总体的研究数据,以得出有关未观测总体的结论。<br />
<br />
五,结合多个研究的估计值(例如<math>P(W|X)</math>)可以提高结论的准确性。<ref name=":1" /><br />
<br />
do演算为移植提供了一个一般性准则:目标变量可以通过一系列不涉及任何“差异”变量(能够区分两个总体的变量)的do运算转换为另一个表达式。<ref name=":1" />有一个类似的规则适用于参与者相对不同的研究。<ref name=":1" /><br />
<br />
<br />
==贝叶斯网络 Bayesian network==<br />
<br />
因果模型可以用贝叶斯网实现。贝叶斯网络可用于提供事件的逆概率(给定结果,反推具体原因的概率是多少)。这就需要准备一个条件概率表,显示所有可能的输入和结果以及相关的概率。<ref name=":1" /><br />
<br />
例如,给定疾病和针对疾病的检验的两变量模型,条件概率表的形式为:<ref name=":1" /><br />
<br />
{| class="wikitable"<br />
<br />
|+特定疾病检测为阳性的概率<br />
<br />
!<br />
<br />
!colspan="2"|Test<br />
<br />
|-<br />
<br />
!疾病<br />
<br />
!阳性<br />
<br />
!阴性<br />
<br />
|-<br />
<br />
|阴性<br />
<br />
|12<br />
<br />
|88<br />
<br />
|-<br />
<br />
|阳性<br />
<br />
|73<br />
<br />
|27<br />
<br />
|}<br />
根据该表,当患者没有疾病时,测试为阳性的可能性为12%。<br />
尽管这对于小问题很容易解决,但是随着变量数量及其相关状态的增加,概率表(以及相关的计算时间)呈指数增长。<ref name=":1" /><br />
贝叶斯网络在商业上可用于如无线数据纠错和DNA分析之类的应用中。<ref name=":1" /><br />
<br />
==不变量/语境Invariants/Context ==<br />
<br />
因果的不同概念涉及不变关系的概念。在识别手写数字的情况下,数字形状决定含义,因此形状和含义是不变量,改变形状会改变含义。其他属性则没有此性质(如颜色)。此不变性对于在各种非不变量所构成语境中生成的数据集都应满足。与其使用汇总的数据集进行学习评估因果关系,不如对一个数据集进行学习并对另一数据集进行测试,这可以帮助将变化属性与不变量区分开。<ref>{{Cite web|url=https://www.technologyreview.com/s/613502/deep-learning-could-reveal-why-the-world-works-the-way-it-does/|title=Deep learning could reveal why the world works the way it does|last=Hao|first=Karen|date=May 8, 2019|website=MIT Technology Review|language=en-US|access-date=February 10, 2020}}</ref><br />
<br />
== 其他词条 ==<br />
<br />
*[[反事实]] – 因果之梯第三层<br />
<br />
==引用==<br />
<br />
{{Reflist}}<br />
<br />
== 源 ==<br />
* {{Cite book|url={{google books |plainurl=y |id=LLkhAwAAQBAJ}}|title=Causality|last=Pearl|first=Judea|date=2009-09-14|publisher=Cambridge University Press|isbn=9781139643986|language=en}}<br />
<br />
== 外部链接 ==<br />
<br />
*{{Cite journal|last=Pearl|first=Judea|date=2010-02-26|title=An Introduction to Causal Inference|journal=The International Journal of Biostatistics|volume=6|issue=2|pages=Article 7|doi=10.2202/1557-4679.1203|issn=1557-4679|pmc=2836213|pmid=20305706}} <br />
* [https://philpapers.org/browse/causal-modeling causal-modeling]<br />
* {{Cite news|url=https://www.quantamagazine.org/how-artificial-intelligence-is-changing-science-20190311/|title=AI Algorithms Are Now Shockingly Good at Doing Science|last=Falk|first=Dan|date=2019-03-17|work=Wired|access-date=2019-03-20|issn=1059-1028}}<br />
* {{Cite web|url=https://bostonreview.net/science-nature/tim-maudlin-why-world|title=The Why of the World|last=Maudlin|first=Tim|date=2019-08-30|website=Boston Review|language=en|access-date=2019-09-09}}<br />
* {{Cite web|url=https://www.quantamagazine.org/to-build-truly-intelligent-machines-teach-them-cause-and-effect-20180515/|title=To Build Truly Intelligent Machines, Teach Them Cause and Effect|last=Hartnett|first=Kevin|website=Quanta Magazine|access-date=2019-09-19}}<br />
*[https://www.facebook.com/iclr.cc/videos/534780673594799 Learning Representations using Causal Invariance (in English), ICLR, February 2020, retrieved 2020-02-10]<br />
<br />
== 编者推荐==<br />
<br />
<br />
===书籍推荐===<br />
[[File:统计因果推理入门.jpg|200px|thumb|right|《统计因果推理入门》封面]]<br />
<br />
*[[统计因果推理入门]] 对应英文[[Causal Inference in Statistics: A Primer]]<br />
<br />
这本书非常适合初学者入门因果科学,这里面涉及到对结构因果模型的详细定义和阐述,非常清晰易懂。<br />
<br />
<br />
*[[为什么-关于因果的新科学]]<br />
<br />
[[File:为什么-关于因果关系的新科学.jpg|200px|thumb|right|《为什么-关于因果关系的新科学》封面]]<br />
<br />
在本书中,人工智能领域的权威专家朱迪亚·珀尔及其同事领导的因果关系革命突破多年的迷雾,厘清了知识的本质,确立了因果关系研究在科学探索中的核心地位。<br />
<br />
<br />
<br />
关于这本书集智俱乐部邀请白楚研究员用100分钟,为大家详细介绍了Judea Pearl绘制的因果科学蓝图,作为一个起点,去拥抱因果革命。可以查看对应的视频分享[https://campus.swarma.org/course/1522 解读《为什么》:攀登因果之梯]<br />
<br />
===课程推荐===<br />
*[https://campus.swarma.org/course/2526 两套因果框架深度剖析:潜在结果模型与结构因果模型]<br />
<br />
::这个视频内容来自[[集智俱乐部读书会]]-因果科学与Causal AI读书会第二季内容的分享,由英国剑桥大学及其学习组博士陆超超详细的阐述了潜在结果模型和结果因果模型,并介绍了两个框架的相互转化规律。<br />
::1. 讲述因果推断的两大框架:潜在结果模型和结构因果模型,讨论他们各自的优缺点以及他们之间的联系,详细介绍他们之间的转化规律。 <br />
<br />
<br />
::2. 与大家一起深入探讨因果推断中最基本的概念、定理以及它们产生的缘由,了解每个概念背后的故事,从而建立起对因果更全面的感知。 <br />
<br />
<br />
::3. 分享它们在不同学科中的具体的应用,包括社会科学、经济学、医学、机器学习等,借助这些应用,进一步启发大家用因果科学思维来思考和解决问题。<br />
<br />
<br />
<br />
*[https://campus.swarma.org/course/1937 如何用信息视角理解现代因果模型框架?]<br />
::这个视频内容来自[[集智俱乐部读书会]]-因果科学与Causal AI读书会第一季内容的分享,这个视频为大家串讲因果推理的相关论文,着眼与因果研究的源头,简单介绍哲学中的因果思考。其次重点是用因果之梯(她的信息视角--回答因果问题需要相应的信息)和一个例子,来理解现代因果建模框架;最后梳理因果推理和 AI 领域的融合,以及Causal AI 的强人工智能之路。<br />
<br />
=== 文章总结 ===<br />
<br />
[https://mp.weixin.qq.com/s/mOdTI0gR9rsRyX4fz3mYNA 因果科学入门读什么书?Y. Bengio博士候选人的研读路径推荐]<br />
<br />
[https://mp.weixin.qq.com/s/TtYsTyyGEX7U1ZOSl-lsPw 前沿综述:因果推断与因果性学习研究进展]<br />
<br />
[https://mp.weixin.qq.com/s/W6PIE211TavEgg_s3adDdg 因果表征学习最新综述:连接因果科学和机器学习的桥梁]<br />
<br />
[https://mp.weixin.qq.com/s/pRgLZFJpbgmAyI7LgOnHug 历时3个月,全球32位讲者,共同讲述因果科学与Causal AI的全景框架!]<br />
<br />
[https://mp.weixin.qq.com/s/f-rI5W6tc6qOzthbzK4oAw 崔鹏:稳定学习——挖掘因果推理和机器学习的共同基础]<br />
<br />
[https://mp.weixin.qq.com/s/l-05jRYabGI-JoXedU-PLA 因果科学:连接统计学、机器学习与自动推理的新兴交叉领域]<br />
<br />
[https://mp.weixin.qq.com/s/ZOUeF_HEFneYVi2BPe8LFg 因果观念新革命?万字长文,解读复杂系统背后的暗因果]<br />
<br />
[https://mp.weixin.qq.com/s/dVxgHcQAz_VjT-HDa2fXgg 周晓华:因果推断的数学基础和在医学中的应用]<br />
<br />
<br />
=== 相关路径 ===<br />
* [https://pattern.swarma.org/path?id=99 因果科学与Casual AI读书会必读参考文献列表],这个是根据读书会中解读的论文,做的一个分类和筛选,方便大家梳理整个框架和内容。<br />
* [https://pattern.swarma.org/path?id=9 因果推断方法概述],这个路径对因果在哲学方面的探讨,以及因果在机器学习方面应用的分析。<br />
* [https://pattern.swarma.org/path?id=90 因果科学和 Causal AI入门路径],这条路径解释了因果科学是什么以及它的发展脉络。此路径将分为三个部分进行展开,第一部分是因果科学的基本定义及其哲学基础,第二部分是统计领域中的因果推断,第三个部分是机器学习中的因果(Causal AI)。<br />
* [https://pattern.swarma.org/path?id=28 复杂网络动力学系统重构文献],这个路径是张江老师梳理了网络动力学重构问题,描述了动力学建模的常用方法和模型,并介绍了一些经典且重要的论文,这也是复杂系统自动建模读书会的主要论文来源,所以大部分都有解读视频。<br />
* [https://pattern.swarma.org/path?id=114 因果纠缠集智年会——因果推荐系统分论坛]关于因果推荐系统的参考文献和主要嘉宾介绍,来源是集智俱乐部的因果纠缠年会。<br />
<br />
<br />
<br />
<br />
<br />
----<br />
本中文词条由[[用户: Sikongpop|Sikongpop]]用户参与编译,[[用户:LFZ|LFZ]]参与审校,[[用户:思无涯咿呀咿呀|思无涯咿呀咿呀]]编辑,欢迎在讨论页面留言。<br />
<br />
<br />
'''本词条内容源自wikipedia及公开资料,遵守 CC3.0协议。'''<br />
<br />
<br />
<br />
[[Category:Causal diagrams]]<br />
[[Category:Causality]]<br />
[[Category:Formal epistemology]]<br />
[[Category:Scientific modeling]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=23603倾向得分匹配2021-06-13T07:22:27Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。</font><br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种用于估计治疗、政策或其他干预的效果统计匹配技术,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
<font color="#aaaaaaa">【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。</font><br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化分配机制意味着每个协变量将在处理组和对照组中呈现类似的分布。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
<font color="#aaaaaaa">【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。</font><br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单地通过对比评估吸烟者和不吸烟者来估计平均处理效果将产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
==Overview ==<br />
<br />
==综述==<br />
<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。</font><br />
<br />
PSM适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的协变量特征进行比较。<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。</font><br />
<br />
在常规的匹配机制中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有显著的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。</font><br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
==General procedure==<br />
<br />
==一般步骤==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
<font color="#aaaaaaa">2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。</font><br />
<br />
2. 依照倾向性评分的估计量进行分层,<font color="#32cd32">检查协变量的倾向性评分的估计量在每层处理组和对照组是否均衡</font><br />
<br />
*Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
<font color="#aaaaaaa">3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:</font><br />
<br />
3. 根据倾向性评分的估计量,将每个处理组个体与一个或多个对照组个体进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score<br />
<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
<font color="#aaaaaaa">4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本</font><br />
<br />
4. <font color="#32cd32">对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
<font color="#aaaaaaa">5.【机器翻译】基于新样本的多变量分析</font><br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
* 如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
<font color="#aaaaaaa">【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。</font><br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
==Formal definitions==<br />
<br />
==形式定义 ==<br />
<br />
===Basic settings ===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55" /> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55" />是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]个体。每个个体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示个体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个个体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。个体编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些个体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有协变量''X''(即:条件非混杂变量)向量,以及对应着控制和处理两种情况的'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp \!\!\!\! \perp Z \mid X </math><br />
<br />
这里<math>\perp \!\!\!\! \perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55" /><br />
<br />
=== Balancing score===<br />
<br />
=== 平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp \!\!\!\! \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分 ===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。倾向性评分匹配将使得处理组和对照组的协变量分布趋同,从而减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后这些倾向性评分即可作为用于逆概率加权方法的权重估计量。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55" /><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55" /><br />
<br />
*The propensity score <math>e(x)</math> is a balancing score.<br />
*倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
*Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
*任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成只有一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数(保留全部维度)。<br />
<br />
*If treatment assignment is strongly ignorable given ''X'' then:<br />
*如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:*It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp \!\!\!\! \perp Z \mid e(X).</math><br />
回归<br />
:*For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>.<br />
:*对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
*Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
*利用平衡得分的样本估计可产生在X上均衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系 ===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小充分统计量。<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<br />
<br />
== Disadvantages==<br />
== 缺点 ==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。</font><br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的直观仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
<font color="#aaaaaaa">【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。</font><br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个协变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维数问题“,即每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。</font><br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
<font color="#aaaaaaa">【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。</font><br />
<br />
Judea Pearl也提出了关于匹配方法的普遍担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它也很容易地手动实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]<br />
<br />
<references /></div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=23602倾向得分匹配2021-06-13T07:16:49Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。</font><br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种用于估计治疗、政策或其他干预的效果统计匹配技术,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
<font color="#aaaaaaa">【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。</font><br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化分配机制意味着每个协变量将在处理组和对照组中呈现类似的分布。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
<font color="#aaaaaaa">【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。</font><br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单地通过对比评估吸烟者和不吸烟者来估计平均处理效果将产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
==Overview ==<br />
<br />
==综述==<br />
<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。</font><br />
<br />
PSM适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的协变量特征进行比较。<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。</font><br />
<br />
在常规的匹配机制中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有显著的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。</font><br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
==General procedure==<br />
<br />
==一般步骤==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
<font color="#aaaaaaa">2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。</font><br />
<br />
2. 依照倾向性评分的估计量进行分层,<font color="#32cd32">检查协变量的倾向性评分的估计量在每层处理组和对照组是否均衡</font><br />
<br />
*Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
<font color="#aaaaaaa">3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:</font><br />
<br />
3. 根据倾向性评分的估计量,将每个处理组个体与一个或多个对照组个体进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score<br />
<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
<font color="#aaaaaaa">4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本</font><br />
<br />
4. <font color="#32cd32">对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
<font color="#aaaaaaa">5.【机器翻译】基于新样本的多变量分析</font><br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
* 如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
<font color="#aaaaaaa">【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。</font><br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
==Formal definitions==<br />
<br />
==形式定义 ==<br />
<br />
===Basic settings ===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55" /> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55" />是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]个体。每个个体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示个体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个个体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。个体编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些个体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有协变量''X''(即:条件非混杂变量)向量,以及对应着控制和处理两种情况的'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp \!\!\!\! \perp Z \mid X </math><br />
<br />
这里<math>\perp \!\!\!\! \perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55" /><br />
<br />
=== Balancing score===<br />
<br />
=== 平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp \!\!\!\! \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分 ===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。倾向性评分匹配将使得处理组和对照组的协变量分布趋同,从而减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后这些倾向性评分即可作为用于逆概率加权方法的权重估计量。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55" /><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55" /><br />
<br />
*The propensity score <math>e(x)</math> is a balancing score.<br />
*倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
*Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
*任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成只有一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数(保留全部维度)。<br />
<br />
*If treatment assignment is strongly ignorable given ''X'' then:<br />
*如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:*It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
回归<br />
:*For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>.<br />
:*对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
*Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
*利用平衡得分的样本估计可产生在X上均衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系 ===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小充分统计量。<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<br />
<br />
== Disadvantages==<br />
== 缺点 ==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。</font><br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的直观仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
<font color="#aaaaaaa">【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。</font><br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个协变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维数问题“,即每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。</font><br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
<font color="#aaaaaaa">【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。</font><br />
<br />
Judea Pearl也提出了关于匹配方法的普遍担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它也很容易地手动实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]<br />
<br />
<references /></div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=23457倾向得分匹配2021-06-10T03:43:00Z<p>Aceyuan:/* 正式定义 */</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。</font><br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
<font color="#aaaaaaa">【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。</font><br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,均衡分配处理组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
<font color="#aaaaaaa">【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。</font><br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比评估吸烟者和不吸烟者会让处理效果产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。</font><br />
<br />
PSM适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似的单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的处理前特征进行比较。<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。</font><br />
<br />
在正常的匹配中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有实质性的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。</font><br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
<font color="#aaaaaaa">2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。</font><br />
<br />
2. <font color="#32cd32">检查协变量的倾向性评分在处理组和对照组是否均衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
<font color="#aaaaaaa">3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:</font><br />
<br />
3. 根据倾向性评分,将每个参与者与一个或多个非参与者匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
<font color="#aaaaaaa">4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本</font><br />
<br />
4. <font color="#32cd32">对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
<font color="#aaaaaaa">5.【机器翻译】基于新样本的多变量分析</font><br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
<font color="#aaaaaaa">【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。</font><br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 形式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]物体。每个物体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。单元编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些物体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有协变量''X''(即:条件非混杂变量)向量,以及对应着控制和处理两种情况的'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
<br />
===平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。通倾向性评分可用于让处理组和对照组的协变量趋同,从而减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后将这些倾向性评分即可作为权重估计量用于逆概率加权方法。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* 倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* 任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成只有一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数(保留全部维度)。<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
* 如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
回归<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
:* 对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
* 利用平衡得分的样本估计可产生在X上均衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小充分统计量。<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<br />
<br />
==Disadvantages==<br />
==缺点==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。</font><br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的见解仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
<font color="#aaaaaaa">【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。</font><br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个些变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维度问题“,即,每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。</font><br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
<font color="#aaaaaaa">【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。</font><br />
<br />
Judea Pearl也提出了关于匹配方法的一般性担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它很容易手工实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22824倾向得分匹配2021-06-02T03:10:24Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。</font><br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
<font color="#aaaaaaa">【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。</font><br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,均衡分配处理组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
<font color="#aaaaaaa">【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。</font><br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比评估吸烟者和不吸烟者会让处理效果产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。</font><br />
<br />
PSM适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似的单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的处理前特征进行比较。<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。</font><br />
<br />
在正常的匹配中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有实质性的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。</font><br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
<font color="#aaaaaaa">2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。</font><br />
<br />
2. <font color="#32cd32">检查协变量的倾向性评分在处理组和对照组是否均衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
<font color="#aaaaaaa">3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:</font><br />
<br />
3. 根据倾向性评分,将每个参与者与一个或多个非参与者匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
<font color="#aaaaaaa">4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本</font><br />
<br />
4. <font color="#32cd32">对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
<font color="#aaaaaaa">5.【机器翻译】基于新样本的多变量分析</font><br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
<font color="#aaaaaaa">【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。</font><br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 正式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]物体。每个物体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。单元编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些物体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有协变量''X''(即:条件非混杂变量)向量,以及对应着控制和处理两种情况的'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
<br />
===平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。通倾向性评分可用于让处理组和对照组的协变量趋同,从而减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后将这些倾向性评分即可作为权重估计量用于逆概率加权方法。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* 倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* 任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成只有一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数(保留全部维度)。<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
* 如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
回归<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
:* 对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
* 利用平衡得分的样本估计可产生在X上均衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小充分统计量。<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<br />
<br />
==Disadvantages==<br />
==缺点==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。</font><br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的见解仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
<font color="#aaaaaaa">【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。</font><br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个些变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维度问题“,即,每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。</font><br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
<font color="#aaaaaaa">【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。</font><br />
<br />
Judea Pearl也提出了关于匹配方法的一般性担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它很容易手工实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22822倾向得分匹配2021-06-02T03:06:02Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。</font><br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
<font color="#aaaaaaa">【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。</font><br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,均衡分配处理组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
<font color="#aaaaaaa">【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。</font><br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比评估吸烟者和不吸烟者会让处理效果产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。</font><br />
<br />
PSM适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似的单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的处理前特征进行比较。<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。</font><br />
<br />
在正常的匹配中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有实质性的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。</font><br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
<font color="#aaaaaaa">2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。</font><br />
<br />
2. <font color="#32cd32">检查协变量的倾向性评分在处理组和对照组是否均衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
<font color="#aaaaaaa">3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:</font><br />
<br />
3. 根据倾向性评分,将每个参与者与一个或多个非参与者匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
<font color="#aaaaaaa">4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本</font><br />
<br />
4. <font color="#32cd32">对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
<font color="#aaaaaaa">5.【机器翻译】基于新样本的多变量分析</font><br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
<font color="#aaaaaaa">【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。</font><br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 正式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]物体。每个物体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。单元编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些物体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有向量协变量''X''(即:条件非混杂变量),以及'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>分别对应着控制和处理两种情况。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
<br />
===平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。通倾向性评分可用于让处理组和对照组的协变量趋同,从而减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后将这些倾向性评分即可作为权重估计量用于逆概率加权方法。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* 倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* 任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成只有一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数(保留全部维度)。<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
* 如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
回归<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
:* 对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
* 利用平衡得分的样本估计可产生在X上均衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小充分统计量。<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<br />
<br />
==Disadvantages==<br />
==缺点==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。</font><br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的见解仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
<font color="#aaaaaaa">【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。</font><br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个些变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维度问题“,即,每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。</font><br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
<font color="#aaaaaaa">【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。</font><br />
<br />
Judea Pearl也提出了关于匹配方法的一般性担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它很容易手工实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22805倾向得分匹配2021-06-01T14:56:52Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。</font><br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
<font color="#aaaaaaa">【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。</font><br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,均衡分配处理组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
<font color="#aaaaaaa">【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。</font><br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比评估吸烟者和不吸烟者会让处理效果产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。</font><br />
<br />
PSM适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似的单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的处理前特征进行比较。<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。</font><br />
<br />
在正常的匹配中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有实质性的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。</font><br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
<font color="#aaaaaaa">2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。</font><br />
<br />
2. <font color="#32cd32">检查协变量的倾向性评分在处理组和对照组是否均衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
<font color="#aaaaaaa">3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:</font><br />
<br />
3. 根据倾向性评分,将每个参与者与一个或多个非参与者匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
<font color="#aaaaaaa">4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本</font><br />
<br />
4. <font color="#32cd32">对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
<font color="#aaaaaaa">5.【机器翻译】基于新样本的多变量分析</font><br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
<font color="#aaaaaaa">【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。</font><br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 正式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]物体。每个物体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。单元编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些物体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有向量协变量''X''(即:条件非混杂变量),以及'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>分别对应着控制和处理两种情况。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
<br />
===平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。通倾向性评分可用于让处理组和对照组的协变量趋同,从而减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后将这些倾向性评分即可作为权重估计量用于逆概率加权方法。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* 倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* 任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数。<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
* 如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
回归<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
:* 对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
* 利用平衡得分的样本估计可产生在X上均衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小统计量。<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<br />
<br />
==Disadvantages==<br />
==缺点==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。</font><br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的见解仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
<font color="#aaaaaaa">【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。</font><br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个些变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维度问题“,即,每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。</font><br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
<font color="#aaaaaaa">【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。</font><br />
<br />
Judea Pearl也提出了关于匹配方法的一般性担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它很容易手工实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22800倾向得分匹配2021-06-01T12:46:20Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。</font><br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
<font color="#aaaaaaa">【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。</font><br />
<br />
出现偏倚的原因可能是某个因素通过决定样本是否接受处理而导致了处理组和对照组的效果(如平均处理效果)差异,而不是处理本身导致了差异。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,均衡分配处理组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是分别从处理组和对照组中分别取样,让两组样本的全部协变量都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
<font color="#aaaaaaa">【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。</font><br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比评估吸烟者和不吸烟者会让处理效果产生偏差,它会受到能影响吸烟行为的因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。</font><br />
<br />
PSM适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似的单元很少; (ii)选择与处理单元类似的对照单元集合很困难,因为必须对一组高维的处理前特征进行比较。<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
<font color="#aaaaaaa">【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。</font><br />
<br />
在正常的匹配中,对一组能够区分处理组和对照组的特征做匹配,以使两组的特征更加相似。但如果这两个组的特征没有实质性的重叠,那么可能会引入实质性的错误。例如,拿对照组最糟的病例和处理组最好的病例进行比较,结果可能倾向于回归均值,这会让对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。</font><br />
<br />
PSM利用观察数据预测样本落入不同分组(例如,处理组与控制组)的概率,通常用Logistic回归方法,然后利用此概率创造一个反事实的群体。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:参与处理(属于处理组),则''Z'' = 1;未参与处理(属于对照组),则''Z'' = 0。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
<font color="#aaaaaaa">2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。</font><br />
<br />
2. <font color="#32cd32">检查协变量的倾向性评分在处理组和对照组是否均衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
<font color="#aaaaaaa">3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:</font><br />
<br />
3. 根据倾向性评分,将每个参与者与一个或多个非参与者匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:在处理单元倾向性评分的一个范围内选取对照单元,范围的宽度通常用倾向性评分的标准差乘上一个比例值<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
<font color="#aaaaaaa">4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本</font><br />
<br />
4. <font color="#32cd32">对比处理组和对照组的匹配样本或加权样本,验证协变量是否均衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
<font color="#aaaaaaa">5.【机器翻译】基于新样本的多变量分析</font><br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个参与者都匹配了多个非参与者,则适当应用非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
<font color="#aaaaaaa">【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。</font><br />
<br />
注意:当一个处理样本有多个匹配时,则必须用加权最小二乘法,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 正式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[[Independent and identically distributed random variables|独立同分布]]物体。每个物体''i''如果接受了处理则响应为<math>r_{1i}</math>,接受控制则响应为<math>r_{0i}</math>。被估计量是[[average treatment effect|平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中也可以不包括那些决定是否接受处理的特征。单元编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分在讨论某些物体的随机行为的时候将省略索引''i''。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有向量协变量''X''(即:条件非混杂变量),以及'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>分别对应着控制和处理两种情况。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说样本是否接受处理分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
<br />
===平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。通倾向性评分可用于让处理组和对照组的协变量趋同,从而减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为,在给定背景变量条件下单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,通过Logistic回归、随机森林或其他方法,利用一组协变量估计倾向性评分。然后将这些倾向性评分即可作为权重估计量用于逆概率加权方法。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* 倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* 任何比倾向性评分更“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数。<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
* 如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
回归<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
:* 对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值之差(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
* 利用平衡得分的样本估计可产生在X上均衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小统计量。<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。通过把混杂变量加入回归的控制变量,或者在混杂变量上进行匹配可以实现后门路径的阻断。<br />
<br />
==Disadvantages==<br />
==缺点==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。</font><br />
<br />
PSM已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的见解仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
<font color="#aaaaaaa">【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。</font><br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个些变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维度问题“,即,每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量。<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
<font color="#aaaaaaa">【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。</font><br />
<br />
PSM的一个缺点是它只能涵盖已观测的(和可观测的)协变量,而无法涵盖潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考量范围。由于匹配过程只控制可观测变量,那些隐藏的偏差在匹配后依然可能存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
<font color="#aaaaaaa">【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。</font><br />
<br />
Judea Pearl也提出了关于匹配方法的一般性担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏差。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏差。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它很容易手工实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22777倾向得分匹配2021-05-31T16:53:19Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对处理单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
之所以可能出现偏倚,是因为处理组和对照组处理结果(如平均处理效果)的差异可能更多受到决定样本是否接受处理的某个因素的影响,而不是处理本身。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,平衡分配处理组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是分别从处理单元和对照单元中各选择一个样本,让它们在所有的协变量上都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。<br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比吸烟者和不吸烟者而评估的处理效果会产生偏差,它会受到任何能影响吸烟行为因素的影响(例如:性别及年龄)。PSM要做的是通过让处理组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与处理组中的类似的单元很少; (ii)选择与处理单元类似的对照单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
在正常的匹配中,区分处理组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自处理组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。<br />
<br />
PSM使用了一种基于观察变量预测样本落入不同分组(例如,处理组与控制组)的概率的方法来创造一个反事实的群体,通常用逻辑回归来预测概率。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:''Z'' = 1, 如果参与处理(属于处理组);''Z'' = 0,如果未参与处理(属于对照组)。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(假设其既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。<br />
<br />
2. <font color="#32cd32">检查协变量在倾向性评分的分层中处理组和对照组是否平衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
3. 根据倾向性评分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:用处理单元的一段宽度的倾向性评分范围选取对照单元,通常用倾向性评分的标准差的一部分确定这一宽度<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本<br />
<br />
4. <font color="#32cd32">验证协变量在跨越处理组和对照组的匹配样本或加权样本中是否平衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5.【机器翻译】基于新样本的多变量分析<br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个参与者都匹配了多个非参与者,则应用适当的非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。<br />
<br />
注意:当一个处理样本有多个匹配时,加权最小二乘法是必需的,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 正式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[独立同分布随机变量|i.i.d]物体。每个物体''i''如果接受了处理则响应为<math>r_{1i}</math>,否则响应为<math>r_{0i}</math>。被估计量是[[平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中可以不包括哪些决定是否接受处理的特征。假设对单元的编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分将省略索引''i'',同时仍会讨论某些物体的随机行为。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有协变量向量''X''(即:条件无混杂),以及一些'''潜在结果‘’‘ ''r''<sub>0</sub>和''r''<sub>1</sub>分别对应着控制和处理两种情况。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说对样本是否接受处理的分配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
<br />
===平衡得分===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
A balancing score b(X) is a function of the observed covariates X such that the conditional distribution of X given b(X) is the same for treated (Z = 1) and control (Z = 0) units:<br />
<br />
平衡得分b(X)是观测协变量X的函数。在给定b(X)时,处理单元和控制单元的X有相同的条件分布:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
最一般的平衡得分函数是<math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
<br />
===倾向性评分===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
A propensity score is the probability of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce selection bias by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment indicator Z, a response variable r, and background observed covariates X. The propensity score is defined as the conditional probability of treatment given background variables:<br />
<br />
倾向性评分是根据协变量观测值计算得出的一个单元(例如:个人,教室,学校)被指配接受特定处理的概率。通过让处理组和对照组的协变量趋同,倾向性评分可用于减少选择偏差。<br />
<br />
假设有一个二值处理标识Z,一个响应变量r,以及被观测的背景协变量X。倾向性评分定义为在给定背景变量条件下,单元接受处理的条件概率:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
In the context of causal inference and survey methodology, propensity scores are estimated (via methods such as logistic regression, random forests, or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with Inverse probability weighting methods.<br />
<br />
在因果推断和调查方法的范围内,使用一些协变量估计倾向性评分(通过逻辑回归、随机森林或其他方法)。然后将这些倾向性评分作为权重的估计量用于逆概率加权方法。<br />
<br />
===Main theorems===<br />
<br />
===主要定理===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:[1]<br />
<br />
以下是Rosenbaum和Rubin于1983年首次提出并证明的:[1]<br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* 倾向性评分<math>e(x)</math>是平衡得分。<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* 任何比倾向性评分“精细”的得分都是平衡得分(即:对于函数''f'',<math>e(X)=f(b(X))</math>)。倾向性评分是最粗粒度的平衡得分函数,因为它把一个(可能是)多维的对象(''X''<sub>''i''</sub>)转换成一维(尽管其他维度显然也存在),而<math>b(X)=X</math>则是最细粒度的平衡得分函数。<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
* 如果对于给定的''X'',处理分配满足强可忽略条件,则:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:* 给定任何的平衡函数,具体来说,给定倾向性评分,处理分配也是强可忽略的:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
:* 对于有相同平衡得分值的处理样本和对照样本,它们响应变量均值的差值(即:<math>\bar{r}_1-\bar{r}_0</math>),可以作为[[average treatment effect|平均处理效应]]的[[Bias of an estimator|无偏估计量]]:<math>E[r_1]-E[r_0]</math>。<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
* 利用平衡得分的样本估计可产生在X上平衡的样本<br />
<br />
===Relationship to sufficiency===<br />
===与充分性的关系===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
If we think of the value of Z as a parameter of the population that impacts the distribution of X then the balancing score serves as a sufficient statistic for Z. Furthermore, the above theorems indicate that the propensity score is a minimal sufficient statistic if thinking of Z as a parameter of X. Lastly, if treatment assignment Z is strongly ignorable given X then the propensity score is a minimal sufficient statistic for the joint distribution of {\displaystyle (r_{0},r_{1})}{\displaystyle (r_{0},r_{1})}.<br />
<br />
如果我们把''Z''的值想成影响''X''分布的群体参数,则平衡得分充当了''Z''的充分统计量。进一步,上述定理指出,如果把''Z''视为''X''的参数,则倾向性评分就是最小充分统计量。最后,给定''X'',如果''Z''是强可忽略的,则倾向性评分是<math>(r_0, r_1)</math>联合分布的最小统计量<br />
<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
===存在混杂变量的图检测方法===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
Judea Pearl has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.[2]<br />
<br />
朱迪亚·珀尔Judea Pearl已经表明存在一个简单的图检测方法,称为后门准则,它可以检测到混杂变量的存在。为了估计处理效果,背景变量X必须阻断图中的所有后门路径。可以通过在回归的控制变量中加入混杂变量,或者在混杂变量上进行匹配来实现阻断。<br />
<br />
==Disadvantages==<br />
==缺点==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
PSM 已经被证明会加剧模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。匹配方法背后的见解仍然成立,但应该与其他匹配方法一起应用;倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
【机器翻译】与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。<br />
<br />
与其他匹配过程一样,PSM也是从观测数据中估计平均处理效应。在引入PSM之时,它的主要优点是,通过使用协变量的线性组合得到一个单一评分,以大量的协变量为基础平衡了处理组和对照组,却不大量损失观测数据。如果在有众多协变量的情况下,对每一个些变量都分别做处理单元和对照单元平衡的话,就需要大量的观测数据来克服”维度问题“,即,每引入一个新的平衡协变量都会在几何上增加最小所需的观测样本数量<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
【机器翻译】PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。<br />
<br />
PSM的一个缺点是它只能概括已观测的(和可观测的)协变量,而无法概括潜在变量。那些能影响处理分配却不可观测的因素无法被纳入匹配过程的考虑范围。由于匹配过程只控制可观测变量,那些隐藏的偏倚在完成匹配后可能依然存在。另一个问题是PSM还要求在大量样本中,在处理组和对照组之间有大量的重叠。<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
【机器翻译】朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。<br />
<br />
Judea Pearl也提出了关于匹配方法的一般性担忧,他认为对可观测变量进行匹配可能会让那些原本处于休眠状态的混杂因素被释放,从而实际上可能加剧隐藏的偏倚。同样,Pearl认为,只有通过对处理、结果、可观测和不可观测的协变量之间的定性因果关系进行建模,才能确保(渐进地)减少偏倚。当试验者无法控制<font color="#32cd32">对独立变量和因变量之间观察到的关系的替代性、非因果性解释时</font>,混杂就会发生。这样的控制应该满足Pearl的“后门准则”。它很容易手工实现。<br />
<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向性评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22748倾向得分匹配2021-05-31T03:42:23Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对试验单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
之所以有可能出现偏倚,是因为试验组和对照组处理结果(如平均处理效果)的差异,可能更多反映了决定样本是否接受处理的某个影响因素,而不是处理效果本身。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,平衡分配试验组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是分别从试验单元和对照单元中各选择一个样本,让它们在所有的协变量上都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。<br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比吸烟者和不吸烟者而评估的处理效果会产生偏差,它会受到任何能影响吸烟行为因素的影响(例如:性别及年龄)。PSM要做的是通过让试验组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与试验组中的类似的单元很少; (ii)选择与试验单元类似的对照单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
在正常的匹配中,区分试验组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自试验组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。<br />
<br />
PSM使用了一种基于观察变量预测样本落入不同分组(例如,试验组与控制组)的概率的方法来创造一个反事实的群体,通常用Logistic回归来预测概率。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:''Z'' = 1, 如果参加试验(属于试验组);''Z'' = 0,如果未参加试验(属于对照组)。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(假设其既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。<br />
<br />
2. <font color="#32cd32">检查协变量在倾向性评分的分层中试验组和对照组是否平衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
3. 根据倾向性评分,将每个试验参与者与一个或多个非试验参与者进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:用试验单元的一段宽度的倾向性评分范围选取对照单元,通常用倾向性评分的标准差的一部分确定这一宽度<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本<br />
<br />
4. <font color="#32cd32">验证协变量在跨越试验组和对照组的匹配样本或加权样本中是否平衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5.【机器翻译】基于新样本的多变量分析<br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个试验参与者都匹配了多个非试验参与者,则应用适当的非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。<br />
<br />
注意:当一个处理样本有多个匹配时,加权最小二乘法是必需的,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 正式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[独立同分布随机变量|i.i.d]物体。每个物体''i''如果做了试验则响应是<math>r_{1i}</math>,否则响应是<math>r_{0i}</math>。被估计量是[[平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中可以不包括哪些决定是否接受处理的特征。假设对单元的编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分将省略索引''i'',同时仍会讨论某些物体的随机行为。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有协变量向量''X''(即:条件非混杂变量),以及一些'''潜在结果'''''r''<sub>0</sub>和''r''<sub>1</sub>分别对应着控制和处理两种情况。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说对样本接受处理的指配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
The most trivial function is <math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
===Main theorems===<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
===Relationship to sufficiency===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
<br />
==Disadvantages==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。<br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。<br />
<br />
<br />
Category:Regression analysis<br />
<br />
类别: 回归分析<br />
<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
<br />
Category:Epidemiology<br />
<br />
类别: 流行病学<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
Category:Observational study<br />
<br />
类别: 观察性研究<br />
<br />
<br />
<br />
Category:Causal inference<br />
<br />
类别: 因果推理<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22746倾向得分匹配2021-05-31T03:37:08Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
【机器翻译】在观察数据的统计分析中,倾向性评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
在观察数据的统计分析中,倾向性评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对试验单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
【机器翻译】出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
之所以有可能出现偏倚,是因为试验组和对照组处理结果(如平均处理效果)的差异,可能更多反映了决定样本是否接受处理的某个影响因素,而不是处理效果本身。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,平衡分配试验组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是分别从试验单元和对照单元中各选择一个样本,让它们在所有的协变量上都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
【机器翻译】例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。<br />
<br />
例如,人们想知道吸烟的后果。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比吸烟者和不吸烟者而评估的处理效果会产生偏差,它会受到任何能影响吸烟行为因素的影响(例如:性别及年龄)。PSM要做的是通过让试验组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
<br />
== Overview ==<br />
<br />
== 综述 ==<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
【机器翻译】PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)对照组与试验组中的类似的单元很少; (ii)选择与试验单元类似的对照单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
【机器翻译】在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
在正常的匹配中,区分试验组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自试验组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
【机器翻译】PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。<br />
<br />
PSM使用了一种基于观察变量预测样本落入不同分组(例如,试验组与控制组)的概率的方法来创造一个反事实的群体,通常用Logistic回归来预测概率。倾向性评分可用于匹配,也可作为协变量,可以单独使用,也可以与其他匹配变量或协变量一同使用。<br />
<br />
<br />
== General procedure ==<br />
<br />
== 一般步骤 ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1. 做Logistic回归:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*因变量:''Z'' = 1, 如果参加试验(属于试验组);''Z'' = 0,如果未参加试验(属于对照组)。<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*选择合适的混杂因素(假设其既影响处理方式又影响处理结果的变量)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
*计算倾向性评分的[[Estimator|估计量]]:预测概率(''p'')或log[''p''/(1&nbsp;−&nbsp;''p'')]。<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. 【机器翻译】检查协变量是平衡的治疗和比较组内的倾向分层。<br />
<br />
2. <font color="#32cd32">检查协变量在倾向性评分的分层中试验组和对照组是否平衡</font><br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
*使用标准化差异指标或者图形来检验分布情况<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. 【机器翻译】根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
3. 根据倾向性评分,将每个试验参与者与一个或多个非试验参与者进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*[[Nearest neighbor search|最近邻匹配]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*卡钳匹配:用试验单元的一段宽度的倾向性评分范围选取对照单元,通常用倾向性评分的标准差的一部分确定这一宽度<br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Mahalanobis distance|马氏度量]] 与PSM配合使用<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*[[Stratified sampling|分层匹配]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*<font color="#32cd32">双重差分匹配(核和局部线性加权)</font><br />
<br />
*Exact matching<br />
<br />
*精确匹配<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4.【机器翻译】验证协变量是平衡的处理和对照组在匹配或加权样本<br />
<br />
4. <font color="#32cd32">验证协变量在跨越试验组和对照组的匹配样本或加权样本中是否平衡</font><br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5.【机器翻译】基于新样本的多变量分析<br />
<br />
5. 基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
*如果每个试验参与者都匹配了多个非试验参与者,则应用适当的非独立匹配样本分析<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
【机器翻译】注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。<br />
<br />
注意:当一个处理样本有多个匹配时,加权最小二乘法是必需的,而不能用普通最小二乘法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
== 正式定义 ==<br />
<br />
===Basic settings===<br />
<br />
===基本设置===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
基本场景<ref name="Rosenbaum 1983 41–55"/>是,有两种处理方式(分别记为1和0),''N''个[独立同分布随机变量|i.i.d]物体。每个物体''i''如果做了试验则响应是<math>r_{1i}</math>,否则响应是<math>r_{0i}</math>。被估计量是[[平均处理效应]]:<math>E[r_1]-E[r_0]</math>。变量<math>Z_i</math>指示物体''i''接受处理(''Z''&nbsp;=&nbsp;1)还是接受控制(''Z''&nbsp;=&nbsp;0)。让<math>X_i</math>代表第''i''个物体处理前观测值(或者协变量)的向量。对<math>X_i</math>的测量发生于处理前,但是<math>X_i</math>中可以不包括哪些决定是否接受处理的特征。假设对单元的编号(即:''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'')不包含任何<math>X_i</math>所包含信息之外的的信息。以下部分将省略索引''i'',同时仍会讨论某些物体的随机行为。<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
===强可忽略处理分配===<br />
<br />
{{See also|Ignorability}}<br />
<br />
{{See also|可忽略性}}<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
设某个物体有协变量向量''X''(即:条件无混杂),以及一些'''潜在结果‘’‘ ''r''<sub>0</sub>和''r''<sub>1</sub>分别对应着控制和处理两种情况。如果潜在结果在给定背景变量''X''的条件下独立于处理举动(''Z''),则可以说对样本接受处理的指配是'''强可忽略'''的。可简洁表述为<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
这里<math>\perp</math>代表[[statistical independence|统计独立]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
===Balancing score===<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
The most trivial function is <math> b(X) = X</math>.<br />
<br />
===Propensity score===<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
===Main theorems===<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
===Relationship to sufficiency===<br />
<br />
If we think of the value of ''Z'' as a [[Statistical parameter|parameter]] of the population that impacts the distribution of ''X'' then the balancing score serves as a [[Sufficient_statistic#Mathematical_definition|sufficient statistic]] for ''Z''. Furthermore, the above theorems indicate that the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] if thinking of ''Z'' as a parameter of ''X''. Lastly, if treatment assignment ''Z'' is strongly ignorable given ''X'' then the propensity score is a [[Sufficient_statistic#Minimal_sufficiency|minimal sufficient statistic]] for the joint distribution of <math>(r_0, r_1)</math>.<br />
<br />
===Graphical test for detecting the presence of confounding variables===<br />
<br />
[[Judea Pearl]] has shown that there exists a simple graphical test, called the back-door criterion, which detects the presence of confounding variables. To estimate the effect of treatment, the background variables X must block all back-door paths in the graph. This blocking can be done either by adding the confounding variable as a control in regression, or by matching on the confounding variable.<ref name="pearl">{{cite book |last=Pearl |first=J. |year=2000 |title=Causality: Models, Reasoning, and Inference |url=https://archive.org/details/causalitymodelsr0000pear |url-access=registration |location=New York |publisher=Cambridge University Press |isbn=978-0-521-77362-1 }}</ref><br />
<br />
<br />
==Disadvantages==<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
【机器翻译】PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。<br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。<br />
<br />
<br />
Category:Regression analysis<br />
<br />
类别: 回归分析<br />
<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
<br />
Category:Epidemiology<br />
<br />
类别: 流行病学<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
Category:Observational study<br />
<br />
类别: 观察性研究<br />
<br />
<br />
<br />
Category:Causal inference<br />
<br />
类别: 因果推理<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22661倾向得分匹配2021-05-29T18:23:44Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
在观察数据的统计分析中,倾向评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
在观察数据的统计分析中,倾向评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果,方法是将协变量对样本“是否接受处理”的影响考虑在内。PSM试图减少由于混杂变量造成的偏倚。这些偏倚一般会在那些只对试验单元和对照单元的结果做简单对比的评估中出现。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
之所以有可能出现偏倚,是因为试验组和对照组处理结果(如平均处理效果)的差异,可能更多反映了决定样本是否接受处理的某个影响因素,而不是处理效果本身。在随机实验中,随机化选择样本可以做到对处理效果的无偏估计,根据大数定律,随机化意味着基于协变量的平均水平,平衡分配试验组和对照组。不幸的是,对于观察性研究来说,研究对象通常不是随机接受处理的。匹配就是要减少对象非随机接受处理产生的偏倚,并模拟随机试验,方法是分别从试验单元和对照单元中各选择一个样本,让它们在所有的协变量上都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。<br />
<br />
例如,人们对吸烟的后果感兴趣。但是随机分配让患者“吸烟”是不道德的,所以需要做一个观察性研究。简单对比吸烟者和不吸烟者而评估的处理效果会产生偏差,它会受到任何能影响吸烟行为因素的影响(例如:性别及年龄)。PSM要做的是通过让试验组和对照组的控制变量尽量相似来达到控制这些偏差的目的。<br />
<br />
<br />
== Overview ==<br />
<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。<br />
<br />
<br />
<br />
== General procedure ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1.返回文章页面 Logit模型:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2.检查协变量是平衡的治疗和比较组内的倾向分层。<br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3.根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*Exact matching<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4.验证协变量是平衡的处理和对照组在匹配或加权样本<br />
<br />
<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5.基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
===Basic settings===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
The basic case<br />
<br />
基本情况<br />
<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
{{See also|Ignorability}}<br />
<br />
<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。<br />
<br />
where <math>\perp</math> denotes [[statistical independence]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。<br />
<br />
===Balancing score===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
<br />
<br />
The most trivial function is <math> b(X) = X</math>.<br />
<br />
<br />
<br />
===Propensity score===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
<br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
<br />
<br />
===Main theorems===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
<br />
Category:Regression analysis<br />
<br />
类别: 回归分析<br />
<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
<br />
Category:Epidemiology<br />
<br />
类别: 流行病学<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
Category:Observational study<br />
<br />
类别: 观察性研究<br />
<br />
<br />
<br />
Category:Causal inference<br />
<br />
类别: 因果推理<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22660倾向得分匹配2021-05-29T17:03:16Z<p>Aceyuan:/* Overview */</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
在观察数据的统计分析中,倾向评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
在观察数据的统计分析中,倾向评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果。它的方法是计算协变量对预测效果的贡献。倾向评分匹配试图减少由于混杂变量造成的偏倚。如果只是通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果,就会出现这些偏倚。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
之所以有可能出现偏倚,是因为影响试验组和对照组处理结果差异(如平均处理效果)的因素,可能更多是影响了试验的分组,而不是试验的结果。在随机实验中,随机试验可以对处理效果进行无偏估计; 根据大数定律,随机试验对于每个协变量都意味着试验组在平均水平上达到平衡。不幸的是,对于观察性研究来说,研究对象通常不是被随机地指定到试验组的。匹配就是要减少指定试验对象时引入的偏倚,模拟随机试验,方法是分别从试验单元和对照单元中各选择一个样本,让它们在所有的协变量上都比较接近。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。<br />
<br />
<br />
<br />
== Overview ==<br />
<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。<br />
<br />
<br />
<br />
== General procedure ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1.返回文章页面 Logit模型:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2.检查协变量是平衡的治疗和比较组内的倾向分层。<br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3.根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*Exact matching<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4.验证协变量是平衡的处理和对照组在匹配或加权样本<br />
<br />
<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5.基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
===Basic settings===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
The basic case<br />
<br />
基本情况<br />
<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
{{See also|Ignorability}}<br />
<br />
<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。<br />
<br />
where <math>\perp</math> denotes [[statistical independence]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。<br />
<br />
===Balancing score===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
<br />
<br />
The most trivial function is <math> b(X) = X</math>.<br />
<br />
<br />
<br />
===Propensity score===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
<br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
<br />
<br />
===Main theorems===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
<br />
Category:Regression analysis<br />
<br />
类别: 回归分析<br />
<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
<br />
Category:Epidemiology<br />
<br />
类别: 流行病学<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
Category:Observational study<br />
<br />
类别: 观察性研究<br />
<br />
<br />
<br />
Category:Causal inference<br />
<br />
类别: 因果推理<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22632倾向得分匹配2021-05-29T04:25:28Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
在观察数据的统计分析中,倾向评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
在观察数据的统计分析中,倾向评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,用来估计治疗、政策或其他干预的效果。它的方法是计算协变量对预测效果的贡献。倾向评分匹配试图减少由于混杂变量造成的偏倚。如果只是通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果,就会出现这些偏倚。保罗·罗森鲍姆Paul R. Rosenbaum和唐纳德·鲁宾Donald Rubin在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
之所以有可能出现偏倚,是因为影响试验组和对照组处理结果差异(如平均处理效果)的因素,可能更多是影响了试验的分组,而不是试验的结果。在随机实验中,随机试验可以对处理效果进行无偏估计; 根据大数定律,随机试验对于每个协变量都意味着试验组在平均水平上达到平衡。不幸的是,对于观察性研究来说,研究对象通常不是被随机地指定到试验组的。匹配就是要减少指定试验对象时引入的偏倚,模拟随机试验,它通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。<br />
<br />
<br />
<br />
== Overview ==<br />
<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。<br />
<br />
<br />
<br />
== General procedure ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1.返回文章页面 Logit模型:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2.检查协变量是平衡的治疗和比较组内的倾向分层。<br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3.根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*Exact matching<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4.验证协变量是平衡的处理和对照组在匹配或加权样本<br />
<br />
<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5.基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
===Basic settings===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
The basic case<br />
<br />
基本情况<br />
<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
{{See also|Ignorability}}<br />
<br />
<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。<br />
<br />
where <math>\perp</math> denotes [[statistical independence]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。<br />
<br />
===Balancing score===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
<br />
<br />
The most trivial function is <math> b(X) = X</math>.<br />
<br />
<br />
<br />
===Propensity score===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
<br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
<br />
<br />
===Main theorems===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
<br />
Category:Regression analysis<br />
<br />
类别: 回归分析<br />
<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
<br />
Category:Epidemiology<br />
<br />
类别: 流行病学<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
Category:Observational study<br />
<br />
类别: 观察性研究<br />
<br />
<br />
<br />
Category:Causal inference<br />
<br />
类别: 因果推理<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuanhttps://wiki.swarma.org/index.php?title=%E5%80%BE%E5%90%91%E5%BE%97%E5%88%86%E5%8C%B9%E9%85%8D&diff=22631倾向得分匹配2021-05-28T11:32:39Z<p>Aceyuan:</p>
<hr />
<div>此词条暂由彩云小译翻译,翻译字数共893,未经人工整理和审校,带来阅读不便,请见谅。<br />
<br />
{{Short description|Statistical matching technique}}<br />
<br />
In the [[statistics|statistical]] analysis of [[observational study|observational data]], '''propensity score matching''' ('''PSM''') is a [[Matching (statistics)|statistical matching]] technique that attempts to [[Estimation theory|estimate]] the effect of a treatment, policy, or other intervention by accounting for the [[covariate]]s that predict receiving the treatment. PSM attempts to reduce the [[Bias (statistics)|bias]] due to [[confounding]] variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among [[Statistical unit|units]] that [[Treatment and control groups|received the treatment versus those that did not]]. [[Paul R. Rosenbaum]] and [[Donald Rubin]] introduced the technique in 1983.<ref name="Rosenbaum 1983 41–55">{{cite journal |last=Rosenbaum |first=Paul R. |last2=Rubin |first2=Donald B. |title=The Central Role of the Propensity Score in Observational Studies for Causal Effects |journal=[[Biometrika]] |year=1983 |volume=70 |issue=1 |pages=41–55 |doi=10.1093/biomet/70.1.41 |doi-access=free }}</ref><br />
<br />
In the statistical analysis of observational data, propensity score matching (PSM) is a statistical matching technique that attempts to estimate the effect of a treatment, policy, or other intervention by accounting for the covariates that predict receiving the treatment. PSM attempts to reduce the bias due to confounding variables that could be found in an estimate of the treatment effect obtained from simply comparing outcomes among units that received the treatment versus those that did not. Paul R. Rosenbaum and Donald Rubin introduced the technique in 1983.<br />
<br />
在观察数据的统计分析中,倾向评分匹配是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
在观察数据的统计分析中,倾向评分匹配Propensity Score Matching (PSM)是一种统计匹配技术,它试图通过计算预测接受治疗的协变量来估计治疗、政策或其他干预的效果。PSM 试图减少由于混杂变量造成的偏倚,这些变量可以通过简单地比较接受治疗的单位和没有接受治疗的单位之间的结果来估计治疗效果。保罗 · 罗森鲍姆和唐纳德 · 鲁宾在1983年介绍了这项技术。<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the [[average treatment effect]]) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In [[randomized experiment]]s, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the [[law of large numbers]]. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. [[Matching (statistics)|Matching]] attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
The possibility of bias arises because a difference in the treatment outcome (such as the average treatment effect) between treated and untreated groups may be caused by a factor that predicts treatment rather than the treatment itself. In randomized experiments, the randomization enables unbiased estimation of treatment effects; for each covariate, randomization implies that treatment-groups will be balanced on average, by the law of large numbers. Unfortunately, for observational studies, the assignment of treatments to research subjects is typically not random. Matching attempts to reduce the treatment assignment bias, and mimic randomization, by creating a sample of units that received the treatment that is comparable on all observed covariates to a sample of units that did not receive the treatment.<br />
<br />
出现偏倚的可能性是因为治疗组和未治疗组之间治疗结果(如平均治疗效果)的差异可能是由预测治疗的因素而不是治疗本身造成的。在随机实验中,随机化可以对治疗效果进行无偏估计; 对于每个协变量,随机化意味着治疗组将按照大数定律在平均水平上达到平衡。不幸的是,对于观察性研究来说,对研究对象的治疗分配通常不是随机的。匹配试图减少处理分配偏差,并模拟随机化,通过创建一个样本单位接受的处理是可比的所有观察到的协变量的一个样本单位没有接受处理。<br />
<br />
<br />
<br />
For example, one may be interested to know the [[Health_effects_of_tobacco#Early_observational_studies|consequences of smoking]]. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
For example, one may be interested to know the consequences of smoking. An observational study is required since it is unethical to randomly assign people to the treatment 'smoking.' The treatment effect estimated by simply comparing those who smoked to those who did not smoke would be biased by any factors that predict smoking (e.g.: gender and age). PSM attempts to control for these biases by making the groups receiving treatment and not-treatment comparable with respect to the control variables.<br />
<br />
例如,人们可能有兴趣知道吸烟的后果。因为随机分配患者接受‘吸烟’治疗是不道德的,所以需要一个观察性研究简单地比较吸烟者和不吸烟者的治疗效果会受到任何预测吸烟的因素的影响(例如:。: 性别及年龄)。PSM 试图通过使接受治疗和不接受治疗的组与控制变量相比较来控制这些偏差。<br />
<br />
<br />
<br />
== Overview ==<br />
<br />
<br />
<br />
PSM is for cases of [[Inductive reasoning#Causal inference|causal inference]] and simple selection bias in [[non-experimental]] settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM is for cases of causal inference and simple selection bias in non-experimental settings in which: (i) few units in the non-treatment comparison group are comparable to the treatment units; and (ii) selecting a subset of comparison units similar to the treatment unit is difficult because units must be compared across a high-dimensional set of pretreatment characteristics.<br />
<br />
PSM 适用于非实验环境中因果推断和简单选择偏差的情况,其中: (i)非处理对照组中与处理单元可比的单元很少; (ii)选择与处理单元类似的比较单元子集很困难,因为必须跨一组高维预处理特征进行比较。<br />
<br />
<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial [[Errors and residuals|error]] may be introduced. For example, if only the worst cases from the [[control group|untreated "comparison" group]] are compared to only the best cases from the [[treatment group]], the result may be [[regression toward the mean]], which may make the comparison group look better or worse than reality.<br />
<br />
In normal matching, single characteristics that distinguish treatment and control groups are matched in an attempt to make the groups more alike. But if the two groups do not have substantial overlap, then substantial error may be introduced. For example, if only the worst cases from the untreated "comparison" group are compared to only the best cases from the treatment group, the result may be regression toward the mean, which may make the comparison group look better or worse than reality.<br />
<br />
在正常的匹配中,区分治疗组和对照组的单一特征被匹配,试图使这些组更加相似。但是,如果这两个组没有实质性的重叠,那么可能会引入实质性的错误。例如,如果只将来自未经治疗的对照组的最差病例与来自治疗组的最好病例进行比较,结果可能是趋中回归,这可能使对照组看起来比实际情况更好或更糟。<br />
<br />
<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from [[logistic regression]] to create a [[Impact evaluation#Counterfactual evaluation designs|counterfactual group]]. Propensity scores may be used for matching or as [[covariance|covariate]]s, alone or with other matching variables or covariates.<br />
<br />
PSM employs a predicted probability of group membership—e.g., treatment versus control group—based on observed predictors, usually obtained from logistic regression to create a counterfactual group. Propensity scores may be used for matching or as covariates, alone or with other matching variables or covariates.<br />
<br />
PSM 使用了一种预测的群体成员概率---- 例如,治疗组与控制组---- 基于观察预测,通常从 Logit模型获得来创造一个反事实的群体。倾向得分可用于匹配或作为协变量,单独或与其他匹配变量或协变量。<br />
<br />
<br />
<br />
== General procedure ==<br />
<br />
1. Run [[logistic regression]]:<br />
<br />
1. Run logistic regression:<br />
<br />
1.返回文章页面 Logit模型:<br />
<br />
*Dependent variable: ''Z'' = 1, if unit participated (i.e. is member of the treatment group); ''Z'' = 0, if unit did not participate (i.e. is member of the control group).<br />
<br />
*Choose appropriate confounders (variables hypothesized to be associated with both treatment and outcome)<br />
<br />
*Obtain an [[Estimator|estimation]] for the propensity score: predicted probability (''p'') or log[''p''/(1&nbsp;−&nbsp;''p'')].<br />
<br />
<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2. Check that covariates are balanced across treatment and comparison groups within strata of the propensity score.<br />
<br />
2.检查协变量是平衡的治疗和比较组内的倾向分层。<br />
<br />
* Use standardized differences or graphs to examine distributions<br />
<br />
<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3. Match each participant to one or more nonparticipants on propensity score, using one of these methods:<br />
<br />
3.根据倾向得分,将每个参与者与一个或多个非参与者进行匹配,使用以下方法之一:<br />
<br />
*[[Nearest neighbor search|Nearest neighbor matching]]<br />
<br />
*Caliper matching: comparison units within a certain width of the propensity score of the treated units get matched, where the width is generally a fraction of the standard deviation of the propensity score <br />
<br />
*[[Mahalanobis distance|Mahalanobis metric]] matching in conjunction with PSM<br />
<br />
*[[Stratified sampling|Stratification matching]]<br />
<br />
*Difference-in-differences matching (kernel and local linear weights)<br />
<br />
*Exact matching<br />
<br />
<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4. Verify that covariates are balanced across treatment and comparison groups in the matched or weighted sample<br />
<br />
4.验证协变量是平衡的处理和对照组在匹配或加权样本<br />
<br />
<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5. Multivariate analysis based on new sample<br />
<br />
5.基于新样本的多变量分析<br />
<br />
*Use analyses appropriate for non-independent matched samples if more than one nonparticipant is matched to each participant<br />
<br />
<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
Note: When you have multiple matches for a single treated observation, it is essential to use Weighted Least Squares rather than Ordinary Least Squares.<br />
<br />
注意: 当你有多个匹配的单一处理的观察,它是必不可少的使用加权最小二乘而不是一般最小平方法。<br />
<br />
<br />
<br />
== Formal definitions ==<br />
<br />
===Basic settings===<br />
<br />
The basic case<ref name="Rosenbaum 1983 41–55"/> is of two treatments (numbered 1 and 0), with ''N'' [Independent and identically distributed random variables|i.i.d] subjects. Each subject ''i'' would respond to the treatment with <math>r_{1i}</math> and to the control with <math>r_{0i}</math>. The quantity to be estimated is the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. The variable <math>Z_i</math> indicates if subject ''i'' got treatment (''Z''&nbsp;=&nbsp;1) or control (''Z''&nbsp;=&nbsp;0). Let <math>X_i</math> be a vector of observed pretreatment measurement (or covariate) for the ''i''th subject. The observations of <math>X_i</math> are made prior to treatment assignment, but the features in <math>X_i</math> may not include all (or any) of the ones used to decide on the treatment assignment. The numbering of the units (i.e.: ''i''&nbsp;=&nbsp;1,&nbsp;...,&nbsp;''i''&nbsp;=&nbsp;''N'') are assumed to not contain any information beyond what is contained in <math>X_i</math>. The following sections will omit the ''i'' index while still discussing about the stochastic behavior of some subject.<br />
<br />
The basic case<br />
<br />
基本情况<br />
<br />
<br />
<br />
===Strongly ignorable treatment assignment===<br />
<br />
{{See also|Ignorability}}<br />
<br />
<br />
<br />
PSM has been shown to increase model "imbalance, inefficiency, model dependence, and bias," which is not the case with most other matching methods. The insights behind the use of matching still hold but should be applied with other matching methods; propensity scores also have other productive uses in weighting and doubly robust estimation.<br />
<br />
PSM 已经被证明会增加模型的“不平衡性、低效率、模型依赖性和偏差”,这与大多数其他匹配方法不同。使用匹配的见解仍然有效,但应该与其他匹配方法一起应用; 倾向得分在加权和双重稳健估计方面也有其他有益的用途。<br />
<br />
Let some subject have a vector of covariates ''X'' (i.e.: conditionally unconfounded), and some '''potential outcomes''' ''r''<sub>0</sub> and ''r''<sub>1</sub> under control and treatment, respectively. Treatment assignment is said to be '''strongly ignorable''' if the potential outcomes are [[statistical independence|independent]] of treatment (''Z'') conditional on background variables ''X''. This can be written compactly as<br />
<br />
<br />
<br />
Like other matching procedures, PSM estimates an average treatment effect from observational data. The key advantages of PSM were, at the time of its introduction, that by using a linear combination of covariates for a single score, it balances treatment and control groups on a large number of covariates without losing a large number of observations. If units in the treatment and control were balanced on a large number of covariates one at a time, large numbers of observations would be needed to overcome the "dimensionality problem" whereby the introduction of a new balancing covariate increases the minimum necessary number of observations in the sample geometrically.<br />
<br />
与其他匹配程序一样,PSM 从观测数据中估计平均处理效果。在引入 PSM 的时候,它的主要优点是,通过使用一个线性组合的协变量作为一个单一的评分,它平衡了治疗组和对照组在大量的协变量上,而不会失去大量的观察数据。如果处理和控制中的单元在大量的协变量上一次平衡,就需要大量的观测数据来克服“维数问题”,即引入新的平衡协变量几何地增加样本中必要的最小观测数据。<br />
<br />
:<math> r_0, r_1 \perp Z \mid X </math><br />
<br />
<br />
<br />
One disadvantage of PSM is that it only accounts for observed (and observable) covariates and not latent characteristics. Factors that affect assignment to treatment and outcome but that cannot be observed cannot be accounted for in the matching procedure. As the procedure only controls for observed variables, any hidden bias due to latent variables may remain after matching. Another issue is that PSM requires large samples, with substantial overlap between treatment and control groups.<br />
<br />
PSM 的一个缺点是它只能解释观察到的(和可观察到的)协变量,而不能解释潜在的特征。影响治疗分配和结果但无法观察的因素不能在匹配程序中说明。由于程序只控制观察变量,任何隐藏的偏见由于潜在变量可能仍然匹配后。另一个问题是 PSM 需要大量的样本,治疗组和对照组之间有大量的重叠。<br />
<br />
where <math>\perp</math> denotes [[statistical independence]].<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
General concerns with matching have also been raised by Judea Pearl, who has argued that hidden bias may actually increase because matching on observed variables may unleash bias due to dormant unobserved confounders. Similarly, Pearl has argued that bias reduction can only be assured (asymptotically) by modelling the qualitative causal relationships between treatment, outcome, observed and unobserved covariates. Confounding occurs when the experimenter is unable to control for alternative, non-causal explanations for an observed relationship between independent and dependent variables. Such control should satisfy the "backdoor criterion" of Pearl. It can also easily be implemented manually.<br />
<br />
朱迪亚 · 珀尔也提出了关于配对的普遍担忧,他认为隐性偏见实际上可能会增加,因为观察变量的配对可能会由于潜在的未观察混杂因素而释放出偏见。同样,珀尔认为,只有通过建立治疗、结果、观察和未观察协变量之间的定性因果关系模型,才能确保(渐近地)减少偏见。当实验者无法控制对独立变量和因变量之间观察到的关系的替代性、非因果性解释时,混淆就发生了。这种控制应满足珍珠的“后门规范”。它也可以很容易地手动实现。<br />
<br />
===Balancing score===<br />
<br />
A '''balancing score''' ''b''(''X'') is a function of the observed covariates ''X'' such that the [[conditional probability|conditional distribution]] of ''X'' given ''b''(''X'') is the same for treated (''Z''&nbsp;=&nbsp;1) and control (''Z''&nbsp;=&nbsp;0) units:<br />
<br />
<br />
<br />
:<math> Z \perp X \mid b(X).</math><br />
<br />
<br />
<br />
The most trivial function is <math> b(X) = X</math>.<br />
<br />
<br />
<br />
===Propensity score===<br />
<br />
A '''propensity score''' is the [[probability]] of a unit (e.g., person, classroom, school) being assigned to a particular treatment given a set of observed covariates. Propensity scores are used to reduce [[selection bias]] by equating groups based on these covariates.<br />
<br />
<br />
<br />
Suppose that we have a binary treatment [[Indicator function|indicator]] ''Z'', a response variable ''r'', and background observed covariates ''X''. The propensity score is defined as the [[conditional probability]] of treatment given background variables:<br />
<br />
<br />
<br />
:<math>e(x) \ \stackrel{\mathrm{def}}{=}\ \Pr(Z=1 \mid X=x).</math><br />
<br />
<br />
<br />
In the context of [[causal inference]] and [[survey methodology]], propensity scores are estimated (via methods such as [[logistic regression]], [[random forests]], or others), using some set of covariates. These propensity scores are then used as estimators for weights to be used with [[Inverse probability weighting]] methods.<br />
<br />
<br />
<br />
===Main theorems===<br />
<br />
The following were first presented, and proven, by Rosenbaum and Rubin in 1983:<ref name="Rosenbaum 1983 41–55"/><br />
<br />
<br />
<br />
* The propensity score <math>e(x)</math> is a balancing score.<br />
<br />
* Any score that is 'finer' than the propensity score is a balancing score (i.e.: <math>e(X)=f(b(X))</math> for some function ''f''). The propensity score is the coarsest balancing score function, as it takes a (possibly) multidimensional object (''X''<sub>''i''</sub>) and transforms it into one dimension (although others, obviously, also exist), while <math>b(X)=X</math> is the finest one.<br />
<br />
* If treatment assignment is strongly ignorable given ''X'' then:<br />
<br />
:* It is also strongly ignorable given any balancing function. Specifically, given the propensity score:<br />
<br />
:::<math> (r_0, r_1) \perp Z \mid e(X).</math><br />
<br />
Category:Regression analysis<br />
<br />
类别: 回归分析<br />
<br />
:* For any value of a balancing score, the difference between the treatment and control means of the samples at hand (i.e.: <math>\bar{r}_1-\bar{r}_0</math>), based on subjects that have the same value of the balancing score, can serve as an [[Bias of an estimator|unbiased estimator]] of the [[average treatment effect]]: <math>E[r_1]-E[r_0]</math>. <br />
<br />
Category:Epidemiology<br />
<br />
类别: 流行病学<br />
<br />
* Using sample estimates of balancing scores can produce sample balance on&nbsp;''X''<br />
<br />
Category:Observational study<br />
<br />
类别: 观察性研究<br />
<br />
<br />
<br />
Category:Causal inference<br />
<br />
类别: 因果推理<br />
<br />
<noinclude><br />
<br />
<small>This page was moved from [[wikipedia:en:Propensity score matching]]. Its edit history can be viewed at [[倾向评分/edithistory]]</small></noinclude><br />
<br />
[[Category:待整理页面]]</div>Aceyuan