Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction
Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction
复制标题
DOI：
10.1111/insr.12170
复制
发表时间：
2016-04
影响因子：
2
通讯作者：
C. Blumberg
中科院分区：
文献类型：
--
作者：
C. Blumberg
Guido Imbens和Don Rubin对因果推理的潜在结果框架进行了深刻的讨论。在这个框架中，目标是定义处理的效果（当然，包括“处理”本身的定义），了解这种对象何时在人群中（或至少在超人群中）可识别，并对该兴趣量以及福利措施或其他政策相关效果等相关量进行估计和统计推断。尽管这个目标是基本的，但最终的答案远未确定，因为文献中提出了许多竞争和/或互补的方法（例如，Heckman和Vytlacil 2007； Manski 2008; Pearl 2009; VanderWeele 2015； Hernán和Robins 2016）。作者使用潜在结果框架来研究应用工作中三个最重要的设置中的因果推理：(1)随机实验，(2)观察性研究中的无混杂治疗分配（假设也称为随机缺失，条件独立，可观察性选择或可忽略性），以及(3)非依从性实验（工具变量方法的特殊情况）。本书的一个关键特点是反复使用真实的经验例子，以及实现所需的最重要计算的详细推导，所有这些都为如何从理论到实践提供了急需的指导。虽然这些特性肯定有助于读者理解不同的方法是如何工作的，但我怀疑如果作者提供自始至终使用的数据和计算机代码，这本书将会更加有效。作者还非常仔细地区分了识别、估计、推理、实现和解释问题。所有这些特点使这本书必读的所有应用研究人员，无论他们的技术背景。本书的第一和第二部分（第1-11章）简要介绍了因果关系和潜在的结果框架，然后对经典的实验分析进行了彻底的讨论。这本书的一部分向读者介绍了各种相互关联的经典思想，包括费舍尔的精确测试，内曼的重复抽样，基于回归的方法和基于贝叶斯模型的方法，这些方法经常在实证工作中遇到，但被视为根本不同。Imbens和Rubin在一个统一的方法论框架内阐明了这些观点，同时也提供了几个说明性的实证应用。任何有兴趣加深对实验分析和解释的理解的人都会发现这些章节是无价的和永恒的参考。第三部分和第四部分（第12-20章）关注的是治疗分配不混杂的情况，也就是说，治疗分配独立于对可观察特征进行调节后的潜在结果。这种类型的条件独立假设是统计性质的，在实证工作中广泛使用，是因果推理中各种估计和推理程序的基础。在本书的这一部分中，作者不仅回顾了他们在该领域的大部分有影响力的工作，而且还介绍和讨论了各种新思想，如模型，调整参数和协变量选择方法。此外，第五部分（第21-22章）介绍了评估潜在关键识别假设的合理性的方法，以及基于边界和其他敏感性分析的方法，这些方法对于理解在可观测选择假设下获得的结果的稳健性至关重要。最后几章，第六部分（第23-25章），重点分析了不完全依从性的实验。这里的重点主要是作者在局部平均处理效果（LATE）的非参数识别方面的广泛影响的工作，尽管也提出了另一种基于模型的估计和推断方法。总之，这本书提出了一个统一的框架，以因果推理的基础上的潜在结果框架，侧重于实验，无混淆的经典分析，和不合规。这本书已经成为因果推理文学中的经典之作，定义广泛，必将指导这一领域的未来研究。所有研究人员都会从仔细研究这本书中受益，不管他们对这个主题有什么具体的看法。展望未来，这本书的一个额外的好处是，它奠定了在因果推理，如多值，动态和适应性治疗制度更先进的主题的系统和统一的统计分析的基础；differencein-differences方法;回归不连续设计；线性和非线性面板或纵向数据；干预、外部性、溢出和对等效应；外部有效性，元分析和反事实政策评估，只是举几个例子。我们只能希望能说服作者们写第二卷。
Guido Imbens and Don Rubin present an insightful discussion of the potential outcomes framework for causal inference. In this framework, the goal is to define the effect of a treatment (including, of course, a definition of “treatment” itself), to understand when such an object is identifiable in the population (or, at least, in the superpopulation), and to estimate and conduct statistical inference about this quantity of interest as well as related quantities such as welfare measures or other policy-relevant effects. Despite this goal being fundamental, a final answer is far from settled as many competing and/or complementary approaches have been proposed in the literature (e.g., Heckman and Vytlacil 2007; Manski 2008; Pearl 2009; VanderWeele 2015; Hernán and Robins 2016). The authors employ the potential outcomes framework to study causal inference in three of the most important settings in applied work: (1) randomized experiments, (2) unconfounded treatment assignment in observational studies (an assumption also known as missing-at-random, conditional independence, selection-on-observables, or ignorability), and (3) experiments with noncompliance (a special case of instrumental variables methods). A key feature of the book is the recurrent use of real empirical examples, and the detailed derivation of the most important calculations needed for implementation, all of which provide much needed guidance on how to go from theory to practice. While these features surely help the reader understand how the different methods work, I suspect the book would be even more effective if the authors provided the data and computer codes used throughout. The authors are also very careful to distinguish between identification, estimation, inference, implementation, and interpretation issues. All these features make the book amust-read for all applied researchers, regardless of their technical background. Parts I and II (Chapters 1–11) of the book give a brief introduction to causality and the potential outcomes framework, and then present a thorough discussion of the classical analysis of experiments. The part of the book introduces the reader to a variety of interconnected classical ideas, including Fisher’s exact testing, Neyman’s repeated sampling, regression-based methods and Bayesian model-based methods, which are often encountered in empirical work but treated as fundamentally distinct. Imbens and Rubin elucidated these ideas within a unifiedmethodological framework, while also offering several illustrative empirical applications throughout. Anyone interested in deepening their understanding of the analysis and interpretation of experiments will find these chapters to be an invaluable and timeless reference. Parts III and IV (Chapters 12–20) focus on settings in which the treatment assignment is unconfounded, that is, where the treatment assignment is independent of the potential outcomes after conditioning on observable characteristics. This type of conditional independence assumption, which is statistical in nature and widely used in empirical work, is the basis for a variety of estimation and inference procedures in causal inference. In this portion of the book, the authors not only review most of their influential work in the area, but also introduce and discuss various new ideas such as model, tuning parameter, and covariate selection methods. Moreover, Part V (Chapters 21–22) presents methods for assessing the plausibility of the underlying key identifying assumptions, as well as approaches based on bounds and other sensitivity analyses that are crucial to understanding the robustness of the results obtained under the selection-on-observables assumption. The final chapters, Part VI (Chapters 23–25), focus on the analysis of experiments with imperfect compliance. The focus here is mostly on the authors’ widely influential work on nonparametric identification of the local average treatment effect (LATE), although an alternative model-based approach to estimation and inference is also presented. In sum, this book presents a unified framework to causal inference based on the potential outcomes framework, focusing on the classical analysis of experiments, unconfoundedness, and noncompliance. The book has become an instant classic in the causal inference literature, broadly defined, and will certainly guide future research in this area. All researchers will benefit from carefully studying this book, no matter what their specific views are on the subject matter. Looking ahead, an added benefit of the book is that it lays out the foundations for a systematic and unified statistical analysis of more advanced topics in causal inference such as multivalued, dynamic, and adaptive treatment regimes; differencein-differences methods; regression discontinuity designs; linear and nonlinear panel or longitudinal data; interference, externalities, spillovers and peer effects; and external validity, meta analysis and counterfactual policy evaluations, just to mention a few examples. One can only hope that the authors can be persuaded to write the second volume.