Statistical Analysis of Causal Mechanisms: Identification, Inference, and Sensitivity Analysis

该奖项是根据2009年美国复苏和再投资法案（公法111-5）资助的。因果推理是大多数社会科学研究的中心目标。近年来，实证研究者越来越依赖实验来提高因果结论的有效性。与此同时，关于因果推理的方法论文献蓬勃发展。尽管有这两方面的发展，但实验面临着一个根本性的弱点，因为它们只提供了因果关系的黑箱视图。治疗的随机化使研究人员能够获得治疗效果的有效估计，从而确定治疗是否会对结果产生因果影响。然而，随机实验并没有提供多少关于产生这种因果效应的潜在因果机制的信息。它们没有回答治疗如何以及为什么会对结果产生因果影响的重要问题。这是一个重要的限制，因为因果机制的识别通常需要测试对因果效应提供不同解释的竞争理论的有效性。拟议研究的目标是通过开发一套新的统计方法来分析因果机制来克服这一限制。具体地说，研究者专注于因果中介分析，其中因果机制是通过估计直接和间接影响来检验的。在这个框架中，通过检查位于治疗和结果变量之间的因果路径中的中间变量的作用，研究了其他因果解释。提出的研究涉及因果中介分析的三个重要方面。首先，识别分析建立了确定直接和间接影响所需的一套最低限度的假设，并澄清了有关特定因果机制的数据提供信息的确切程度。其次，参数估计器和非参数估计器的发展将使从观测数据推断成为可能，并允许研究人员计算不确定性估计和点估计。第三，敏感性分析对因果中介分析至关重要，这需要一个强有力的识别假设。提出的敏感性分析方法将允许应用研究人员检查他们的经验发现对可能违反关键识别假设的稳健性。

This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).Causal inference is the central goal of most social science research. In recent years, empirical researchers are increasingly relying on experiments in order to improve the validity of causal conclusions. At the same time, the methodological literature on causal inference has flourished. Despite these two developments, experiments confront a fundamental weakness since they only provide a black-box view of causality. The randomization of treatment makes it possible for researchers to obtain a valid estimate of treatment effect and thereby determine whether the treatment causally affects the outcome. However, randomized experiments do not provide much information about the underlying causal mechanisms that have brought about such causal effects. They do not answer the important questions of how and why the treatment causally affects the outcome. This is an important limitation because the identification of causal mechanisms is often required to test the validity of competing theories that offer different explanations about causal effects.The goal of the proposed research is to overcome this limitation by developing a set of new statistical methods for the analysis of causal mechanisms. Specifically, the investigator focuses on causal mediation analysis where causal mechanisms are examined by estimating direct and indirect effects. In this framework, alternative causal explanations are investigated by examining the roles of intermediate variables that lie in the causal path between the treatment and outcome variables. The proposed research addresses three important aspects of causal mediation analysis. First, identification analysis establishes a minimum set of assumptions that are required to ascertain direct and indirect effects and clarifies the exact degree to which the data are informative about particular causal mechanisms. Second, the proposed development of parametric and nonparametric estimators will make inference from the observed data possible and allow researchers to compute uncertainty estimates as well as point estimates. Third, sensitivity analysis is essential for causal mediation analysis which requires a strong identifying assumption. The proposed methods for sensitivity analysis will allow applied researchers to examine the robustness of their empirical findings to the possible violation of the key identifying assumption.