Identification of Causal Mechanisms from Randomized Experiments: A Framework for Endogenous Mediation Analysis

Practice- and Policy-Oriented Abstract Experimental research often focuses on the overall treatment effect and the heterogeneity therein. Whereas this type of research allows us to understand the strength and direction of the treatment effect under different conditions, it does not directly speak to the generative mechanisms, namely, why and how the effect arises. A standard procedure to identify the mechanisms underlying a treatment effect is mediation analysis, but extant mediation analysis frameworks either have no causal interpretation or require the mediators to be unconfounded. Because mediators typically cannot be preassigned beforehand, their endogeneity remains a serious concern even in randomized experiments. This paper presents a flexible endogenous mediation analysis framework that still has causal interpretation when the mediator is endogenous. We discuss the identification conditions for different types of endogenous mediators, including unobserved or partially observed ones, under this framework. We show that endogenous mediation models can be parametrically identified without an instrumental variable when the generating process of the mediator is nonlinear. We further examine how the identification strengths of these models vary with a series of factors. Finally, we provide guidelines on when and how to use endogenous mediation analysis. We offer an R package that implements the proposed models.

• 저널Information Systems Research · 34(1) · 67–84
• 토픽Advanced Causal Inference Techniques · Statistics and Probability
• DOI10.1287/isre.2022.1113
• 저자Jing Peng