Causal Mediation and Effect Modification: A Unified Framework

• In Event: Causal Inference: Mediation, Experiments, and Sensitivity

Abstract

Causal mediation and moderation are often contrasted as two distinct concepts for describing the role of a third variable in a cause-effect relationship. Although both concepts are broadly considered to be related to the notion of mechanisms, they are typically analyzed via different causal models -- potential outcomes and sufficient component causes -- in the causal inference literature. In this paper, we provide a unified perspective where both mediation and moderation are defined as evidence of the same underlying causal mechanism. In doing so, we introduce the concept of "switch," or the effect modifier that has no effect on any mediator of the treatment effect except a particular mediator that represents the mechanism of interest. We show that the existence of such a switch is sufficient for the nonparametric identification of the average natural indirect effect. Our proposed framework also resolves an important limitation of Pearl's causal direct acyclic graph (DAG) framework in representing interactive relationships.

Authors

• Teppei Yamamoto, Massachusetts Institute of Technology

• David Yeager, Stanford University