Causal Inference
Graphical insights into exposure-induced mediator–outcome confounders in causal mediation analysis Etsuji Suzuki* Etsuji Suzuki Suzuki Suzuki Okayama University
As has been well appreciated in the causal mediation literature, the total effect can be decomposed into natural direct and indirect effects in the counterfactual framework. If certain assumptions about confounding are met, we can use the so-called mediation formula to identify these effects under the nonparametric structural equation models with independent errors (NPSEM-IE). However, natural direct and indirect effects have been criticized because these rely on a specific cross-world quantity, and the so-called cross-world independence assumption is not empirically verifiable. In this context, previous studies have highlighted challenges posed by exposure-induced mediator–outcome confounders for identifying natural direct and indirect effects. In this presentation, we discuss the implications of exposure-induced mediator–outcome confounders for causal mediation analysis, providing visualized explanations using causal directed acyclic graphs (DAGs). To achieve this, we incorporate potential outcomes and error terms into the causal DAGs under NPSEM-IE. Then, we visually demonstrate that the absence of the exposure-induced mediator–outcome confounder is a sufficient but not a necessary condition for identification of the natural direct and indirect effects. Indeed, if there is an exposure-induced mediator–outcome confounder, the natural direct and indirect effects are not generally identified irrespective of whether data are available on it, except under strong assumptions, such as no interaction between the exposure and mediator at the individual level. Furthermore, we identify additional settings in which the natural direct and indirect effects can be identified even in the presence of an exposure-induced mediator–outcome confounder. Our visual illustration provides clear insight into the implications of exposure-induced mediator–outcome confounders for causal mediation analyses.