G-Computation Demonstration in Causal Mediation Analysis with Multiple Mediators

Abstract

Parametric and semiparametric estimation methods are available for the case of multiple mediators in causal mediation analysis. However, in the presence of multiple exposure-mediators and mediator-mediator interactions, it is challenging to study the direct and indirect effects of multiple mediators. In this paper, we use G-computation to derive a decomposition of the overall effect that unifies the mediating and interacting effects when multiple mediators are present, including natural direct and indirect effects, direct effects of standard and random controls, and reference and mediating interaction effects. The final regression of potential outcomes using the exposure intervention variables was used to calculate the point estimates and obtain their confidence intervals (CI). The methods presented here accommodate exposure-mediator interactions and, to some extent, mediator-mediator interactions. These methods are applicable to binary, continuous mediators and binary, continuous, or count results in the case where the exposure is discrete. This framework can also be extended to complex multivariate and longitudinal mediator settings.