Nonparametric Regression Methods for Causal Mediation Analysis

Abstract

Causal mediation analysis is a popular approach to quantify the mechanisms through which an exposure operates on an outcome. Using causal mediation analysis, one can decompose the total effect (TE) of an exposure on an outcome into the pathway that operates indirectly through an intermediate (mediator) variable and the pathway that is independent of the mediator variable or that operates directly from the exposure to outcome. Researchers' understanding of the pathways operating through an intermediate variable is crucial for policy recommendations to reduce the potentially harmful impact of the exposure(s) and/or uneven burden of disease. With the growing interest in causal mediation analysis, new methods are needed to estimate direct and indirect effects with complex data. In this dissertation, I propose three novel approaches to estimate mediation effects using Bayesian nonparametric regression models. These methods allow for data with the following complexities: (Chapter 1) the exposure is binary and nonmanipulable and a normality assumption for the mediator variable is not suitable; (Chapter 2) the joint effect of multiple exposures (mixture) is of interest; and (Chapter 3) the exposure of interest is a mixture and the mediator and outcome variables are latent constructs composed of multiple measurements. For each method, I discuss how my approach addresses gaps in the literature and demonstrate how my proposed approach preforms compared to current methods via simulation and data application.