Bayesian Mediation Analysis With Bayesian Additive Regression Trees

Abstract

Understanding how environmental exposures influence health outcomes through biological pathways is critical for environmental risk assessment. We propose a Bayesian mediation analysis framework that integrates Bayesian Additive Regression Trees (BART) to model complex, nonlinear, and high-dimensional relationships among exposures, mediators, and outcomes. The method accommodates multiple mediators, interaction effects, and uncertainty quantification through posterior inference. We introduce two estimation strategies for direct and indirect effects and develop an algorithm for computing the Deviance Information Criterion (DIC) for model evaluation. An open-source R package, bmabart, implements the approach with visualization tools. Simulation studies demonstrate robustness in identifying significant mediators under challenging correlation structures. We apply the method to Louisiana Triple Negative Breast Cancer (TNBC) data, investigating microRNAs that mediate the association between air pollution burden, measured by the Environmental Justice Index, and cancer stage at diagnosis. Results highlight key biomarkers linking environmental exposures to disease progression, offering insights into mechanisms underlying health disparities. This framework provides a flexible and reproducible tool for environmental health research where complex mediation structures are common.