Bayesian mediation analysis for zero-inflated mediator

Mediation analysis is a powerful tool for exploring the causal relationships between exposures and outcomes that are mediated by intermediate variables. In this paper, we propose a flexible Bayesian mediation analysis framework to accommodate zero-inflated mediators, compatible with a wide range of outcome distributions. This novel technique employs Bayesian models for both the mediator and the outcome, utilizing Markov Chain Monte Carlo algorithms for parameter estimation. While addressing the challenges posed by an excess of zeros, we further decompose the mediation effects into components influenced by either the probability of zero or the mean of the non-zero distribution in the mediator. An associated R package mediationBayes (https://github.com/jhcuibst/mediationBayes.git) has been developed to facilitate the application of this framework. Through comprehensive simulation studies, we demonstrate that our method outperforms alternatives in terms of the accuracy of point estimates, coverage probabilities, and the precision of the mediation effects decomposition. We further illustrate the practical applicability of our method by conducting an analysis on the REasons for Geographic And Racial Differences in Stroke Study to investigate the mediating influence of smoking pack-years on the association between educational levels and incident hypertension, where mediation effects are quantified on a risk ratio scale for binary outcomes.