A More Flexible Bayesian Nonparametric Approach to Causal Mediation

• Activity Number: 58 - Advanced Bayesian Topics (Part 1)
• Type: Contributed
• Date/Time: Sunday, August 8, 2021 : 3:30 PM to 5:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Woojung Bae* and Michael J. Daniels
• Companies: University of Florida and University of Florida
• Keywords: Bayesian non-parametric modeling; Causal inference; Cluster; Enriched Dirichlet process mixture model

Abstract

We propose a new Bayesian non-parametric (BNP) method to estimate the causal effects of mediation. We specify an enriched Dirichlet process (EDP) to model the joint distribution of the observed data (outcome, mediation, treatment, and confounders). The proposed BNP model allows more confounders-clusters than clusters for outcome and mediator. For identifiability, we consider the standard sequential ignorability presented in Imai et al. (2010). The observed data model along with the causal assumptions allows us to estimate and identify causal effects of mediation, the natural direct, and indirect effects. Simulation studies in a variety of settings are presented to examine the performance of this approach. We used this approach to evaluate the mediation effect in a physical activity promotion trial.