B0730
Title: A Bayesian nonparametric approach for principal causal effects
Authors: Chanmin Kim - SungKyunKwan University (Korea, South) [presenting]
Abstract: In estimating the causal effect, principal stratification analysis is a method for interpreting the effect of treatment on the outcome based on the relationship between treatment and post-treatment (intermediate) variables. In general, modeling of an intermediate variable is required when the intermediate variable is continuous, but existing parametric modeling methods are difficult to fully capture the complex relationship between variables. Furthermore, estimating the outcome model and intermediate model separately makes it difficult to account for the uncertainty introduced by each model estimation in the final causal effect estimation. Using the Bayesian additive regression trees model, we propose a fully Bayesian method. All intermediate, outcome, and propensity score models are flexibly estimated using Bayesian nonparametric models, unlike other flexible methods in the literature. Also, the proposed method is very useful in both a specific confounding situation (referred to as targeted selection) and a broad confounding situation. A simulation study is used to demonstrate this. With the proposed method, we examine the impact of the sulfate abatement device (scrubber) installed in US coal-fired power plants on surrounding PM2.5 concentrations from various perspectives, based on the relationship between the scrubber and SO2 emissions.