Bayesian nonparametric adjustment of confounding.

The article focuses on a Bayesian nonparametric method using Bayesian additive regression trees (BART) for confounder selection and causal effect estimation in observational studies with high-dimensional covariates. This approach jointly models exposure and outcome with linked priors on variable selection probabilities, prioritizing adjustment variables consistent with confounder selection principles—specifically the "disjunctive cause criterion without instruments"—while allowing for complex nonlinear relationships and accounting for uncertainty in confounding. Simulation studies demonstrate that the proposed method, particularly the single-model variant, outperforms existing methods in bias, mean squared error, and coverage across various scenarios. The method is applied to estimate the causal effect of sulfur dioxide emissions from coal-fired power plants on ambient fine particulate matter (PM2.5) concentrations in the Eastern United States, showing consistent and more efficient causal effect estimates across consecutive years compared to alternative approaches.