Bayesian nonparametric adjustment of confounding.

Area of Science:

• Causal Inference
• Bayesian Statistics
• Environmental Epidemiology

Background:

• Observational studies face challenges in selecting relevant covariates for causal effect estimation.
• Ignorable treatment assignment is a key assumption requiring careful confounder adjustment.
• High-dimensional covariate sets complicate confounder identification and analysis.

Purpose of the Study:

• To propose a Bayesian nonparametric method for simultaneous confounder selection and causal effect estimation.
• To account for uncertainty in the nature and role of confounding variables.
• To improve the accuracy and efficiency of causal inference in complex observational data.

Main Methods:

• Utilizing multiple Bayesian additive regression trees (BART) models.
• Employing a common prior distribution to assign posterior selection probabilities to covariates.
• Prioritizing covariates based on their association with both exposure and outcome.

Main Results:

• Simulations show the proposed method outperforms existing approaches in various scenarios.
• The method was applied to assess the causal effect of SO2 emissions from coal-fired power plants on air pollution.
• Results demonstrated more efficient and consistent causal estimates compared to alternatives.

Conclusions:

• The Bayesian nonparametric approach effectively addresses confounder selection and estimation challenges.
• The method provides robust causal estimates, accounting for confounding uncertainty.
• This approach strengthens evidence for the causal link between SO2 emissions and ambient particulate pollution.