A Bayesian nonparametric approach to causal inference on quantiles.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Causal inference is essential for understanding treatment effects and observational data.
• Traditional methods often struggle with high-dimensional confounders and restrictive assumptions.
• Bayesian nonparametric (BNP) methods offer flexibility but require specialized approaches for causal inference.

Purpose of the Study:

• To develop and evaluate a BNP approach for causal inference on quantiles with many confounders.
• To address limitations of parametric assumptions in causal modeling.
• To provide a flexible and robust framework for analyzing complex health data.

Main Methods:

• Utilized Bayesian additive regression trees (BART) for propensity score modeling.
• Employed a Dirichlet process mixture (DPM) of normals model for potential outcomes.
• Defined causal quantities within the BNP framework to ensure valid inference.

Main Results:

• The proposed BNP approach demonstrated robust performance in simulations.
• It effectively handled a large number of confounders, reducing bias.
• Comparison with existing Bayesian and frequentist methods showed competitive or superior results.

Conclusions:

• The developed BNP method provides a powerful tool for causal inference on quantiles, particularly in high-dimensional settings.
• This approach mitigates bias from parametric assumptions, enhancing the reliability of findings.
• The method is applicable to real-world clinical questions, such as analyzing electronic health records for acute kidney injury.