10th U6+ Consortium of African Universities International Conference

Propensity Score Matching (PSM) is widely used to estimate causal treatment effects from observational data by reducing confounding bias arising from non-random treatment assignment. Conventional propensity score estimation based on logistic regression may, however, inadequately capture complex nonlinear relationships and interactions among covariates, potentially leading to suboptimal covariate balance and biased treatment effect estimates. This study developed a Bayesian Propensity Score Matching framework that integrates Random Forest (RF) and Bayesian Additive Regression Trees (BART) for enhanced propensity score estima-tion and causal effect inference. The proposed framework was evaluated using the Lindner cardiovascular dataset. The predictive performance of the machine learning models was assessed using the Area Under the Receiver Operating Characteristic Curve (AUC), Log-loss, and Brier Score, while covariate balance was evaluated using standardized mean differences before and after matching. Treatment effects were estimated using the Average Treatment Effect on the Treated (ATT), with bootstrap confidence intervals for the RF model and posterior credible intervals for the Bayesian BART model. The results demonstrated that Bayesian BART outperformed Random Forest in propensity score estimation, achieving a higher AUC (0.7250 versus 0.5813), lower Log-loss (0.5422 versus 0.7875), and lower Brier Score (0.1816 versus 0.2265), indicating superior discrimination and calibration of treatment assignment probabilities. Bayesian BART also produced better covariate balance after matching, reducing the maximum absolute standardized mean difference to 0.1991 compared with 0.2782 for the Random Forest model. Furthermore, the Bayesian framework yielded a more precise treatment effect estimate, with an ATT of 0.3900 and a posterior standard deviation of 0.0734, compared with an ATT of 0.2900 and a bootstrap standard deviation of 0.1786 for the Random Forest model. The posterior 95% credible interval for Bayesian BART (0.2486, 0.5367) indicated a statistically meaningful positive treatment effect. These findings demonstrate that integrating Bayesian machine learning into propensity score matching improves propensity score estimation, enhances covariate balance, provides principled uncertainty quantification, and strengthens causal effect estimation in observational studies.