Journal of Scientific Research and Reports

Propensity score matching (PSM) is widely used to improve comparability between treatment groups in observational studies, but its performance depends on treatment-model specification, covariate overlap, and post-matching balance. This study compares Random Forest (RF) and Bayesian Additive Regression Trees (BART) as alternative propensity score estimators using the public Lindner percutaneous coronary intervention dataset (n = 996; 698 abciximab-treated and 298 control observations). Propensity scores were estimated with RF and BART and used for nearest-neighbour matching. Descriptive predictive diagnostics included the area under the receiver operating characteristic curve (AUC), log-loss, and Brier score; balance was assessed using standardised mean differences (SMDs). BART yielded an AUC of 0.7250, log-loss of 0.5422, and Brier score of 0.1816, compared with 0.5813, 0.7875, and 0.2265 for RF. Matching retained 280 treated-control pairs for RF and 269 pairs for BART. Across measured baseline covariates, the maximum absolute post-match SMD was 0.2782 for RF and 0.0935 for BART; the propensity-score distance diagnostic was excluded from the baseline-covariate maximum. The reported matched-sample outcome contrasts were 0.2900 for RF and 0.3900 for BART. Because the RF interval was obtained by ordinary non-parametric bootstrap after nearest-neighbour matching, it is treated as exploratory rather than as formal matching-based inference. The BART interval summarises variation across posterior propensity-score design draws and is not interpreted as a complete posterior interval for the causal effect. The results show that flexible treatment models can differ materially in predictive fit and covariate balance, while also demonstrating the need to separate predictive performance, balance, overlap, and uncertainty when interpreting matched observational analyses.