Propensity Score Estimation for Causal Inference in Observational Healthcare Studies

The application of propensity score (PS) methods is fundamental for causal inference in observational studies, particularly within the healthcare sector. This study, devoted to observational healthcare studies, compares traditional approaches such as Logistic Regression (LR), Boosting and Bayesian Additive Regression Trees (BART) with the Balanced Super Learner (BSL), an ensemble method that integrates both LR, Boosting and BART. We examine datasets from UK Biobank and the Atlantic Causal Inference Conference (ACIC) 2016. While BSL was demonstrated improved performance, it was hindered by Boosting, which had the worse performance on ACIC 2016 dataset. LR on the other side perform comparatively well on most datasets, except on the extreme imbalance. These findings highlight the potential robustness of LR and highlights the critical role of model selection in PS-based analysis for healthcare-related causal inference tasks.