Machine learning versus logistic regression for propensity score estimation: a trial emulation benchmarked against the PARADIGM-HF randomized trial.

Machine learning (ML) algorithms are increasingly used to estimate propensity score with expectation of improving causal inference. However, the validity of data-driven ML-based approaches for confounder selection and adjustment remains unclear. In this study, we emulated the device-stratified secondary analysis of the PARADIGM-HF trial among U.S. veterans with heart failure and implanted cardiac devices from 2016 to 2020. We benchmarked observational estimates from three propensity score approaches against the trial results. (1) logistic regression with pre-specified confounders (2), generalized boosted models (GBM) using the same pre-specified confounders, and (3) GBM with expanded covariates and automated feature selection. Logistic regression-based propensity score approach yielded estimates closest to the trial (HR = 0.93, 95% CI 0.61-1.42; 23-month RR = 0.86, 95% CI 0.57-1.24 vs. trial HR = 0.81, 95% CI 0.61-1.06). Despite better predictive performance, GBM with pre-specified confounders showed no improvement over the logistic regression approach (HR = 0.97, 95% CI 0.68-1.37; RR = 0.96, 95% CI 0.89-1.98). Moreover, GBM with expanded covariates and data-driven automated feature selection substantially increased bias (HR = 0.61, 95% CI 0.30-1.23; RR = 0.69, 95% CI 0.36-1.04). Our findings suggest that ML-based propensity score methods do not inherently improve causal estimation possibly due to residual confounding from omitted or partially adjusted variables and may introduce overadjustment bias when combined with automated feature selection. These results underscore the importance of careful confounder specification and causal reasoning over algorithmic complexity in causal inference.