Contribution of machine learning methods to propensity score estimation: A Real World Comparative Analysis of Two Treatment Routes
Objectifs
We used a methodology combining causal inference approaches with machine-learning-based propensity score (PS) matching to assess the impact of the route of administration - Route A versus Route B - on treatment persistence.
Conclusion
Integrating causal inference with propensity score matching based on ML models offers a robust framework for evaluating treatment effects in real-world data. In our analysis, more complex ML models tended to produce less stable propensity scores, probably due to overfitting, a limited prevalence of treatment route A, and reduced overlap of propensity scores, which translated into higher ASAM values and a reduced number of matched pairs in small cohorts. Logistic regression achieved better covariate balance across the populations. These results underline the need for careful model selection and suggest that hybrid strategies could improve matching performance and reliability, although their applicability remains context-dependent and must be confirmed by future larger-scale studies.