Estimating treatment effects in real-world data is challenging due to potential bias arising from non-random treatment assignment and confounding variables. The propensity score (PS) is commonly used to correct for such bias, but its accuracy depends on properly modelling the treatment–covariate relationships. This study proposes estimating PS via Bayesian Networks (BNs), offering a flexible and often superior alternative to logistic regression, especially in complex dependency settings. The BN-based PS is applied to construct two estimators of the Average Treatment Effect (ATE): the Horvitz-Thompson (HT) and Hajek (H) types. When the PS model is correctly specified, both are asymptotically equivalent; under misspecification, the H-type performs better. Extensive simulations confirm the robustness and effectiveness of the proposed approach.
21 Novembre 2025, ore 12.00
Paola Vicard
Roma Tre University
In person: Room 34 (4th floor) building CU002 Scienze Statistiche
Webinar: https://uniroma1.zoom.us/j/83625004899?pwd=bXCtz0mp759PUh2lkqT0BUoVa0Uegg.1
ID riunione: 836 2500 4899
Passcode: 123456