Model-Averaged Propensity Scores for Estimating the Average Treatment Effect under Treatment Effect Heterogeneity

Sammanfattning

Estimating the Average Treatment Effect using Propensity Score Weighting is highly sensitive to the choice of propensity score model, especially when there is heterogeneity in treatment effects and limited overlap in covariates. In practice, the true treatment assignment mechanism is rarely known, and ATE estimates based on a single propensity score model can vary substantially across models. We propose a propensity score method that averages across multiple candidate models and selects weights to directly achieve covariate balance. The resulting model-averaged propensity score is used to estimate the ATE within the inverse probability weighting framework. Simulation studies show that, although the proposed method is not uniformly superior to single-model methods, it produces more stable ATE estimates across various propensity score models.