Robust nonparametric estimation of average treatment effects: A propensity score-based varying coefficient approach

We present a novel nonparametric approach for estimating average treatment effects (ATEs), addressing a fundamental challenge in causal inference research, both in theory and empirical studies. Our method offers an effective solution to mitigate the instability problem caused by propensity scores close to zero or one, which are commonly encountered in (augmented) inverse probability weighting approaches. Notably, our method is straightforward to implement and does not depend on outcome model specification. We introduce an estimator for ATE and establish its consistency and asymptotic normality through rigorous analysis. To demonstrate the robustness of our method against extreme propensity scores, we conduct an extensive simulation study. Additionally, we apply our proposed methods to estimate the impact of social activity disengagement on cognitive ability using a nationally representative cohort study. Furthermore, we extend our proposed method to estimate the ATE on the treated population.