Primary Submission Category: Heterogeneous Treatment Effects
The Parachuted Hybrid CATE Estimator with Bootstrap Methods for Inference
Authors: Xianlin Sun, Stephen Man Sing Lee,
Presenting Author: Xianlin Sun*
In our study, we introduce a novel estimator for the Conditional Average Treatment Effect (CATE), termed the parachuted estimator, notable for its double robustness. This means it remains consistent if either the propensity score model or the conditioned expected outcome model is correctly specified, and uniquely, it still performs reliably even if both models are misspecified, albeit at a slower convergence rate similar to non-parametric estimators. Our method combines parametric techniques from Augmented Direct Learning with non-parametric kernel estimation strategies, achieving optimal convergence rates and maintaining consistency.
A pivotal achievement of this research is the derivation of the asymptotic distribution for this hybrid estimator. This distribution is characterized by having its mean aligned with the estimated target—the true CATE—and its variance describable through a closed-form expression. Moreover, our investigation extends into the statistical inference concerning our parachuted estimator for CATE via established bootstrap techniques. Through rigorous theoretical analysis grounded in the work of Chatterjee and Bose (2005), we provide substantive proof affirming that these bootstrap methods yield consistent estimations within our specified framework, subject to certain regular constraints.