A doubly robust learner for regression and inference with right-censored outcomes

This paper introduces a general framework for doubly robust nonparametric regression with right-censored outcomes, adapting the strategy of the 'DR-learner' for heterogeneous treatment effects to the censored data setting. Our method generalizes censoring unbiased transformations to generate pseudo-outcomes, which serve as inputs for a second-stage nonparametric regression with sample-splitting or cross-fitting. We derive a novel representation of the conditional bias of these pseudo-outcomes; this representation is central to establishing the estimator's asymptotic properties, including rate double robustness and oracle efficiency when the second-stage regression is a linear smoother. Simulation studies demonstrate the method's favorable finite-sample performance, and we showcase its practical utility by applying it to estimate a causal effect within a regression discontinuity design with right-censored outcomes.