Generalizing Conditional Average Treatment Effects From Nested Randomized Trials to Trial-Eligible Individuals

Randomized controlled trials often enroll participants whose characteristics differ from those of a target population. When the distribution of effect modifiers differs between trial participants and the target population, treatment effects estimated within the trial may not be directly applicable to all trial-eligible individuals. Although substantial progress has been made in generalizing average treatment effects from randomized trials to target populations, clinical and policy decisions are often informed by conditional average treatment effects (CATEs) that characterize heterogeneity across selected effect modifiers in the target population. In this work, we provide a unified approach for estimating CATEs in a target population of trial-eligible individuals under a nested trial design. Our approach first estimates nuisance functions using data-adaptive methods and constructs pseudo-outcomes from conditional influence functions. These pseudo-outcomes are subsequently regressed on pre-specified effect modifiers using local linear (kernel) regression to obtain a smoothed estimate of the CATE function. We establish asymptotic properties under sample splitting and cross-fitting, and provide conditions under which first-stage nuisance estimation has negligible impact on the final estimator. We assess the finite-sample performance of the proposed estimator via simulation studies, and illustrate its practical use through an application to the Coronary Artery Surgery Study (CASS).