Randomized clinical trials are the gold standard for evaluating interventions, but their average treatment effects can hide important heterogeneity. Estimating conditional average treatment effects (CATEs) is a natural step, yet restrictive eligibility criteria and differential participation often limit how well these effects generalize to real-world target populations.
The authors build on existing DR-Learners to propose the gDR-Learner, which generalizes RCT CATEs to a target population using cross-fitted machine-learning nuisance estimation. They derive excess-risk bounds showing that nuisance errors enter only as products, allowing the estimator to achieve oracle-rate convergence even when individual nuisance components converge more slowly. The method also extends to two incomplete-covariate settings: when some covariates are measured only in the trial (such as biomarkers), and when participation-driving covariates are measured only in a subsample of the target population.
In simulations, the gDR-Learner achieved the lowest estimation error when all covariates were observed, outperforming standard T- and DR-Learners and an inverse-probability-weighted T-Learner; it also beat the latter under incomplete covariate observation. The authors illustrate the approach by reanalyzing the Coronary Artery Surgery Study, examining treatment effect heterogeneity by prior myocardial infarction and baseline ejection fraction. They further construct weighted conformal prediction intervals for individual treatment effects, accounting for covariate shift between the trial and target population.