DR-Learners for Generalizing CATEs from a Randomized Controlled Trial
Abstract
Randomized clinical trials are essential for evaluating interventions, yet average treatment effects can obscure treatment effect heterogeneity, making conditional average treatment effects (CATEs) a natural target. However, restrictive eligibility criteria and differential trial participation can limit the generalizability of these effects to broader target populations. Building on existing DR-Learners, we develop the gDR-Learner for generalizing RCT CATEs to a target population, using cross-fitted machine-learning nuisance estimation. We derive excess-risk bounds in which nuisance errors enter only as products, showing that the estimator achieves oracle-rate convergence despite slower individual nuisance estimation. The variables needed to generalize are rarely measured across the entire target population. We therefore extend the methods to two settings with incomplete covariate information: when certain covariates are measured only in the trial (e.g., biomarkers), and covariates driving trial participation measured only in a subsample of the target population, such as a representative cohort or registry. We construct weighted conformal prediction intervals for individual treatment effects, accounting for covariate shift between the trial and target population. In simulations, the gDR-Learner achieved the lowest estimation error when all covariates were observed relative to standard T- and DR-Learners and an inverse-probability-weighted T-Learner, and also outperformed the latter under incomplete covariate observation. We illustrate the methods by reanalyzing the Coronary Artery Surgery Study, examining treatment effect heterogeneity by prior myocardial infarction and baseline ejection fraction.
Related
- Model selection in linear mixed-effects analysis of craving outcomes in a randomized controlled trial of medications for opioid use disorders
- A Bayesian hierarchical push-forward method for cluster Randomized Controlled Trial power calculations in complex humanitarian settings
- Evaluating an LLM-assisted workflow for clinical documentation: A pilot randomized controlled trial on time and quality
- Global Sensitivity Analysis for Studies Extending Inferences From a Randomized Trial to a Target Population
- Leveraging Large Language Models to Improve Precision in Randomized Controlled Trials
BibTeX
@misc{rodriguez2026drlearners,
title = {{DR-Learners for Generalizing CATEs from a Randomized Controlled Trial}},
author = {Vanessa Rodriguez and Brieuc Lehmann and Karla Diaz-Ordaz},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.25405},
eprint = {2610.25405},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.25405},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - DR-Learners for Generalizing CATEs from a Randomized Controlled Trial
AU - Vanessa Rodriguez
AU - Brieuc Lehmann
AU - Karla Diaz-Ordaz
PY - 2026
DA - 2026/10/18
PB - arXiv
DO - 10.48550/arxiv.2610.25405
UR - https://arxiv.org/abs/2610.25405
AB - Randomized clinical trials are essential for evaluating interventions, yet average treatment effects can obscure treatment effect heterogeneity, making conditional average treatment effects (CATEs) a natural target. However, restrictive eligibility criteria and differential trial participation can limit the generalizability of these effects to broader target populations. Building on existing DR-Learners, we develop the gDR-Learner for generalizing RCT CATEs to a target population, using cross-fitted machine-learning nuisance estimation. We derive excess-risk bounds in which nuisance errors enter only as products, showing that the estimator achieves oracle-rate convergence despite slower individual nuisance estimation. The variables needed to generalize are rarely measured across the entire target population. We therefore extend the methods to two settings with incomplete covariate information: when certain covariates are measured only in the trial (e.g., biomarkers), and covariates driving trial participation measured only in a subsample of the target population, such as a representative cohort or registry. We construct weighted conformal prediction intervals for individual treatment effects, accounting for covariate shift between the trial and target population. In simulations, the gDR-Learner achieved the lowest estimation error when all covariates were observed relative to standard T- and DR-Learners and an inverse-probability-weighted T-Learner, and also outperformed the latter under incomplete covariate observation. We illustrate the methods by reanalyzing the Coronary Artery Surgery Study, examining treatment effect heterogeneity by prior myocardial infarction and baseline ejection fraction.
ER -