Cross-Validation Model Averaging Under Covariate-Adaptive Randomization

Le résumé fourni par la source

Covariate-adaptive randomization (CAR) leverages covariates at the design stage to enhance comparability between treatment groups, but typically allows only a limited set of covariates to be incorporated.We therefore study covariate adjustment at the analysis stage to achieve more efficient estimation of the treatment effect under CAR.Existing work on covariate adjustment under CAR typically relies on a single, pre-specified model.Although this complies with regulatory guidance that the adjustment model be specified before the trial, a single-model approach can be vulnerable to misspecification and may be less efficient when the covariate-outcome relationship is uncertain.To address this issue, we propose a model averaging framework for covariate adjustment under CAR that accommodates a collection of pre-specified candidate models.The averaging weights are selected via K-fold cross-validation to minimize the asymptotic variance of the treatment effect estimator.We establish that the resulting estimator is asymptotically normal and achieves asymptotic variance optimality within the candidate model class under mild regularity conditions.Simulation studies demonstrate that the proposed approach yields substantial variance re-Statistica Sinica: Newly accepted Paper duction compared with single-model adjustment.Overall, this paper introduces a theoretically grounded and flexible framework for treatment effect estimation under CAR that complies with regulatory pre-specification requirements while effectively handling uncertainty in covariate-outcome relationships.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.