Primary Submission Category: Design of Experiments
Dynamic Adaptive Rerandomization for Efficient Sequential Trials Under Budget Constraints
Authors: Kateryna Husar,
Presenting Author: Kateryna Husar*
We introduce the Dynamic Adaptive Rerandomization (DAR) framework to improve the statistical efficiency of the Average Treatment Effect (ATE) estimator in resource-constrained sequential randomized controlled trials (RCTs). DAR combines two linked Bayesian procedures for concurrent adaptive learning and robust estimation. In each batch, a Thompson Sampling-based Multi-Armed Bandit (MAB) policy selects covariates with the highest posterior predictive importance while respecting the measurement budget. Measured covariates are then used in rerandomization to minimize the variance of the batch-specific ATE estimator ($hattau_k$). The overall Average Treatment Effect ($tau$) is estimated sequentially using a Bayesian weighted averaging approach. The $hattau_k$’s are combined using inverse-variance weighting, where uncertainty in the variance estimates ($sigma_k^2$) is explicitly modeled. This structure naturally assigns greater weight to batches rerandomized on more predictive covariates, as these estimators exhibit smaller sampling variances. Covariate predictive utility is subsequently reassessed to update the importance prior, closing the adaptive loop. This dual-path mechanism allows for identification of influential covariates, optimizes resource use, and yields a precise, fully quantified posterior for the ATE.