Response-Adaptive Rerandomization

Abstract

Randomized experiments are the gold standard for estimating treatment effects, as randomization potentially removes the differences of covariates across treatment groups. However, covariateimbalanced allocation often occurs in a single-shot assignment and thus dampens the estimation precision. To remedy this issue, Morgan and Rubin (2012) proposed rerandomization (RR) for twoarm experiments, which repeatedly randomizes enrolled units and terminates the process when the Mahalanobis distance of covariates across groups meets a predefined threshold. This scheme is shown to well control the covariate balance and thus improve the statistical efficiency.Since the publication of the seminal paper by Morgan and Rubin (2012), many works have been conducted to generalize RR practically or theoretically. For example, Morgan and Rubin (2015)

PCA Rerandomization

“…Wang, Wang & Liu (2021) studied the statistical properties of stratification with rerandomization. Zhang & Yin (2021) incorporated the response information into rerandomization for ethical concerns in clinical trials. Yang, Qu & Li (2021) proposed rerandomization for survey experiments, using asymptotic theories.…”