Randomization is an essential component of experiments because it balances covariates between the treatment groups. However, randomization only balances covariates in expectation, and in any one particular allocation, covariates may be unbalanced simply by chance. We propose forcing balance on any covariates available at the time of randomization by rerandomizing if covariate balance is not acceptable, according to pre-specified criteria. Here we discuss how to conduct rerandomization in a principled way to improve covariate balance and power, while still maintaining desired frequentist properties such as unbiasedness and Type I error rates.

Register