Rerandomization to Improve Covariate Balance by Minimizing the MSE of a Treatment Effect Estimator


We present a new approach to treatment assignment in (field) experiments for the case of one or multiple treatment groups. This approach, which we call the minimizing Mean Squared Error (MSE) approach, uses sample characteristics to obtain balanced treatment groups. Compared to other methods, the min MSE procedure is attrition tolerant, offers greater flexibility, is very fast, it can be conveniently implemented and balances different moments of the distribution of the treatment groups. Additionally, it has a clear theoretical foundation, works without parameter being specified by the researcher and allows multiple treatments. The information used for treatment assignment can be multivariate, discrete or continuous and may consist of any number of variables. In this paper, we derive the underlying theoretical selection criteria, which we then apply to various scenarios and datasets. Our proposed method performs better than, or comparably to, competing approaches, such as matching, in most of the commonly used measures of balance. We provide implementations in Stata, R and Python.

Older version (A New Approach to Treatment Assignment for One and Multiple Treatment Groups), was published in 2017 as: Courant Research Centre: Poverty, Equity and Growth - Discussion Papers 228,