2019/05/08 by Adam Kapelner, Abba Μ. Krieger, Kapelner, Adam +6
Mathematics · #Advanced Causal Inference Techniques #FOS: Computer and information sciences #Methodology (stat.ME) #Statistical Methods and Inference #Statistical Methods in Clinical Trials
paper · pdf · doi:10.48550/arxiv.1905.03337
openalex publication_date 2019/05/08 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
We present an optimized rerandomization design procedure for a non-sequential\ntreatment-control experiment. Randomized experiments are the gold standard for\nfinding causal effects in nature. But sometimes random assignments result in\nunequal partitions of the treatment and control group visibly seen as imbalance\nin observed covariates. There can additionally be imbalance on unobserved\ncovariates. Imbalance in either observed or unobserved covariates increases\ntreatment effect estimator error inflating the width of confidence regions and\nreducing experimental power. "Rerandomization" is a strategy that omits poor\nimbalance assignments by limiting imbalance in the observed covariates to a\nprespecified threshold. However, limiting this threshold too much can increase\nthe risk of contracting error from unobserved covariates. We introduce a\ncriterion that combines observed imbalance while factoring in the risk of\ninadvertently imbalancing unobserved covariates. We then use this criterion to\nlocate the optimal rerandomization threshold based on the practitioner's level\nof desired insurance against high estimator error. We demonstrate the gains of\nour designs in simulation and in a dataset from a large randomized experiment\nin education. We provide an open source R package available on CRAN named\nOptimalRerandExpDesigns which generates designs according to our algorithm.\n