Semiparametric Allocation of Subjects to Cohort Strata

Abstract

Background: We illustrate a method for stratum assignment in small cohort studies that avoids modeling assumptions. Methods: Off-the-shelf software (rgenoud) made stratum assignments to minimize a loss function built on within-stratum and population-adjusted Euclidean distances. Results: In 100 trials using simulated data of 300 records with a binary treatment and four dissimilar covariate treatment predictors, minimizing a loss based on Euclidean distance reduced covariate imbalance by a median of 99%. Stratification by propensity score and weighting records by the inverse of their probability of treatment reduced imbalance by 76%–89% and 83%–94%, respectively. Loss minimization applied to a cohort of 361 children undergoing immunotherapy achieved nearly complete elimination of covariate differences for important treatment predictors. Conclusion: With the availability of semiparametric stratum-assignment algorithms, analysts can tailor loss functions to meet design goals. Here, a loss function that emphasized covariate balance performed well under limited testing.