Abstract
Randomization is a key component to scientific inquiry as it facilitates unbiased estimation of treatment effects via balancing of measured and unmeasured prognostic variables across treatment groups. Recent reports have noted that randomization is lacking in animal studies, threatening internal validity. Animal studies often involve rodents (mice or rats) sent in small batches to laboratories or bred on site in litters. Randomizing half of each batch to treatment and half to control (simple randomization) is a viable strategy to implementing randomization in animal studies, however experimenters may be concerned about chance imbalances, given the smaller sample sizes utilized in animal studies, in key prognostic variables, eg, baseline weight. Constrained randomization, wherein key prognostic factors are balanced within each batch, may offer benefits over simple randomization, especially if it were sequential, ie, could take balance of previous batches into account when randomly assigning treatment in current batch. Adjusting for prognostic variables in a statistical model is a way to address imbalances, independent of choice of randomization scheme. In simulations designed to mimic realistic scenarios, all methods of randomization tested led to unbiased treatment effect estimation, with model adjustment reducing standard errors and improving statistical power in all scenarios. Treatment effects in unadjusted and adjusted models were nearly an order of magnitude closer to each other in sequentially constrained randomization compared to simple randomization, yielding more robust findings.