A simulation study to evaluate the effect of constrained randomization for the design and analysis of stepped wedge cluster-randomized trials

Abstract

In this study, we conducted simulations to evaluate the effect of constrained randomization on testing the treatment effect in terms of type I error and power with data generated from a stepped wedge cluster-randomized design, under the presence of cluster-level covariates. We considered two cases, one with a single binary co- variate and the other with a mixture of continuous and categorical covariates. For case one we used stratified randomization to achieve perfect covariate balance whereas for case two we constrained the randomization space by setting scores based on different balance criteria. For each case we consider eight different scenarios, and apply model-based and permutation-based inference to estimate and test for the treatment effect, both adjusted and unadjusted for covariates in the analysis phase. We found that the type I error is close to the nominal level most of the time except for permutation inference with constrained randomization and unconstrained analysis, where it drops down towards zero. In general, we see that constrained randomization can slightly increase power when covariates are also included at the analysis phase, and such increase is more visible in case one with a single binary covariate than case two with multiple covariates. Overall, although we discovered some advantages of doing constrained randomization in terms of power, such gain is only marginal and its impact in practice is likely to be much smaller than under traditional cluster-randomized design. Therefore, controlling for covariates in the analysis phase is still considered to be a more effective way to attain higher testing power under stepped wedge setting.