A simulation study of outcome adaptive randomization in multi-arm clinical trials

Abstract

Randomizing patients among treatments with equal probabilities in clinical trials is the established method to obtain unbiased comparisons. In recent years, motivated by ethical considerations, many authors have proposed outcome adaptive randomization, wherein the randomization probabilities are unbalanced, based on interim data, to favor treatment arms having more favorable outcomes. While there has been substantial controversy regarding the merits and flaws of adaptive versus equal randomization, there has not yet been a systematic simulation study in the multi-arm setting. A simulation study was conducted to evaluate four different Bayesian adaptive randomization methods and compare them to equal randomization in five-arm clinical trials. All adaptive randomization methods included an initial burn-in with equal randomization and some combination of other modifications to avoid extreme randomization probabilities. Trials either with or without a control arm were evaluated, using designs that may terminate arms early for futility and select one or more experimental treatments at the end. The designs were evaluated under a range of scenarios and sample sizes. For trials with a control arm and maximum same size 250 or 500, several commonly used adaptive randomization methods have very low probabilities of correctly selecting a truly superior treatment. Of those studied, the only adaptive randomization method with desirable properties has a burn-in with equal randomization and thereafter randomization probabilities restricted to the interval 0.10–0.90. Compared to equal randomization, this method has a favorable sample size imbalance but lower probability of correctly selecting a superior treatment. In multi-arm trials, compared to equal randomization, several commonly used adaptive randomization methods give much lower probabilities of selecting superior treatments. Aside from randomization method, conducting a multi-arm trial without a control arm may lead to very low probabilities of selecting any superior treatments if differences between the treatment success probabilities are small.