Comparison of response adaptive randomization features in multiarm clinical trials with control

We investigate multiple features of response adaptive randomization (RAR) in the context of a multiple arm randomized trial with control, where the primary goal is the identification of the best arm for use in a broader patient population. We maintain constant control allocation and vary the length of time until RAR is started, interim frequency, the underlying quantity used to calculate the randomization probabilities, and a threshold resulting in temporary arm dropping. We evaluate the designs on five metrics measuring benefit to the internal trial population, the future external population, and statistical estimation. Our results indicate these features have minimal interaction within the space explored, with preference for earlier activation of RAR, more frequent interim analyses, randomizing in proportion to the probability each arm is the best, and aggressive thresholding for temporarily dropping arms. The results illustrate useful principles for maximizing the benefit of RAR in practice.