Adaptive Randomization Ratios in Multi-arm Clinical Trials

Abstract

Ethics and economics are two of many motivations for streamlining the process of discovering new therapies for the treatment of human disease. Population ethics and economics dictate that any true benefit be demonstrated as quickly as possible to allow effective therapies to be marketed and distributed to patients in need, while individual ethics favor the protection of the human subjects actually enrolled in a specific research study. Adaptive designs address both of these issues by using accumulating results to modify the trial design during the trial. One statistical aspect of multi-arm clinical trial design that can be adaptively modified is the randomization ratio, such as has been proposed in the I-SPY 2 and BATTLE trials. This thesis explores the statistical properties of adaptively modifying randomization ratios to favor accrual to experimental arms whose interim results appear most promising. Biases and inflation of type I error that can arise from modifying randomization ratios, in particular in the presence of secular time trends in patient characteristics, are described and demonstrated with simulations. In addition, we show that these issues can largely be mitigated by properly accounting for randomization ratios through stratified analyses.