Statistical Issues in Platform Trials with a Shared Control Group

Platform trials evaluating multiple treatment arms against a shared control are an efficient alternative to multiple two-arm trials. Motivated by a randomized clinical trial of the effectiveness of two neuroprotection devices during aortic valve surgery against a control, this dissertation addresses two open questions in the optimal design of these trials. First, to explore whether multiplicity adjustments are necessary in a platform design, simulation studies evaluating the operating characteristics of platform designs relative to independent two-arm trials were conducted. Under the global null hypothesis, relative to a set of two-arm trials, we found that platform trials have slightly lower familywise error; however, conditional error rates for an experimental treatment being declared effective given another was declared effective are above the nominal alpha-level. Adjusting for multiplicity reduces familywise error, but has little impact on conditional error. These studies show that multiplicity adjustments are unnecessary in platform trials of unrelated treatments. Second, to determine the optimal approach for comparing delayed entry arms to the shared control, five methods for incorporating historical controls into two-arm trials were applied to the analyses of simulated open platform trials and compared to pooling all controls. We found that when response rates are constant, pooling yields the lowest error and most precise, unbiased estimates. However, if drift occurs, pooling results in type I error inflation or deflation depending on the direction of drift, as well as biased estimates. Although superior to naive pooling, none of the alternatives explored guarantee error control or unbiased estimates in the presence of drift. Thus, only concurrent controls should be used as comparators in the primary analysis of confirmatory studies. Finally, these findings were applied to assess the design and analysis of the neuroprotection trial.