Statistical approaches for adding or switching hypotheses in multi-armed clinical trials.

摘要

Background. As treatments become ready for testing at staggered times, it is desirable to have clinical trials that can accommodate different entry and exit times without prematurely discarding potentially efficacious treatments. In a group sequential multi-armed clinical trial, one treatment arm could be found to be superior or inferior at an interim analysis while the remaining arm is inconclusive. Existing methods address dropping an inferior arm from further study, but do not fully address the handling of an early finding of overwhelming superiority of one arm ( scenario 1). We consider an approach to transition from a multi-armed superiority trial to a two-armed non-inferiority trial after superiority for a single arm has been determined. Additionally, the literature does not address statistical methods for adding another treatment arm into an ongoing trial (scenario 2). Methods. For these novel scenarios, potential adaptive and non-adaptive analytical approaches for pairwise comparisons of a difference in means in independent normal populations are proposed with emphasis on controlling the type I error rate strongly. Statistical operating characteristics are compared via Monte Carlo simulation. An example is given using Parkinson's disease data (NET-PD FS1 and FS-TOO). Results. For scenario 1, all methods performed similarly, but power was highest when using an inverse chi-square adaptive test. For scenario 2, in the presence of a cohort effect, when data were pooled from before/after the design change, the type I error rate was inflated and power was reduced. The alternative approaches given were more powerful and controlled the type I error rate. Conclusions. When two treatment arms are equally efficacious, it is likely that one, but not the other, will be found efficacious at an interim analysis. When transitioning into a non-inferiority trial, the adaptive methods allow for a reduction in total sample size with increased power for testing non-inferiority (compared to a non-adaptive approach). Both adaptive and non-adaptive analytical methods are possible when a new treatment arm is added mid-study. When these methods are applied to real Parkinson's disease trial data, the conclusions support the primary trial findings.