Type I Error Control of an Adaptive Endpoint Selection Procedure

In clinical trials, a single primary endpoint or a primary endpoint family is traditionally pre-specified before trial initiation. While in general the primary endpoint(s) of a trial should not be altered thereafter, there may be situations where the pre-specified primary analysis no longer makes sense. For example, when endpoint sensitivity is markedly less than expected. To investigate the operating characteristics of endpoint adaptation, we present two approaches to tackle this problem: primary endpoint adaptation based on (a) unblinded or (b) blinded estimate of the coefficient of variation from the control group. We show that the adaptation method based on blinded data introduced in this article has strong control of the family-wise Type I error, while the adaptation based on unblinded data does not. We examine the asymptotic conditions under which the adaptation based on blinded data correctly chooses the most powerful endpoint. We also compare the performance of the blinded adaptation with other potential approaches. We present a case study illustrating how the proposed blinded adaptation could have been used in a real clinical trial. Finally, we provide recommendations on how to implement the adaptation in practice.