Accounting for the Presence of Surrogate Data in Adaptive Clinical Trials

Abstract

Some adaptive designs for randomized clinical trials (RCTs) allow for flexibility in modifying the sequential sampling plan using results from unblinded interim analyses. However, care must be taken to ensure that desired statistical operating characteristics are preserved when allowing for such adaptations. Approaches to analyzing results from RCTs with adaptations have been proposed in the literature, such as methods by Bauer & Köhne (1994), Proschan & Hunsberger (1995), Fisher (1998), and Cui, Hung, & Wang (1999), but these generally assume the use of adjustments that incorporate sufficient statistics derived from the data available at the time of adaptation. Bauer & Posch (2004) noted that this assumption may be violated in settings involving time-to-event data, where observed surrogate outcomes that are potentially informative about future events are not reflected in the adjustment for adaptation. For instance, in an analysis of overall survival, the adaptive analysis typically does not account for the way disease progression in censored subjects might be informative for the eventual death times of those subjects. The impact of falsely assuming unavailability of partial knowledge regarding future data needs to be explored. Via simulation, we found that under an extreme scenario the type I error might be inflated from 0.025 to 0.205 when surrogate data can be used to predict future event times. Factors contributing to this inflation include minimal spending of error before adapting, early adaptations, minimal restrictions on sample size modifications, and stopping patient recruitment after the time of adaptation. Interestingly, in some situations where surrogate outcomes are informative but not accounted for, use of the standard methods for adaptation (e.g., Cui, Hung, & Wang) can behave worse than ignoring adaptation completely. Generally, we find that adjusting for surrogate data in the analyses does not control type I error unless strong or conservative assumptions are made. It is not immediately clear that adaptive designs which correctly account for surrogate data hold any significant advantages to comparable group sequential designs, and thus we explore situations where the methods to adjust for this surrogacy might decrease the efficiency so much as to suggest that unblinded adaptations be avoided.