Seamless Phase II/III designs are a natural vehicle for the randomized dose optimization called for by the US Food and Drug Administration's Project Optimus initiative. Their extension to group sequential repeated testing of a time-to-event endpoint after end-of-Phase-II dose selection, however, remains largely unestablished. Existing proposals index the boundaries for the post-selection looks by the overall information fraction of the primary endpoint---the correct index for the log-rank statistics of an unadapted trial, but not for the combination statistics these designs actually test---and support family-wise error rate (FWER) control by simulation. We instead work from the joint distribution of the combination statistics themselves, calibrating group sequential boundaries separately for each hypothesis in the closure and applying closed testing at every look; a general lemma establishes that this controls the FWER in the strong sense, for any number of hypotheses and analysis times. Specifically, under patient-wise separation and an inverse-normal combination test, the interim and final statistics for an individual hypothesis are bivariate normal, with a correlation determined by how the events at each look divide between the pre- and post-selection cohorts rather than by the overall information fraction, so that their boundaries follow from standard group sequential software; the corresponding statistics for the intersection hypothesis, formed from a Hochberg-adjusted Stage-1 $p$-value, are not multivariate normal, and we obtain their boundaries by a dedicated numerical integration. Head-to-head against a recently proposed incremental-Dunnett procedure, the two were near-identical when the design projections were correct, but ours was more robust when the combination weights were based on a misprojected division of the Stage~1 events either side of the interim.

翻译：暂无翻译