Stat Colloquium: Dr. Demissie Alemayehu

Title: Anytime-Valid Inference and Bias Correction in Adaptive Dose-Ranging Trials in Drug Development

Abstract: In drug development, Phase II dose-ranging trials routinely inspect accumulating data to inform go/no-go decisions. This practice creates two distinct problems: the maximum of several noisy estimates is optimistic because it was selected for being large, and repeated looks across multiple doses and stages inflate the Type I error rate. The usual remedies are unrelated, namely, bias correction for the first issue and multiplicity adjustment for the second. However, neither fixes the other, since a multiplicity-adjusted test still reports an inflated effect, and a bias-corrected estimate still does not license repeated looks.This talk introduces a framework in which both issues are addressed by a single predictable quantity: the selection premium, which measures the expected gain from being free to re-select the current leader. Charging this premium before each interim look yields an e-process, resulting in a decision rule that remains valid at whatever moment the trial is actually read out. The rule takes the form of a ledger: the observed best effect must clear a prespecified clinical margin after paying a selection charge for reporting a winner and a monitoring margin for choosing when to look. Numerical results illustrate the performance of the procedure, and extensions to other design options are discussed.