A1825
Title: Exact Poisson two-stage designs for landmark survival endpoints
Authors: John Rice - University of Michigan (United States) [presenting]
Abstract: Landmark analyses of time-to-event endpoints are increasingly used in phase II clinical trials to enable timely efficacy assessment under fixed follow-up. Common two-stage designs rely on large-sample normal approximations or simulation-based calibration, which can lead to distorted type I error and inefficient early stopping in small to moderate samples. A class of exact two-stage designs for landmark survival endpoints based on a Poisson event-count framework is proposed. Modeling the number of events within a fixed window as a Poisson process with known exposure, exact interim and final decision boundaries are derived that control the one-sided type I error without asymptotic approximations. The designs allow early stopping for futility while preserving exact error control. Closed-form expressions are obtained for operating characteristics, including rejection probabilities, probability of early termination, and expected sample size under null and alternative hypotheses. This enables efficient design evaluation without extensive simulation. The approach is illustrated using landmark progression-free survival as a motivating example. These designs provide a transparent and computationally efficient alternative to approximate methods for landmark survival endpoints in early-phase trials.
