Sample size calculations for clinical trials typically aim to achieve a specified power, based on a number of events among all subjects who are enrolled in the trial. The timing of data analysis is thus often determined by accrual of events during a study until a "landmark" event occurs. Significant resources are often linked to the correct estimation of this event, including scheduling of interm meetings or promises of final results. Further, conducting analyses using a different number of events than planned can adversely affect Type I error.
Bagiella and Heitjan (2001) and Ying et al. (2004) explored a variety of parametric and nonparametric models to estimate landmark event times under the assumption that the treatment arm assignments are known. However, for many clinical trials, the treatment arms are "masked" from the project team. The manuscript below considers a parametric mixture model for event time prediction, where treatment arm assignment is unknown. A fully Bayesian model is used, with estimation and inference for the Tth landmark event obtained from its posterior predictive distribution.