Methods for the prediction of endpoint-occurrence times in clinical trials

Abstract

This research is motivated by the need to anticipate the calendar time of landmark numbers of events in a clinical trial with a time-to-event monitoring endpoint. In such trials, the observed number of events usually determines the statistical information. Thus the timing of accumulating events in the trial factors strongly into the schedule of interim analyses and Data and Safety Monitoring Board reviews, as well as the overall length of the study. Once a trial is underway, patient accrual, event, and dropout rates may deviate from those anticipated during the design of the study. Then current data from the trial itself may provide the most relevant indicators of the timing of future event occurrences. Our goal is to develop methods to predict the future calendar time of the Nth event in an ongoing clinical trial using data collected between the onset of the study and an interim data cutoff time. We describe two methods for this type of conditional event-time prediction. The first is a fully parametric method, based on maximum likelihood estimation and inference. The second is a semiparametric approach that allows for more flexible modeling of baseline hazards of patient events and loss-to-follow-up, as well as the inclusion of baseline covariate information. For each method, we discuss event-time point prediction and the construction of prediction intervals. Simulation studies are used to demonstrate the performance of the prediction methods in a variety of possible scenarios. This research was motivated by monitoring considerations in the HIV Prevention Trials Network 052 Study of the prevention of HIV-1 infection with early antiretroviral therapy, and data from the trial is used to demonstrate the proposed methods. The dissertation concludes with a discussion of possible extensions, refinements, and further applications of this research.