Bayesian Predictive power: the bathtub phenomenon

Abstract

In a clinical trial with a time-to-event primary endpoint, it is generally of interest to estimate timepoints when a pre-specified number of events will be reached, e.g., timepoints for planned interim or the final analysis. We summarize a simple frequentist framework how to do such event predictions based on either an assumed or estimated survival function. The approach proposed by Fang and Su (2011), that estimates the survival function by Kaplan-Meier up to a formally detected changepoint and uses an Exponential tail fit beyond that changepoint, will be used to do the predictions. We apply the methodology to a real clinical trial in oncology and provide a confidence interval around predicted timepoints via bootstrapping of survival data. We share our experience in how expectations from broader teams on such type of predictions can be managed.

Date
Location
Basel, Switzerland