Dynamic Prediction of Analysis Timing in Clinical Trials Using Joint Models of Longitudinal and Time-to-Event Data

In confirmatory clinical trials with a time-to-event endpoint, the timing of the analysis is often determined based on the observed number of events. In such event-driven trials, the analysis timing is inherently uncertain, making its accurate prediction crucial for optimizing trial costs, personnel allocation, and resource management. Although various methods for predicting analysis timing have been extensively studied under different assumptions, these approaches predominantly rely on data pertaining to the primary endpoint when estimating the time-to-event distribution. To the best of our knowledge, no existing method incorporates longitudinally measured data, such as prostate-specific antigen (PSA) levels in prostate cancer trials. We developed a method for predicting analysis timing within the framework of dynamic prediction using joint models that account for longitudinally measured covariates and time-to-event outcomes. Furthermore, we assessed the prediction performance by comparing it against two alternative methods based on the degree of covariate data: (1) a method without covariates, (2) a method using only baseline covariates, and (3) the proposed method using baseline and longitudinally measured covariates. Numerical experiments demonstrated that the proposed method enhances the accuracy of analysis timing predictions compared with methods that do not utilize longitudinally measured covariates. The proposed method is valuable for accurately predicting analysis timing when longitudinally measured covariates are available.