Prediction of event times in the REMATCH Trial

In clinical trials with time-to-event outcomes, it is common to design the interim analysis plan around the occurrence of designated target numbers of events. As the trial progresses, one may wish to use the accumulating data to predict the calendar times of these events as an aid to logistical planning.

To demonstrate three models for the prediction of event times using the accumulating data of the Randomized Evaluation of Mechanical Assistance for the Treatment of Congestive Heart Failure (REMATCH) Trial.

We apply three prediction models--an exponential model, a Weibull model, and a nonparametric model--to the evolving REMATCH data. Using Bayesian simulation methods, we predict the times of the designated landmark events and the end-of-study treatment effect, and calculate the predictive power.

The models were practical to apply from an early stage of the trial, and gave largely similar predictions. Intervals from the nonparametric model were generally wider and more responsive to surges and droughts in events. Predictions made early in the trial were sensitive to the assumed prior on the accrual rate.

The use of badly calibrated priors can lead to poor predictions.

Predictions of landmark event times in REMATCH were accurate and responded deftly to a strong shift in the treatment effect that occurred midway through the trial. The method can provide reliable guidance for clinical trial planning.