Comparison of parametric and hybrid methods for estimating mean survival time in clinical study.

Abstract

The mean survival time (MST) is usually estimated as the area under the curve of the estimated survival function obtained using the Kaplan-Meier method. However, when the maximum observed survival time is censored, the MST cannot be estimated because the survival function does not reach zero. In such cases, parametric and hybrid methods are used to estimate the MST. The parametric method assumes a probability distribution throughout the entire time and has been evaluated in several studies. The hybrid method combines two approaches: it first applies the Kaplan-Meier method up to a specified time point and then extrapolates the survival curve beyond this point using a parametric distribution. Evaluation of the performance of the hybrid method is limited to a few data-generating mechanisms and analysis models. This study evaluated the performance of the parametric and hybrid methods through numerical experiments, assuming nine probability distributions for the data-generating mechanism and 16 analysis models. The bias and root mean square error of the generalized gamma model and the Royston-Parmar models with the log(-log) link function tended to be smaller than those of the other analysis models, even when the assumed probability distribution of the analysis model was inconsistent with that of the data-generating mechanism when the sample size is relatively large. Overall, the performances of the parametric and hybrid methods were comparable across all the data-generating mechanisms.