On the predictive performance of two Bayesian joint models: a simulation study

Abstract

Joint modeling of longitudinal and survival data is becoming increasingly popular. It can properly handle multiple issues commonly encountered in longitudinal studies, such as endogenous time-dependent covariates and informative missingness. There are several statistical packages for conducting Bayesian joint modeling analysis, two of which are JMbayes, which has been applied in biomedical research of various fields, and rstanarm, which has been developed recently. Both packages are very flexible in specifying different association structures, and are capable of handling multivariate outcomes. However, no studies have ever been conducted to compare their performance. In this study, we conducted simulation studies to compare the performance of the two packages, with a focus on their prediction. We found that rstanarm often had better predictive performance than JMbayes, but its computation was more intensive and increased dramatically with larger sample sizes. In contrast, JMbayes was fast in model fitting and prediction, and in some cases, its performance could be improved by using larger sample sizes or longer periods of longitudinal data. Model mis-specification seemed to have a greater influence on the predictive performance of rstanarm than JMbayes.