Bayesian model for joint longitudinal and survival outcomes in the presence of subpopulation heterogeneity
Elizabeth Slate, PhD Florida State University
Biomedical studies often monitor subjects using a longitudinal marker that may be informative about a time-to-event outcome of interest. An example is periodic monitoring of CD4 cell count as the longitudinal marker and time to death from AIDS. By modeling these two outcomes jointly, there is potential to improve the precision of inference for each. Brown and Ibrahim (2003, Biometrics) adopted the common approach of incorporating the mean longitudinal trajectory as a predictor for the event time hazard, and enhanced flexibility by using a Dirichlet process prior for the coefficients of the trajectory. We generalize this model to accommodate non-Gaussian longitudinal outcomes and emphasize that the Dirichlet process enables discovery of subgroups of subjects with distinct behavior for the joint outcome. Our formulation, developed using the multivariate log-gamma distribution, offers greater flexibility in the longitudinal model and computational advantage for Markov chain Monte Carlo estimation. We illustrate with simulation and application.