Combining longitudinal and survival information in Bayesian joint models: When are treatment estimates improved?

In studying a biological process, investigators may repeatedly measure features of the process (longitudinal data) and also measure the time to an important state change (survival data). For example, a clinical trial may measure symptom severity and time until death. The most popular class of joint models for simultaneously analyzing longitudinal and survival data uses latent variables to link longitudinal and survival submodels. To compare such joint modeling to modeling the data separately, we construct a Gaussian joint model and focus on the effects of a treatment on longitudinal and survival outcomes. With vague (improper) priors on the treatment effects and known variances, treatment effect posteriors are the same in joint and separate models. To whom correspondence should be addressed.