Abstract

In risk prediction for cardiovascular disease (CVD), where a risk factor such as systolic blood pressure (SBP) is volatile, giving high within-subject variability, correctly modelling that variability could offer further improvements compared to the classic joint model for SBP and CVD. Motivated by this example, joint models are proposed for the survival outcome (time to CVD) as well as both the mean and variance of the longitudinal outcome (SBP). These models are linked via heterogeneous random effects sharing the same distribution, allowing us to capture the pairwise associations between the three outcomes through the random effects covariance matrix. Both the modified Cholesky and Hypersphere decompositions are considered to reparameterise the conditional covariance of the longitudinal response and employ a Bayesian approach for estimation. The performance of the proposed approach is demonstrated via simulation and application to the Systolic Blood Pressure Intervention Trial (SPRINT) dataset.