Efficient and scalable Bayesian inference for joint modelling longitudinal and event history data

Guangquan Li

Applied Statistics and Data Science Lab, School of Engineering, Physics, and Mathematics, Northumbria University, UK

Joint models of longitudinal and event history data provide a power statistical framework for analysing complex health data from clinical studies and longitudinal studies of health. However, they are computationally expensive to fit due to the complex joint likelihood and the need to estimate many individual-level random effects. In this talk, we will introduce fastBJM, an efficient algorithm for fitting Bayesian joint models using Markov chain Monte Carlo. The algorithm updates model parameters via the Metropolis-within-Gibbs scheme, allowing us to exploit the structures of the full conditionals to derive efficient sampling. The results from our extensive simulation studies have demonstrated the estimation accuracy of fastBJM and its faster fitting compared to existing methods. The simulations also highlight its scalability in handling complex models as well as large datasets. Our application uses fastBJM to fit a suite of joint models to data extracted from the Survey of Health, Ageing and Retirement in Europe (SHARE), a large-scale multi-country longitudinal survey on health. We will discuss the results on the value of using body mass index as a tool to dynamically monitor the risks of hypertension and stroke.

Keywords: Joint modelling, Bayesian computation, Longitudinal health survey