Repository logo

Dynamic predictions using flexible joint models of longitudinal and time-to-event data.

Published version

Change log


Joint models for longitudinal and time-to-event data are particularly relevant to many clinical studies where longitudinal biomarkers could be highly associated with a time-to-event outcome. A cutting-edge research direction in this area is dynamic predictions of patient prognosis (e.g., survival probabilities) given all available biomarker information, recently boosted by the stratified/personalized medicine initiative. As these dynamic predictions are individualized, flexible models are desirable in order to appropriately characterize each individual longitudinal trajectory. In this paper, we propose a new joint model using individual-level penalized splines (P-splines) to flexibly characterize the coevolution of the longitudinal and time-to-event processes. An important feature of our approach is that dynamic predictions of the survival probabilities are straightforward as the posterior distribution of the random P-spline coefficients given the observed data is a multivariate skew-normal distribution. The proposed methods are illustrated with data from the HIV Epidemiology Research Study. Our simulation results demonstrate that our model has better dynamic prediction performance than other existing approaches.



P-splines, random effects, shared parameter models, survival analysis

Journal Title

Statistics in Medicine

Conference Name

Journal ISSN


Volume Title


John Wiley & Sons Inc.
MRC (MR/L501566/1)
Medical Research Council (MR/K014811/1)
Medical Research Council (G0700463)
Medical Research Council (MR/L003120/1)
MRC (unknown)
British Heart Foundation (None)
Medical Research Council (G0701619)
Medical Research Council (MR/M025152/1)
Medical Research Council (G0700463/1)
Medical Research Council. Grant Numbers: G0902100, MR/K014811/1 Unit Programme. Grant Numbers: U105261167, U64-CCU10675 US Centers for Disease Control and Prevention