Joint modelling of longitudinal and event time data

Abstract

Abstract The relationship between the development over time of a quantitative response variable and a parallel process recording events in time can be investigated through a joint model for the two outcome types. Interest in this topic arose from two separate backgrounds: the possibility and consequences of informative dropout in longitudinal trials, and measurement error problems for survival studies with time‐varying covariates. Modeling strategies include combining a marginal distribution for one response type with a conditional distribution for the other, or adopting a correlated random effects model with otherwise conditional independence. Naive methods can lead to significant bias, yet more sophisticated modeling approaches are often sensitive to assumptions, some of which cannot be checked from observable data. When simple summaries such as treatment effect estimators are of interest, sensitivity analyses can bring some reassurance as to whether important conclusions are maintained as assumptions are allowed to change. Overall, the topic is relatively new and research is continuing in many directions.