Predicting Longitudinal Measures Taking Risk of Time-to-Event Data Into Account Using Joint Modelling

Abstract

Analyses of longitudinal and time-to-event processes might lead to biased results if those are linked through an common underlying process. In order to get reliable estimates it is recommended to use a joint model. Joint models are composed of two submodels: a mixed-e_ects model for the longitudinal outcome and a Cox hazard model for the event outcome. This article fo- cuses on a HIV study where CD4+ T-cell counts were repeatedly measured over time and the occurence of tuberculosis serves as the time-to-event out- come. In order to _nd the best-_tting longitudinal model, this article seeks to compare di_erent models in an extrapolation study and, based on this, a joint model will be used to predict survival probabilities of tuberculosis. Predictions from the mixed-e_ects model will be weighted with the survival probabilities to account for the association between the processes.