Séminaire de Biostatistique

Non-linear effect of shared components in joint longitudinal-survival modeling

by Denis Rustand (King Abdullah University of Science and Technology (KAUST))

Salle Mann (ISPED)

Salle Mann



Speaker: Denis Rustand from King Abdullah University of Science and Technology (KAUST)

Abstract: The simultaneous analysis of longitudinal and survival data have received important attention in statistical literature recently. Indeed, when examining repeated measurements of a biomarker alongside an event of interest, an inherent association between these two outcomes often exists. The risk of the event can be influenced by the longitudinal biomarker, with biomarker measurements typically being truncated by the occurrence of the event. Joint models, integrating longitudinal and survival outcomes, have become indispensable for capturing the intricate interplay between time- varying endogenous variables and survival times. However, these models traditionally assume linearity in the effects of shared components (i.e., the shared component from the longitudinal submodel is scaled by a single parameter in the survival submodel). This may not be flexible enough to capture the true underlying relationship between the outcomes. 
In this talk, we present a novel approach to relax this assumption where we allow for a non-linear dependence through the use of Bayesian smooth splines. This both avoids parametric assumptions on the shape of the association and overfitting (i.e., non-realistic brutal changes in the shape of the association). We demonstrate the implementation of our methodology with an application to data from a cancer clinical trial where we investigate the possible non-linear effect of tumor size on the risk of death. We further evaluate the frequentist properties and performances of the proposed approach through simulation studies. This methodology is implemented in the R package INLAjoint and allows to include simultaneously the non-linear effect of multiple longitudinal markers in multiple survival submodels (e.g., multi-state). Moreover, the shared component can have multiple forms, such as shared random effects, current value or current slope parametrizations. By relaxing the linearity assumption, our methodology provides a more nuanced understanding of the relationship between longitudinal markers and survival times, offering valuable insights for clinical research and a more powerful tool for individual dynamic predictions.


Calendar subscription link for the complete seminar series:


Program of the Biostatistics seminars:


Subscribe to the seminar mailing list:


Former e-seminars on our YouTube channel (mostly in French): https://www.youtube.com/channel/UCURp-hEQL7k23UzGfqgEurA/videos


Biostatistics seminar series from the Department of Public Health from the University of Bordeaux and the Bordeaux Population Health UMR 1219 research center


Organized by

Boris Hejblum