A0555
Title: Two-stage variable selection in joint models for multiple longitudinal biomarkers and competing risk events
Authors: Taban Baghfalaki - The University of Manchester (United Kingdom) [presenting]
Reza Hashemi - Razi University (Iran)
Christophe Tzourio - University of Bordeau (France)
Catherine Helmer - University of Bordeau (France)
Helene Jacqmin-Gadda - Universite de Bordeaux (France)
Abstract: Understanding the relationship between longitudinal biomarkers and time-to-event outcomes is key in clinical and epidemiological research. However, joint modeling becomes computationally intensive and prone to convergence issues as the number of markers grows. A novel two-stage Bayesian approach is proposed for efficient variable selection in joint models, demonstrated through a practical application. In the first stage, individual joint models are fit for each longitudinal marker associated with the event. This allows for accurate trajectory predictions while mitigating bias from informative dropout. In the second stage, a proportional hazards model is used that incorporates the expected current values of all markers as time-dependent covariates. To perform variable selection, continuous and Dirac spike-and-slab priors are implemented within a Markov chain Monte Carlo (MCMC) framework. The method addresses the challenges of high-dimensional joint modeling, enabling improved parameter estimation and risk prediction. Through simulation studies, the approach shows strong performance. It is also applied to predict dementia risk using data from the French three-city (3C) study, a longitudinal cohort focused on aging. This two-stage Bayesian strategy provides an effective and scalable solution for variable selection, enhancing the utility of joint models in complex longitudinal research settings.
