Bayesian joint modelling of longitudinal and time to event data: a methodological review

Area of Science:

• Statistical modeling
• Biostatistics
• Clinical research methodology

Background:

• Increasing interest in joint modeling of longitudinal and time-to-event data in clinical research.
• Joint models reduce bias and increase efficiency in statistical inference.
• Need for a review of Bayesian estimation for joint models due to their growing popularity.

Purpose of the Study:

• To conduct a comprehensive review of Bayesian univariate and multivariate joint models.
• To focus on outcome types, model assumptions, association structures, estimation algorithms, dynamic prediction, and software implementation.
• To provide recommendations for future research in Bayesian joint modeling.

Main Methods:

• Systematic review of 89 articles (75 methodological, 14 applied).
• Focused on key components: outcome types, assumptions, association, estimation, prediction, and software.
• Identified common modeling approaches and estimation algorithms.

Main Results:

• Linear mixed-effects models with proportional hazards were most common for joint outcomes.
• Random effect association structures were frequently used to link sub-models.
• Markov Chain Monte Carlo (MCMC) algorithms were used in 93% of studies for parameter estimation.
• Only six articles focused on dynamic predictions.

Conclusions:

• While methodologies exist for various data types, research is limited for time-varying associations and determining association structures without prior knowledge.
• Joint modeling improves dynamic prediction accuracy, but validation tools are lacking.
• Future research should address time-varying associations, structure determination, and prediction validation.