A two-stage joint modeling approach for multiple longitudinal markers and time-to-event data

Area of Science:

• Biostatistics
• Clinical Data Analysis
• Longitudinal Data Modeling

Background:

• Joint modeling of longitudinal markers and time-to-event data is crucial in clinical studies.
• Increasing marker numbers lead to computational challenges and convergence issues in standard joint models.

Purpose of the Study:

• Propose a novel two-stage Bayesian approach for joint modeling of multiple longitudinal markers and time-to-event outcomes.
• Address computational intractability and bias due to informative dropout in complex clinical datasets.

Main Methods:

• A two-stage Bayesian approach: Stage 1 estimates one-marker joint models, predicting marker trajectories and avoiding informative dropout bias.
• Stage 2 fits a proportional hazards model using predicted marker values/slopes as time-dependent covariates.
• Employs multiple imputation to handle uncertainty in first-stage predictions for survival model estimation.

Main Results:

• The proposed method successfully handles a large number of longitudinal markers, overcoming computational limitations of conventional multi-marker joint models.
• Validated through simulation studies and application to the PBC2 dataset.
• Demonstrated effectiveness in predicting dementia risk using a dataset with seventeen longitudinal markers.

Conclusions:

• The novel two-stage Bayesian method offers a computationally feasible solution for joint modeling with numerous longitudinal markers.
• Facilitates the development of robust prediction models in complex clinical research settings.
• An R package, TSJM, is available to support the practical application of this approach.