Bayesian variable selection for joint models of heterogeneous longitudinal variables and a binary outcome

Area of Science:

• Biostatistics
• Health Informatics
• Epidemiology

Background:

• Biomedical research frequently involves mixed-type longitudinal data for binary health outcomes.
• Analyzing irregularly collected, high-dimensional predictors linked to outcomes presents a significant challenge.
• Existing joint models (JM) lack formal variable selection, hindering predictor identification.

Purpose of the Study:

• To introduce novel Bayesian variable selection strategies for joint models with high-dimensional longitudinal and binary outcome data.
• To develop a framework (JM1 and JM2) for selecting predictors of binary outcomes and longitudinal trajectories simultaneously.
• To enhance the utility of joint models for identifying relevant predictors in complex biomedical datasets.

Main Methods:

• Developed two structured Bayesian variable selection strategies (JM1 and JM2) for joint models.
• Employed shrinkage priors to manage high-dimensional predictors and interactions, preventing overfitting.
• Extended false discovery rate (FDR) rules for variable selection in multi-part joint models, accommodating grouped categorical predictors.

Main Results:

• Applied the novel methods to the Women's Health Initiative and Life and Longevity After Cancer study data.
• Identified key predictors of post-treatment insomnia in breast cancer survivors.
• The proposed model detected a significant predictor missed by conventional analytical methods.

Conclusions:

• The introduced Bayesian variable selection framework offers a robust and interpretable approach for joint models.
• This method significantly enhances statistical power and addresses a critical gap in analyzing complex longitudinal biomedical data.
• The framework provides a computationally efficient solution for high-dimensional variable selection in health outcome prediction.