Bayesian quantile regression joint models: Inference and dynamic predictions.

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Survival Analysis

Background:

• Traditional joint models for longitudinal and time-to-event data often rely on normality assumptions for the longitudinal component, which are frequently violated in practice.
• Linear mixed models, commonly used, focus on the conditional mean, limiting their utility when interest lies in other quantiles or when data is non-normal or contains outliers.

Purpose of the Study:

• To introduce and advocate for the use of linear quantile mixed models within the joint modeling framework for longitudinal and time-to-event outcomes.
• To address the limitations of traditional models by providing a flexible, distribution-free approach robust to non-normality and outliers.

Main Methods:

• Developed a Bayesian methodology utilizing the location-scale representation of the asymmetric Laplace distribution for the linear quantile mixed model.
• Applied the proposed method to a prospective study of Huntington's disease, using longitudinal motor scores and covariates to predict disease risk.

Main Results:

• The linear quantile mixed model provides a robust and flexible alternative to traditional linear mixed models in joint analyses.
• The developed Bayesian approach effectively models longitudinal data at various quantiles, improving prediction accuracy for time-to-event outcomes.
• Demonstrated the utility of the approach for subject-specific dynamic survival probability predictions.

Conclusions:

• Linear quantile mixed models offer a powerful tool for joint modeling when standard normality assumptions are not met.
• This methodology is particularly valuable for predicting disease progression and risk in complex conditions like Huntington's disease.
• The Bayesian framework facilitates robust inference and personalized predictions in joint modeling scenarios.