Multivariate longitudinal data analysis with mixed effects hidden Markov models.

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Health Outcomes Research

Background:

• Multiple longitudinal responses are frequently collected to infer underlying health states.
• These responses often reflect a hidden disease process that is not directly measurable.
• Existing methods may not fully capture the complex interdependencies between such responses.

Purpose of the Study:

• To develop and validate a statistical modeling approach for multivariate longitudinal responses.
• To specifically model these responses as originating from a latent disease process.
• To apply and evaluate the proposed methodology in the context of smoking cessation clinical trials.

Main Methods:

• Proposed a class of models utilizing a hidden Markov model (HMM).
• Incorporated separate but correlated random effects for multiple longitudinal responses.
• Employed Bayesian inference with Markov chain Monte Carlo (MCMC) methods.
• Compared bivariate response models against separate univariate models.

Main Results:

• The bivariate hidden Markov model approach effectively captures the dynamics of correlated longitudinal responses.
• Demonstrated superior performance in modeling smoking behavior compared to univariate models in a clinical trial dataset.
• Simulation studies confirmed the robustness and properties of the proposed methodology.

Conclusions:

• The developed hidden Markov model provides a robust framework for analyzing multivariate longitudinal data influenced by a latent process.
• This approach offers enhanced insights into complex health behaviors and disease progression.
• The methodology is particularly valuable for clinical trials involving multiple, correlated outcome measures.