Empirical-likelihood-based criteria for model selection on marginal analysis of longitudinal data with dropout...

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Longitudinal Data Analysis

Background:

• Longitudinal data with missing outcomes are prevalent in clinical and observational studies.
• The weighted generalized estimating equation (WGEE) approach is standard for marginal analysis under missing at random assumptions.
• Current model selection criteria for WGEE have limitations, including separate selection for mean and correlation structures and poor performance in small samples.

Purpose of the Study:

• To develop novel joint information criteria for simultaneous model selection of marginal mean and correlation structures in WGEE.
• To address the limitations of existing information criteria in longitudinal data analysis.
• To enhance the accuracy and efficiency of statistical modeling in the presence of missing data.

Main Methods:

• Embedding empirical likelihood within the weighted generalized estimating equation (WGEE) framework.
• Proposing two joint empirical information criteria: a joint empirical Akaike information criterion and a joint empirical Bayesian information criterion.
• Conducting extensive simulation studies to evaluate the performance of the proposed criteria against existing methods.

Main Results:

• The proposed empirical-likelihood-based criteria effectively select both marginal mean regression variables and correlation structures simultaneously.
• These novel criteria demonstrate robustness, flexibility, and superior performance compared to weighted quasi-likelihood under independence criterion, missing longitudinal information criterion, and joint longitudinal information criterion.
• Theoretical justification for the proposed criteria is provided, along with practical illustration using two real data examples.

Conclusions:

• The developed joint empirical information criteria offer a significant advancement for model selection in WGEE analysis of longitudinal data.
• These criteria provide a unified approach to selecting both mean and correlation structures, improving analytical precision.
• The proposed methods are recommended for robust and accurate analysis of longitudinal studies with missing outcomes.