Association models for periodontal disease progression: a comparison of methods for clustered binary data

Area of Science:

• Biostatistics
• Epidemiology
• Clinical Trials

Background:

• Periodontal disease progression is a significant concern in dental health.
• Longitudinal studies are crucial for understanding disease trajectories.
• Clustered binary logistic regression is often employed in clinical trials with correlated data.

Purpose of the Study:

• To compare population-averaged (PA) and cluster-specific (CS) association estimation methods.
• To evaluate the impact of non-informative clusters on logistic regression inferences.
• To assess generalized estimating equations (GEE), conditional likelihood (CL), and mixed effects (ME) models.

Main Methods:

• Analysis of a longitudinal clinical trial on periodontal disease.
• Application of generalized estimating equations (GEE) for PA models.
• Utilized conditional likelihood (CL) and mixed effects (ME) for CS models.
• Investigated the influence of non-informative clusters on statistical inferences.

Main Results:

• Conditional likelihood (CL) methods produced smaller test statistics compared to mixed effects (ME) methods when non-informative clusters were present.
• CL estimates demonstrated reduced efficiency relative to ME estimates under specific conditions.
• Observed consistency between population-averaged and cluster-specific parameter estimates, even with non-informative clusters.

Conclusions:

• The choice of statistical method impacts inferences in clustered binary logistic regression, particularly with non-informative clusters.
• Conditional likelihood methods may be less efficient but yield comparable parameter relationships.
• Findings support the robustness of established relationships between PA and CS parameters in complex clustered data scenarios.