Comparing Generalized Estimating Equation and Linear Mixed Effects Model for Estimating Marginal Association with...

Area of Science:

• Biostatistics
• Ophthalmology
• Statistical Modeling

Background:

• Generalized Estimating Equations (GEE) and Linear Mixed-Effects Models (LMEM) are statistical tools for analyzing clustered continuous outcomes.
• Bivariate continuous outcomes are frequently encountered in ophthalmic research, necessitating appropriate analytical methods.
• Comparing the performance of GEE and LMEM is essential for accurate estimation of exposure-outcome associations in eye studies.

Purpose of the Study:

• To compare the statistical performance of GEE and LMEM for bivariate continuous outcomes.
• To evaluate bias, coverage probability, and power in estimating marginal exposure-outcome associations.
• To determine the most suitable model for analyzing clustered continuous outcomes in ophthalmology.

Main Methods:

• Simulations were conducted using both parametric and non-parametric approaches.
• GEE models with independent, exchangeable, and unstructured working correlation structures were assessed.
• LMEM with random intercept only and random intercept and slope models were evaluated.
• Data distributions were based on a real-world study of ocular structure-visual function relationships in retinitis pigmentosa.

Main Results:

• The random intercept LMEM demonstrated superior coverage probability of 95% confidence intervals (CI) compared to GEE exchangeable models, particularly with small sample sizes.
• GEE exchangeable models showed higher power for detecting exposure-outcome associations but exhibited inflated Type I error rates.
• The random intercept LMEM maintained Type I error rates closer to the nominal 0.05 level, though occasionally under 0.05.
• GEE independent models performed poorly, and LMEM with random intercept and slope faced convergence issues.

Conclusions:

• The random intercept LMEM is recommended for estimating marginal exposure-outcome associations with bivariate continuous outcomes due to its robust coverage probability and accurate Type I error rates.
• While GEE models may offer higher power, LMEM provides more reliable confidence intervals and error control.
• Researchers should consider potential limitations of LMEM, such as lower power and wider CIs in small sample sizes or with low inter-eye correlation.