Evaluation of Statistical Methods for Clustered Eye Data with Skewed Distribution

Area of Science:

• Ophthalmology and vision research
• Biostatistics
• Statistical modeling

Background:

• Ophthalmology and vision research frequently involve analyzing skewed, correlated data from a subject's two eyes within the same group.
• Evaluating the performance of different statistical methods for such data is crucial for accurate analysis.

Purpose of the Study:

• To assess the effectiveness of various analytical approaches for skewed, correlated ocular data.
• To compare the type 1 error rate and statistical power of different statistical tests under various simulation conditions.

Main Methods:

• A simulation study was conducted using skewed correlated data with varying intereye correlation, sample sizes, and mean differences.
• Nine different analysis methods were evaluated, including t-tests, generalized estimating equations (GEE), and linear mixed-effects models (LMM), with and without rank-based normalization.
• Real data from the Dry Eye Assessment and Management (DREAM) study, specifically tear break-up time (TBUT) data, were analyzed to validate simulation findings.

Main Results:

• Standard t-tests and Wilcoxon tests ignoring intereye correlation inflated the type 1 error rate.
• Generalized estimating equations (GEE) Wald test also showed inflated type 1 error rates with small sample sizes.
• For skewed data without normalization, GEE score and LMM demonstrated lower statistical power compared to clustered Wilcoxon methods.
• After normalization, GEE score and LMM achieved comparable or slightly superior statistical power to clustered Wilcoxon methods.
• Analysis of TBUT data from the DREAM study corroborated the simulation results.

Conclusions:

• Clustered Wilcoxon methods are suitable for comparing skewed correlated eye measures between groups.
• Normalizing skewed data followed by GEE score or LMM analysis offers an alternative approach.
• Normalization with GEE score or LMM may provide slightly higher statistical power and greater flexibility for adjusting covariates.