Repeated measures discriminant analysis using multivariate generalized estimation equations.

Area of Science:

• Multivariate Statistics
• Biostatistics
• Longitudinal Data Analysis

Background:

• Traditional discriminant analysis methods for repeated measures designs often assume multivariate normality.
• These normality assumptions are frequently violated in practice, especially with binary, ordinal, or mixed response types.
• This limitation hinders accurate population discrimination in complex longitudinal studies.

Purpose of the Study:

• To evaluate the classification accuracy of repeated measures discriminant analysis (RMDA) utilizing the generalized estimating equations (GEE) framework.
• To compare the performance of GEE-based RMDA against traditional maximum likelihood estimators (MLE)-based RMDA.
• To assess these methods across various simulation conditions, including different response types and correlation structures.

Main Methods:

• Employed Monte Carlo simulations to compare classification accuracy.
• Investigated RMDA based on generalized estimating equations (GEE) and maximum likelihood estimators (MLE).
• Varied simulation parameters: number of measurement occasions, number of responses, sample size, correlation structures, and response distribution types.

Main Results:

• RMDA based on GEE demonstrated higher average classification accuracy than RMDA based on MLE.
• The superiority of GEE-based RMDA was particularly pronounced in scenarios with multivariate non-normal distributions.
• The study explored classification accuracy under diverse longitudinal data conditions.

Conclusions:

• GEE-based RMDA provides a more robust and accurate approach for classification in multivariate repeated measures designs, especially when normality assumptions are not met.
• This framework is valuable for analyzing longitudinal data with mixed or non-normal response types.
• The findings support the application of GEE-based RMDA in fields like epilepsy research, as demonstrated by the case study.