A Comparative Study On The Selection Of Work - Related Matrices In

In this article, we compare two modified Gaussian pseudolikelihood criteria (GPCs) with existing Gaussian pseudolikelihood criterion and empirical likelihood based cri-teria to choose the working correlation matrix in generalized estimating equations ap-proach. These criteria are also extended to multi-group longitudinal data to choose the working correlation matrices simultaneously. Rich simulation studies are conducted to investigate the performance of these criteria under a range of model settings for both continuous and discrete data. The results show that the modified criteria outperform the original GPC and empirical likelihood based criteria in most cases in terms of selection accuracy. Empirical likelihood based criteria perform better to identify exchangeable structure in data with binary response. To choose working correlation matrices simulta-neously in multi-group longitudinal data, modified Gaussian pseudolikelihood criteria still have robust advantages to choose the underlying structures and the advantages have been magnified in these cases. The goodness of empirical likelihood based criteria in binary data still exists in some cases. These criteria are also applied to Epilepsy Seizure and Madras Longitudinal Schizophrenia Study clinical data sets analysis. Another AD-HD data set is investigated to choose working correlation matrices simultaneously.