A PRESS Statistic for Working Correlation Structure Selection in Generalized Estimating Equations

• Activity Number: 37
• Type: Contributed
• Date/Time: Sunday, August 9, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): A.H.M. Mahbub Latif and John Preisser*
• Companies: University of Dhaka and The University of North Carolina
• Keywords: Correlated binary responses ; Model selection ; Deletion diagnostics ; Bias-correction ; Longitudinal data

Abstract

A Predicted Residual Sums of Squares (PRESS) statistic is considered for model selection in generalized estimating equations. A computational formula is proposed that generalizes PRESS for multiple linear regression. It is based on matrix weights for the cluster-level residual vectors that are functions of the cluster leverage matrix. The introduction of PRESS completes a unified, simple and elegant approach to model selection, one-step deletion diagnostics, and bias-corrected covariance estimation for regression parameters in marginal models for correlated data. In a simulation study involving marginal models for longitudinal binary data, PRESS performed better than other criteria for selection of the working correlation matrix. An analysis of 15-year trends in smoking in a cohort of young adults illustrates its use.