Ill-Conditioned Generalized Estimating Equations

Louisiana State University, Baton Rouge LA

Investigators

Abstract

This project will investigate the behavior of ill-conditioned regressors with correlated response generalized estimating equations (GEE) framework. This ill conditioning occurs when the Fisher's Information matrix is nearly singular, a situation that leads to traditional problems of 'multicollinearity'. The resulting effect is poor prediction, large variability in the population parameters that are being estimated and poor testing. The Principal Investigator will subdivide the alternative estimators and will then propose well-defined estimations techniques for the subclasses. The estimators are used to understand regression under ill-posed conditions. The procedures to be developed under this grant are primarily for the purpose of understanding the fundamental aspects, but also will have application to large longitudinal data sets and provide a unifying framework for generalized models and the expectation/maximization algorithm, chemometrics and genomics. The Principal Investigator will enable many students from a large number of disciplines to participate in his education/research efforts.