A Power Analysis of Misspecified Types of Time-Varying Covariates for Generalized Method of Moments Applied to Binary Responses

• Activity Number: 353
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Trent Lalonde*+ and Jeffrey R. Wilson
• Companies: University of Northern Colorado and The University of Utah
• Keywords: Generalized Linear Model ; Method-of-Moments ; Longitudinal ; Time-Dependent Covariates ; Binary Response ; Correlated Observations

Abstract

Correlated observations are commonly encountered in longitudinal studies involving binary responses. A marginal model has shown to be best modeled with GEE (Zeger and Liang (1986)). Both the GLM and GEE assume that the covariates are time-independent; however, neither of these methods is suitable when any covariates are time-dependent. For time-dependent covariates, a generalized method-of-moments has been developed (Lai and Small (2007)), in which the estimating equations are determined by the type of time-varying covariate.

This paper applies the techniques of Lai and Small to a data situation with a binary response, compares the GMM to the GLM and GEE, that both ignore information about the time-dependence of covariates, and also addresses questions of misspecification of time-dependent covariate types, and the effects on parameter estimates, standard errors, and test statistics.