Title
Generalized Estimating Equations Analysis for Correlated Ordinal Response Data
Speaker
Dr. Fangyi Luo, Proctor and Gamble, Pharmaceuticals Division
Abstract
Generalized estimating equations (GEE) have been used for the analysis of longitudinal or clustered response data. GEE requires the specification of a working correlation structure for the correlated responses within clusters. Choosing the working correlation structure close to the actual one can improve the efficiency of the parameter estimates. Existing working correlation structures for the GEE analysis of categorical responses include independence, exchangeable, m-dependent and unstructured.
A new class of correlation structures, which allows for different correlation parameters in different components of the multinomial responses, is proposed and incorporated into the GEE analysis of correlated ordinal response data using the One Step Gauss-Newton (OSGN) procedure for proportional odds model. This class of correlation structures requires the estimation of only d different correlation parameters for responses with (d+1) categories.
Under mild regularity conditions and given that the link function of the response probabilities is correctly specified, the OSGN procedure for estimation of GEE model parameters provides consistent and asymptotically normal estimates of regression parameters and consistent estimates of intra cluster correlation parameters.
Through a simulation study, it is shown that the proposed method, i.e., GEE assuming the new working correlation structure, provides correct Type I error rate for large samples. For correlated ordinal response data generated having the new correlation structure within clusters, the proposed method provides more power, i.e., is more efficient, as compared to the GEE methods assuming independent or exchangeable working correlation structure.
The proposed method is illustrated on datasets collected from clinical trials.
