The Performances Of Several Modified CIC Criteria For Working Intra-cluster Correlation Structure Selection In GEE Analysis

The method of generalized estimating equations(GEE) is widely used for analyzing correlated response data, where selection of the working intra-cluster correlation structure(ICS) is pertinent because an improper selection sometimes results in inefficient parameter estimates(Fitzmaurice, 1995; Wang and Carey, 2003, 2004). In this paper, several modified versions of the well-known correlated information criterion(CIC) by Hin and Wang(2009) are proposed for working ICS selection from different points of view. Performances of these proposed modified criteria are examined and compared to the CIC criterion via simulations. Regardless of the response is Gaussian, binary or Poisson, the modified criteria are demonstrated to have higher detection rates when the true ICS is exchangeable, while the CIC would perform better when the true ICS is AR(1). Besides, under the assumption of unstructure correlation structure, the performance of the several criteria also shows. In addition, an application of the proposed modified criteria is also made to the epilepsy dataset available from Thall and Vail(1990) and the unbalance Madras dataset.