A simulation experiment was conducted to assess the performance of the generalized es- timating equations (GEE) approximated cluster Cook’s Distance relative to its fully-iterated exact version. Specific interest was in the eect of cluster-deletion on the overall fit of a model for the marginal mean and on the overall fit of a model for the marginal association, in the context of alternating logistic regressions (ALR). The experiment demonstrated that one-step approximated cluster Cook’s Distance statistics successfully identify clusters having the great- est influence, as measured by their fully iterated exact counterparts. The regression diagnostic identified the cluster with the greatest influence on the fit of the marginal association model 88% of the time. The cluster with the greatest influence on the fit of the marginal mean model was identified 60% of the time, and the probability of detection increased with increasing size of the diagnostic. Cook’s Distance statistics for GEE and recently proposed analogous measures for ALR are useful for the analysis of clustered binary data.
Biostatistics | Longitudinal Data Analysis and Time Series | Numerical Analysis and Computation | Statistical Models
Preisser, John; By, Kunthel; and Qaqish, Bahjat, "Performance of One-Step Approximation Relative to Exact Cluster Cook's Distance for GEE" (July 2008). The University of North Carolina at Chapel Hill Department of Biostatistics Technical Report Series. Working Paper 8.