Introduction: With generalized estimating equation (GEE) modelling often a uniform correlation structure or independent structure is used when they are only weakly appropriate. Moreover while a Huber-White correction makes the result more robust, it also makes it less powerful even in a large database like Pediatric Health Information System (PHIS) with 5,825 hospitalized patients and 23,521 visits for sickle cell disease. Of interest were effects of Psychiatric Diagnoses on Hospital Admissions for Pain in Children with Sickle Cell Disease.
Method: A GEE model with independent, uniform and Toeplitz(2) and an unstructured correlation structure was used to model the Length of stay considering patients as hospitalization clusters occurring during the year. A gamma distribution with log link was used. Covariates were used to improve the model. Result: The choice of correlation structure affected the results. The models were compared using the GEE fit criteria (QIC) in order to provide the best model.
Conclusion: With longitudinal data choosing an appropriate link function and a correlations structure is essential.
|