Appropriate Statistical Treatment of Frequent Emergency Department Visits in Health Services Research

Abstract

I read with interest the article by Hunt et al1Hunt K.A. Weber E.J. Showstack J.A. et al.Characteristics of frequent users of emergency departments.Ann Emerg Med. 2006; 48: 1-8Abstract Full Text Full Text PDF PubMed Scopus (362) Google Scholar on the characteristics of frequent users of emergency departments (EDs). In my opinion, there might be 2 potential caveats in their data analysis regarding statistical treatment of ED visits. On multivariate analysis, the authors applied the logistic regression model using 4 ED visits as a cutoff point, and the cutoff point was determined based on the distribution of ED visits. First, logistic regression is most powerful when dealing with dichotomous outcomes; therefore transforming count variable to dichotomous variable using an arbitrary cutoff may not be appropriate. This has been addressed by the authors in the limitation section. Second, their analysis did not take into account intraindividual dependence, ie, ED visits are more likely to recur in some individuals than others. If logistic regression is chosen, then appropriate statistical methods accounting for intraindividual dependence should accompany, such as the General Estimating Equations method.2Zeger S.L. Liang K.Y. Longitudinal data analysis for discrete and continuous outcomes.Biometrics. 1986; 42: 121-130Crossref PubMed Scopus (6183) Google Scholar In view of these 2 potential limitations, I felt that negative binomial regression model might be a better choice for identifying the characteristics of frequent ED users in this particular analysis. The negative binomial regression model is a variant of the Poisson regression model, and in both models, there is no need to specify an arbitrary cutoff for count data.3Cook R.J. Lawless J.F. Analysis of repeated events.Stat Methods Med Res. 2002; 11: 141-166Crossref PubMed Scopus (100) Google Scholar In addition, with a greater flexibility of distribution assumptions, negative binomial regression model has been shown to substantially better fit to this type of count data (eg, numbers of ED visits) than Poisson model by accounting for intraindividual dependence and resulting overdispersion.4Glynn R.J. Buring J.E. Ways of measuring rates of recurrent events.BMJ. 1996; 312: 364-367Crossref PubMed Scopus (157) Google Scholar, 5Sturmer T. Glynn R.J. Kliebsch U. et al.Analytic strategies for recurrent events in epidemiologic studies: background and application to hospitalization risk in the elderly.J Clin Epidemiol. 2000; 53: 57-64Abstract Full Text Full Text PDF PubMed Scopus (44) Google Scholar Moreover, it can accommodate excess zeros (individuals with no ED visits) in the data analysis. Those patients who had no ED visits did provide significant statistical information as a baseline group in data analysis but somehow were not included in this article. It is my suspicion that the results might be substantially changed if the appropriate statistical method is applied. Due to its statistical complexity, appropriate analysis for recurrent events has been underused in health services research; however, it should no longer be neglected by researchers. In replyAnnals of Emergency MedicineVol. 49Issue 3PreviewWhile we appreciate the possibility that negative binomial regression mode may have provided additional information, we disagree with the unsubstantiated assertion that our “results might be substantially changed” if negative binomial regression mode analysis was used. A variety of statistical techniques were applied to the data, all of which produced consistent results. It is very unlikely that additional multivariate manipulations would affect our most important results: 8% of frequent users with 4 or more visits were responsible for 28% of adult emergency department (ED) visits and the vast majority of frequent ED users had health insurance and a usual source of care (84% and 81%, respectively). Full-Text PDF