Evaluation of readmision to the hospital using regression methods

As in many fields, regression analysis is frequently used in health sciences. Different models are used depending on the structure of the dependent variable examined. Poisson and negative binomial regression are the most commonly used regression models when the dependent variable is discrete. Poisson and negative binomial distributions are used to model the probability of counts. Poisson and negative binomial regression require the dependent variable to be discrete and countable but there is no restriction for the independent variable. In some cases, the dependent variable has so many zero counts that it disrupts the fit to the poisson and/or negative binomial distribution. In such cases, zero-inflated regression models may fit better. Re hospitalization of patients who have been discharged from the hospital within a certain period is defined as readmission (re-admission). It is important to evaluate the factors affecting the number of hospitalizations of patients in order to reduce health expenditures, improve the quality of patient care, increase the efficiency of health services and quality in health. The aim of this thesis study is to evaluate the regression methods and regression models to select the most suitable one for the patients' readmission status data obtained from the hospital database and identify the factors affecting the readmission. The evaluation of the Vuong test applied to AIC, BIC and LL obtained from the backward model selection to poisson, negative binomial and their zero-inflated regression models, result in favor of zero-inflated negative binomial model. For which the constant, Age, Pathology, Blood product and Surgery group were found to be significant in the Count Component whereas Duration of Hospitalization (days), Pathology and Blood product variables were found to be significant in the Zero Component. Keywords: Regression analysis, Readmission, Poisson Regression, Negative Binomial Regression, Zero inflated model