Two Ways of Modeling Hospital Readmissions: Mixed and Marginal Models

• Activity Number: 251
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Hui Fen Tan*+ and Ronald Low and Shunsuke Ito and Raymond Gregory and Van Dunn
• Companies: and New York City Health and Hospitals Corporation and New York City Health and Hospitals Corporation and New York City Health and Hospitals Corporation and Metroplus Health Plan
• Keywords: hospital readmissions ; longitudinal data ; marginal model ; generalized estimating equations ; mixed model ; random effects

Abstract

The Center for Medicare and Medicaid Services is progressively reducing reimbursement for some hospital readmissions. Understanding factors associated with readmission is increasingly important. Using 13 years of admissions data from New York City's public hospitals, we develop models to predict if a pneumonia patient will be readmitted to a hospital within 30 days of discharge, using covariates such as hospital conditions, patient medical history and demographics, weather, and pollution levels. As a patient could return to the hospital several hundred times over 13 years or never return at all, we use two types of models to account for correlation between observations: a marginal model estimated using generalized estimating equations, and a mixed model with random effects for patient and hospital. The latter model has a higher prediction accuracy of 89.47%. Gender, insurance status, and past medical history were significant predictors. Having a history of serious medical diseases increases readmission risk greatly. However, the two models ask different questions, with the former model being perhaps more relevant for hospital administrators who wish to know the effect of covariates on the population average instead of individual patients. We illustrate the similarities and differences between the empirical results of both models.