A Semiparametric Analysis of Health Care Utilization in Patients with Heart Failure
*Zhuokai Li, Indiana University School of Medicine
Keywords: bivariate splines, exponential family, health care utilization, multivariate mixed model
Hospital admissions and emergency department (ED) visits are important markers of disease exacerbation in patients with heart failure. They are also significant contributors of health care cost. In health service research, it is often of interest to identify factors associated with increased utilization of inpatient and ED care. Traditionally, linear models have been used to identify correlates of increased utilization, and inpatient and ED visits have been analyzed separately, thus ignoring the synergistic relationship between these different forms of utilization. In this research, we propose a unified semiparametric analytical framework that treats different forms of utilization as multivariate outcomes and accommodates the concurrent nonlinear influences of important explanatory variables. The model is developed for data of different types under the exponential family of distributions. The simultaneous modeling of multiple outcomes permits comparison of independent variable effects across the outcomes. Within-subject correlations across the multivariate outcomes and among the repeated measurements are modeled through shared random effects. We use thin plate regression splines to capture the potentially nonlinear joint effects of two independent variables. Penalized likelihood method is used for model estimation and a data visualization tool is developed for the depiction of the estimated bivariate effects. We then conducted a simulation study to evaluate model performance. Using the proposed model, we analyzed health care utilization data from 694 patients with a diagnosis of heart failure, as recorded in an electronic medical record system. In this heart failure cohort, we studied the joint effects of B-type natriuretic peptide (BNP) and ejection fraction (EF), two most frequently used bedside assessments for heart functioning, on the annual counts of hospital admissions and ED visits. The estimated bivariate effect surfaces were visualized through contour plots. We found that both BNP and EF were associated with the number of hospital admissions, while BNP had a more dominant positive effect on the number of ED visits.
Important Dates & Deadlines
- October 9 - 11, 2013