Online Program

Enhancing the Hospital Compare Mortality Model Using a Bayesian Framework

*Jeffrey H. Silber, University of Pennsylvania/Children's Hospital of Philadelphia 

Keywords: Hospital Compare, Random Effects, Bayesian Model

Controversy surrounds Medicare's Hospital Compare (HC) mortality model, in part because of the absence of hospital characteristics in the random effects model used for public reporting. The HC model assumes hospital quality is statistically independent of hospital attributes, such as volume and nurse staffing, with the consequence that small hospitals are almost invariably reported as “no different from the national average.” In this report, we construct a Bayesian model that mimics the HC model, and then enhance it to depart from the initial HC assumptions. Nesting the HC model as a special case, our enhancement allows the data to determine whether including hospital characteristics in the model is preferable to the HC assumption of randomly drawn hospital effects from the same population (constant mean and variance). The report also compares direct and indirect standardization using this Bayesian framework. To make these concepts more transparent, we examine posterior distributions of hospital mortality both from the HC model and from a proposed enhanced model. Using hospitals in Chicago, hospital comparisons under the HC model are contrasted with comparisons under the new model.