Classifying hospitals on process performance measures using flexible random-effects models
*Yulei He, Harvard Medical School 

Keywords: hospital profiling, Medicare, nonnormality, random effects, pay for performance, semiparametric Bayesian

Process measures, binary indicators of what was done to or for a patient, are collected to quantify hospital compliance with evidence-based guidelines. Often, hospitals are subsequently tiered into categories based on compliance estimates and possibly financially rewarded. A common approach uses Bayesian shrinkage estimators that accommodate between-hospital variation in the true rates, and assume normal distribution for the random (hospital) effects. The classification accuracies of this approach, however, are sensitive to the actual distributions of the random effects. We consider a class of flexible random-effects models using Dirichlet process priors. To improve the classification accuracy, we adopt the triple-goal estimates to overcome the underdispersion of the posterior mean estimates in estimating the empirical distribution function of the hospital performance. We use Monte Carlo simulation studies to assess the performance of the proposed approach. We present an illustrative example using process measures data for patients treated for a heart attack, heart failure, and pneumonia from Medicare's national Hospital Compare database.