| Hospital reporting1 is an important, complex and yet controversial issue in terms of evaluating and monitoring health care performance. A typical hospital report card presents hospital-specific summaries for certain conditions and procedures, typically focusing on outcomes such as mortality, infection rate, and length of stay. The statistical models generating these summary measures often called "risk models" which adjust for pre-existing risk factors pertinent to patients. However, hypothetical and real examples show that hospital performance may vary among patient types. Hence, the goal of this dissertation is to build a risk model that enables an individualized hospital report card based on accurate patient-level prediction. An individual patient can thus choose hospitals based on the accurate predicted hospital-specific probabilities.; We propose the use of hierarchical Bayes models that incorporate the multilevel structure of hospital data to facilitate individualized hospital report cards. We explore the idea that even with basic UB-92 data, patient-level predictive models can take advantage of a very large set of predictors, thereby improving predictive performance. The reasons for choosing hierarchical Bayes multilevel models are mainly two fold. First, multilevel modeling accounts for the important aspect of multilevel structure of hospital data. Second, Bayesian methods treat all parameters as random quantities with probability distributions thereby simplifying interpretation.; Results from real data showed that the proposed methods indeed increased the predictive performance in discrimination (measured by c-index) compared to their simple model counterpart. We also propose two approaches to improve the predictive performance of the risk models. The first is by incorporating informative priors into models. The results of a real example showed that incorporating informative priors could increase the predictive performance of the models. The second approach is to implement the relaxed Lasso for variable selection. Although the relaxed Lasso and Lasso have the same predictive performance, the relaxed Lasso produced a sparser model than the Lasso. At last, we demonstrated the practicality of the proposed methods by building an interactive webpage that implemented the proposed method for individualized hospital report cards.; 1Hospital reporting is also referred as "hospital comparison" or "hospital profiling". All terms are used interchangeably in this dissertation. |
