Selection of Related Multivariate Means: Monitoring Psychiatric Care in the Department of Veterans Affairs

Quantitative performance monitoring of health care providers has become an increasingly prominent activity in health care delivery. As a consequence, multidimensional "report cards" comparing the quality of services provided by diverse health care organizations have begun to appear. We present a case study examining the performance of 22 geographically defined health care systems, known as Veterans Integrated Service Networks; involving the treatment of more than 37,000 veterans discharged from general psychiatry program beds in fiscal year 1998. We describe multilevel multidimensional latent variable models and fit these to a set of related mixed discrete and continuous outcomes to identify networks providing inadequate care to its patients. We use validation sample to assess model fit by comparing observed statistics to the distribution of the statistics across replicated datasets. We propose using posterior tail probabilities associated with functions of the latent variables to select networks with poor outcomes, and finally compare our results with those from current methods.