Health care quality surveys in the United States are administered to individual respondents (hospital patients, health plan members) to evaluate performance of health care units (hospitals, health plans). For better understanding and more parsimonious reporting of dimensions of quality, we analyze relationships among quality measures at the unit level. Rather than specifying a full parametric model for the observed responses and the non-response patterns at the lower (patient) level, we first fit generalized variance-covariance functions that take into account nonresponse patterns in the survey responses. We then specify a likelihood function for the unit mean responses using these generalized variance-covariance functions, letting us model directly the quantities we want to report. Because the response scales are bounded we assume that the unit means follow a truncated multivariate normal distribution. We calculate maximum likelihood estimates using the EM algorithm or drawing directly from Bayesian models using Markov-chain Monte Carlo. Finally factor analysis is performed on the between-unit covariance matrices obtained from the fitted models. Using posterior draws we assess posterior distributions of the number of selected factors and the assignment of items to groups under conventional rules. We compare maximum likelihood estimates of this factor structure to those from several Bayesian models with different prior distributions for the between-unit covariance. Results are presented using data from the Consumer Assessment of Healthcare Providers and Systems (CAHPS R) survey of Medicare Advantage health plans. (January 2008)
Journal of the American Statistical Association
2008
http://www.stat.columbia.edu/~gelman/stuff_for_blog/omalley.pdf