Abstract:
As health care markets become more competitive, both consumers and health plans are increasingly demanding information on the quality of health care providers. Private organizations, provider groups, and state and federal agencies now produce a wide range of provider "report cards" that are used to guide choice among providers, as a contracting tool, and as part of quality improvement efforts. However, there has been considerable skepticism to date about the validity and accuracy of the quality measures being used in such report cards. Several barriers exist to the creation of good measures of quality for health care providers. These include: (1) the multidimensional nature of the quality of medical care, (2) high levels of noise due to small sample sizes for high-risk patient populations, and (3) bias across providers due to differences in patient mix.
In this paper, we apply a new method for measuring quality that overcomes the key limitations of available quality measures and apply it to estimating the quality of care for a high-risk patient population, infants with very low birth weights. Infants with very low birth weights (under 1500 g at birth) account for only 1% of births in the United States each year, but 46% of infant deaths. The typical neonatal intensive care unit (NICU) has only 80 patients per year, generating high levels of noise in provider performance. We use data from the Vermont Oxford Network, which contains a 10-year panel of abstracted medical records for very low birth weight infants, and includes an excellent set of risk adjusters. The Network contains data from 40% of all of the neonatal intensive care units in the United States and covers half of all very low birth weight births in the country each year.
Our approach to the methodological problems in quality measurement is based on multivariate signal extraction models. This approach optimally combines information from all available current and past quality indicators in order to more accurately estimate each provider's quality level. The "filtered" quality estimates that are generated extract the quality signal from the noise in the data. The resulting estimates are considerably smoother and more stable than unfiltered estimates. The method is able to predict and forecast differences in patient outcomes across hospitals remarkably well - far better than existing methods. For instance, it permits reliable identification of top performing units. The method also identifies which quality measures carry the quality signal. This implies that a minimum set of quality indicators could be developed to monitor the quality of neonatal intensive care units on an ongoing basis. The research further demonstrates that there are large differences in the quality of care across NICU providers that are persistent over time.
