We find that controls for quality and patient burden of illness can have a nontrivial impact on the inferences derived from SFA about hospital performance.
The AHRQ QIs have the advantage of taking the multidimensional nature of hospital quality into account. As the coefficients on the AHRQ QIs show, measures of hospital quality can have conflicting effects on hospital costs. A single measure that combines these effects into one variable offers less insight into hospital performance. Indeed, the perspective it provides may be misleading because the measure is taking so many factors into account.
However, the AHRQ QIs, as well as some of the other controls employed in this analysis, may not be accessible to all researchers. Our results indicate that the impact on both mean estimated hospital inefficiency and the ranking of hospitals of accounting for outcome measures of quality is small once controls for patient burden of illness and teaching status have been employed. Therefore, researchers estimating average institutional inefficiency for aggregate groups of hospitals can be less concerned about how they control for outcome measures of quality than researchers attempting to use hospital cost functions to understand the trade-offs between cost and quality.
The negative coefficients on some of the comorbidity variables were unexpected. However, we are inclined not to attach too much importance to the sign of the coefficients. The measures generated by the Comorbidity Software reflect secondary diagnoses that are unrelated to the principal reason for admission. So, a person admitted for a broken leg (who is also depressed) will probably be much less costly to treat than a person admitted for a stroke (who is also depressed). Thus, while we may not put too much trust on the signs of the coefficients, the variables as a group might be capturing some previously unexplained patient burden of illness that had been masquerading as inefficiency.
Our results suggest that users of SFA who want to measure overall hospital inefficiency or to investigate inefficiency differences among groups of hospitals may want to use the Comorbidity Software variables. If they do not use these measures, they might want to view the overall inefficiency measure they generate as an upper bound. They also might want to be aware that differential differences in patient burden of illness across hospitals could be influencing their results. Analysts seeking to investigate the cost impacts of control variables might want to be more cautious about using the comorbidity variables because of multicollinearity concerns.
Our estimated mean cost-inefficiencies ranged from 14.8 to 17.3 percent. Before our study, Zuckerman, Hadley, and Iezzoni (1994)
used the most extensive set of outcome and product descriptor variables in hospital SFA. Their estimates dropped from 18.8 percent in a basic model to 13.6 percent in the model with the most control variables, a range very similar ours. The estimates in both studies fall in the middle of estimates obtained from national SFA studies of hospitals that ranged from 10.8 to 25.5 percent. The largest estimated mean inefficiency was obtained from a study (Rosko 1999
) that did not include mortality rates and used relatively few product descriptor variables. Our analysis suggests that this is the type of study where we would expect to have output heterogeneity masquerading as inefficiency, thereby inflating inefficiency estimates. This reinforces our contention that it is important to include comorbidities and other product descriptors in the cost function. Our analysis of the literature suggests that SFA estimates over time and across methods have been quite stable, especially among the more completely specified models.
We are unaware of other SFA applications have attempted to control for the multiple concepts that comprise patient burden of illness, especially important comorbidities that are unrelated to the principal diagnosis. Thus, our findings indicate that SFA is a powerful analytic technique that is still in its developmental stage. This, of course, does not preclude it from being used to analyze categories of hospitals (e.g., for-profit and non-for-profit institutions). Even researchers who contend that SFA generates estimates that are insufficiently robust to be used to reward individual hospital performance believe that it can be used to assess the efficiency of categories of hospitals (e.g., Folland and Hofler 2001
Thus, the results of this study combined with those from Rosko and Mutter (2008)
who found that cost-inefficiency estimates are not very sensitive to assumptions about the composed error, quality measures, and structure of production suggest that SFA can be a very useful tool for analyzing the cost-inefficiency of different groups of hospitals when a fully specified model is used.