Objective.  Numerous researchers have called for better controls for hospital quality and patient burden of illness in inefficiency measurement generally and in frontier techniques specifically.    This research assesses the impact of employing a variety of controls for hospital quality and patient burden of illness on the mean estimated inefficiency and relative ranking of hospitals generated by stochastic frontier analysis (SFA). Such a robustness check is useful because estimates of hospital inefficiency are likely to have more perceived practical value if their robustness is demonstrated.
Data and Methods. Hospital data for 1,290 hospitals operating in 20 states in the United States in 2001 were taken from the American Hospital Association (AHA) Annual Survey and the Medicare Cost Reports.  A variety of controls for hospital quality and patient burden of illness were employed.  Among the variables used were a subset of the quality indicators generated from the application of the Patient Safety Indicator (PSI) and Inpatient Quality Indicator (IQI) modules of the Agency for Healthcare Research and Quality (AHRQ) Quality Indicator (QI) software to the Healthcare Cost and Utilization Project (HCUP) State Inpatient Databases (SID).  Measures of a component of patient burden of illness came from the application of the Comorbidity Software to HCUP data.
Hospital inefficiency estimates using SFA were generated using a maximum likelihood estimation procedure in Frontier 4.1.  On the basis of likelihood ratio tests, we selected a translog cost function with a composed error having a truncated normal distribution as the preferred model. The cost function included the capital costs per bed and average wages as input price variables and admissions, outpatient visits and patient days in non-acute care units as output variables. Output heterogeneity was controlled by the inclusion of cost function variables for inpatient and outpatient case-mix, teaching activities, and presence of high-level technological services.  We estimated various models with (and without) different quality and patient burden of illness measures.
Principal Findings.  Estimated cost inefficiency for the various models ranged from 13.9% to 17.3%.  Pearson correlation coefficients ranged from 0.71 to 0.99.  The most dramatic changes to mean estimated inefficiency and correlations between the models came with the addition of comorbidity variables. 
Conclusions. Choices about controlling for quality and patient burden of illness can have a non-trivial impact on mean estimated hospital inefficiency and the relative ranking of hospitals generated by SFA.  Indeed, the measures produced by the Comorbidity Software appear to account for variations in patient burden of illness that had previously been masquerading as inefficiency.  Although we found that outcome measures of quality can provide insight into a hospital’s operations, our estimates suggested that they have little impact on estimated inefficiency once controls for structural quality and patient burden of illness have been employed.  We also found that outcome measures of quality that take the multi-dimensional nature of hospital quality into account yield more insight into hospital performance than a measure that combines the often conflicting effects of hospital quality into a single variable.