When comparing outcomes of medical care, it is essential to adjust for patient risk, including severity of illness. A variety of severity measures exist, but perceptions of outcomes may differ depending on how severity is defined. We used two severity-adjustment approaches to demonstrate that comparisons of outcomes across subgroups of patients can vary dramatically depending on how severity is assessed. We studied two approaches: model 1 was the admission MedisGroups score; model 2 was computed from age and 12 chronic conditions defined by diagnosis codes. Although common summary measures of model performance (R-squared and C) both suggested that model 1 is a better predictor of in-hospital death than model 2, the weaker model consistently produced more accurate expectations by payer class and age group. Using model 1 for severity adjustment suggested that Medicare patients did substantially worse than expected and Medicaid patients substantially better. In contrast, use of model 2 found Medicare patients doing as expected, but Medicaid patients faring poorly.