Our study also has a number of limitations. First, we used claims data from a group of hospitals that voluntarily submit data to Premier Healthcare Informatics for the purposes of quality improvement. Claims data are subject to biases that might occur as a result of variation in documenting and coding across hospitals, and findings in hospitals that participate in Perspective may not be generalizable to hospitals nationwide. Second, we used ICD-9-CM codes, blood cultures, and treatment with antibiotics to identify patients with sepsis. While this is likely to have increased the likelihood that the patients included in our study actually had sepsis, we recognize that some patients with sepsis may not routinely undergo blood culturing. Third, our mortality and costs models used treatments provided to patients early during the admission, rather than direct physiologic measurements, as indicators of patient severity. Nevertheless, these variables, including the use of vasopressors and mechanical ventilation, were strongly associated with mortality, and other mortality prediction models rely on a similar framework.35 It is possible that hospitals with higher spending achieve better outcomes that were not measured in this study, such as improved quality of life, functional status, or higher patient satisfaction. However, previous work on regional-level costs and outcomes has not found a relationship between higher costs and patient preferences or patient satisfaction.8 We examined in-hospital rather than 90-day mortality, and hospitals with differing rates of discharge to hospice or nursing care could have affected our results in unpredictable ways. However, the percentage of patients discharged to hospice was exceedingly small, and previous work has considered in-hospital mortality to be an effective aggregate measure of structure and process for high-mortality conditions.36- 38 We excluded transferred patients, and although we adjusted for rates of transfer in our model, our findings may not be generalizable to patients who underwent transfer during their hospitalization. Although we had information on actual costs for 75% of hospitals, we used cost to charge ratios for 25% of the hospitals. Finally, we were not able to account for variability of care within hospitals. Hospitals that use a greater number of invasive interventions for all patients may harm lower-acuity patients while simultaneously benefiting higher-acuity patients, blurring the beneficial impact of higher spending for higher-acuity patients. However, this scenario is unlikely in a real-world setting because the most common invasive interventions for severe sepsis (ie, mechanical ventilation, vasopressors) are generally not administered to patients who are not in shock and respiratory failure because the risk associated with these interventions is very high.