|
Abstract:
|
In public reporting and payment programs, hospital performance is routinely assessed on patient outcome measures, such as mortality and readmission rates, which reflect the underlying quality of the hospital. To make fair and credible comparisons across hospitals, risk adjustment is used to remove variation in patient characteristics, such as comorbid conditions, that could otherwise explain differences in outcomes. Techniques include regression and matching—for instance, template matching creates a patient panel that is virtually identical across hospitals so that their outcomes can be compared. Regression can be aided by machine learning methods, such as targeted learning, which yields double robust inferences; in our study, we use this property to account for differences in patient characteristics across hospitals and the relationship between clinical risk factors (e.g., comorbidity) and the patient’s outcome. We apply template matching and targeted maximum likelihood estimation to assess their methodological performance when used to compare hospital mortality and readmission rates in the Nationwide Readmission Database from the Healthcare Cost and Utilization Project.
|
