Using Medicare claims data to assess provider quality for CABG surgery: does it work well enough?

Abstract

Objectives: To assess the relative abilities of clinical and administrative data to predict mortality and to assess hospital quality of care for CABG surgery patients.

Data sources/study setting: 1991-1992 data from New York's Cardiac Surgery Reporting System (clinical data) and HCFA's MEDPAR (administrative data). STUDY DESIGN/SETTING/SAMPLE: This is an observational study that identifies significant risk factors for in-hospital mortality and that risk-adjusts hospital mortality rates using these variables. Setting was all 31 hospitals in New York State in which CABG surgery was performed in 1991-1992. A total of 13,577 patients undergoing isolated CABG surgery who could be matched in the two databases made up the sample.

Main outcome measures: Hospital risk-adjusted mortality rates, identification of "outlier" hospitals, and discrimination and calibration of statistical models were the main outcome measures.

Principal findings: Part of the discriminatory power of administrative statistical models resulted from the miscoding of postoperative complications as comorbidities. Removal of these complications led to deterioration in the model's C index (from C = .78 to C = .71 and C = .73). Also, provider performance assessments changed considerably when complications of care were distinguished from comorbidities. The addition of a couple of clinical data elements considerably improved the fit of administrative models. Further, a clinical model based on Medicare CABG patients yielded only three outliers, whereas eight were identified using a clinical model for all CABG patients.

Conclusions: If administrative databases are used in outcomes research, (1) efforts to distinguish complications of care from comorbidities should be undertaken, (2) much more accurate assessments may be obtained by appending a limited number of clinical data elements to administrative data before assessing outcomes, and (3) Medicare data may be misleading because they do not reflect outcomes for all patients.