We agree with Drs. Baron and Barrett that including procedure codes or linking claims records from multiple encounters may enhance the sensitivity of claims data for clinical conditions. The "much better results" they mention, however, have yet to be shown. Recently, Roos and colleagues [1] compared claims data from a single hospitalization with data collected from all medical encounters during the previous 6 months in Manitoba for 677 patients having prostatectomy. Despite the universal nature of the Canadian claims system, the additional data in the "claims history" did not significantly improve its sensitivity for clinically identified conditions, increasing sensitivity from 0.43 to 0.51 for cardiovascular disease and from 0.32 to 0.37 for respiratory disease. In our study, we chose to examine the accuracy of single-encounter data, the form used in most research in this area. We also agree with Drs. Baron and Barrett that claims data are suitable for certain types of epidemiologic studies.

We agree with Dr. Huff's concern about generalizability of the findings at Duke to other settings, and we specifically noted this limitation in our paper. During our study, we tried to find data that compared medical record coding at Duke with that at other institutions. The current minimal quality control of claims data and the lack of available information about the result of such quality control from different institutions raise further concerns about the use of hospital claims data to specify illness severity. We do not have reason to believe that medical record coding at Duke differs substantially from that at other institutions. The hospital follows standard guidelines and procedures for medical record coding [2]. In our study, we used hospital claims data containing up to 30 diagnosis codes rather than the standard Medicare data set that was truncated at 5 diagnoses. Most of the episodes of "intensified review" at Duke resulted from errors in billing procedures rather than in coding, such as claims for dental extractions, a procedure not reimbursed by the North Carolina Medicare intermediary. We also agree that patients having cardiac catheterization may have fewer comorbid illnesses due to selection bias. If claims data have a low specificity for comorbidity in these patients, however, we are still concerned about their use in patients with potentially more comorbid illnesses. Regarding Dr. Huff's concern about the "claimed riches" from the Duke data, we refer him to the article by Hlatky and colleagues [3], in which the findings from the major randomized trials of coronary bypass surgery were duplicated in the Duke data.

The article by Drs. Green and Wintfeld extends our work, showing that coding errors in California claims may result in misclassification of more than 50% of hospitals identified as "outliers" [4]. If these investigators had access to concurrently collected clinical data rather than only to repeat chart abstraction, we suspect that the magnitude of the hospital mortality estimate bias would have been even greater. Dr. Prochazka's observations also support the need for improved clinical data systems specifically designed for treatment comparisons.

Finally, we feel that the burden of proof concerning the accuracy of claims data rests with those who use such data to test hypotheses and to draw conclusions. Our study simply points out the results of one such examination relative to our database.