Comparison of pharmacy-based and diagnosis-based comorbidity measures from medical administrative data.

PURPOSE: Health status is sometimes quantified by chronic condition (CC) scores calculated from medical administrative data. We sought to modify two pharmacy-based comorbidity measures and compare their performance in predicting hospitalization and/or death. The reference was a diagnosis-based score.
METHODS: One of the two measures applied an updated approach linking specific ATC codes of dispensed drugs to 22 CCs; the other used a list of 37 drug categories, without linking them to specific CCs. Using logistic regressions that took repeated measures into account and hospitalization and/or death the following year as the outcome, we assigned weights to each CC/drug category. Comorbidity scores were calculated as the weighted sum of the 22 CCs/37 drug categories. We compared the performance of both measures in predicting hospitalization and/or death with that of a diagnosis-based score based on 30 groups of long-term illnesses (LTIs), a status granted in France to exempt beneficiaries with chronic diseases from copayments. We assessed the predictive performance of the scores with the quasi-likelihood under the independence model criterion (QIC), the c statistic and the Brier score.
RESULTS: The two pharmacy-based scores performed better than the LTI score, with lower QIC and Brier scores and higher c statistics. Their predictive performance was very similar.
CONCLUSIONS: While there is no clear consensus or recommendations about the optimal choice of comorbidity measure, both pharmacy-based scores may be useful for limiting confounding in observational studies among general populations of adults from health insurance databases.