Background: Comorbidity scores are increasingly used to reduce potential confounding in epidemiological research. Our objective was to compare metrical and practical properties of published comorbidity scores for use in epidemiological research with administrative databases.
Methods: The literature was searched for studies of the validity of comorbidity scores as predictors of mortality and health service use, as measured by change in the area under the receiver operating characteristic (ROC) curve for dichotomous outcomes, and change in R(2) for continuous outcomes.
Results: Six scores were identified, including four versions of the Charlson Index (CI) which use either the three-digit International Classification of Diseases, Ninth Revision (ICD-9) or the full ICD-9-CM (clinical modification) code, and two versions of the Chronic Disease Score (CDS) which used outpatient pharmacy records. Depending on the population and exposure under study, predictive validities varied between c = 0.64 and c = 0.77 for in-hospital or 30-day mortality. This is only a slight improvement over age adjustment. In one study the simple measure 'number of diagnoses' outperformed the CI (c = 0.73 versus c = 0.65). Proprietary scores like Ambulatory Diagnosis Groups and Patient Management Categories do not necessarily perform better in predicting mortality. Comorbidity indices are susceptible to a variety of coding errors.
Conclusions: Comorbidity scores, particularly the CDS or D'Hoore's CI based on three-digit ICD-9 codes, may be useful in exploratory data analysis. However, residual confounding by comorbidity is inevitable, given how these scores are derived. How much residual confounding usually remains is something that future studies of comorbidity scores should examine. In any given study, better control for confounding can be achieved by deriving study-specific weights, to aggregate comorbidities into groups with similar relative risks of the outcomes of interest.