ICD-9-CM–based comorbidity indices (Charlson and Elixhauser). The Charlson Score includes 19 conditions. Each condition is assigned a weight from 1 to 6, which was originally derived from relative risk estimates of a proportional hazards regression model (5). The Charlson Score is the sum of the weights for all conditions. The Elixhauser method involves creating separate indicator variables for 30 different diagnoses (9). As an alternative to the original Elixhauser method, we also summed all of the indicators to create a total Elixhauser score. The RxRisk-V was calculated using pharmacy data obtained from local VA electronic medical records. This score involves the creation of 45 indicators for general drug categories (i.e., human immunodeficiency virus drugs, diabetic drugs). These indicator variables are summed to create the total RxRisk-V score. Within the VA health care system, medication copayments are low ($3 at the time of this study) and equal for all drugs. In addition, the VA covers the costs of some over-the-counter medications such as ibuprofen, acetaminophen, and aspirin. Therefore, VA patients have a high incentive to obtain their medications within the VA health care system, and the VA pharmacy database provides an excellent reflection of patients' medication use. We examined 3 health service use variables: number of physician visits, number of prescription drugs, and whether patients were hospitalized. We included all prescription drugs and visits to any type of physician, because we were interested in examining the ability of the comorbidity scores to predict overall health care use. All outcome variables were calculated using data from October 1999 to September 2000 (the fiscal year following creation of the comorbidity indices). Using the national VA outpatient clinic file, we calculated the total number of physician visits during the year. Local (Durham VA Medical Center) electronic pharmacy records were used to calculate the total number of prescriptions filled during the year. The national VA inpatient treatment file was used to examine hospitalizations. Because very few participants had more than 1 hospitalization during the year, we simply examined whether patients were hospitalized (rather than the number of hospitalizations). In the first step of the analysis, the proportion of participants with each drug class/disease indicator was examined for each of the comorbidity indices. Correlations between comorbidity indices were also calculated. In the next step, 4 separate models, each using a different comorbidity index (Charlson, Elixhauser total summed score, Elixhauser with individual disease indicators, and RxRisk-V) as a predictor variable, were fit to each of the 3 health service outcome variables. Negative binomial regression models were used for models of the number of physician visits and number of prescriptions. Initially, we applied Poisson regression models, but these models did not fit the data well. For the binary hospitalization outcome, logistic regression models were used. Model selection was used to identify the best comorbidity index to use for predicting each of the health service outcomes. A model selection criterion, the Akaike Information Criterion (AIC) (16), was applied to the 4 models for each outcome, and the model with the lowest AIC was selected as the best model. AIC is computed from the log likelihood of a model and the number of parameters in a model. In general, the rule of thumb for identifying a model or set of models as “better” is a difference