Predicting resource utilization in a Veterans Health Administration primary care population: comparison of methods based on diagnoses and medications.

BACKGROUND: Valid methods of predicting resource utilization in primary care populations are needed. We compared the predictive validity of a method based on diagnoses from administrative data (Adjusted Clinical Groups [ACGs]) and a method using medication profiles (Chronic Disease Index [CDI]).
METHODS: This retrospective cohort study included 31,212 primary care patients in a Veterans Health Administration (VA) network who received outpatient medication prescriptions in 1999 and who had VA utilization in 1999 and 2000. ACG and CDI classifications were determined using 1999 data. Analyses compared the predictive validity with respect to outpatient clinic visits and days of hospital care.
RESULTS: Both ACGs and CDI explained a higher proportion of the variance in outpatient visits than demographic data alone. However, explained variance was higher for ACGs. For example, ACGs explained 30.2% of the variance in total visits in 1999, compared with 8.8% for the CDI. Results were similar for 2000, although the explained variance declined for both methods (eg, 16.3% and 5.7%, respectively, for total visits). Results were similar in analyses examining the discrimination of the 2 methods to predict hospital use; for example, c statistics for ACGs and CDI scores were 0.86 versus 0.70, respectively (P <0.05), for 1999 and 0.72 and 0.65, respectively (P <0.05), for 2000.
CONCLUSION: Among VA patients, ACGs had superior predictive validity than the CDI, a newer nonproprietary method based on pharmacy data. The findings suggest that diagnosis-based measures could be preferable for ambulatory case-mix adjustment and are valid across a wide range of populations.