Session number: 1030
Abstract title: How Well Does the CDPS Predict Utilization: a Comparison to Other Case-Mix Measures
Author(s):
M Montez - Bedford VA, HSR&D Field Program
C Rakovski - Bedford VA, HSR&D Field Program
S Loveland - Bedford VA, HSR&D Field Program
C Christiansen - Bedford VA, HSR&D Field Program
D Berlowitz - Bedford VA, HSR&D Field Program
A Rosen - Bedford VA, HSR&D Field Program
Objectives: While Adjusted Clinical Groups (ACGs) and Diagnostic Cost Groups (DCGs) are frequently used as case-mix measures to predict utilization, few studies have evaluated the Chronic Disability Payment System (CDPS) in the VA. Because the CDPS was developed on the Medicaid population, a group with chronic conditions, we hypothesized that it would perform especially well in predicting utilization among patients with spinal cord disorders, an important VA subgroup.
Methods: We examined a 60% random sample of all veterans who used VA healthcare in FY’97. We developed models on 40% of the patients (n=1,046,803) and validated them on 20% (n=524,461). The outcome variable was days of care: total inpatient and outpatient days of utilization in FY’97 (concurrent) and FY’98 (prospective). We ran 3 weighted least squares regression models that had age and gender plus: 1) ACG/ADG model: 32 Adjusted Diagnostic Groups (ADGs); 2) DCG/HCC model: 116 Hierarchical Condition Categories (HCCs); and 3) CDPS model: 56 CDPS subcategories. We compared R-squares across models. To evaluate the fit for a subgroup with chronic conditions, we calculated predictive ratios (PRs) for patients with spinal cord disorders. PRs close to 1.0 indicate that expected utilization for a subgroup is close to its actual utilization.
Results: For the overall population, the DCG/HCC model had higher predictive ability than the ACG/ADG or CDPS models. Concurrent R-squares were 0.315, 0.232, and 0.256; prospective R-squares were 0.175, 0.137, and 0.154, respectively. Validated R-squares were similar to development R-squares, indicating model stability. Among the patients with spinal cord disorders, the CDPS model had consistently better PRs compared to the ACG/ADG and DCG/HCC models (concurrent PRs: 0.96, 0.72, and 0.93; prospective PRs: 0.94, 0.68, and 0.90, respectively).
Conclusions: Although all three case-mix measures performed comparably in predicting overall utilization in the VA, as hypothesized, the CDPS did better in predicting utilization for a subgroup with chronic conditions.
Impact statement: These results suggest that selecting a case-mix measure depends upon the purpose for which it will be applied. The VA needs to develop a set of criteria that can be used for selecting a case-mix measure that is appropriate for selected subgroups and purposes.