Session number: 1039
Abstract title: How Well Do Standard Risk Adjustment Models Predict Utilization for Patients with a Major Functional
Author(s):
G Warner - Center for Health Quality, Outcomes and Economic Research
ARosen - Center for Health Quality, Outcomes and Economic Research
MMontez - Center for Health Quality, Outcomes and Economic Research
CRakovski - Center for Health Quality, Outcomes and Economic Research
HHoenig - Durham VAMC
Objectives: Patients with a major functional limitation may have different utilization patterns then the overall population. This analysis evaluated how well diagnostic-based risk adjustment models predicted resource use and then bolstered models with additional information in a sample of patients diagnosed with spinal cord dysfunction (SCD).
Methods: A 40% random sample of veterans who used inpatient and outpatient services in FY’97 was obtained from VA databases (N=1,046,803). The number of service days was used to measure resource utilization. Previously validated ICD-9 codes for traumatic spinal cord injury, multiple sclerosis, quadriplegia, paraplegia and other pathologies leading to possible paraplegia/quadriplegia were used to identify a subgroup of patients with SCD (N=7,761). Weighted least squares regression models compared the performance of two risk adjustment systems (Adjusted Clinical Groups (ACG) and Diagnostic Cost Groups (DCG)). Additional markers (e.g., SCD subgroup, high risk comorbidities, and comorbidity-subgroup interaction terms) were added to basic risk adjustment models to improve model predictability. Measures of fit were calculated. These included R-squared and root mean square errors (RMSE).
Results: The DCG model explained more total variance and had less error for both the overall sample and for the SCD subgroup (R-squared =.315, RMSE=32.32 and R-squared =.221, RMSE= 57.8 respectively) than the ACG model (R-squared=.232, RMSE=34.12 and R-squared=.112, RMSE= 65.3). Additional markers minimally improved predictive ability and decreased mean errors for the overall sample (DCG model: R-squared=.317 and RMSE= 32.18) or the SCD subgroup (R-squared=.243, RMSE= 60.3).
Conclusions: For people with SCD, DCG models predicted total resource utilization better than the ACG model. Adding other markers only minimally improved the model. Given available inpatient/outpatient administrative information, the DCG model appears to be a better risk adjustment system than ACG for this group.
Impact statement: The VA provides services for a large proportion of people with SCD, and has identified this subgroup as an important subpopulation to monitor because of their high service utilization. To ensure that adequate resources are allocated for subgroups such as these it is important to maximize the predictive accuracy of risk-adjustment methods. To further improve prediction models, additional clinical and functional status information may useful.