Risk Adjustment to Predict Healthcare Cost and UtilizationObjective: The Military Health System (MHS) is examining new approaches to financing the care of covered beneficiaries, including prospective payment. Since population health care costs depend on illness burden (case-mix), efficient financing should recognize variations in the illness burden of populations seen by different providers. Several mature risk adjustment systems that extract illness burden profiles from computerized encounter records are now used by Medicare, Medicaid, the Veteran’s Health Administration, commercial insurers in the US and international stakeholders to understand and manage health care delivery systems. We conducted side-by-side testing of a simple age/sex method and four claims-based risk-adjustment models: Adjusted Clinical Groups (ACGs), Chronic Disease and Disability Payment System (CDPS), Clinical Risk Groups (CRGs), and Diagnostic Cost Groups (DCGs). All risk models use a person’s age, sex and the morbidities recorded (in ICD-9-CM codes) during a year to predict total costs (inpatient + outpatient + pharmacy) the next year. In addition, the CRG model used limited information on dates of service, place of service and procedures. We applied each model to the MHS, measuring its overall ability to predict future costs and the concordance between model-measured needs and actual health care spending in policy-relevant subgroups of TRICARE Prime enrollees.
Results: The four risk adjustment models all performed well. It is difficult to say which model performed best, as there is no one perfect test for model fit. All four models properly “priced” sicker subgroups. Decisions about which risk adjustment model to use depend on what criteria are most important. A model used for resource allocation, or to monitor provider quality or efficiency, would adjust for differences in medical risk but not for inefficient practice patterns. If a part of the delivery system is inefficient, a good model will reveal this by producing a predictive ratios of <1.0. If the goal is budget planning, to make good estimates of what resources are used and how much money is going to be spent within various parts of the system next year, then the preferred model would adjust for both medical risk and past patterns of care or cost, and have predictive ratios closer to 1.0. Other important factors "Across a wide range of demographic, geographic, and service-related subgroups, all four models predicted differences in average costs across such subgroups well; the CRG model’s predictions were closest to actual. If we have adequately adjusted for differences in medical risk, then PRs that differ from 1.0 across subgroups defined by service system factors (such as, patients who use different facilities) can provide important information for system monitoring. For these subgroups, PRs ranged from a low of 0.63 to a high of 1.21. We also examined how these risk models performed, separately for people with prior low, intermediate or high costs. The risk models somewhat over-predicted costs for the lowest cost subgroups and under-predicted costs for the highest cost ones, but much less than the age/sex model. At the ends of the cost spectrum, the DCG model is closest to perfect accuracy; in the middle, the CDPS model is closest. Predictive power for all four “off-the shelf” risk models is far better than what can be achieved with the age/sex models currently used for TRICARE Prime beneficiaries, and compares well to their performance in other populations."