CMS uses a risk adjustment model to determine its payments to plans based on the expected healthcare costs of each plan’s enrollees. This process is known as risk adjustment. Avalere finds that CMS’s risk adjustment model under-predicts costs for individuals with multiple chronic conditions by $2.6 billion on an annual basis. CMS last updated the model in 2014 and has indicated that it will make changes to the model intended to improve its accuracy for certain Medicare-Medicaid “dual eligibles” in 2017.
“The risk adjustment model used by CMS since 2014 does not adequately pay plans for individuals with multiple chronic conditions or for high-cost enrollees,” said Tom Kornfield, vice president at Avalere. “While CMS is considering changes to the risk model to address underpayments for low-income beneficiaries, the agency may also want to conduct research on methods to address underpayments for individuals with chronic conditions.”
Avalere assessed the accuracy of the 2014 model for beneficiaries in the traditional Medicare program with certain common chronic conditions by comparing predicted healthcare costs to actual healthcare costs. These beneficiaries include individuals with multiple chronic conditions and some specific health conditions. Among chronic conditions, rheumatoid arthritis (RA) and osteoarthritis have the largest percentage difference between predicted and actual expenditures (15 percent, or $2.3 billion, for RA and 12 percent, or $4.4 billion, for osteoarthritis) as noted below.
In addition, Avalere examined the highest spending individuals within 55 chronic conditions. As noted below, the model substantially under-predicts costs for those individuals with the largest disease burden. We find that the model under-predicts expenditures for chronic, non-cancer pain ($14.3 billion), osteoarthritis ($13.4 billion), depression ($8.9 billion), and RA ($5.3 billion).
On October 28, CMS announced proposed changes to the MA risk model that the agency believes will improve its predictive power for low-income beneficiaries. Specifically, CMS intends to further refine the model by accounting for both dual eligible/low-income subsidy (LIS) eligible and disabled status. However, it is not clear that these proposed changes will address underpayments for low-income beneficiaries with multiple chronic conditions as shown in this analysis.
“Payment accuracy in Medicare Advantage is critical so that the appropriate incentives exist for plans to treat the chronically ill. Ensuring adequate payment levels encourages broader program participation and robust coverage options for seniors,” said Caroline Pearson, senior vice president at Avalere.
The full report is available here.
America’s Health Insurance Plans (AHIP) provided funding for this research. Avalere maintained full editorial control.
To perform this analysis we used the 2012 and 2013 Five Percent Limited Dataset (LDS) Standard Analytic Files (SAFs), Medicare claims for a random 5 percent sample of Medicare fee-for-service (FFS) beneficiaries, and version 22 of the CMS-HCC model software. Conditions are identified using 2012 claims, and actual expenditures are computed from 2013 claims. We adopted the same approach to beneficiary selection and expenditure computation as CMS uses for model calibration. Furthermore, to calculate the applicable use of the institutional versus community segments of the CMS-HCC model (the institutional segment generally predicts somewhat lower program expenditures than would the community segment, given identical risk factors) we estimated when during 2013 the person was a nursing home resident. Since there is no indicator in the available administrative data for when a beneficiary resided in an institutional setting, we imputed this using an algorithm described in a peer-reviewed article by Koroukian, et al. from 20081 that uses physician claims for nursing home evaluation and management (E&M) services.
We then constructed several subgroups of beneficiaries, mostly defined by having one or more chronic conditions. We selected chronic conditions from three sets of definitions: 1) CMS Chronic Conditions Warehouse (CCW), 2) the set of criteria for Part D Medication Therapy Management (MTM) programs, and 3) lower levels of chronic kidney disease (CKD) as defined by HCC version 21 (these HCCs were excluded from version 22). In addition, we stratified beneficiaries by dual eligible/low-income subsidy (LIS) status and also by specific stages of chronic kidney disease (CKD). With the exception of CKD, we selected a set of conditions that aligned with HCCs, only by happenstance, in order to understand how well the model performs for groups with costs that the model did not intentionally attempt to identify.
For this analysis, we focus on the predictive ratio, the measure of predictive accuracy computed as the ratio of predicted versus actual expenditures. This measure indicates whether the model predicts total expenditures on average. This is also the measure that CMS principally focuses on for assessing predictive accuracy of subgroups of Medicare beneficiaries. For ease of interpretation, we converted these predictive ratios into estimated percentage over/underpayment. We also include confidence intervals in our model to test for statistical significance.
1. Koroukian SM, Xu F, Murray P. Ability of Medicare Claims Data to Identify Nursing Home Patients. Medical Care 2008 Nov; 46(11): 1184–7.