This study examined the accuracy of the current Medicare Part D risk adjustment approach in predicting 2006 Part D drug expenditures and plan liability. Accurate risk adjustment is important for providing fair payments to health plans and discouraging deliberate selection of patients with more favorable risk profiles. The results from this study suggest that incorporating information on prior drug use and/or costs into the risk adjustment approach would substantially improve the accuracy of payments to Part D plans.
In a cohort of Medicare Advantage beneficiaries continuously enrolled in 2005 and 2006, expenditures for Part D-covered drugs were relatively stable during consecutive years. Not surprisingly, the performance of the current Part D risk adjustment approach is improved by using prior year drug information in addition to using only diagnoses and demographic information to predict future drug costs. Including information on whether patients had specific types of drug use (i.e., any use of a drug within a therapeutic class) improves prediction substantially over the current Part D risk adjustment approach: in the case of predicting total Part D drug expenditures in 2006, the percent of variation explained increased from 12% to 29%. Including drug cost information further improves prediction, e.g., increasing the percent of variation explained to 39%. Better accounting for the curvilinear relationship between prior year and current year expenditures by including a quadratic term further improves the performance of cost-based risk-adjustment measures. Specifically it decreases the level of overprediction at low expenditure levels and underprediction at high expenditure levels. In short, approaches that combine prior year diagnosis and drug information, and allow for curvilinear expenditure relationships perform the best among the approaches we examined in predicting future year expenditures, with 43% and 46% of variation in all drug expenditures and plan liability, respectively, explained. The existing Part D risk adjustment approach, of course, could not use drug information in 2006 because prior drug information was not available to CMS before the introduction of the drug benefit.
Part D drug information is now available for beneficiaries who enrolled in Part D in 2006 through the Prescription Drug Event files and would permit development of a revised system for future payments. Both of these approaches should be operationally feasible, and have precedents in the Medicare hospital payment program where the Diagnosis-Related Group (DRG) system depends in part on procedures performed.
Risk adjustment approaches cannot and need not perfectly predict costs because some costs will be unpredictable and plans cannot select based on unpredictable variation.10
Drug expenditures, however, tend to be stable from year-to-year, and are more predictable than other types of medical costs. Ignoring past spending when making payments thus results in preventable misallocation of dollars, and creates strong incentives for selection. At this point in time, plans have detailed information on enrollees' drug use and expenditures, and so can predict future costs substantially better than CMS does in its current risk adjustment score. Inadequate risk adjustment offers plans an incentive to select the most profitable beneficiaries through strategies that limit enrollment of beneficiaries with certain types of conditions, cost-profiles, or drug use.11
These strategies include subtle adjustments of formularies or to drug prices, or policies that restrict patient access to more expensive medications, such as prior authorization or fail-first requirements.
Importantly, as the risk corridors start to widen in 2008 and eventually disappear, the incentives for favorable selection increase. Additional work is needed to examine such potential strategies, especially relative to plan bids, and the implications for spending under Parts A and B of the Medicare program.
Balancing Concerns about Divergent and Perverse Incentives
Traditionally, the Medicare program has not used payment adjustments based on prior utilization because of concerns that this creates perverse incentives for overuse (i.e., generating higher costs in the current year leads to higher payments next year). Use of standardized or reference costs for each of the Part D therapeutic classes rather than actual past costs would decrease this perverse incentive to the degree it exists, albeit potentially at the expense of increasing incentives to select favorable risks.12
How the Part D market might respond to these competing incentives is unknown.
In addition to the divergent incentives for adverse selection and for overuse, approaches that exclude drug use information also create perverse incentives for underuse (e.g., stinting on delivery of necessary medications). This is particularly relevant for stand-alone PDP plans, which do not bear any inpatient or outpatient costs incurred as a result of under-treatment with prescription drugs. Incorporating information on drug use into the risk adjustment scheme would decrease such incentives.
In this study, we used information from a single integrated delivery system in a single region of the country. Actual expenditures and risk adjuster performance in other plans will vary by prescribing practice patterns, patient cost-sharing structures, drug formularies, pharmacy management, and contracts with drug suppliers including any rebates. For example, the health system that provided data for this study historically has had a higher level of generic drug use compared with use across the general U.S. population. Although these differences could affect the relative performance of the alternative risk adjustment approaches in other settings, they are unlikely to affect the overall study finding that incorporating information on past drug use substantially improves risk adjustment.
Moreover, CMS could use drug information from all plans when refining the risk adjustment approach. This study was also limited to beneficiaries who were continuously enrolled over 24-months through 2005 and 2006, and these beneficiaries may have different cost profiles. Our findings were similar, however, in sensitivity analyses that included subjects who died (3.2%) during 2006.