Risk Adjustment Basics and The Controversy Over Medicare RADV

A few readers have sent me messages and asked me to detail some basics of risk adjustment (RA) — how it works, its benefits, and its challenges — and the controversy surrounding risk adjustment data validation (RADV) in Medicare Advantage (MA) specifically. RA is a complex world, but here is my best effort to keep the overview simple and then move to the coming RADV conflagration.

While not practiced in the employer world because of the penetration of self-insured funds under the Employee Retirement Income Security Act (ERISA, where the employer shoulders the entire burden of costs and insurers are not at risk), risk adjustment has become an important and common practice in MA, Medicaid managed care, and Exchange managed care.

Quite simply, risk adjustment is critical to ensuring that health plans are compensated fairly to cover the costs of a given individual as well as the population as a whole in a given region or product. Risk adjustment is meant to ensure that insurers avoid “cherry picking” populations during the enrollment process, maintaining members’ coverage as long as they need it, and meeting the needs of an individual in terms of access to care in an efficient and quality fashion. This is especially important for those who may be more adverse than the average of a given population. It is illegal to cherry pick, remove members, and withhold care, but RA is an aligned incentive that provides equity in terms of government reimbursement and is meant to incentivize proper plan behavior.

What are the key features of an RA system?

• Risk adjustment factors have both a demographic and clinical component. The demographic component can represent a rate region, age, and sex, but also a variety of other factors that may help predict cost in the government program. These can include:
  • Medicaid and dual eligible status (full vs. partial)
  • Frailty
  • Community vs. institutional vs. other groupings
  • Disability or aged status
• In some cases, not all enrollees may be risk scored using the model.
• As well, there are both pharmacy and medical models. Some agencies may not apply both, but often they do. Different codes (National Drug Codes (NDC) vs. vs. ID-10 Diagnoses) may apply between pharmacy and medical models, but in Medicare Advantage pharmacy risk scores are actually tied to medical service diagnoses.
• Usually, the overall population risk score is set to a 1.0 average. A plan is paid more than a per-person base amount or rate for those who are sicker or more adverse based on the various components. A plan is paid less for those who are healthier or more beneficial.
• Risk adjustment often is used against a set amount of money based on overall enrollment in the program/region. Payments are reconciled against a blend of 1.0 times the base amount times the number of enrollees. For example, this is usually the case for Medicaid programs. But this is not always the case. In Medicare, both fee-for-service (FFS) and MA enrollees receive risk scores. The overall risk score may not remain equal to 1.0 in given years and thus monies paid out are not set from year to year based just on risk scores. Critics of MA argue that MA plans overcode to derive the most revenue and have artificially high risk scores compared with FFS. If you buy that argument, an overpayment to MA plans would occur whether the budget was fixed or not. But the way the current system works, critics can easily argue that overpayments in MA are huge.
•  There are various risk adjustment models, but often a model is refined for the specific characteristics of the population in a given program. As an example, the Medicare Hierarchical Condition Category (HCC) Risk Model is calibrated specifically to Medicare populations. There is a different Health and Human Services (HHS) HCC Risk Model that is used in the Exchange program, which focuses more on child and adult populations that are not aged or disabled. There are a variety of other risk models, including many used by state Medicaid agencies for Medicaid managed care. The most prominent is the Chronic Illness and Disability Payment System (CDPS) risk model, which is specific to Medicaid. Another point, these same models may be used internally at a health plan to identify and stratify members for risk, interventions, and care management. This is true of HCC, CDPS, Adjusted Clinical Groups (ACG), Clinical Risk Groups (CRG), and Diagnostic Cost Group (DxCG).
• Most risk models use diagnoses to determine the clinical risk component. A diagnosis is tied to a given disease state and then a relevant factor (based on costs and risks in the program) is applied. These then factor into the aggregate clinical risk component for an individual. It is also true that not every diagnosis is recognized in a risk model.
• Plans must document each diagnosis in a claim or encounter submitted to the applicable regulatory agency.
• Risk models can be structured differently. Some are fully additive, where risk is calculated based on each diagnosis/disease state/condition reported. Others may determine risk by taking only the disease state with the greatest cost or risk. But often there are numerous hybrid approaches.
• The risk model seeks to predict an enrollee’s likely medical use and costs. Models can be either concurrent (where the current year diagnoses are used to predict the current year’s healthcare costs/payments) or prospective (where the current year demographics and diagnoses predict the following year’s or a future year’s costs/payments with applicable trending). In a concurrent model, because all diagnoses may not be known until after conclusion of a given year, often times actual payments can be reconciled in the next or later years on an aggregate plan-enrollment basis based on the total population. Thus, in the Medicaid and Exchange world, there can be so-called winners and losers when a given year is reconciled. This is not the case in Medicare Advantage. Sometimes, the reconciliations can surprise plans if their estimates are not accurate.
• A further complication of risk adjustment in the Exchanges is that risk scores differ by so-called metal tiers because of the different actuarial values of Bronze (60%), Silver (70%), Gold (80%), and Platinum (90%) plans. The richer the benefit, the higher the risk score is for given disease states because plans pay a bigger share of the costs.

Criticisms and challenges

In general, RA is widely accepted and supported as it seeks to reimburse plans based on differing risks of members. But the predictive accuracy of RA models is sometimes questioned. I will not get into the predictive accuracy of specific models, but the literature indicates that even the most honed and accurate risk adjustment models explain less than half of costs (and it may be as little as 30 percent). In some ways, this makes sense. Studies have found that issues unrelated to underlying disease states (such as social determinants of health – socio-economic status, employment/income, housing stability, transportation availability, food security, health literacy, and more) are a greater predictor of overall medical costs.

We do know that risk models tend to overpredict costs and risks for people healthier than the average and underpredict costs and risks for people who are sicker than the average.

It is also true that predictive accuracy has many nuances. There are different accuracy rates between models. A risk model may be more inaccurate as the number of concurrent disease states increase. The risk model may be inaccurate for certain groups of individuals within the model as well.

Risk adjustment is also criticized for reimbursing based on the risk of the member alone and not taking into account costs related to improving and maintaining health/outcomes.

Other concerns and recommendations for improvement

Risk models are constantly refined to improve accuracy as trends change. This is a good thing, but can also be controversial. For example, CMS refines the Medicare HCC model every few years. The most recent changes will be phased in over three years beginning in 2024. The changes include the following:

• Migration to an ICD-10 diagnosis-based risk model
• Changes in coefficient factors
• A greater number of HCC grouping that will drive risk payments
• But fewer diagnoses that qualify for HCC groupings, which takes away from risk payments
• In some cases, there are new codes that tie to an HCC category that did not exist before.

CMS says that some of the changes were to address the so-called gaming in the MA industry. When fully phased in, the risk adjustment model changes (version 28) represent a several point reduction alone in revenue. This is billions of dollars to MA plans.

Accuracy of risk models is hobbled by a number of things:

• The quality of data: Encounter and claims submissions can be inaccurate. In addition, some of the data comes from manual chart reviews of physician records. This can lead to prioritization of certain submissions at the expense of overall data submission accuracy and completeness. Part of the recommendations here include moving fully to digital capture of risk information (claims and Electronic Medical Records) and eliminating so-called chart chasing for diagnoses to drive risk scores. Many argue that home visits (which are not tied to actual care) should be eliminated as well as it skews results.
• The capture of true functional risk: While certain demographic factors capture some aspects of this, there are concerns that models do not adequately capture the full functional risk of a member with given disease states, which can have a major bearing on overall costs. Recommendations here are to expand the capture of functional status via coding and weighting.
• Social determinants’ impact: In addition, as alluded to above, social determinant barriers are not adequately captured in risk models. Demographic components that are included can only go so far to compensate. Various pilots could lead to solutions here.
• Health equity: Separate from the social determinant issues noted just above, there is also concern that there is an overall health equity imbalance in risk modeling. Predictive cost accuracy tend to vary based on ethnic groups or subgroups. There are new health equity experiments that could lead to valid factors for risk models in the future.

The RADV rule and controversy

CMS already adjusts rates downward by putting a coding intensity adjustment on MA rates each year. But, to further attack the perceived issue of overpayments, it added a new and controversial RADV rule and process that was finalized in 2023. The rule stemmed from earlier pilot audits done by CMS and the HHS Office of Inspector General (OIG).

In these new individual audits, the government will send a sample to a plan. Plans will need to provide documentation of diagnoses submitted via encounter data or supplemental submissions. CMS will extrapolate penalties based on a sample audit’s errors. What does this mean? Based on the error rates in the sample, CMS and the HHS OIG (both can conduct audits under the rule) will calculate penalties across the entire membership and revenue as if the errors occurred throughout the enterprise. So, poor risk adjustment policies at a plan could lead to huge recoveries – billions across the industry.

The rule also gives CMS significant flexibility on who and how they audit. Audits could be focused on contracts at the highest risk for improper payments. Diagnoses and types of beneficiaries (subpopulations) may be singled out if they are at the highest risk for improper penalties (extrapolation would be narrowed to these groups as well if undertaken).

Numerous due process, selective enforcement, regulatory scope, and other legal issues are ripe here and it led Humana to file a lawsuit once the rule was finalized. Among the issues to be litigated will be:

• Does the scope exceed CMS’ authority?
• Does CMS have the ability to extrapolate penalties?
• Can CMS apply such rules retroactively (in this case to 2018)?
• CMS’ flexibility may violate due process, adequate notice, and other legal tenets.
• CMS argues that RADV’s requirement to have MA plans submit documentation will be a payment integrity tool. But FFS claims are accepted without this type of documentation. Save for some fiscal intermediary oversight, there is little oversight in the FFS system. Why shouldn’t there be a similar process in FFS claims submission given all the talk of rampant fraud, waste, and abuse?
• The FFS adjuster in the RADV formula was eliminated. It was in CMS’ 2012 RADV methodology and deemed critical. The adjuster accounts for potential errors in FFS data as well as the difference in how coding is validated between FFS and Medicare Advantage.
• CMS claims FFS providers submit fewer diagnoses; therefore no adjuster is needed. But in MA, there is a risk score coding intensity adjustment applied to the MA rates. Are MA plans then being penalized twice, once with coding intensity and then on RADV?

Let me be clear. I do think MA plans need to clean up their act and not be so aggressive on some of its risk adjustment practices. In most cases, it is sheer sloppiness; based on some whistle-blower complaints, in others it may be willful and fraudulent. Each plan needs a comprehensive and accountable submission and validation process. The American taxpayer deserves this.

But at the same time, MA plans should not apologize for having a business incentive to accurately document diagnoses for revenue when a diagnosis exists. FFS providers simply do not have the incentive as they remain captured primarily in a transaction system. This alone requires that any rule that eventually goes into effect is fair and equitable.

#riskadjustment #raf #ra #radv #medicare #medicaid #exchanges #medicareadvantage #hcc #cdps #overpayments #oig #hhs #cms

— Marc S. Ryan