Kip Sullivan on Risk Adjustment in the Medicare Advantage Program
December 2019
In a few days, Medicare's open enrollment period will be over. CMS will then calculate how much to pay each Medicare Advantage (MA) plan for its new pool of enrollees for the year 2020 (presumably this will all be done in the last two weeks of December). CMS will attempt to adjust its 2020 per-enrollee payments (aka capitation or premium payments) downward to reflect the better health of the MA plan enrollees, but CMS will fail as it has every year since 1972 when Congress passed legislation allowing HMOs into the Medicare program.
CMS will fail in two respects: It will fail to lower average payments to each plan sufficiently to prevent overpayments to the plans; it will overpay for healthier enrollees and underpay for sicker (and poorer) enrollees, and that in turn will incentivize plans to avoid the sick and to deny services to the sick they cannot avoid. This will go on every year as long as there is a Medicare Advantage program (or any other program that pays any entity on a per-enrollee/capitation basis).
In this three-part series, I'll explain how MA plans enroll healthier beneficiaries, and why it's impossible to "risk adjust" premium/capitation payments to the plans downward sufficiently so that payments reflect the lower cost of their enrollees.
Part I of my three-part series on how CMS overpays Medicare Advantage (MA) plans.
Here's a summary of what I'll tell you:
(1) MA plans enroll Medicare beneficiaries who are healthier than the average Medicare beneficiary by cherry-picking and lemon-dropping. This is done covertly, not overtly.
(2) CMS's method of adjusting capitation payments to the MA plans has been extremely crude since 1972 when Congress passed legislation allowing HMOs to participate in Medicare. Between 1972 and 1985, it appears there was no risk adjustment at all. Between 1985 and 2000, CMS used only demographic data (age, sex, Medicaid status, and nursing home status) to predict what each beneficiary would cost the HMOs. That demographic risk adjuster predicted only 1 percent of the variation in spending on Medicare beneficiaries. In 2008 (after a phase-in period that began in 2000), CMS began to use diagnostic data in addition to demographic data, but this new method (known as the Hierarchical Condition Categories, or HCC, method) can only explain 12 percent of the variation. I'll illustrate what it means to say only 1 or 12 percent of the variation in spending is "explained" or "predicted."
(3) The miserable performance of the HCC is due to these three factors:
* the HCC uses less than half of the diagnoses used in modern societies; 40-45 percent of Medicare beneficiaries do not have a diagnosis covered by the HCC, which means the cost of those beneficiaries must be estimated using only the pathetic demographic risk adjuster;
* there is tremendous variation in spending on beneficiaries within each of the diagnoses included within the HCC (for example, the sickest 20 percent of beneficiaries with diabetes cost four times what the healthiest 20 percent cost); and
* income and other socioeconomic factors explain much of the variation, but CMS does not collect data on socioeconomic factors.
(4) CMS's HCC risk adjuster will never be more accurate than it is now.
I won't discuss upcoding in this series (that's what the plans will do to ensure they are overpaid regardless of how accurate CMS's risk adjuster becomes). Nor will I discuss other factors that contribute to MA plan overpayments such as the county-level "floors" CMS uses, the bonuses CMS gives plans for "quality," and a little known law that guarantees that hospitals have to give MA plans the same discounts they give FFS Medicare.
Over the two decades that I have been studying this issue, I have encountered no other writers who attempt to do what I'm doing here -- explain why risk adjustment will remain crude forever. In other words, I have no mentors. I welcome any suggestions any of you might have about how to explain these issues more accurately and persuasively.
Part II of my three-part series on how CMS overpays Medicare Advantage (MA) plans.
In this part, I explain how MA plans enroll healthier beneficiaries, and I illustrate what it means to say CMS's Hierarchical Condition Categories (HCC) risk adjuster explains (or predicts) very little of the variation in spending between beneficiaries.
How MA plans achieve favorable selection
By the early 1980s, studies were warning Congress that CMS (then HCFA) was overpaying Medicare HMOs (non-HMO insurance companies were not allowed into Medicare until 1997). Prior to the implementation of CMS's current HCC adjuster, estimates of the overpayments were as high as 40 percent.
HMOs, and today all Medicare Advantage plans, achieve "favorable selection" (enrollment of healthier beneficiaries) by cherry-picking and lemon dropping. They cherry-pick by:
(1) advertising in a manner that appeals to the healthy;
(2) setting coverage in a manner that discourages the sick from enrolling (for example, by requiring high copayments for drugs the sick will use); and
(3) excluding from their networks doctors and clinics that treat an above-average percent of the sick (e.g., addiction clinics).
A fourth factor that aggravates cherry-picking is beneficiary behavior. Sicker beneficiaries tend to stay away from MA plans while healthier beneficiaries are more willing to give up choice and autonomy for (the appearance of) lower costs for better coverage.
MA plans lemon-drop by denying services to sicker enrollees. Frustrated plan enrollees then disenroll and return to the FFS Medicare program to have their hip replaced or whatever it was they couldn't get from their MA plan.
What it means to say CMS's risk adjuster explains only 1 or 12 percent of the variation in spending
Between 1985 and 2004, CMS used only age, sex, Medicaid and nursing home status to adjust capitation (per enrollee) payments to MA plans and their predecessors. These factors, often called "demographic factors," explained only 1 percent of the variation in spending on Medicare beneficiaries. Between 2000 and 2008, CMS phased in a new risk adjuster that added diagnoses obtained from claim forms to the demographic factors. This risk adjuster is called the Hierarchical Condition Categories (HCC) adjuster. Between 2008 and 2012, the HCC was able to explain only 11 percent of the variation. CMS made a small adjustment for the year 2013 that raised the accuracy to 12 percent (see the "R2" values in Table 2-1 p. 12 here https://www.cms.gov/Medicare/Health-Plans/MedicareAdvtgSpecRateStats/Downloads/RTC-Dec2018.pdf )
I will illustrate what these 1- and 12-percent figures mean -- how inaccurate they are -- two ways: With a hypothetical, and then some real life data.
Illustration One: Let's assume the average cost of insuring a Medicare beneficiary is $10,000 a year (it's closer to $12,000 today). Let's say Tendercare HMO enrolls beneficiaries who on average would have cost $9,000 to insure if they had remained in the traditional FFS Medicare program. The overpayment is, therefore, $1,000. An accurate risk adjuster would tell CMS to pay Tendercare only $9,000. But in the days when CMS used only the demographic adjuster, CMS would reduce that $1,000 overpayment by only 1 percent, that is, by a mere $10; CMS would pay Tendercare $9,990 per enrollee instead of $9,000. The result: Tendercare made $990 on average off each enrollee for no reason other than it excelled at cherry-picking and lemon-dropping.
Similarly, since 2013 CMS's HCC risk adjuster only recoups 12 percent of the overpayment -- or $120. CMS would pay Tendercare HMO $9,880, and Tendercare would make an unearned $880 on average on each of its enrollees.
Illustration Two: In their June 2014 report to Congress, the Medicare Payment Advisory Commission (MedPAC) published the data in the table below to illustrate how inaccurate the HCC is (MedPAC examined the post-2012, 12-percent version of the HCC http://medpac.gov/docs/default-source/reports/jun14_ch02.pdf?sfvrsn=0 ) MedPAC divided traditional (FFS) Medicare beneficiaries into quintiles (fifths) according to how healthy or sick they were, and then compared the average cost of each quintile with the average for all beneficiaries in the FFS program.
They found that the HCC overpays for the healthiest quintile by 62 percent, the next healthiest quintile by 30 percent, and the middle quintile by 10 percent, and underpays for the sickest quintile. MedPAC found that the HCC underpays by 21 percent for the sickest 1 percent. As bad as that 21-percent underpayment is, it's a huge improvement over the old demographic adjuster, which underpaid by 83 percent for the sickest 1 percent! (see "age/sex" column in Table 1, p. 127 file:///C:/Users/Owner/Documents/2004_Pope_etal_HCFR%20CMSHCC.pdf )
_________________________________________________________________
Percentile Payment relative to average cost
<20th (healthiest) 1.62
20-40 1.30
40-60 1.10
80-95 0.86
95-99 0.82
>99 (sickest) 0.79
Medicare Payment Advisory Commission, Report to Congress, June 2014, Table 2-1, "Standard model" column, p. 30 http://medpac.gov/docs/default-source/reports/jun14_ch02.pdf?sfvrsn=0 .
__________________________________________________________________
This is what we get with the current HCC -- vast overpayment for the healthy and vast underpayment for the sick, despite a quarter-century of research.
As I'll demonstrate in Part III, it is not possible to improve the accuracy of the HCC by even a few percentage points past its current 12-percent level using just demographic and diagnostic data. It might be possible to improve the accuracy by a few points if CMS were to incur the immense cost of gathering useful socioeconomic data for all 60 million Medicare beneficiaries, or at minimum the 40 million who remain in FFS Medicare. But that's not going to happen.
Part III of my three-part series on how CMS overpays Medicare Advantage (MA) plans.
In this part I explain why the accuracy of CMS's current risk adjuster cannot be improved by more than a few percentage points.
As I said I Part I, there are three reasons why CMS's Hierarchical Condition Categories (HCC) risk adjuster is so crude (it can only explain 12 percent of the variation in spending):
* the HCC uses only 20 percent of the available diagnoses;
*there is great variation in spending within diagnoses the HCC does use; and
*socioeconomic factors explain much of the variation, but CMS does not collect data on socioeconomic factors.
I will discuss each of these three reasons in more detail. We will see that the first two problems are not fixable. The third problem could be ameliorated to some degree, but only at great expense.
The HCC uses only 20 percent of all diagnoses
By the 1990s, Congress had been repeatedly warned that overpayments to the Medicare HMOs were raising Medicare's costs. "Research showed that the managed care program increased total Medicare expenditures because its enrollees were healthier than FFS enrollees …." is how CMS put it in this 2018 report on its HCC (p. 11).
In the 1997 Balanced Budget Act, Congress ordered CMS to improve its risk adjuster by adding diagnoses to the demographic data CMS had been using for decades. CMS's first step was to turn to the International Classification of Diseases (ICD), published by the World Health Organization, for a list of diagnoses used throughout the Western world. At that time, the ninth revision of the ICD was in effect. ICD 9 contained 15,000 diagnoses.
(Today CMS uses ICD10, which has 68,000 diagnoses. The explosion in diagnoses has not changed the HCC, nor its gross inaccuracy. For those who want to check that out, you can compare this 2004 report on how the HCC was constructed with CMS's 2018 explanation of the HCC.
CMS mapped every one of the 15,000 diagnoses in the ICD9 to one of 189 HCCs, or groups of related diagnoses (heart attack and unstable angina, for example, were grouped together). Using a list of ten guiding principles, CMS then excluded all but 70 of the 189 HCCs. These 70 HCCs contained only 3,000 of the 15,000 diagnoses in the ICD9 (see page 129 in the 2004 report referenced above). (CMS currently uses about 80 HCCs, but those additional HCCs have added less than a percent to the HCC's accuracy.)
The reasons for excluding the other 12,000 diagnoses varied, but the overarching reason was the failure of the excluded HCCs to predict future expenditures. So, for example, osteoarthritis was excluded because the diagnosis is relatively subjective ("discretionary") compared with rheumatoid arthritis (the latter can be confirmed by tests); not "medically significant" ("muscle strain"); or "transitory" (appendicitis).
Great variation in spending exists among patients with the same diagnosis
Let's take diabetes as an example. The average Medicare beneficiary with diabetes costs 1.5 times the average for all Medicare beneficiaries. You might think you could greatly improve the accuracy with which CMS could predict spending on diabetics by simply taking 1.5 times per capita spending on all Medicare beneficiaries. That's better than doing nothing, but not by much. Diabetics in the sickest quintile of diabetics cost 6 times as much per diabetic as diabetics in the healthiest quintile; the sickest 1 percent of diabetics cost 13 times the average cost of diabetics in the healthiest quintile.
Since CMS implemented the HCC, the MA plans have figured out how to cherry-pick and lemon-drop not just within the total pool of Medicare beneficiaries, but within disease categories. For example, Medicare beneficiaries with diabetes who join MA plans are $1,072 less expensive than diabetic beneficiaries who remain in traditional Medicare (see Figure 4 here.
Socioeconomic factors explain much of the variation in spending
Many of you have probably heard it said that only 20 percent of illness can be explained by access to medical care. I don't know if that statement is accurate, but I have no doubt that factors outside the control of doctors and hospitals play an enormous role in contributing to differences in the need for medical care. Yet aside from "Medicaid status" (which is an almost totally worthless measure), CMS has never attempted to adjust its payments to MA plans with an algorithm that included a measure of income, education, food insecurity, presence of a partner at home, and numerous other social and economic factors that affect health and, therefore, the cost of treatment.
Factors like income and education are not collected by doctors and hospitals either. Imagine the cost to CMS, or the nation's clinics and hospitals, of asking every one of Medicare's beneficiaries for just that information. We know from the Medicaid program that verifying income alone is an expensive process. It's probably the main reason the overhead of unprivatized Medicaid programs is 4-5 percent compared with Medicare's 2 percent.
The ultimate measure of HCC's futility
In their 2004 report on the HCC, Pope et al. published a graph that all those who promote capitation of anything -- MA plans, HMOs, ACOs, integrated delivery systems -- should tape to their refrigerators. I was unable to copy it, so you'll have to click on the link and go to p 128.
You'll see a graph with HCC accuracy on the vertical axis (measured in R2 values) and number of HCCs on the horizontal axis. You see the R-squared value climbs steeply from a mere 1.69 (which is the accuracy of the demographic variables plus disability) as the ten most predictive HCCs are added on. But then you see the line level off at about 10-20 HCCs and climb very slowly toward 11, which was the maximum predictive power of the HCC in 2004.
Pope et al. reported that the congestive heart failure HCC was the most powerful predictor, then came the COPD HCC. They stated that the five most predictive HCCs alone "accounted for 61 percent of the maximum explanatory power of the full (101 HCC) model," and the first 30 accounted for 90 percent. In other words, after 30 HCCs, there is very little to be gained by adding more HCCs. "The incremental R2 from adding a diagnostic category is … 0.05 percentage points at 30 HCCs," Pope et al. reported. You can see why CMS decided to limit the number of HCCs to 70 (and the total diagnoses to just 20 percent of all diagnoses).
Conclusion
CMS spent a quarter-century and millions of dollars to develop the HCC. The final product is worthless. No amount of tinkering is going to make it any easier for CMS to risk adjust accurately the premium/capitation payments this time next year, or this time 50 years from now.
In this email I want to explain a statement that seems too weird to be true: Improved risk adjustment only worsens the overpayments to MA plans (and all other capitated organizations). Not only is the statement true, it explains why enrollment in MA plans soared after 2004, which is when CMS began to phase in its "improved" HCC risk adjuster.
In a subsequent email I'll address other issues, including your statement that you disagree with MedPAC that adding prior expenditures to CMS's risk adjuster reduces MA plans' incentive to hold costs down.
Improving risk adjustment worsens overpayments
In 2011 Don McCanne reported on one of the earliest studies to demonstrate the truth of this statement. The "digest" of the study that Don quotes first is quite readable.
A second study published in 2014 confirmed the first one. You can read a digest of that study here and the entire study here.
Here's what these studies reported: The MA plans figured out how to cherry-pick the healthiest of patients with expensive diagnoses and lemon-drop the sickest of them. The new HCC, with its use of diagnostic codes, made cherry-picking and lemon-dropping WITHIN DISEASE GROUPS even more profitable than it was when CMS used only demographic data. This in turn shifted money away from the FFS Medicare program (with its sicker and poorer and more rural beneficiaries) and to the MA program with its healthier, wealthier, and more urban enrollees.
Here's the metaphor used by the second study: Before CMS introduced the HCC, MA plans "went fishing" within the entire pool of Medicare beneficiaries for the healthiest, but after the HCC was introduced they went fishing within the pool of all diabetes patients, all cancer patients, etc. for the healthiest diabetes patients, the healthiest cancer patients, etc.
Because the average payment CMS makes to MA plans for diabetes, cancer etc. patients is so much higher than it is for beneficiaries without these diagnoses, and because variation in spending is immense within disease groups, plans make out like bandits by finding the healthiest of the beneficiaries within these pools.
Moral of story.
Don ended his 2011 qotd saying we must get rid of Medicare Advantage "crooks." Amen! I would add we must get rid of capitation. The crooks are only getting away with robbery because nearly the entire world, including many within the single-payer movement, either celebrate or silently tolerate capitation.
There is no way to improve risk adjustment substantially, and even if there were, there's no way to stop entities from continuing to cherry-pick and lemon-drop. There is, therefore, no way to stop overpayments to entities paid on a per-head basis. It doesn't matter if we obscure what's going on by calling the per-head payments capitation payments instead of premiums, or calling the capitated entities HMOs, ACOs, IDS's, instead of insurance companies. The result will be the same: vast overpayment to the capitated entities, and a shift of resources from the have-less to the have-more.