Five Areas of Concern

Comments on risk adjustment under the ACA

Roy Goldman

My comments assume the reader is familiar with the structure and overall process underlying risk adjustment under the Affordable Care Act (ACA). I recommend reading “Insights on the ACA Risk Adjustment Program,” published by the American Academy of Actuaries (Academy) in April 2016,1 and the Centers for Medicare and Medicaid Services (CMS’s) discussion paper for the “March 31, 2016 HHS-Operated Risk Adjustment Methodology Meeting,”2 along with its written answers to the questions posed at that meeting (published June 6, 2016)3. We all benefit greatly from the transparency shown by the Center for Consumer Information and Insurance Oversight (CCIIO). Its model is a good one for all of CMS to follow.

One cannot overstate the importance of risk adjustment for the successful implementation of the new insurance system for the individual and small group markets under the ACA. Clearly, if all insurers must enroll all applicants, which I believe is a good national policy, then a mechanism is needed that efficiently transfers risk among the insurers. Ideally, the transfers should be predictable, timely and truly reflect the costs of the underlying risks. We have not yet achieved this ideal.

I am not going to repeat all the good ideas and initiatives discussed in the documents cited earlier (e.g., changes being made for partial-year enrollees). Rather, I want to emphasize five areas of concern.

10 Percent of Premium is Too Much

In its “Summary Report on Transitional Reinsurance Payments and Permanent Risk Adjustment Transfers for the 2015 Benefit Year,”4 CCIIO reported risk adjustment is working as anticipated in transferring funds from plans with low paid claim amounts to those with high paid claim amounts. It also reported that the absolute value of risk adjustment payments and receipts averaged 10 percent of premium in the individual marketplaces. The percentage for the 2014 benefit year was the same. In my opinion, this percentage is too high.

Remember, 10 percent was just the mean among all plans! The spread between the 25th and 75th percentiles was [−12 percent, 8 percent] for plans with greater than 120,000 member months; [−10 percent, 21 percent] for plans with member months between 12,000 and 120,000; and [−36 percent, 35 percent for plans with fewer than 12,000 member months.

There is scarce information available for a plan to judge its relative risk 24 months in advance of when it learns the magnitude of its risk transfer payables or receipts. Plans need to build expected transfers into their pricing for rate filings due by June, year Y−1, and they learn their transfer amount in June, year Y+1. But 10 percent, 20 percent, 30 percent or more is a lot for a plan to deduct from expected claims if it anticipates receiving risk adjustment payments. If it is wrong, it has underpriced. On the other hand, plans thinking they have lower risks than average are nervous about adding too much to expected claims to cover anticipated risk adjustment payments and, thereby, becoming uncompetitive. In addition, such a plan might need to defend to CMS or state actuaries why one’s premium rate is “reasonable” since a 10 percent increase is considered, prima facie, “unreasonable.”

Year-to-Year Volatility

Of course, my last statement about premium increases would not be a concern if there were consistency in transfer payments from year to year. Then, theoretically, if a plan expects to pay, say, 20 percent of premium each year into the risk pool, at some point that amount is built into the base premium rate, and one does not need to raise rates another 20 percent the following year. However, my analysis shows huge fluctuation in transfer payments for nearly all plans in 2015 versus 2014. As explained in the section subtitled “Volatility Analysis,” even after removing potential outliers in the data and only considering plans in the eight largest marketplaces that were in the same payer or receiver status in both years, we found:

  • 47 percent of the plans were “worse” off in that they paid more or received less in 2015 than 2014; 53 percent of the plans fared “better.”
  • Only 9 percent of plans had transfer amounts in 2015 within 10 percent of 2014.
  • Only 40 percent of plans had transfers amounts in 2015 within 50 percent of 2014.
  • 33 percent of plans had transfers in 2015 that were more than 100 percent greater or less than in 2014.
  • The median transfer amount in 2015 was 65 percent of the amount in 2014—greater or less.
  • When all 113 plans in the eight largest states were analyzed, the median transfer amount was 114 percent of the prior year’s transfer amount.

I would hope that volatility would reduce in the future when the ACA becomes fully accepted by most Americans, all eligible enrollees are covered and changes in methodology have been made. But right now the fluctuations are huge.

Concurrent Versus Prospective Model

Given the volatility, how can we improve the risk adjustment process to increase predictability for the plans and accuracy of the model while increasing timeliness of knowing the transfer amounts?

Clearly, there is a tradeoff among these three elements. Almost by definition, a concurrent model based on year Y’s diagnoses is going to be a better predictor of year Y’s costs than a prospective model based on, say, diagnoses in Year Y-1. Obviously, when ACA was first implemented, there were no prior-year diagnoses to use for members, and membership does change from year to year. But even as membership grows and becomes more stable, CCIIO states that “greater accuracy” is their reason for preferring a concurrent model. Yet, Medicare uses a prospective model; however, there are five main differences in risk adjustment for Medicare Parts C and D versus the ACA:

  1. Nearly everyone eligible for Medicare enrolls and stays in the Medicare program until death.
  2. Medicare’s risk adjustment payments are not restricted to be a zero-sum game among insurers by state.
  3. Medicare does not have an unpredictable, expensive risk class such as births.
  4. Medicare Part D program has reinsurance and risk corridor protection in addition to risk adjustment.
  5. Medicare has a true-up process six months into the plan year and then eight months after the end of the plan year.

As the number of insureds continues to grow in ACA plans, I think eventually the ACA risk adjustment process needs to move to a prospective model. With a prospective model, plans would have more knowledge of risks of their insured and could be given more knowledge of how their average risks compare to those of other plans in the state. The settlement of risk adjustment payments could occur much more quickly, at least on a preliminary basis.

I would expect that each individual’s set of diagnoses remains with that individual as he or she moves throughout the health care system, whether to another individual payer, to Medicaid, or to a small group. Each insured’s risk profile would be given to his or her insurer upon enrollment. To supplement the known prospective risks, I would add two concurrent features to determine a plan’s risk profile in a given year:

  1. Retain the concurrent system for newborns.
  2. Add large claim pooling to account for current year, very high, unpredictable costs due to accidents, new procedures or drugs, or for those members new to the marketplace.

I recommend one other change to the risk adjustment model, which is to switch from calibrating the risk scores from a database of claims and demographic information from employer groups to using a database of experience based on members who joined marketplace plans. I do not know if this change would materially alter the relative risk scores of the disease conditions in the model, but it would improve the acceptability of the model among actuaries and other users of the risk adjustment model.

Obviously, there are a lot of details to consider, and these may not even be the best ideas. But I think the actuarial community should form a task force to determine an improved approach that gives plans a better sense of their risks compared to the other plans in the state at the time they are pricing the plans (year Y−1) and determining risk adjustment liabilities or assets (Year Y).

Prescription Drugs

It seems clear from the literature that predictive accuracy of either a prospective or concurrent model can be improved by adding prescription drug information. CMS has a good point about drugs that are used for off-label purposes. But I think CMS is overly concerned about plans’ abilities to induce physicians to overprescribe drugs with the aim of increasing a plan’s risk score to cover the costs of paying for, presumably, unnecessary drugs. I believe this is a red herring. CMS gives health plans too much credit (or discredit!). Even in an integrated health care system, health plans do not have such control over providers. As the Academy’s insight paper points out, a plan’s insureds in the marketplace generally are only a small percentage of a provider’s patient base. Further, if all plans did this, their relative risk might not improve at all.

In their published Q&A from the March 31, 2016 meeting,5 CMS-CCIIO stated, “We intend to propose to incorporate a small number of prescription drug classes as predictors in the HHS risk adjustment methodology for the 2018 benefit year to impute missing diagnoses and to indicate severity of illness.” They go on to say, “We intend to limit the number of prescription drug classes included as predictors to only those drug classes where we are confident that the risk of unintended effects on provider prescribing behavior is low.” I urge them to be more expansive in their thinking by including a wider class of drugs than they are currently considering, as long as there is a meaningful increase in R2 from adding the imputed diagnosis.

In addition to looking at specific drugs, CMS also should consider adding diagnoses that appear in the Medicare Part D Rx risk adjustment methodology that are not included under the Part C model. The disease hierarchies from the Medicare Part C model were input into the development of the ACA risk adjustment model, but there are categories of disease in the Part D model that are not in the Part C model—presumably because they affect drug costs more than medical costs. For example, members with mental health conditions, such as depression, schizophrenia and bipolar disorder, are recognized as being more costly members under Part D, but not under Part C or under the current ACA risk adjustment model. Because the Part D model uses diagnoses codes rather than drug claims, health plans could be compensated more adequately for these risks by including some Part D diagnoses without adding these drugs to the model.6

Credibility

Another way to improve the predictability of risk adjustment transfers is to incorporate some type of credibility based on member months. Given the wide spread of transfer payments for very small plans,7 it is unrealistic to expect these plans to accurately predict and build in enough margin to cover possible risk adjustment transfers. I understand CMS’s concern that if credibility were applied to a plan’s risk score, the plan may not be “fully compensated” for its risks. But how can one say with certainty what the relative risk is for a plan with, say, fewer than 12,000 member months? And what about the very small plan that learns in year Y+1 that has to pay 35 percent or more of its premium into the transfer pool?

Designing a credibility-weighted methodology may require some subsidization by larger plans in order to retain a zero-sum game. But there are lots of subsidies within the ACA program now with respect to age, metallic tier, geography (risk adjustment transfers from lower-cost areas occur at the average premium for the state) and income (premium and benefits for low-income members). Credibility is used in determining minimum loss ratios under the ACA, and credibility is part of Medicare’s risk adjustment methodology. I urge consideration of credibility to reduce the year-to-year volatility for very small plans.

Conclusion

Risk adjustment was made a permanent feature of the ACA for good reason: In order to have a level playing field, insurers should be compensated based on the risks they assume and should have no incentive to unfairly discriminate among insured. It is incumbent on the actuarial community to develop the best program possible that transfers risk accurately, predictably and in a timely manner. It is impossible to know the exact future of the marketplaces. The details will change over time, but risk adjustment needs to remain a permanent feature. In my opinion, if the country is going to continue to reduce the number of uninsured, and do so in a cost-effective manner, private health plans will be needed to manage the risks, even if they are not ultimately the insurers of the risk. In either scenario, the risks of their underlying membership must be calibrated as equitably as possible.

Appendix: Development of Volatility Analysis

We analyzed data for all plans in the eight states (California, Florida, Texas, Pennsylvania, Georgia, North Carolina, Illinois and New York) with greater than 5 million member months (Figure 1). After excluding a few very small plans, there were 113 plans in those states with risk adjustment transfers in the individual market in both 2014 and 2015.

Figure 1: Risk Adjustment Transfers in Plans with Greater Than 5 Million Member Months
Payer both years 31 27%
Payer 2014; receiver 2015 8 7%
Receiver 2014; payer 2015 19 17%
Receiver both years 55 49%
Total 113

We then looked at the magnitude of the payments. Ideally, one would want to know each plan’s payments as a percent of premium, but CMS’s summary reports for 2014 and 2015 only give the dollar amounts. We calculated the difference in transfer amount from 2014 to 2015 by Diff = (P15-P14) / |P14|, where Pi is the plan’s payment in year i (negative for a payer; positive for a receiver) and | | indicates absolute value.

When Diff is negative, a plan pays out more or receives less in 2015 than in 2014. Among all 113 plans, 52 percent were negative. A plan was considered an “outlier” if the transfer amount in either year was less than $1 million. Removing those plans left 71 plans.

Our goal was to analyze the distribution of |Diff| for those plans that were in the same payer or receiver status in both years. Because |Diff| > 100 percent for all plans that switched from payer to receiver status or vice versa, by excluding these plans from the analysis, we understate the overall volatility as these 13 plans had a median |Diff| of 240 percent.

Of the remaining 58 plans, 47 percent of these plans were worse off in 2015. The distribution of |Diff| is shown in Figure 2 for the three sets of data. The median |Diff| for all 113 plans was 114 percent. It was 80 percent for the 71 plans, and 65 percent for the 58 plans. A summary by state is given in Figure 3.

Figure 2: Percentage Difference Between 2014 and 2015

Figure 3: State-by-State Summary
State Member Months (Millions) Median |Diff| (Payers or Receivers Only)* Number of Payers or Receivers Only Median |Diff| All Plans Total Number of Plans
California 24.3  68%   9   72%   13
Florida 16.9  66%   9   93%   20
Texas 14.1  63%   5 318%   16
Pennsylvania   6.9  92%   9 100%   20
Georgia   6.2 159%   7 261%   10
North Carolina   6.0  64%   2 197%     7
Illinois   5.7  64%   3 125%     9
New York   5.3  45% 14   83%   18
Total 85.4  65% 58 114% 113
*Note: Illinois and North Carolina were combined for this analysis.
Roy Goldman, Ph.D., FSA, CERA, MAAA, is a retired health care executive.