Overcoming the challenges caused by risk adjustmentWeb Exclusive
For both of the first two years of the ACA risk-adjusted individual market, company-to-company transfers have averaged approximately 10 percent of marketplace premiums. On average, each issuer will transfer (either receive or pay) more than $10 million per market. For 2015 alone, more than $5.6 billion will be transferred within the individual market, and an additional $2.2 billion will be transferred within the small group market.
With the elimination of underwriting, risk adjustment becomes each carrier’s first—and sometimes only—line of defense against anti-selection. Just as actuaries used to incorporate the impact of underwriting into their pricing and forecasting, they now must be able to predict these risk adjustment transfers. Our ability to understand, explain and forecast risk adjustment is critical to our ability to create stable markets and maintain our professional credibility.
In addition, the ACA risk adjustment mechanism has significant implications for sustainability beyond whether we can forecast these transfers. Specifically, a critical element of a sustainable insurance system is that it is at least resilient, as defined by Nassim Taleb in Antifragile.1 Here, our experience with this risk adjustment mechanism is in its infancy, but there are reasons for concern.
THE PREDICTION PROBLEM
When evaluating whether a risk-adjusted market is more or less predictable than what preceded it, it is important first to understand the mechanisms in context, and to understand where complexity has been added and removed.
The Complexity Trade-off: Underwritten Versus Risk-Adjusted Markets
In the underwritten markets of yesteryear, actuaries would need to incorporate good predictions on all of the following:
- The efficacy of underwriting
- Whether the pool of applicants was remaining consistent across time
- The underwriting wear-off curve
- The impact of any contract reserves, if applicable
In addition to these complexities, there was an inherent fragility in the pricing process. Specifically, if a carrier was unable to previously implement actuarially-justified rate increases on a particular policy form, then the subsequently-needed rate increase could include an anti-selective death spiral, further deteriorating the block. In response to these types of concerns, different states developed different rules to smooth out rate increases, establishing more complexity by creating regional differences in product structures.
In an ACA market today, underwriting as a complicating variable essentially is eliminated.2 In its place, however, we have added the complexity of the risk adjustment model and the transfer equation.
The CMS Risk Adjustment Model
Concerns about the risk adjustment model already have risen in various contexts: Should it be a concurrent model? Should it include drug data? Should it be calibrated to Exchange data, and, if so, how? All of these are valid technical questions; however, I believe the prediction and sustainability problems associated with these and similar questions can be summarized by examining difficulties that generally arise when the models are imperfectly calibrated and, surprisingly, the problems that occur even when they are perfect.
If the risk models themselves are imperfect, prediction and sustainability problems are intuitive. As an example, the risk model implemented by the Centers for Medicare & Medicaid Services (CMS) originally used ICD-9 codes. In October 2015, ICD-10s were implemented, requiring a mapping between the classification systems. This mapping is complex, and the resulting coding practice changes can affect the incidence rate of various diseases. Others have questioned the risk calibration for hepatitis, given the high cost of new pharmaceuticals.3 When these calibration concerns arise, carriers with disproportionate enrollment of individuals with these conditions will clearly be at a disadvantage. In short, this sort of error dampens or eliminates the value of risk adjustment itself and creates market fragilities.
Less obvious, significant problems can arise even when the models are perfectly calibrated.
By construction, the CMS risk adjustment model is calibrated to a national data set based on historical data, but its results are applied to future state-specific markets. Given large regional variation in physician practice patterns and the incidence of disease, this means we should expect patterns of winners and losers to differ in different parts of the country, and this can change over time. This also means that in most cases, there may not be singular broad-sweeping conclusions we can draw from historical observations. Put simply as a purely hypothetical example, it could be true that individuals diagnosed as diabetic are insufficiently adjusted in Florida, but over-adjusted in California (although they will be properly adjusted, on average, nationally).
Further, CMS’s risk adjustment models currently are calibrated strictly from data that was collected without risk adjustment in mind. As diagnosis data is repurposed to be a funding source for carriers, a natural carrier response will be to ensure complete and accurate coding. CMS will require detailed claims review as part of their auditing process. As a significant byproduct of these efforts, the incidence rate and intensity of coding will undergo an evolution. This coding intensity also will be affected by the adoption of electronic health record (EHR) systems. Therefore, even if the current risk models are perfectly calibrated, they will be imperfectly executed. This will create new administrative expenses, and it will add noise to an already complicated health care system.
As the risk models are recalibrated, patterns of winners and losers will change over time; what is true in 2014 may no longer hold true in 2017 or 2020. These same changes will affect not just issuer-level totals; they will also have area-specific and benefit-specific pricing implications because the risk adjustment model has area and metal tier components.
These complications don’t mean data analytics aren’t important; they just mean that actuaries will need to be especially careful when interpreting data and open-minded when seeing results. As a profession, we also will need to be cautious about generalizing from any limited set of observations. We also need to be aware that these are new fragilities that have been introduced in the system; what currently is not knowable is how these fragilities relate in magnitude to the problems of the system that preceded it.
Transfer Equation Issues
Our ability to predict and explain results in ACA markets hinges not just on the risk adjustment model, but also the way that model creates a financial asset or liability via the transfer equation. To this end, the predictability of statewide average risk scores and premiums are critical.
For 2015, the individual market risk score increased by less than 4 percent nationally, superficially suggesting some predictability, especially because we typically think of the 2014–2016 period as one of instability in these markets. However, this change in scores may be relatively low for the foreseeable future. Beginning in 2016, the model coefficients will be updated, with more changes in store for 2017 and 2018. Beginning in 2017, many states will no longer offer so-called transition relief. This may change the risk characteristics of their insured. The two years of data released (shown in the chart below) show the impact of transitional relief may be larger in the individual market, which is intuitive.4
Actuaries therefore are confronting a fluid situation for the foreseeable future.
Further, even though risk scores nationally were fairly stable, state-specific variation is what makes prediction difficult. To that end, 23 states had changes greater than 5 percent (either positive or negative) and the swings ranged from a low of a 12 percent drop for Nevada and an increase of 17 percent for Alabama.
Given the above instability, it probably comes as no surprise that there were issuer-market combinations with large dollar swings. Specifically, more than 20 percent of the issuer-markets saw swings greater than $10 million, in either direction, as shown in Figure 1.5
|Figure 1: Percentage of Issuers with Large Swings|
|Losing more than $10 million||9.4%|
|Gaining more than $10 million||11.9%|
The point here isn’t to explore whether this is the appropriate amount of stability; the point is to illustrate that significant changes in status are observed in the data, and these changes must be managed.
Pre-ACA underwriting had at least two distinct components. First, it moved the average cost levels by charging calculated surcharges to price for individual-specific risk, and by using policy exclusions, riders and denials to lower average costs. Second, it controlled tail risk. Even if a condition might not have the highest expected cost, if it carried with it an extreme tail that a carrier needed to avoid, underwriting could be used to protect against that tail risk.
Risk adjustment models, by construction, adjust for average costs, albeit average costs that are conditional upon granular slices of data. Among individuals who receive a given hierarchical condition category (HCC), there will be a mix of severities, with a minority of high-cost individuals being insufficiently risk-adjusted. Good case management might increase the proportion of profitable, risk-adjusted members both by making sure that lower-cost individuals are recognized properly as having that condition, and by helping to keep the average costs down, conditional on having the disease. Regardless, because products are guaranteed issue, carriers will be exposed to the high-cost tail risk, even after risk adjustment is applied.
Worse, individuals that are high cost may actively seek out certain carriers relative to others. Physicians may do likewise. This problem illustrates that risk adjustment, by itself, does not entirely compensate for the inability of carriers to underwrite.
One solution is to continue to create ever more granular models, with ever-finer severity levels. This would result in more complexity and increasingly fragile models. Another solution is to augment risk adjustment with some form of actual or virtual reinsurance, which currently is being considered.
The issue of tails also is likely to arise in a very different way with the risk adjustment validation processes that will have financial implications beginning with the 2016 benefit year. Although scientifically designed sampling within those processes will reduce the likelihood of a carrier being mis-identified as having systemic coding errors, given that hundreds of insurance companies will be sampled every year and given the skewed distribution of individuals with impactful HCCs, it is likely that some company at some time will be penalized erroneously.
In summary, ACA markets present significant predictive challenges, and carriers must be prepared to deal with new exposure to tail risks. Next, we will explore what these challenges may say about the sustainability of the marketplaces.
No market structure will be perfect. The key to sustainability is whether the markets can be made robust to modeling error or bad luck.
Issue 1: When Predictions Fail, What Happens?
ACA commercial risk adjustment differs from the Medicaid and Medicare risk-adjusted markets in one critical operational aspect: the size of the risk adjustment transfer depends on the concurrent level of market premium, as well as the concurrent risk scores in the market. Neither are known by the pricing actuary at the time of rate filing; in fact, the pricing actuary doesn’t even know those variables for the experience period used in rating.
This is a significant challenge. To illustrate, consider a market that currently is underpriced on average. This means risk transfers currently are too small. Over time, as pricing corrections get implemented, the market average increase will be applied within the payment transfer formula, causing corresponding changes in risk transfers. This is likely very difficult for actuaries to predict with precision, because members can change carriers, plans or exit the market altogether, and other carriers will be joining and exiting markets.
In the long run, markets may retain their resilience in the face of these problems. Once carriers get closer to establishing the correct levels of market premiums, the absolute magnitude of transfers should become easier to project. Some of the pricing volatility observed since 2014 ironically may help predict member churn within the market. And, lastly, as long as risk scores are positively correlated with claims, then there is an inherent dampening mechanism with forecast errors: underpredicting claim costs will result in higher than expected risk adjustment, dampening future year corrections. This is in stark contrast to the pre-ACA situation, where underpredicting claim costs could result in even higher rate increases.
Further, policies could be enacted to improve stability. Risk adjustment results could be published earlier and rates filed later, allowing carriers to move to equilibrium rates sooner. Risk adjustment transfers could be a function of prior year rate levels, or a published benchmark. Rate review needs to allow for the uncertainties in the risk adjustment process to be reflected adequately in rates.
Given these issues, at a minimum what we know is that achieving a stable market funding level will be a new and complex challenge, but it will take many years to establish whether the new structures are inherently fragile.
Issue 2: Introduction of Systemic Risks
Risk adjustment in each ACA market is funded from within the market. This creates a series of new systemic risks, because each carrier now is directly dependent on other carriers within the market.
- If one carrier underprices the market, this reduces market average premium and, therefore, will reduce the magnitude of risk adjustment transfers between other carriers.
- If one carrier is unable to make risk adjustment payments, this will reduce funding available to other carriers.
- If one carrier is able to significantly improve its coding intensity, then other carriers will have their risk adjustment transfers negatively impacted.
- There is also upside risk; if one carrier fails to pass its CMS EDGE validation audit or is unable to validate its diagnosis coding through chart review, then other carriers will have their risk adjustment transfers increased.
Many of these systemic risk issues should be greatest in the initial years of the ACA markets, as carriers search for appropriate pricing levels. It would be possible for a federal or state government to guarantee risk adjustment transfers to mitigate some of this systemic risk.
Issue 3: Other Fragilities
When thinking about whether the ACA markets are sustainable, it is worth acknowledging that risk adjustment can only smooth results within a market that is already at a sufficient funding level; it cannot make an unsustainable model sustainable.
Ultimately, sustainability will be governed by whether subsidies are sufficiently low to taxpayers and sufficiently high for consumers. The CBO’s latest projections indicate the cost of marketplace tax credits will rise to $70 billion in 2026.6 How will taxpayers react to this increased cost?
These tax credits, in turn, make marketplace plans affordable for consumers. Beginning in 2018, the ACA included a failsafe clause: if tax credits exceed .504 percent of GDP, they will be reduced. Although current projections are that we will stay below this threshold for some time,7 premium growth has historically exceeded GDP growth. Unless that trend reverses, at some point consumer subsidies are likely to be at risk either through this failsafe provision or from direct cost-driven pressure to modify the law.8
We are just beginning to collect the kind of data that will help us understand how to operate within a fully risk-adjusted market, and determine whether its imperfections are manageable. Risk adjustment does create some fragilities that must be overcome, but they are perhaps not significantly worse than challenges that existed previously. As a profession, we are well positioned to advise policymakers in ways they can make risk adjustment more resilient and, by extension, the new markets more sustainable.
- 1. If an insurance system is “fragile,” then it will not be able to recover from stresses caused by mispricings, anti-selection or other shocks. In contrast, Taleb created the term “antifragile” to mean the opposite of fragile, whereas “resilient” refers to things that resist damage from shocks. Taleb, Nassim Nicholas. Antifragile: Things That Gain from Disorder, 2014. ↩
- 2. This is an admitted over-simplification because the nature and enforcement of open and special enrollments can be significant. Here, however, I am concentrating on the relative competitiveness among carriers and am assuming that these rules are applied uniformly and affect all carriers equally. ↩
- 3. CMS has evaluated this particular problem extensively. See the discussion in its White Paper: “March 31, 2016, HHS-Operated Risk Adjustment Methodology Meeting: Discussion Paper,” March 24, 2016. ↩
- 4. Risk transfers are a function of much more than just risk scores; the Allowable Rating Factors (ARF) and Actuarial Values (AV) didn’t vary by as much in the data. Further, the point here is that variation will create pockets of differential impacts by state; the fact that there also is complexity in the ARF and AV in addition to what is discussed here doesn’t alter that argument.
I excluded Vermont from the data because it has a merged market. These averages are averages of the state averages from the CMS reports. California has a large membership at a very low risk score; using member-weighted averages would enlarge these differences somewhat, especially for 2014. ↩
- 5. The data released in June doesn’t contain premium data, so the absolute dollar transfers are the closest indicator we have for stability. ↩
- 6. “Federal Subsidies for Health Insurance Coverage for People Under Age 65: 2016 to 2026,” Table 2. March 2016. ↩
- 7. Projected tax credits in 2026 would be .251 percent of GDP, based upon the GDP projections from “The Budget and Economic Outlook: 2016 to 2026,” January, 2016. ↩
- 8. Page 29 of Douglass Elmendorf’s March 30, 2011, testimony to Congress has a good discussion of the sustainability tension between taxpayers, consumers, and the failsafe provision. Congressional Budget Office. “Statement of Douglas W. Elmendorf Director: CBO’s Analysis of the Major Health Care Legislation Enacted in March 2010,” Page 29. March 30, 2011. ↩