Claims and claims based risk adjustment


Claims based risk adjustment is on my mind a lot this week. 

I’m thinking that my dissertation is likely to be a Defense Against the Dark Arts dissertation on the ways that insurers and other risk bearing entities respond to incentives that are skewed. Risk adjustment is critical to aligning insurer incentives with societal incentives. I think we have a problem.  I need to think some things through.

Claims based risk adjustment is quite common.  It is used to move money to insurers that attract and insurer individuals with high, expected and plausibly predictable future medical expenses.  Some claims based risk adjustment systems are zero sum like the ACA where the money moves from insurers with low coded risk to insurers with high coded risk.  Some risk adjustment systems are externally funded like in Medicare Advantage where CMS pays out risk adjustment.  Risk adjustment is critical for any guaranteed issue, community rated(ish) products otherwise insurers will only compete on being as unattractive as possible to people with likely medical expenses and attractive to people with few if any probable expenses.  This means narrowing of networks, increased gatekeeping, prior authorization and administrative burden run-around for everyone as the game to play is hot potato with short fuse hand grenades.

Claims based risk adjustment requires claims. Claims are a function of utilization.  Utilization is a function of actual medical acuity, access to medical care (cost-sharing, transportation, administrative burdens are some potential barriers that vary based on both individual, cultural and geographic basis), and willingess to use medical care.  At a given level of acuity some people are more able and willing to go to a clinical environment for care than others for all sorts of reasons.  Any given moment of formal utilization is a non-zero chance that a risk adjustable diagnosis is entered into the formal bureaucratic system of accounting and creditation.  Any moment of non-formal utilization or outright care avoidance (rub some dirt on it medicine…) is purely a zero probability chance of creating a risk adjustable diagnosis.

Claims based risk adjustment relies on diagnoses.   A claim can have 25 diagnoses (Dx) on it.  Each Dx slot can either be empty or filled with a Dx Code.  Conditional on it being filled, there is some probability that the code is a risk adjustable code and some probability that it is not a risk adjustable code.   The number of codes that are filled in is a function of the interaction of the individual patient, the clinician, and the broader, local medical systems.  Some patients may be deemed to be more credible in reporting their symptoms that leads to a diagnosis being entered onto a claim while others may not be able to either report at all or have their reports discounted.  Some clinicians will code everything and anything, some will only code the immediate problem.  Most are somewhere in between.  Coding environments like Critical Access Hospitals will have different coding practices than tertiary academic medical centers.  The distribution of coding environments is likely not random can could be associated with various metrics of social vulnerability.

Claims based risk adjustment relies on the accumulation of novel diagnosis groups. Any diagnosis that is risk adjustable only adds incremental value to the risk adjustment transfer if it is unique.  The probability that any new diagnosis slot is filled with a unique to the individual/time period dyad diagnosis is a function of everything above plus the number of interactions.  An individual with no utilization who then has one visit will have a fairly high probability that they are adding to risk adjustment value.  An individual with 100 clinical encounters in the year having one more encounter is very unlikely to have a new diagnosis that has not been coded before on the 101st claim of the year.

If we think about the construct of “Predictable Healthcare Spending” as quasi-latent and each claim is a quasi-random draw into that well, some draws will come up with no risk adjustable diagnosis, a lot of draws for a lot of people will come up with diagnoses that were previously drawn before, and a few draws will come up with new stuff.  The new stuff is likely to happen early rather than late for prevalent conditions and quasi-randomly for new conditions.  As I’ve been thinking about this, I just keeping on thinking about the Rubber Ducky Game at school carnivals where every rubber duckie has a prize but almost all of them are small prizes but every now and then there is a rubber ducky with a big prize…. risk adjustment is not quite this random as people go to specialists because there is a prior belief that there is something big/weird happening but this is how I’m visualizing the draw process at the moment.

So where am I going with this?

If we think that the probability of a risk adjustment score is a cumulative interaction of the probabilities of an actual problem (or at least a medically defensible codeable problem) plus the probability of a patient actually having at least one clinical encounter plus the probability that a risk adjustable diagnosis is coded plus the probability that the diagnosis grouping is unique for the patient/period dyad then we need to think that these are socially constructed scores.  People who are likely, all else equal, to seek care and who are likely to get a risk adjustable diagnosis are probably meaningfully different than people who either are not able/willing to see care or conditional on seeking care, less likely to get a diagnosis that is risk adjustable for the same fundamental condition.

Risk adjustment as an economic and actuarial process moves money to insurers whose population codes as sick.  Coding as sick is both a technocratic and a social process.    Insurers will want to seek out individuals whose net revenue (premiums +  risk adjustment) is greater than expected expenses.  If we think that risk adjustment is socially skewed, it overweights people who can get coded “well”/”heavily” and underweights people who either have low utilization or have low probabilities of their underlying medical conditions being coded aggressively.




The post Claims and claims based risk adjustment appeared first on Balloon Juice.

High Deductible Health Plans, Shadow Prices and IBNR

Last week, my son had an asthma exacerbation.  We tried to drown it in albuterol but that was not working as well as he needed so we took him to the urgent care where he got a COVID test, a chest x-ray to check for pneumonia and then a scrip for some good generic drugs that have significant reduced the inflammation to the point where he could go to school this morning with just a case of the Mondays.  That night, after I dropped him off at home and while I was waiting at the pharmacy for the prescription to be filled, I thought about health insurance and incurred but not reported claims (IBNR) as this is what I do.  We’ve talked about IBNR before and I’m thinking about this for a potential project.

When I was working full time at Duke, the family was covered under my employer’s insurance.  It was damn good insurance on a narrow network (anything with “DUKE” in its name was in network, anything else, YOLO) with very low cost-sharing.  The entire episode would eventually cost me $75 in co-pays.  When I became a student, my insurance switched to the student plan while my wife and kids went to her employers’ insurance.  They now are covered by a national PPO high deductible health plan (HDHP) with a Health Savings Account (HSA).  This means that the entire contract rate for the visit and services will be paid by us until we hit a deductible.  The fundamental concept is that lower actuarial value plans with significant first dollar spending requirements will make us better and more discerning shoppers that reduce utilization for unneeded care and drive down prices for needed care.  Yeah, my son likes to breathe.  We like him to breathe.

As I was waiting at the pharmacy, I knew that I would get told exactly what I owed at the point of sale (~$15).  That is information I can theoretically use to think about my incentives for the rest of the year. My family’s  functional deductible is BIG NUMBER –  $15.  There was just enough lag in the system for me to convince myself that I really wanted the big bag of peanut butter M&Ms.

I have no idea what the actual visit will cost.  I’m not worried, we have money in the HSA and given the procedures and location, it won’t be tooooo expensive (this is privileged as fuck, I know).  But my family’s functional deductible at this moment is BIG NUMBER – $15 – [some unknown number centered on a distribution of $250 with a distribution skewed to the right].  Given that I think about insurance way too much, I am very confident that this equation produces a number well above zero.  However, there is a slight chance (<1%) that my family will hit their deductible for the year.

I think that we can effectively act as if we still have a significant deductible left, but I am not certain.  I am not certain if that deductible is big enough to make us be 100% responsible for an uncomplicated broken wrist or if it is small enough that another PCP visit or two for school yard crud diseases would eat it up.  Given that it resets on January 1, I am assuming that the deductible’s incentive effects are going to be a near constant.  If one of my kids or my wife needs medical services in the next thirty three days, we could face a shadow price that differs from the actual price if our deductible is fully met.  I don’t think that is the case, but there is a possibility that is the case.

But we are in a world of uncertainty due to IBNR.  We know the claims are out there from both a personal, experienced level, and from a societal collective statistical level.  We know that the claims will eventually be paid. We just don’t know when any particular claim will be submitted.  This creates a moment of fuzz and uncertainty as some of the incentive shaping structures fundamentally assume rapid claims payment or at least fully transparent and coherent pricing information that integrates very nicely with accumulators of previous claims payments.  Price transparency rules are likely to not be harmful to this problem.  But even full price transparency at the point of service without immediate claims submission and claims processing in the amount of time needed to impulse buy something with both chocolate and peanut butter in it only does so much.  Even in a much lower friction universe, there is significant IBNR challenges.


The post High Deductible Health Plans, Shadow Prices and IBNR appeared first on Balloon Juice.

Risk Adjustment comment letter

The Center for Medicare and Medicaid Services (CMS) published a big risk adjustment technical considerations paper in October.

I’m commenting as I’m done with most of my classes and I have the head space to engage with CMS’s thinking.  Risk adjustment is critical to creating markets where insurers want to cover sick and expensive people because those people are fundamentally profitable versus creating markets where insurers are told to cover sick and expensive people but all of the rules and systems in the market make profitability far easier if insurers avoid likely risk.  I think CMS is overthinking and overvaluing the value of a model that fits well on low cost enrollees who drive very few claims.

My letter starts below:

To whom it may concern:

I would like to offer my perspective and thoughts on the recent CMS HHS-Operated Risk Adjustment Technical Paper on Possible Model Changes published on October 26, 2021.

I am an individual health insurance policy researcher with significant expertise in the operations of the ACA individual health insurance marketplace.  My research has focused on automatic re-enrollment,1,2 dominated plan choices,2,3 consumer responses to zero premium health plans,4,5 plan availability,6 and information availability and impact in these markets.7–9   Risk adjustment is critical to all of these aspects as risk adjustments shapes profitability possibilities and thus strategies of insurers.

As you are well aware, risk adjustment serves to reduce insurer incentives to risk select.  Risk adjustment is a group level adjustment to gross revenue of an insurer so that the net revenue should, on average be sufficient to make insurers risk agnostic.10  All risk adjustment programs are, at the individual level, imperfect.  Residuals will be created where individuals with certain diagnostic categories will have notably lower or higher actual expenditures relative to the risk adjusted predicted expenditure.  These residuals can create incentives for insurers to either actively seek out and select risk or screen for risk if and only if these residuals are both reasonably predictable and of significant magnitude to be worth the administrative investment.11–13

Risk adjustment should only be used to correct for anticipated adverse selection as outlined in Section 1.1.2 “The purpose of the risk adjustment program is to reduce the influence of risk selection on plan premiums as well as to reduce the incentive for plans to avoid enrolling higher-than-average risk enrollees.”

Adverse selection is fundamentally a challenge of information asymmetry.14  Acute, unanticipated events are not subject to adverse selection.  Only when an individual has knowledge of a disease conditions and severity of a given condition at the point of the decision to purchase or not purchase health insurance does the possibility of adverse selection exist.   This perspective is in conflict with the inclusion of acute medical events like HCC-02 which can drive up significant costs (p.21)

The most expensive enrollees tended to have severe acute illness HCCs. These enrollees were often hospitalized, received ICU care, and were frequently among individuals exceeding the $1 million high-cost risk pool claim threshold.47 For example, we found that 50 percent of enrollees reaching the $1 million claims threshold have HCC 2 (Septicemia, Sepsis, Systemic Inflammatory Response Syndrome/Shock).

Very few, if any, insurance purchasers have pre-knowledge that they will be septic in the upcoming policy year.  Including acute HCCs in risk adjustment rather than re-insurance or high cost risk pooling systems perverts the purpose of risk adjustment to minimize adverse selection.  Any other policy objective should not be pursed by risk adjustment.  Other tools such as reinsurance and regulation should be utilized in these circumstances.

CMS should consider implementing the interacted HCC counts and HCC-contigent EDF model as discussed in Chapter 4 while abandoning efforts to improve model fit for non-HCC enrollees if there is any trade-off in model accuracy for individuals who drive the vast majority of plan claims liability.

Improving the Predictive Accuracy for the Very Highest Risk Enrollees

 Chapter 4 is critical to the functionality of the markets.  The highest risk individuals are likely to have high variance in medical spend relative to averages that are derived from fairly small samples.  A few outliers can significantly alter the mean. Outliers, either positive or negative, are often fairly predictable from the combination of claims history and non-claims clinical information.  In the context of hemophilia, outlier status can be predicted with a high degree of reliability if an individual beneficiary has documented inhibitions to blood clotting factor infusions.  From this reliable predictor, insurers will alter strategies to avoid individuals with high net of risk adjustment and catastrophic high cost risk pool payment costs.  If only one chapter is to be improved upon, it is critical that it is this chapter.

“enrollees with 7 or more HCCs account for only 0.2 percent of adult enrollees, they make up more than 8 percent of adult silver plan liability”

One factor that may be relevant to the exponential growth of costs for individuals with numerous HCCs and serious illness is that plans selected by individuals with known high cost, high complexity chronic illness may differ from plans that are selected by individuals with low expected medical utilization.  Recent research has found in California, using the IHA claims data warehouse, that individuals with serious illness history were far more likely to choose PPO plans, which have out of network benefits and potentially more expansive and expensive provider networks than individuals without serious illness.25  Some component of the differences in cost may be a matter of the price level per unit of service in the networks that seriously ill patients select relative to the lower unit prices in networks that are attractive to low utilizing no HCC members.  Broad networks and PPOs are likely to be more attractive to seriously ill individuals and these plan features are likely to be more expensive on a per unit basis due to either unique attributes of the star hospitals or the lack of a credible threat by the insurer to exclude high cost providers.

The interaction terms for severe illness and transplant history status are a logical, and reasonable adjustment to the model.  Significant level and number of illness can produce clinically meaningful complexity of care.

P.58 highlights the model improvement:

In particular, we found that the adoption of the proposed interacted HCC counts and proposed HCC-contingent enrollment duration factors approach in the adult models would improve prediction for enrollees at the highest percentiles of plan liability, particularly in the 10th decile, 5%, 1%, and especially 0.1%. For example, using the adult silver plan model in the 2018 enrollee-level EDGE dataset, the adoption of the proposed interacted HCC counts and proposed HCC-contingent enrollment duration factors improves the PR for the top 0.1 percent of adult silver plan enrollees from 0.91 to 0.98 and the PR for enrollees with 10 or more HCCs from 0.83 to 1.00.

 While I have reservations with the use of the predictive ratio (PR) method of model assessment as discussed below, improving model fit for individuals at the right hand tail of the distribution is a critical policy goal.

Policy Recommendation:  Continue to improve risk adjustment for the individuals with high HCC counts and focus model improvement efforts on these populations.

Appropriateness of Predictive Ratio Measures

Section 1.4 makes the claim that the appropriate measure of model performance is the predictive ratio measure:

The predictive accuracy of a risk adjustment model is typically evaluated using predictive ratios (PRs), calculated as the ratio of predicted to actual weighted mean plan liability expenditures. The predictive ratio represents how well the model has done on average at predicting plan liability for that subpopulation. If prediction is perfect, mean predicted expenditures will equal mean actual expenditures, and the PR will be 1.00.

I disagree that this is a pragmatically useful measure.  Insurers are profit seeking and will seek to attract members whose net of risk adjustment payments and premiums are greater than realized costs and insurers will use multiple mechanisms such as network design and formulary design to minimize enrollment of individuals whose net revenue is less than realized costs.11,13,15  As we know, the spending distribution in the United States for medical care has an extreme right hand skew.16  The modal individual in the United States has no reported claims while the average spending for individuals in the bottom half of the US national healthcare expenditure distribution is under $500 per year (<$42 per member per month (PMPM)).  The top decile of spending per MEPS in 2014 had a mean of $31,203 (~$2,600 PMPM).  Some of these costs are from one-off acute conditions such as an extraordinarily complicated pregnancy, or prescription of a Hepatitis-C antiviral.  However, many high-cost individuals are persistently high cost over time and thus are predictable and potentially avoidable by unscrupulous insurers. 17

Per data from ongoing work in progress, the first six deciles of HCC scores are composed of individuals who have no HCCs and thus their risk score is composed of only metal specific demographic co-efficient. These individuals are likely to have low costs.  An individual with $200 in expected expenses and $400 in actual expenses will have an individual predictive ratio of 0.5.  The risk adjustment system would be underpaying this 0 HCC individual by $16.67 PMPM. If there is significant signal in claims, enrollment, geographic and external data, insurers may balance the cost of the risk adjustment miss with the cost of strategically avoiding enrollment individuals who are statistically similar enough to this individual.  However, this is a broad category and external predictors are highly unlikely to be specific and actionable enough for insurers to select this precisely.

However, if an individual has hemophilia (HCC 066) (2021 silver HCC co-efficienct 69.705 and average national premium in 2020 of $578/month) their expected annual risk adjustment transfer before geographic and actuarial value factors is $483,474.  Hemophilia has highly variable costs.  Some individuals will have a good response to factor replacement therapy and good luck so that their realized costs are low.  However, some individuals with the same incremental risk adjustment value will have inhibitors to standard treatments and could readily cost millions of dollars.18  Even excluding the extreme example of an individual with multiple million dollar months, an individual with hemophilia who experiences  $966,947 in claims has the same predictive ratio of 0.5 as the very low cost individual in the paragraph above.

P.21 notes the underprediction of the highest cost individuals:

Very Highest-Risk Enrollees. The current models also underpredicted plan liability for the very highest-risk enrollees (that is, those in the top 0.1 percent risk percentile and those enrollees with the most HCCs). As seen in Figure 1.2 above, the current models underpredicted adult silver plan enrollees’ plan liability in the top 0.1 risk percentile by 9 percent.

A single high cost individual or a small cluster of high cost individuals may be readily predictable through the combination of claims and non-claims data.13  Given the large swings in profitability of an insurer successfully identifying the small cohorts who are likely to be significantly underpredicted within a group that has high costs, the incentives to screen by network exclusions, formulary exclusions or hassle costs are high and actionable.

The predictive ratio is scale-less.  It equally weighs small misses in overpayments and small misses in underpayments.  It equally weighs misses in the same direction of the same percentage basis on a small PMPM base and a large PMPM base despite the fact that the dollar value of the miss could differ by orders of magnitude.  The predictive ratio has is a group level metric where the average of the entire group is considered even through individual level experience within the group may be highly variant.

P.29 reinforces this incorrect view:

 Overall, we considered this to be an acceptable trade-off because across all age and sex factors, most PRs were within a tolerable threshold of +/– 5 percent (e.g., 0.95 to 1.05) and as seen Figure 2.5, the two-stage weighted approach had the major benefit of more accurately predicting the age-sex factors for the enrollees without HCCs, which is a much larger population than the enrollees with HCCs

This view prioritizes enrollment count weight rather than claims incurred weight.  Risk adjustment should prioritize claims incurred weights rather than enrollment weights.  The use of Mean Absolute Error in chapter 5 is a modest improvement to model appropriateness.

A better metric should be used.  And we have such a metric.  We can assess the PMPM standard deviation of each percentile or subpercentile slice and then seek models to minimize the standard deviations for the entire population.  This is a model which will place more attention at accurately predicting or at least minimizing the absolute variance for high cost and high HCC scoring individuals.

Policy Recommendation:  Adopt a measurement process that incorporates information at the individual level and reflects the PMPM difference in the cost of an error at various points in the HCC and expenditure distribution.

No HCC Enrollees and Lowest Risk Enrollees

Chapter 2 of the white paper is concerned about predicting risk for individuals with no HCC co-efficient other than age-sex demographic characteristics and enrollment duration factors.  I think that this is misguided.

Risk adjustment should only be concerned about mitigating adverse selection.  Adverse selection is the presence of information on one side of the contract and concealment of that information to the other side of the contract.  Individuals with full year enrollment, no HCCs, much less individuals with no chronic HCCs have no information to conceal.  Individuals with no HCCS but short enrollment spans may have hidden information.  This may lead to adverse selection which may justify risk adjustment as a mitigating policy.

More importantly, CMS has identified significant concerns about narrowness of networks, significant pre-authorization requirements, and the high cost sharing of specialty drugs in other documents.  These are symptoms of a risk adjustment system, combined with a price linked subsidy system that currently rewards insurers for minimizing premium by taking actions to actively avoid risk and claims.  Marginal active shopping enrollees are highly likely to purchase plans only on the basis of premium.19  If CMS is to take further steps that further make low risk enrollees more profitable by increasing their relative weight in risk adjustment, insurers will reasonably respond to this market signal by further narrowing networks, and designing plans that are fundamentally repugnant to individuals with significant risk all in the quest to lower premiums.

The policy justification offered in the white paper to increase relative HCC scores for individuals with no RxC or HCC factors is also misguided:

By addressing the underprediction of costs associated with lowest-risk enrollees (enrollees without HCCs and low-cost enrollees) in the risk adjustment models, we expect to encourage retention and entry into the individual and small group (including merged) markets by plans that enroll a higher proportion of this subpopulation of enrollees.

 In 2017 and 2018, insurers left the marketplaces due to the combination of significant accumulated losses and policy uncertainty.20  That trend has notably changed.6  Insurers have returned to the marketplaces.  The combination of silver loading and the temporary decrease in applicable percentages for low income enrollees has dramatically reduced the net premiums that marginal, healthy buyers face.21,22  Many individuals are now exposed to zero premium Bronze plans.4,23  Low net premiums are relevant to individuals with low risk as they are fundamentally benefit design indifferent. Insurers have reported low MLRs for the 2018-2020 time period and profitability is not a current concern.24

Figure 2.3 is extraordinarily concerning.  The two-stage weighted model markedly improves predictive value for individuals with 0 HCC who have low PMPM while making the predictive ratio no better and often marginally worse for individuals with two or more HCC.  Individuals with two or more HCC have significant claim dollar weight.  Table 4.1 shows that the bottom 7 deciles of risk scores constitute ~16.3% of total claims liability.  Adjusting the models to shift resources to these populations that are primarily purchasing plans on the basis of premium will further encourage insurers to compete by avoiding risk even more aggressively.

Enrollment Duration Factors

 While we are not certain why enrollment duration is only inversely related to monthly costs among adults with HCCs, the most plausible hypothesis is that medical treatment for many HCC diagnoses likely has a “fixed cost” element that does not vary with the number of months of enrollment.

 There may be two factors at work that may explain the observation that very short duration enrollments with HCCs may have higher PMPM in addition to the very logical proposal that most medical expenses are incurred in short time windows.26  First, many individuals may be automatically re-enrolled into plans that in the first year had zero net premium and then a positive sum net premium in the second year.  This may create an administrative friction that leads to disenrollment after medical expenses are incurred in the first month. 5  Secondly, the inclusion of acute conditions in the model may dissipate some of the enrollment duration factors as these acute conditions may absorb most of the variance that the model needs to make a stable estimate.

Common Data Resource

In section 3.4 Obtaining Diagnosis Codes for Partial-Year Enrollees  there is a concern that insurers have differential abilities to collect and submit relevant diagnosis on the basis of enrollment duration interacting insurer sophistication, size or business model.  CMS may be able to lower the barrier to entry to markets by creating a list of individuals with truly chronic conditions (HCC-018 for instance) where insurers will receive credit for the medical history of individuals with these conditions even if claims are not submitted.  This would require significant revamping of the data sharing systems and would not immediately lower barriers to entry nor immediately increase competition but over the long run, assuming privacy concerns can be addressed, this would improve the markets’ functionality.

Scale of Administrative Carve-outs

Per the 2018 NBPP:

Reducing the statewide average premium: We finalized a 14 percent reduction to the statewide average premium in the state payment transfer formula, beginning with the 2018 benefit year, to reflect the portion of administrative costs that does not vary by claims

Insurers have distinct business models and plans that vary in scale.  Large insurers have higher degrees of administrative complexity and control than insurers with low membership.  Some elements such as exchange user fees and risk adjustment fees are constant across all insurers.  However, other operational considerations such as the cost of maintaining a claims payment system or contracting such a system out to a third-party vendor will vary by insurer size and sophistication.   The reduction in state wide average premium may be disproportionally beneficial to insurers that seek to minimize their exposure to risk by narrowing networks, instituting prior-authorization requirements and offer restrictive formularies as well as offering low premium plans with the lowest allowable actuarial value for each metal band.27

Policy Recommendation:  Re-analyze the size and scope of the reduction of average state wide premium by insurer size and market competitiveness.  Adjust this on a state or regional basis as justified by the re-analysis.

Risk Adjustment and Section 1332 Waivers

 One area that was not mentioned in the white paper is the interaction of risk adjustment and reinsurance waivers.  This is a critical challenge as Section 1332 reinsurance waivers have proliferated since the initial approvals for the 2018 plan year.28  Reinsurance has been critical in reducing gross premiums.29,30  States have adopted several methods.  Most notably some states have adopted disease specific models where the reinsurance pool pays claims for individuals with certain, pre-specified diseases that are likely to be high cost.  This model reduces right hand tail risk.  Most states that have adopted a reinsurance waiver has instead used a caliper model where the state reinsurance funds are used to pay some percentage of claims between an attachment point and a ceiling.   Both of these models likely cover significant portion of costs that are currently accounted for in the risk adjustment model and this may lead to double counting.

CMS has partially recognized the challenge of double counting with the operation of the national high cost catastrophic reinsurance program where CMS pays a portion of an individual claims in excess of $1,000,000.  CMS has truncated the incremental value of these claims in the calculation of disease category coefficiencts so that insurers are compensated for the portion of risk that they bear rather than the gross risk.

However with state reinsurance programs, there is no state specific models which accounts for the portion of predictable high cost claims that the insurer receives credit for in terms of risk adjustment transfers but also receives funding from the state reinsurance pool.  This problem is likely to be particularly acute in states, like Colorado and Georgia,31 with multiple reinsurance models operating concurrently over various sub-state geographies.

Policy Recommendation:  CMS should require states that operate 1332 reinsurance programs to adjust disease specific coefficients to minimize double credit of both risk adjustment and reinsurance payments.










  1. Drake C, Anderson DM. Association Between Having an Automatic Reenrollment Option and Reenrollment in the Health Insurance Marketplaces. JAMA Intern Med. 2019;179(12):1725-1726. doi:10.1001/jamainternmed.2019.3717
  2. Anderson DM, Rasmussen PW, Drake C. Estimated Plan Enrollment Outcomes After Changes to US Health Insurance Marketplace Automatic Renewal Rules. JAMA Health Forum. 2021;2(7):e211642. doi:10.1001/jamahealthforum.2021.1642
  3. Rasmussen PW, Anderson D. When All That Glitters Is Gold: Dominated Plan Choice on Covered California for the 2018 Plan Year. Milbank Q. 2021;n/a(n/a). doi:10.1111/1468-0009.12518
  4. Drake C, Anderson DM. Terminating Cost-Sharing Reduction Subsidy Payments: The Impact Of Marketplace Zero-Dollar Premium Plans On Enrollment. Health Aff (Millwood). 2020;39(1):41-49. doi:10.1377/hlthaff.2019.00345
  5. Drake C, Cai ST, Anderson D, Sacks DW. Financial Transaction Costs Reduce Benefit Take-Up: Evidence from Zero-Premium Health Plans in Colorado. Social Science Research Network; 2021. doi:10.2139/ssrn.3743009
  6. Anderson DM, Griffith KN. Increasing Insurance Choices In The Affordable Care Act Marketplaces, 2018–21. Health Aff (Millwood). 2021;40(11):1706-1712. doi:10.1377/hlthaff.2020.02058
  7. Shafer PR, Anderson DM, Aquino SM, Baum LM, Fowler EF, Gollust SE. Competing Public and Private Television Advertising Campaigns and Marketplace Enrollment for 2015 to 2018. RSF Russell Sage Found J Soc Sci. 2020;6(2):85-112. doi:10.7758/RSF.2020.6.2.04
  8. Shafer PR, Anderson DM, Baum L, Franklin Fowler E, Gollust SE. Changes in Marketplace Competition and Television Advertising by Insurers. Am J Manag Care. 2021;27(8):323-328.
  9. Anderson D, Shafer P. The Trump Effect: Postinauguration Changes in Marketplace Enrollment. J Health Polit Policy Law. Published online 2019. doi:10.1215/03616878-7611623
  10. McGuire TG, Schillo S, van Kleef RC. Reinsurance, Repayments, and Risk Adjustment in Individual Health Insurance: Germany, the Netherlands, and the US Marketplaces. Am J Health Econ. 2019;6(1):139-168. doi:10.1086/706796
  11. McGuire TG, Schillo S, van Kleef RC. Very high and low residual spenders in private health insurance markets: Germany, The Netherlands and the U.S. Marketplaces. Eur J Health Econ. 2021;22(1):35-50. doi:10.1007/s10198-020-01227-3
  12. Farid MS. Tail People: The Extremely Under and Overcompensated in Individual Health Insurance Markets. undefined. Published online 2019. Accessed October 3, 2021.
  13. Geruso M, Layton T, Prinz D. Screening in Contract Design: Evidence from the ACA Health Insurance Exchanges. Am Econ J Econ Policy. 2019;11(2):64-107. doi:10.1257/pol.20170014
  14. Handel BR, Kolstad JT, Spinnewijn J. Information Frictions and Adverse Selection: Policy Interventions in Health Insurance Markets. Rev Econ Stat. 2018;101(2):326-340. doi:10.1162/rest_a_00773
  15. Shepard M. Hospital Network Competition and Adverse Selection: Evidence from the Massachusetts Health Insurance Exchange. National Bureau of Economic Research; 2016. doi:10.3386/w22600
  16. Mitchell EM. Concentration of Health Expenditures and Selected Characteristics of High Spenders, U.S. Civilian Noninstitutionalized Population, 2016 – Abstract – Europe PMC. Agency for Healthcare Research and Quality. Published May 14, 2019. Accessed May 14, 2021.
  17. Figueroa JF, Zhou X, Jha AK. Characteristics And Spending Patterns Of Persistently High-Cost Medicare Patients. Health Aff (Millwood). 2019;38(1):107-114. doi:10.1377/hlthaff.2018.05160
  18. Bryan B. Hemophiliac Iowa Teenager Costs $12 Million Insurance, Obamacare. Insider. Published June 1, 2017. Accessed May 24, 2021.
  19. Marquis MS, Buntin MB, Escarce JJ, Kapur K. The role of product design in consumers’ choices in the individual insurance market. Health Serv Res. 2007;42(6 Pt 1):2194-2223; discussion 2294-2323. doi:10.1111/j.1475-6773.2007.00726.x
  20. Griffith K, Jones DK, Sommers BD. Diminishing Insurance Choices In The Affordable Care Act Marketplaces: A County-Based Analysis. Health Aff (Millwood). 2018;37(10):1678-1684. doi:10.1377/hlthaff.2018.0701
  21. Norris L. How the American Rescue Plan Act will boost marketplace premium subsidies. Published March 5, 2021. Accessed May 24, 2021. /blog/how-the-american-rescue-plan-act-would-boost-marketplace-premium-subsidies/
  22. Anderson D, Abraham JM, Drake C. Rural-Urban Differences In Individual-Market Health Plan Affordability After Subsidy Payment Cuts. Health Aff (Millwood). 2019;38(12):2032-2040. doi:10.1377/hlthaff.2019.00917
  23. Branham DK, DeLeire T. Zero-Premium Health Insurance Plans Became More Prevalent In Federal Marketplaces In 2018. Health Aff (Millwood). 2019;38(5):820-825. doi:10.1377/hlthaff.2018.05392
  24. Keith K. ACA Round-Up: Record-High Medical Loss Ratio Rebates, Pass-Through Funding, Preventive Services | Health Affairs Blog. Published November 17, 2020. Accessed April 2, 2021.
  25. Anderson D, Yanagihara D, Japinga M, Saunders R. Gauging the Experiences of the Seriously Ill in California: Analyzing Serious Illness Beyond Medicare Fee-for-Service. Am J Accountable Care. 2020;8(2). Accessed May 20, 2021.–the-seriously-ill-in-california-analyzing-serious-illness-beyond-medicare-feeforservice
  26. Chen S, Shafer PR, Dusetzina SB, Horný M. Annual Out-Of-Pocket Spending Clusters Within Short Time Intervals: Implications For Health Care Affordability. Health Aff (Millwood). 2021;40(2):274-280. doi:10.1377/hlthaff.2020.00714
  27. Anderson Da. How Silver Loading Impacts Insurance Markets Depends On State And Insurer Decisions. Health Affairs Blog. Published June 27, 2019. Accessed November 21, 2021.
  28. Tracking Section 1332 State Innovation Waivers. Kaiser Family Foundation. Published November 1, 2020. Accessed August 25, 2021.
  29. Sloan C, Rosacker N. State-Run Reinsurance Programs Reduce ACA Premiums by 16.9% on Average. Avalere Health. Published October 29, 2019. Accessed May 24, 2021.
  30. State Innovation Waivers: State-Based Reinsurance Programs.; 2021. Accessed October 3, 2021.
  31. Colorado’s reinsurance program has been lauded as a way to reduce health care costs. Here’s the fine print. – The Colorado Sun. Accessed November 18, 2019.




The post Risk Adjustment comment letter appeared first on Balloon Juice.

COVID, disparities and risk adjustment

Risk adjustment is critical to the functionality of health insurance markets with guaranteed issued and community rating.  These markets include the ACA individual market, Medicare Advantage, Medicaid managed care and provider based alternative payment models like the Accountable Care Organizations (ACOs) that are increasingly common.  Risk adjustment moves money to insurers that are covering a sicker population that is expected to use a lot of services and cost a lot of money.  Bad risk adjustment or no risk adjustment means that insurers are primarily competing on their ability to identify and avoid people who are likely to be expensive.  Good enough risk adjustment means that insurers are competing on service and ability to control costs while being either risk agnostic or risk seeking.

Most risk adjustment systems rely on claims.  Claims have diagnosis and service codes that are entered by the treating clinician that describe what is happening with a patient at a point in time.  There is an underlying dynamic that for risk adjustment to work, there have to be claims and for claims to be legit, there has to be people going to the doctor and hospital.

We know that people go to the doctors at different rates.  Individuals with really good insurance with broad networks and low cost sharing are more likely to go to the doctor than individuals with high cost-sharing insurance and narrow networks.  We know that people who are used to navigating complex bureaucratic systems can navigate the medical system better than people who are disenfranchised, marginalized and who face significant administrative burden and barriers to care.

If we need claims to fuel risk adjustment and the probability of a claim for a given condition is a function of a person’s health, socio-economic status, power, insurance characteristics, and embedded community dynamics, then we should expect, all else being equal, for risk adjustment to be a bit heavy on claims from high status individuals and a bit light on claims from low status individuals who face a lot of barriers to care so that they don’t present as often as the path to getting to the doctor’s office has a lot of potholes, detours and dead ends.

And this is an assumption that we might make in “normal-ish” times.

Yesterday, I was at a seminar where the presented showed us some descriptive and not yet published data.  They were looking at a large claims universe pre and post COVID.  Much like the work led by Paul Shafer that I was on where we looked at North Carolina Medicaid enrollment by Social Vulnerability Index(SVI)  pre and post-COVID, these researchers also looked at SVI.  They looked at how SVI interacted with post-COVID utilization stratified by a bunch of reasonable demographic cuts.  (I’m trying to be useful but vague to respect their publication probabilities)  

Unsurprisingly,  utilization for everyone in their sample dropped dramatically in April 2020.  They then looked at how service utilization bounced back by SVI and demographic cuts.  Controlling for a bunch of reasonable covariates, they found that individuals from low SVI areas had much lower and slower service bounce backs than people from high SVI areas.

This is fascinating on multiple levels.  The area that I immediately dug into was the disparities of risk adjustment.  We’re putting into a system that is already somewhat skewed to better connected individuals an even larger skew because the baseline trend differences in service utilization by SVI and thus claims generation by SVI is even more skewed by connectiveness and privilege in the post-COVID utilization universe.

So what does this mean (besides a potential dissertation question for me as suggested by one of the department’s senior professors)?

IF my inkling is right, then risk adjustment that uses anything from 2020 (and potentially 2021) will be more variant from actual spending with larger residuals than previous year risk adjustment models.  Larger residuals between what individuals with a certain coded profile deliver an insurer in revenue through the combination of premium and risk adjustment flows and what they actually cost increases the incentives for insurers to find ways to cherry pick.  Insurers will more aggressively screen for people who have a big risk adjustment number but relatively low costs.  Insurers will more aggressively try to find ways to be ugly to people who have small risk adjustment numbers but comparatively high costs.

Getting these market incentives right is, in the best of times, a big challenge.  Usually we should aim for good enough.  However, assuming the data at the seminar is correct and my inkling that diagnosis codes collected are also skewed because of utilization disparities that were exacerbated by COVID, the incentives won’t be as good as we’ve gotten used to over the past couple of years.

The post COVID, disparities and risk adjustment appeared first on Balloon Juice.

Standard Plans and simplifying the choice space

My friend, colleague and frequent co-author Dr. Petra Rasmussen and her co-author Dr. Erin Taylor have a recent opinion piece in JAMA Health Forum on a matter of intense personal interest to me: How can the choice space for health insurance on be simplified?

While Americans value choice, research in behavioral economics and psychology has shown that people struggle with the decision-making process when they encounter too many choices and end up relying on biases and quick tricks, also known as heuristics, to cull their options and help them make a decision. This can lead individuals to choose a plan that may not be the best for them. With health insurance, the decision can be particularly challenging, and mistakes can be costly…
Without greater standardization, the use of metal-tier labels alone cannot successfully convey helpful information to consumers about the plans…

HHS should learn from what states like California and federal agencies such as CMS have done to standardize Marketplace and Part D plans. To limit choice overload and ease the decision-making process for enrollees, HHS should consider replicating what Medicare and Covered California have done by requiring plans sold by the same insurer to be meaningfully different from one another.

I think that they are heading down the right path. I am a bit reluctant to embrace meaningful difference as the appropriate pathway though as the Obama Administration definition of meaningful difference was one difference on seven different vectors. The difference could be a $100 change in deductible or eligibility for a Health Savings Account. These differences are real, but small. Are they truly decision altering differences or are they means where a cynical plan designer can tweak their product offerings to dominate virtual and mental shelf space? I inkle on the later.

Instead, I think there is another approach that might allow for meaningful variation and differentiation. We have strong evidence that the marginal enrollees are buying basically on premium. Quality, network, customer service, brand recognition are, at most, secondary concerns to the marginal enrollee. So let’s embrace that.

I would offer that insurers when rolled up to the corporate level should be able to offer whatever they want in each metal tier but there must be a 3% or greater difference in baseline premium for each plan that is offered. Under this schema, an insurer that really wants to offer a no deductible but high co-pay and coinsurance plan can do so but they can’t offer a scratched mirror version of that plan which costs eighty seven cents more per month. Instead they can offer a barebones gold, a middle gold and a high gold plan at 76, 79 and 82% actuarial value. An insurer could also offer a skinny network and a big network with no gatekeeping where the benefits are the same but the network is the premium difference driver.

I think doing this will minimize the number of counties with 150 or more choices while still allowing insurers to offer plans that actually have value beyond throwing decision making for a loop.

The post Standard Plans and simplifying the choice space appeared first on Balloon Juice.