Commentary on: The effects of coding intensity in Medicare advantage on plan benefits and finances

One of the more surprising health policy developments in the decade since the Affordable Care Act (ACA) was enacted in 2010 has been the steady growth in the market share of the Medicare Advantage (MA) program. In 2010, around a quarter of Medicare beneficiaries were enrolled in MA; by November 2020, 43% of eligible beneficiaries were. ¹ (Eligibility for MA requires enrollment in both Parts A and B; the MA market share among all Medicare beneficiaries is somewhat lower.)

When the ACA was being drafted in 2009‐2010, MA reimbursement was viewed as excessive, and so reductions in MA reimbursement became one source of financing for the ACA’s insurance expansion. In scoring the ACA, the Congressional Budget Office estimated the ACA’s statutory reductions in MA reimbursement would save $136 billion over the following decade, about a seventh of the projected cost over that period. ² These reductions led both the Centers for Medicare & Medicaid Services (CMS) actuaries and the Congressional Budget Office to predict that MA’s market share would fall by nearly a factor of two over the next decade, one of the more spectacularly wrong predictions of the post‐ACA future. ³ , ⁴

Why did enrollment rise instead of fall? For several years, Richard Kronick has pointed to coding effects as a potential reason. ⁵ , ⁶ For readers not well versed in the details of MA reimbursement, I give a bit of background. MA reimbursement for a beneficiary is a function of the annual cost of treating a similar patient in Traditional Medicare (TM). When the MA program began in 1985, patient similarity was established by classifying TM beneficiaries into cells using characteristics known as risk adjusters, and the average cost of beneficiaries in each cell was computed. At the time CMS had a sparse list of adjusters to establish similarity, with most of the work being done by age and sex. As a result, if an MA plan could enroll beneficiaries who were healthier than the TM beneficiaries within an age‐sex cell—for example enrolling healthy females 65‐69 years of age—it could make money. As an antidote to the resulting favorable selection into MA, the Balanced Budget Act of 1997 mandated that CMS include health status among its risk adjusters. Responding to that legislation, in 2000 CMS first put a toe in the water by adding an adjustment for diagnoses recorded on inpatient claims but only with a 10% weight. In 2004, however, it began a 4‐year transition to incorporate diagnoses appearing on outpatient claims as well. By the time the transition was complete in 2007, diagnoses had replaced age and sex as doing most of the work in risk adjustment, that is, adjusting MA reimbursement for dissimilarity between TM and MA patients.

Although it was conceptually obvious that patients with a disease such as cancer were costlier to treat than those with no disease so that beneficiaries in an age‐sex‐diagnosis cell were more comparable than those in an age‐sex cell, there was a problem. By the late 1990s, inpatient diagnoses were relatively completely coded on claims because they were the cornerstone of hospital reimbursement under the Diagnosis Related Group (DRG) system. Diagnosis, however, played no role in the reimbursement of physician services. As a result, diagnoses recorded on physician claims, especially comorbidities, were often incomplete for patients who had only made office visits and not been hospitalized. And such patients were the great majority of patients.

Under the new MA reimbursement system, coding additional diagnoses—including those not associated with an inpatient admission—generally raised MA reimbursement, thereby giving MA plans a strong incentive to ensure their enrollees had all diagnoses coded. Indeed, CMS did not even require an office visit to record a diagnosis; if a nurse making a home visit could establish a diagnosis, that counted for purposes of adjusting MA reimbursement. As a result, the less intensive coding in Traditional Medicare meant the same patient in MA might be equally costly to treat as in TM but would appear as more costly simply because the TM patient would not have as many comorbidities recorded on claims. This problem was not overlooked by the drafters of the ACA, who specified in statute annual reductions in MA reimbursement for the more intensive MA coding that was anticipated. These adjustments began as a 3.4% annual reduction in 2010 and have now grown to a 5.9% annual reduction.

This statutory reduction in MA reimbursement for coding intensity raises two policy questions. The first is whether it is large enough to compensate for the greater coding intensity. The second assumes it has not been large enough and asks how the resulting additional reimbursement has been ultimately split between beneficiaries and plans. In economic terms, the second question concerns the incidence of any additional reimbursement. (In principle, the answer to the second question would apply to a reduction in reimbursement from an excessive coding adjustment as well.) The answer to the incidence question turns on the competitiveness of the local MA market; at one extreme, a monopolist would appropriate all of the increase while at the other extreme in a perfectly competitive market all of the increase would pass through to consumers. In this issue, Jacobs and Kronick take up the second question. ⁷

The methodological problem in addressing both questions is to distinguish any true change in the health status of MA beneficiaries relative to TM beneficiaries from more intensive coding in MA. The former would presumptively affect plan costs, while the latter would simply increase plan profit (net of any costs incurred to code more intensively). A similar issue arose when Medicare implemented the Inpatient Prospective Payment System in 1983 using DRGs; previously Medicare had simply reimbursed a hospital its share of the hospital's total cost, but in October 1983 it began to reimburse most hospitals according to their Medicare patients’ diagnoses—and coding additional diagnoses often raised hospital reimbursement. As a result, just as in the MA reimbursement scheme, hospitals had an incentive to code diagnoses on claims more completely. When more diagnosis codes began to appear on hospital claims, the question was whether this was simply an artifact of more intensive coding or whether the mix of Medicare inpatients really was sicker and hence more costly to treat. At that time, a study used medical chart review as a gold standard to estimate that about two‐thirds of the observed change in the diagnoses on claims was true change, presumably because care for some relatively less sick patients such as those undergoing cataract surgery was shifting out of the hospital. ⁸

Jacobs and Kronick is the fourth study to take up the question of incidence of MA reimbursement changes. The four studies have adopted three different approaches to the methodological problem that more diagnoses recorded on claims could reflect either more intensive coding or sicker patients. Song et al used diagnoses on lagged claims to predict current risk scores. ⁹ Duggan et al and Cabral et al exploited exogenous legislative changes that increased MA reimbursement in certain local areas. ¹⁰ , ¹¹ Jacobs and Kronick use a risk score predicted from drug claims and standard demographic variables such as age and sex. Because drugs are not used in MA risk adjustment, there is no incentive for more intensive coding of drugs—and virtually all drugs purchased are recorded in Part D claims in any event. The drugs a person purchases, however, are a function of the person's diagnoses and hence predict the cost of treatment.

Although their estimates have some variation depending on the exact model, Jacobs and Kronick find that less than half of the increase in plan revenue is passed through to beneficiaries in the form of reduced premiums and cost sharing or as additional covered benefits such as dental care. Song et al and Cabral et al estimate that about half the increase in reimbursement from a higher risk score is passed through to beneficiaries, whereas Duggan et al find a much smaller percentage is passed through to beneficiaries, only 13 percent.

Consistent with standard economic theory, all four studies find that more of a reimbursement change is passed through to beneficiaries in more competitive local markets. Although not surprising, this is an important observation because MA depends on plan competition to serve beneficiaries well. Indeed, the Song et al study was used by the Department of Justice in successfully opposing the proposed 2016 merger of Aetna and Humana. The Department argued that Song et al’s finding that only half the change in reimbursement passed through to beneficiaries showed the market was already insufficiently competitive and that the proposed merger would exacerbate the problem.

Unfortunately, none of the four studies can answer the first question, how much more, if at all, should MA reimbursement now be reduced to adjust for coding effects? After accounting for the statutory 5.9% reduction in reimbursement in 2019, MedPAC estimated that in 2019 payments to plans were 3.2% too high because of coding. ¹ Whether they are still too high after the 5.9% reductions in 2020 and 2021 is unknown.

Kronick earlier found that from 2004 to 2014 risk scores of MA beneficiaries increased 1.7% more than those of TM beneficiaries annually, a difference he attributed to differential coding intensity. ⁶ Both Kronick and MedPAC assume that the excess increase in MA risk scores over TM risk scores comes from greater coding intensity in MA. Although the excess could reflect coding, it could also reflect a true effect from regression to the mean of those enrolling in MA. In particular, in their first year of enrollment, age‐sex‐Medicaid–adjusted mortality rates of cohorts enrolling in MA are well below the rates of those enrolling in TM. Over time, however, they rise toward the mortality rate in TM, suggesting that over time the Medicare Advantage cohort may be getting sicker relative to the TM cohort. ¹² In short, more intensive coding and increasing sickness could both be at play.

Two other pieces of evidence are suggestive that not all of the difference in the rate of increase between MA and TM risk scores is attributable to differential coding intensity. The first is that the coding increases have not subsided. Coding additional diagnoses that a patient actually has is entirely legal, but short of fraud there is a limit to how many diagnoses can be coded for any given patient, and there is money on the table for MA plans to code all diagnoses as soon as possible. As a result, one would expect coding effects to asymptote rather quickly. For example, in 2007, CMS replaced the DRG system with the MS‐DRG system, giving hospitals a new incentive to code more intensively; this new incentive appeared to cause an increase of about 5% in the risk scores of hospitalized patients insured by Traditional Medicare, but that increase took place over just two years and then subsided. ¹³ The MA risk score increases have been not only much larger—from 2008‐2019 they cumulated to 111%—they have extended over a much longer period of time and show no sign of abating. ¹

The second piece of evidence is the variation in the growth of risk scores across states shown by Jacobs and Kronick and the even greater variation across MA contracts shown by MedPAC. All plans have an incentive to code intensively, but the change in risk scores in 2018 across contracts relative to the local Traditional Medicare reimbursement varied in reasonably continuous fashion from 10 percent less to 40 percent more. ¹⁴ By 2018, MA plans had had more than a decade to adjust to the new coding system, so to attribute this amount of variation to plan differences in coding aggressiveness does not seem plausible and suggests that some of the difference in the growth of risk scores across states and plans could be from differences in the true change in MA beneficiary health relative to TM beneficiary health.

In sum, Jacobs and Kronick, building on Kronick's earlier work, have brought coding intensity by MA plans to the fore as a first‐order policy issue, with major implications for the entire federal budget. Although the adequacy of the current adjustment for coding remains uncertain, Jacobs and Kronick do demonstrate that historically the MA market has been imperfectly competitive. Entry into MA, however, has been increasing rapidly; the average number of beneficiary‐weighted plan choices rose from 18 in 2017 to 32 in 2021. ¹ The degree to which this increased level of competition will increase the share of reimbursement changes that accrue to beneficiaries rather than plans remains on the table for future research.

Newhouse JP. Commentary on: The effects of coding intensity in Medicare advantage on plan benefits and finances. Health Serv Res. 2021;56:175–177. 10.1111/1475-6773.13639