Just a quick note about something I’ve been thinking about… saving the Big One (my manifesto, if you will) that gets into similar topics in a lot more depth for tomorrow. Or perhaps later in the week — just depends on how productive I can be.
For anybody that doesn’t know (I do have some readers outside the US, weirdly enough), the US does have two “public” insurers, Medicare which primarily covers people over 65 and Medicaid which covers very low-income people. The Medicare Advantage program, which began in 1999 — although it wasn’t named Medicare Advantage until 2003 — essentially amounts to a privatization of Medicare. Medicare Advantage allows private insurers to offer all the same benefits that traditional Medicare offers, as well as some other stuff that isn’t covered by the core components of traditional Medicare. Enrollment in Medicare Advantage plans really started to take off after the passage of the Affordable Care Act in 2010, and has been growing especially quickly in the last few years. By 2018, the share of Medicare beneficiaries enrolled in Medicare Advantage private plans had risen to 37%, to 48% in 2022, and finally to 54% of all Medicare beneficiaries in 2024, some 33 million people.
The Medicare Advantage “space” is dominated by a small number of private insurers, with Humana and United Healthcare insuring nearly half (47%) of all Medicare Advantage enrollees in the country. I am going to write more about the behemoth that is United Health Group (which has an insurer, United Health Care, and another business unit, Optum, which aggregates a number of different businesses under its aegis), but briefly: United Health Group has grown via aggressive acquisitions over the past 10-20 years to make up a huge portion of the health care industry itself, and its profitability depends in large part on paying itself via deals between its different business subunits, a practice euphemistically termed “flywheeling” in the business press. More on this later — again, perhaps tomorrow. For now, I’m going to focus on the insurance side and Medicare Advantage.
The majority of enrollment in United Health Care has come from Medicare Advantage in the last 10 years, and Medicare Advantage is extremely profitable. According to this post, United Health Care “now takes in nearly twice as much revenue from the 7.8 million people enrolled in that program as it does from the 29.6 million enrolled in its commercial insurance plans in the United States.”
One of the ways it’s able to squeeze twice as much revenue out of ~8 million Medicare Advantage enrollees as from ~30 million commercial enrollees this is by hosing the federal government for money by exploiting a weird rule in Medicare Advantage. In other words, by fraud. Typically, a diagnosis is made by a doctor. For some inscrutable reason, known only to the cost-benefit sages who administer the criminal cartel known as “The United States’s health care system,” Medicare Advantage permits insurers to add their own diagnoses to patients’ records, on top of the diagnoses that doctors make.
How’s that going? Well, about how you’d think. A Wall Street Journal investigation published in July of this year (all the figures cited below come from this investigation) documented that, over the years 2019-2021, Medicare paid private insurers $50 billion (that’s a plosive Carl Sagan “b”) for diagnoses added by insurers. Of United Health Care’s $17 billion in 2021 revenue, $8.7 billion (that’s 51.2% for your fractions nerds out there) came from these phantom insurer diagnoses, for which patients never received any kind of treatment or follow-up. (The investigation also notes that in September of last year, health insurer Cigna paid a paltry $172 million in fines to the Department of Justice last year to settle a civil fraud case over exactly this practice — what a joke.) The types of diagnoses that insurers add are for things that Medicare pays a lot for — diabetic cataracts, for example, which Medicare pays insurers roughly $2900 for, compared to old-age cataracts (for which Medicare pays nothing). The list of fraudulent diagnoses also include lucrative ones like HIV, Parkinson’s disease, and schizophrenia. It’s literally just fraud, committed in plain sight, and it’s a huge moneymaker. Private insurers and health care conglomerates like United Health Group pay out billions in campaign contributions every cycle, so is it any wonder that neither Congress nor the DOJ seems interested in punishing this practice more than symbolically, let alone actually curtailing it? Everyone wins, except, of course, for Medicare enrollees.
This has all been context for what I really want to say (this is one of the most vexing things about the Cthulhu of American health insurance — one has got to wade through so much pointless complexity and detail just to get to the extremely obvious points about it that one is trying to make). What I really want to say is something fairly simple, and ultimately not that important, about knowledge production.
Despite the etymology of the word (from dare, “to give”), data are not simply given; they are actively constructed. A good first-year statistics teacher will caution you against applying any models until you can form some kind of picture in your mind of the “data-generating process.” (For example, hospital length-of stay data tend to correspond to a classic Poisson process, a big hump around 0-1 days with a long tail trailing off to the right.) That data-generating process includes not just whatever the causal mechanism is that produces the actual numbers, but the whole social process of encoding them as numbers in the first place. A huge amount of research in the US uses diagnosis codes for billing as a proxy for underlying prevalence of various disease conditions; the data-generating process in these cases includes everything that influences whether and how a diagnosis gets written down — whether a person saw a doctor to get a diagnosis in the first place, whether the practice they visited engages in “upcoding,” which ICD-10 codes are available to describe the patient’s condition.
As a brief detour and illustration, I wrote my dissertation about a perinatal health construct called “severe maternal morbidity” (SMM). This is supposed to capture something like a “near-miss” for maternal death; the WHO and many other countries use some version of a “near-miss” definition. In the US and Canada, a lot of this research is done using administrative (i.e., billing) data, which was the impetus for the creation of a diagnosis code-based definition of a similar construct — hence SMM. One of the classic potentially life-threatening complications of labor and delivery that this construct needs to capture is obstetric hemorrhage. Receiving 4 or more units of blood while a person is in labor is an excellent indicator of an obstetric hemorrhage. However, the volume of blood transfused is only listed on the medical record. The administrative or billing record indicates only whether a transfusion was performed or not, not the volume of blood transfused. Using the algorithm to identify SMM based on diagnosis codes, then, incorrectly identifies a significant number of people as having had an SMM event (a hemorrhage) when in fact, they didn’t — receiving a smaller amount of blood during labor and delivery is pretty common and not necessarily indicative of a serious complication. I did part of my dissertation research using data from Medicaid — administrative data — and was thus limited to using this diagnosis code identification algorithm and locked in to a certain amount of intractable inaccuracy in my exposure assessment (for that question, SMM was the exposure I was interested in — I was investigating whether an SMM event in the perinatal period was associated with increased risk of cardiovascular complications later on). This is one example of how the “data generating process” around medical billing affects the knowledge that can be produced. National estimates of the prevalence of SMM, which form the basis of public health and hospital-based planning, are derived from administrative databases, using the same diagnosis code-based identification algorithm.
Phantom insurer-authored diagnoses in Medicare Advantage is another example of what I’m talking about, with an even more evil edge. It’s an example of how increasing privatization, the market and profit-seeking incentives that introduces, and fraudulent practices among gigantic health care companies distort the actual public health evidence base on which so much depends. There are a lot of reasons, public health and otherwise, why researchers might want or need to know the prevalence of, say, HIV among Medicare beneficiaries. The data-generating process here involves fraud, which has ramifications for more than just the consistency properties of the estimators used to do the statistical analysis.
United Health Group is exemplary here not because its practices are unique but because it is so huge that it distorts the entire field and practice of health care around it, like a black hole, a bowling ball on a mattress. The foregoing is not the most morally or otherwise compelling case against the existence of United Health Group, with its aggressive vertical integration strategy and dirty-dealing flywheeling, or against private health insurance generally, but it is one case. It shows how knowledge is socially constructed, for one; for another, it shows how profit-seeking bends, like Joe Magarac with a steel beam — not the “cost curve,” as health insurance execs are fond of saying — but the entire knowledge process and the entire process of medical care.