By Lambert Strether of Corrente
Readers will recalll I regularly lambaste liberal Democrats for their love of complex eligibility requirements (which function (a) as a Jobs Gaurantee for credentialed gatekeepers, and (b) as an opportunity for endless moralizing about who is worthy of government assistance, and how much). Well, today I get to do a happy dance, because an interesting paper — sadly, a year old — from the National Bureau of Economic Research crossed my Twitter timeline (via; source): from Amitabh Chandra, Evan Flack, and Ziad Obermeyer, “The Health Costs of Cost-Sharing,” Working Paper 28439. Turns out that complex eligibility requirements are even worse than I imagined they would be (and liberal Democrats more culpable).
This will be a very short and simple post. I will first present a portion of the methodology of the paper (it involves a natural experiment), and then summary results as the authors worked them out. (Since the statistics involved are w-a-a-a-y above my paygrade, the presented methodology is, alas, more a “proof of seriousness” than anything else. Also, the paper is written in the idiom of mainstream economics. That is not my idiom, so please forgive and correct any errors). Then I’ll make a few remarks on the relevance of the paper to contemporary politics.
The authors develop their “main sample” with a natural experiment, and then refine it with machine learning. From pages 3-4:
Direct harms to health resulting from cost-sharing, though, have been difficult to detect. This is not for lack of study. A central challenge is that the prices patients face are not random, and typically depend on prior utilization, creating spurious correlations between prices and health….. So it is perhaps unsurprising that studies have largely found null effects (Choudhry et al., 2011), or effects only on proxies for health outcomes, largely utilization of hospital or emergency care, or self-reported health (Chandra et al., 2010; Finkelstein et al., 2012; Geruso et al., 2020).
Here we explore the impact of cost-sharing on mortality, using data from Medicare’s prescription drug program. , we use a strategy introduced by Aron-Dine et al. (2015) and Kaplan and Zhang (2017): . This variation flows from a quirk in Medicare’s drug benefit structure, and specifically the annual spending thresholds that shift the price of drugs. Every January, beneficiaries start the year paying 25% out-of-pocket for drugs; but when they reach approximately $2500 of total drug consumption, they pay 100% out-of-pocket for the next drug. By themselves of course, these price changes are not exogenous: they depend on prior utilization. But critically, plan thresholds are not pro-rated in beneficiaries’ first calendar year of enrollment, and eligibility for enrollment begins in the month beneficiaries turn 65. So those born in later months of the year enroll in later months of the year, and in turn have less time to reach thresholds, meaning they face lower prices on average. Thus birth month creates exogenous variation in prices, by influencing enrollment month in the program. We focus on end-of-year prices and outcomes, specifically the month of December following Einav et al. (2015) and Einav et al. (2018). This strengthens our instrumental variables strategy: the later in the year, the more differences in measured spot price across enrollment months correspond to differences in full future price of a drug, before prices reset for all enrollment months on January 1.
So far we have relied heavily on the existing literature for our identification strategy, but have not addressed a major unsolved problem: used alone, the enrollment month instrument lacks the precision needed to detect health effects. This is because while enrollment month shifts December prices, it is a blunt tool. Within a given month, there is still wide variation in year-end prices, as a function of a beneficiary’s particular spending trajectory. For example, consider two February enrollees. One spends $500 per year, meaning despite her early-year enrollment, she remains far from spending thresholds, and pays only 25% of her drug costs at year-end. Another spends $5,000 per year, meaning she enters the ‘donut-hole’ coverage gap, and pays the full 100% of her drug costs at year-end. Instrumenting for price with enrollment month alone will assign both the same (average) year-end price, over-estimating the price for the first, and under-estimating it for the second. The resulting imprecision in the first stage creates problems for detecting rare health outcomes in the second. This could be avoided if we could condition on spending trajectories: ideally, we would like to compare beneficiaries on similar trajectories, who face different prices solely because they enroll in different months, instead of comparing large groups with heterogeneous prices. But naturally, we cannot use realized year-end spending, which is endogenous to enrollment month-driven variation in cost-sharing.
. To do so, we draw on data from a separate sample of Medicare beneficiaries to generate predictions on spending. The sample is similar to the one we study, but unaffected by enrollment month or cost-sharing: ‘dualeligible’ 65-year old Medicare enrollees, on Medicaid or other low income subsidies, who have the same enrollment criteria for Medicare Part D but face minimal cost-sharing. With these data in place, the task of predicting 12-month spending in the absence of cost-sharing is a straightforward ‘prediction problem’ (Kleinberg et al., 2015). We use machine learning tools to fit a function in the dual-eligible sample, and apply it to generate ‘counterfactual’ predictions in our main sample: ?
There’s a lot more, all of which, as I said, is severely above my paygrade. Perhaps some statisticians/economists would care to weigh in. (When you think about it, it’s more than a little problematic that highly skilled economists have to go through enormous gyrations like this to determine the mortality effects of health care policy implementation (ffs)).
From the Conclusion (page 36):
We find that . Cutbacks are widespread, but most striking are those seen in patients with the greatest treatable health risks, in whom they are likely to be particularly destructive. It is difficult to affirmatively establish that we have identified behavioral hazard, in the precise sense of a systematic failure to balance the cost with the benefit of care. But we emphasize that the size of the mortality increase cannot be reconciled with any current understanding of the value patients place on life.
We emphasize that our results do not capture the total impact of cost-sharing on health. We estimate only mortality, not morbidity, and only how December price changes affect 65-year-olds’ December mortality: a very specific setting, and a very short time period. But patients face costsharing throughout the year, and the life-span. If they respond with cutbacks similar to the ones we observe here, they would experience similar increases in mortality in many other settings and over longer time periods. While these effects are as-yet undetected, there is no reason to think that they are not present and equally large. Indeed, because our estimates are formed on largely healthy 65-year olds, effects in the larger (older) Medicare population may be quite different, and potentially larger, if the benefit of drugs is increasing in the underlying mortality hazard (e.g., older patients, nursing home patients, dementia patients), and if drug benefits cumulate over time horizons longer than one month. Understanding the range of health consequences of cost-sharing, and developing new policies to limit harms, is an urgent need.
And from the Abstract:
We use the design of Medicare’s prescription drug benefit program to demonstrate three facts about the health consequences of cost-sharing. First, we show that an as-if-random increase of 33.6% in out-of-pocket price (11.0 percentage points (p.p.) change in coinsurance, or $10.40 per drug) causes a 22.6% drop in total drug consumption ($61.20), and (0.048 p.p.). Second, we trace this mortality effect to cutbacks in life-saving medicines like statins and antihypertensives, for which clinical trials show large mortality benefits. We find no indication that these reductions in demand affect only ‘low-value’ drugs; on the contrary, those at the highest risk of heart attack and stroke, who would benefit the most from statins and antihypertensives, cut back more on these drugs than lower risk patients. Similar patterns exist for other drug–disease pairs, and irrespective of socioeconomic circumstance. Finally, we document that when faced with complex, high-dimensional choice problems, patients respond in simple, perverse ways. Specifically, price increases cause 18.0% more patients (2.8 p.p.) to fill no drugs, regardless of how many drugs they had been on previously, or their health risks. This decision mechanically results in larger absolute reductions in utilization for those on many drugs.
So I think “lethal” in the headline is more than fair. (It would be interesting to do a similar study on the “tax on time” Medicare’s neoliberal infestation has produced; I would imagine the effects are similar, if smaller.)
The authors, as one might expect, propose making eligibility requirements even more complex, rather in the manner of a Ptolemaic philosopher adding yet another epicycle:
One way to do so would be via value-based insurance design (VBID), where proven treatments (e.g. anti-hypertensives) are given zero (or even negative) copayments, while treatment with ambiguous benefit (e.g. proton pump inhibitors) are given high copayments
The obvious solution is, of course, to blow the entire system to smithereens with a single payer system, eliminating the lethal co-pays entirely. (One will, of course, have to pry complex eligibility requirements from the cold, dead hands of liberal Democrats, but a man can dream.) Trump showed the power of just writing checks with the CARES Act. Trump also set up a vaccination system that was free at the point of care. Get the credentialed gatekeepers ouf of the way.
I recently ran across a new acronym from the right: RAGE. From Vanity Fair’s recent article on Peter Thiel’s
blood bags new right wing ecosystem:
And the way conservatives can actually win in America, [Curtis Yarvin] has argued, is for a Caesar-like figure to take power back from this devolved oligarchy and replace it with a monarchical regime run like a start-up. As early as 2012, he proposed the acronym RAGE—Retire All Government Employees—as a shorthand for a first step in the overthrow of the American “regime.” What we needed, Yarvin thought, was a “national CEO, [or] what’s called a dictator.” Yarvin now shies away from the word dictator and seems to be trying to promote a friendlier face of authoritarianism as the solution to our political warfare: “If you’re going to have a monarchy, it has to be a monarchy of everyone,” he said.
Now, the concept of “a monarchical regime run like a start-up” makes me want to scream and run; it used to be conservatives who warned that it’s always possibile to make things worse. I am in no sense a fellow traveler with Yarvin. But whacking every part of the health care system that implements complex eligibility requirements is most definitely RAGE-adjacent. And I would bet there are millions of voters whose level of irritation is similar to mine (especially when we’ve seem vaccines free at the point of care. Why not everything?). Democrats — faced with an existential crisis as they are — might do well to consider that if they don’t RAGE, others will; the sclerosis and dysfunction is that obvious. And if Theil and his merry men do the RAGEing, there’s won’t even be any government employees left to cut checks. Why would there be? The peasants can wait for coaches to pass, and pray for some gold coins. So we can do this the easy way, or the hard way….