Class 8 — Adverse selection

Tuesday, September 22

This recap comes from a recording of class, transcribed with Whisper and summarized with Claude, with a light edit from me. It’s meant as a convenience rather than a careful document, so the phrasing will sometimes be a little off. If something here looks wrong, please tell me and I’ll fix it.

What we covered

We started with homework questions, then moved from one person buying insurance to a market. Special enrollment periods as the motivating example, the demand, average cost, and marginal cost graph, the three cases (full insurance, partial unraveling, full unraveling), the numerical example, the Paul Ryan clip, and a first pass at policy solutions. We ran out of time partway through the policy section and will pick it up Thursday.

Key points

Reading the graph is three steps, and the quiz will ask for them. Pick a price, go across to the demand curve to find the quantity, go up from that quantity to the average cost curve, and compare the two. Practice it until it’s automatic.
Insurance is unusual because an insurer’s costs are tied to the demand curve itself. In most markets what people will pay and what it costs to serve them are separate things.
Whether a market is in a “death spiral” is the wrong question. There is some unraveling in any market with adverse selection and a common price. The question is how much, and whether the people left out are the healthiest, which we can live with, or the sickest, which we can’t.
A subsidy paid to you shifts demand. It does not move the average cost curve, because your health hasn’t changed. What changes is who else buys, so observed average cost moves even though the curve doesn’t. A subsidy paid to insurers is the one that pulls the cost curve down.
A subsidy and a penalty do the same thing. Twenty dollars off the price and twenty dollars owed if you don’t buy both shift demand out by twenty.
Zero profit is a convenience, not a claim about insurers. We assume it away so the only thing left moving the market is selection.

The life insurance question

Someone asked why we don’t let insurers price each person individually in health insurance when other insurance markets do, and I said I didn’t have a good answer. Here’s a better one.

Life insurance underwrites you individually, but only once. A level-term or whole-life policy fixes the premium when you buy it, and the insurer can’t raise it later when your health changes. So life insurance is already protecting you against being repriced. It just does it through the contract rather than through a rule.

That points at what health insurance regulation is actually protecting, and it isn’t the within-year risk we drew on Tuesday. It’s the risk that your premium jumps in the year you get sick and stays there. Health insurance is bought over and over across a lifetime, and a string of individually priced one-year policies leaves that risk completely uninsured, even though every one of those policies is fairly priced. Community rating is a substitute for the long-term contract this market won’t write, since people change jobs, the product can’t be specified decades ahead, and a healthy enrollee can always walk away.

Auto insurance is the case that really is different, and the reason is that a driving record is mostly a choice. Pricing on it changes behavior. Health status mostly isn’t a choice, so pricing on it just moves money away from people who drew badly. That’s also why the rating factors states restrict in auto insurance, things like credit score and ZIP code, tend to be the ones with the least to do with how you actually drive.

More on the insurer-side programs

I went quickly over this in class and it’s worth separating out, because “subsidize insurers” in the ACA meant several different programs and they ended in different ways.

Risk adjustment is permanent and still running. Money moves between insurers in the same state and market based on the relative risk of who actually enrolled, so an insurer that draws a sicker pool is compensated by one that draws a healthier pool. This is the one I acted out in class with two of you as the insurers, and it’s the same idea as the Medicare Advantage version we saw earlier in the term.

Reinsurance was temporary, 2014 through 2016, and covered the individual market. It paid insurers for enrollees with very high claims. It expired on schedule.

Risk corridors were also temporary, 2014 through 2016. They compared an insurer’s actual costs against a target, with the government collecting from insurers who came in more than three percent under and paying insurers who came in more than three percent over. This is the one that didn’t get paid. Losses far exceeded gains, and an appropriations rider first enacted in December 2014 barred the agency from paying out more than it collected, so insurers received 12.6 percent of what they were owed for 2014. What came in for 2015 and 2016 went toward the 2014 shortfall, so those years paid nothing. In 2020 the Supreme Court held 8 to 1 that the government still owed the money, and insurers were eventually paid through the courts.

Cost-sharing reduction payments are a separate program, not one of the three above. Insurers are required to lower deductibles and copays for silver-plan enrollees under 250 percent of poverty, and these payments reimbursed them for doing it. They were stopped in October 2017. Insurers responded by concentrating premium increases in silver plans, which raised the benchmark plan, which raised the subsidy everyone received. The cut ended up costing the federal government more rather than less.

The economics is the same in each case. All of these pull the average cost curve down, which is the insurer-side policy we were drawing at the end of class. Taking one away pushes it back up.

Thursday and the homework

Panel plus a short quiz, which finishes the insurance section. Bring questions on adverse selection and we can start with them.

Homework 1 is due Friday. One thing I said after class that everyone should hear. The New Jersey historical rate file is a scanned PDF, and asking a model to read it will produce more mistakes than doing it by hand. Build that table yourself. Everything else on the assignment is fair game for the agent.