Macroeconomics of Pensions & Retirement Financing Conference

Transcription

1 Macroeconomics of Pensions & Retirement Financing Conference The Welfare Effects of Long-Term Health Insurance Contracts Presenter: Ben Handel Ben Handel University of California, Berkeley Igal Hendel Northwestern University Michael D. Whinston MIT March 29, th Street Southeast Minneapolis, MN hhei.umn.edu

2 Ben: It's a pleasure to be here. I think this is the first conference I've ever presented at with macroeconomics in the title, so I'm very excited about that. It's kind of a new thing for my CV, which is very cool. And I think this is also hopefully a relevant paper and a paper that you will be interested in. So this is joint work with Igal Hendel, at Northwestern, and Mike Winston at MIT. The paper is designed to think about different market design a different kind of style of market design in insurance markets from what we see today. What are the potential efficiency gains of thinking about playing around with the duration of contracts and leading to more formal insurance in a long run sense? So, just to set the stage, in a prior paper that Igal, Mike, and I had worked on, we studied the effects of different health exchange risk rating rules. What that paper did was it investigated year-to-year contracts, so there was no change in duration of these contracts. But the goal was to compare rules that allow for risk rating to those that don't. So what that means in the health insurance context is in the Affordable Care Act, the main thrust of the rules is community rating, and that means you can't charge different prices to different people based on health status. Risk rating, which we cited in that paper, allows you to price people differently. You can have different risk rating rules, some of which allow some degree of differential pricing, some of which allow full differential pricing. With studying a range of risk rating regulations, what we were studying in that paper was how can we quantify the trade-off as you allow more risk rating between reduced adverse selection, so as you allow more risk rating, there's less adverse selection, at least not theoretically. It doesn't have to be true, but it's generally true empirically. But there s more reclassification risk, as well. When you allow risk rating, you reduce adverse selection, but you increase reclassification risk. What reclassification risk means is the premium risk consumers face over time from the fact that they're in these year-to-year contracts, and their premiums are changing around because of risk rating. For that paper, we developed a lifecycle welfare model that was investigating the impact of these different regulations. We found that on net, when we looked at these regulations with a lot of different kinds of

3 extensions and things like that, we found that the reclassification risk effect was worse. So the risk you faced from being reclassified in a system with a risk rating was worse than the negative impacts of adverse selection on equilibrium year-to-year insurance coverage. So, think about that paper as basically we're studying these tradeoffs in static insurance markets, but how do they manifest from a welfare perspective over a long run horizon? In this paper, we're looking at long-term contracts with health status based pricing, and thinking, "Can these long-term contracts improve on the community rating we see in the Affordable Care Act? Or can they make contracts with the risk rating more attractive because now you're able to insure people over the long run with respect to this reclassification risk?" So this paper is really taking the last paper and saying, "Let's think about long-term contracts and what value they can add in these in the system." So, what are we going to do? Just to set the stage, we're going to perform the following comparisons between these five different models of insurance markets or first-best. We're going to develop the equilibrium properties and the welfare properties and the tradeoffs inherent in each of these benchmarks or market structures, and then we're going to compare the long run welfare impacts of each of these structures. So, we're going to have three benchmarks and two models that we're going to study. The first benchmark we're going to study is first-best. First-best here is going to mean full insurance and full inter-temporal consumption smoothing. The second benchmark is annual spot contracts with risk rating. The second benchmark means insurers offer contracts, which are one year at a time. They can be fully risk rated. That's it. The third model which we're going to focus on for most of the talk today, are long-term contracts with one-sided commitment and risk rating. So these kinds of contracts are the contracts that you typically see, for example, in life insurance markets. The insurer commits to the contract and so that means the insurer commits to a time path of premiums, for example, in the contract. Consumers don't have to commit to this contract, and consumers can lapse from the contract at any time and go sign up with another insurer. The third benchmark is full insurance, but with no inter-temporal consumption smoothing. The reason we're going to have this is as a

4 third benchmark is because in the first-best you're generating value, relatively speaking, both from insurance in a health sense, and from inter-temporal consumption smoothing. Having this third benchmark with the first benchmark allows us to say, "Well, how much are the benefits or none benefits of these contracts are coming from each of these two dimensions?" Finally, we're also going to compare these cases to the case of annual contracts with community rating and age-based pricing. You can think of this is as an ACA-like market that we studied in the last paper. Audience Question: So which one is the pre-aca closet to the pre-aca? Ben: The honest answer is none. I hesitate to say it, but the closest one from a regulatory standpoint is probably three. Because insurers did have a decent bit of leeway to do risk rating, but it's not quite in the sense we're modeling it here. They did it more through coverage denials than through explicit, long-term contracts. There were guaranteed renewable contracts that many states regulated, which have a similar spirit to the contracts we're going to talk about in terms of the dynamic properties, but are not exactly the same. I guess I would say the individual market prior to the ACA is such a mess, and I'll talk about this as we get towards the end of the talk, that it's probably actually not that similar to any of these models. In some sense, I would say there is risk rating in coverage denials allowed, but you didn't see these long-term contracts, generally speaking. Audience Question: I include the employer-based insurance market. Would you say that's closer to three, in a sense? Effectively, the firm is contracting on your behalf, and so you can, of course, quit the firm and move somewhere else. So for that part of the market is three perhaps a reasonable assumption? Ben: I would say probably five. The reason is that if you stay with your employer, it's essentially like you're getting an annual contract with community rating and if you leave the employer Audience Question: It s evident that wages seem to depend on things like your health status and so on and so forth. So, there may be other ways that your employer effectively has you paying part of that payment.

5 Ben: Yes, and that's called a differential wage pass through as a function of health status, and the evidence is mixed on this. There are some studies that do show that. I'm not sure I fully believe them Audience Question: So somewhere between three and five? Ben: but it makes sense. Yes. Somewhere between three and five, and I do want to make sure I take some time at the end because I'm going to talk a lot about these long-term contracts. I'm going to talk a lot about number three. At the end, one of the things I really want to spend time talking about is well, why didn't we see these contracts in the pre-aca environment? Why don't we see them now, and how could we potentially see them? I think that's an important part of this paper. First, we're going to characterize these optimal dynamic contracts. When we do that, it's really going to follow this literature on contracts with one-sided commitment. Similar to Harris and Homes in a wage sense, and similar to their prior work by Igal and Alessandro Lizzeri on life insurance markets. Then we're going to examine their structure and performance empirically. How we're going to do that is we have data from a large firm with micro foundations that we think are important for assessing the implications of different insurance market designs. And from those data, we're going to get the following items. We re going to get expected cost given your health status, so what are your costs as a consumer given your health status? What is your health status in an expected sense or in a diagnostic sense? And how does that transition over time? We re going to have stochastic transitions of health status that we get from the data over time. We're going to observe income paths in the firm that we're studying. We're going to use those to get a sense of how do these different market designs or regulatory environments relate to the steepness or flatness of income profiles over time? Finally, we're going to think, of course, about risk aversion and preferences for consumption smoothing. Audience Question: So, I'm I understanding that very soon, these structures will contract themselves and have more effects on health dynamics, medical expenditure

6 Ben: That's right. First, we re assuming no moral hazard with response to cost sharing, which there will be cost sharing in the ACA contracts because of adverse selection in that marketplace. Second, with the long run contracts, there is some benefit we're not capturing of encouraging insurers to invest in preventive health curative high-cost drugs other aspects that longer run relationships can facilitate. We're going to miss both of those elements right now in this paper. Audience Question: The argument becomes stretched when you think about the income, right? Because if it's long-term contract for a better insurance maybe, you know, I look for a different job. Ben: That's true, and I think that we're going to take these empirical moments with respect to income. One way you could think about it is, which we don't have empirically here, is if there is some endogenous impact of the insurance on the job you take and that changes your income path, you could put that into the framework we've developed. We don't have that here empirically. Audience Question: Would you have savings? Ben: We do have an extension with precautionary savings, but not borrowing. I've already alluded to some of the papers that this relates to. There's a theoretical literature and an empirical literature. Theoretically, it relates to literatures on contracts with one-sided commitments where there's symmetric learning over time. Harris and Holmstrom, which I'm sure many of you are familiar with, is a paper that this clearly builds on and relates to. There's a paper by John Cochrane, from 1995 that investigates long-run insurance with the idea that basically, both the insurer and the consumer can fully commit to contracts that fully insure long-run risk. That's a first-best as long as there's full commitment, so there's two-sided, full commitment. Empirically, there's been papers on commitment from commitment and long-run relationships and life insurance contracts. There's been studies in the long-term care market and the German private insurance market that study how consumer lapsation behavior

7 relates to health status? Generally, finding that the healthier you are, the more likely you are to lapse from a long-term contract, which is what a rational model would predict. The dataset we're going to use here is individual-level panel data from a large employer. The employer has 10,000 employees and covers about 25,000 lives. The way I think about this is I think this is a detailed dataset that gives us an ability to characterize these moments that we think are important for studying these questions. I should be clear up front that I don't think about this as externally valid or representative of some sample that we care about from a policy sense, necessarily. I should also mention, though I'm not going to present this today, that we're adding a component to this paper where we use another dataset which is All Payer Claims Database from Utah that covers two million people, but with a similar style of data. We're projecting that will have better external properties, but the same model and empirical exercise more or less. Data have plan choices from a menu of multiple plans. It has consumer claims, which is very detailed information on consumer health diagnostics. What do they pay? How do those diagnoses change over time? One of the key things we're going to take away here is because we observed a panel with consumers over five years where we observed all of their diagnostic information. We use a sophisticated software program developed at Johns Hopkins, to map these medical diagnoses in a year into a measure of expected health spending, which captures how their risk as based on the data from the last year projects into their expected spending in the upcoming year. Using that measure, we're going to think about health status transitions over time. So you should think about each of these health status measures as indicative of a distribution of health spending for the next year, a probability distribution. And then you should think about transitions between those distributions in a typical Markovian sense, that s what we're going to think about here. Audience Question: So the way you say the score here, is this type of score actually used by insurance companies when they do risk rating in practice? Ben: Yes. This kind of score is used in the industry to do risk rating. Similar scores are used by regulators to do risk adjustment where you arrange transfers between insurers based on enrolling health status. So,

8 this is a well-known algorithm. There's many of them. There's probably 5 or 10 of them, all which I think are reasonably similar. Audience Question: And what's the financial data that you have? This is medical and financial. Just what they owe or you know something more about them? Ben: No, that's just what they owe. It's what they pay for each health claim and what they pay for deductible co-insurance, co-payment, things like that. We also observe their income, which is financial data. We don't observe more than that, which would be something like investment assets, savings, we don't observe that. This gives just a sense of one of the moments we're going to be looking at, which is health expenditures and the properties of health expenditures over time. Here broken down, you can see all ages and then ages in five-year bins. Today, we're going to be focusing on the population 25 to 65 and we're going to be thinking about individuals going into an individual market, not families, which is a separate issue which we can talk about later. The first column is mean health expenditures as a function of age. The second column is standard deviation and the standard deviation is saying it's a global standard deviation, so it's taking one year of one person in this bucket, what's the distribution look like? The third column and fourth column are giving you some sense of the predictable risk, year-to-year, versus the idiosyncratic risk year-to-year. The first column is saying, "What's the standard deviation of your expected health spending one year at a time?" It's saying, if everyone here has an expected level of health spending for the next year, what's the standard deviation of that expectation? The final column here is saying, "Given that expectation, what's the idiosyncratic risk that materializes from that point on?" You can see that these are comparable, but on an average, I'd say about 30% to 40% of risk is what I would call predictable in a forecasting sense, and 60% to 70% is idiosyncratic. Audience Question: So these are yearly expenses? Ben: Yes. Audience Question: And they include the premium?

9 Ben: No. Audience Question: Is this just out of market? Ben: I should be clear. This is total health spending, and so this is what the consumer spends and what the insurer spends. This isn't quite right, but you should think about it as if you're uninsured and this is just what you pay for health spending. It's what everyone pays for this one person in the year for health spending. Think about it as what the providers or drug manufacturers are paid for this person. You can see the population health state. Let me tell you what's here. First, on the rows, we have five-year age bins. On the columns, we have septiles of the ACG score, so of this risk score. One means you're in the healthiest one-seventh of people, from a quantile perspective in a given year from a predictive standpoint, and seven means you're in the sickest. What you can see here is as you go across the ages, the empirical evidence is that people are sicker and sicker. As you get to 60 and 65, more than half of the people are in bins 6 and 7, which means they're expected to spend among the highest amounts of people in the population overall. But if you look at 25 to 30, about 60% of people are in bins 1 and 2, which are the healthiest 2 bins. The next table here gives you a sense of health status transitions. Lambda T is on the rows, and lamda T plus one, which is your health status in the next year and the following year, is on the columns. Here, you can see what these transitions look like in the data. On a diagonal here would mean that you're staying in the same health status over time. What you can see is that health status is reasonably persistent in the sense that if you're healthy, you're pretty likely to stay healthy, and if you're sick, you're pretty likely to stay sick. That's the way you should think about this. Audience Question: If I break my leg, I'm I unhealthy because I spent above? Ben: Yes, so typically speaking, if you break your leg but you have no other kind of chronic conditions, you'll be counted as healthy here. Audience Question: Through the Johns Hopkins score. Ben: Yes, and that's why that adds value over just past spending. That's

10 the way to think about it. Of course, it's not perfect, depending on when that happens Anyway, that's the idea. Okay. And finally, this table shows the present value of expected expenses at various ages, starting off at a given health statuses at age 30. Here the idea is, if you start out in a given septile at age 30, so you're healthy, you're sick, what do your health expenditures look like in the next year over the next 5 years, over the next 10 years, and then over a lifetime? The idea to get across here is basically that there's a reversion to the mean. If you're sick now, you're going to be spending a lot more than a healthy person over the next five years, but over time, these things broadly even out. One thing to point out here is in our data, we observed an individual going over five years and then we're linking these things over time with a Markov Transition Matrix. One thing we're missing here is if there's kind of a specific time path beyond five years where there's a certain type of person that's always going to be sick for 30 years, that's something we're going to be missing here. Audience Question: Did you get people that transitioned into disability? You can have a job. Can you start off with a job and transition to disability? Ben: It's a big question. I don't know if we observe that. I don't think we do, actually. I put data together piece by piece. I don't remember that. Audience Question: How do we think about these tail events? We know that their expenditure is highly concentrated. Ben: Yeah, and disability in programs like that are clearly relevant to what we're talking about, even if we're not going to talk about it. As well as churn across different markets, like you brought up, with employer markets versus individual insurance market. Okay, so let me talk about the model. Audience Question: What will be the sense of how much of this cost is owned by the consumer and how much is owned by the insurer? Ben: In these data here, generally speaking, consumers are bearing about 20% of the costs, and insurers are paying about 80% of the cost.

11 And in general, in the U.S. right now overall, that's about typically about 20% of costs are paid by individuals. That's on average. Healthier individuals pay a higher percentage. Sicker individuals pay a lower percentage just because the shape of the contracts. Audience Question: You only look at the people practicing from the private health insurance. So, how often do they go out of this sample, like getting a job and getting off from employer providing insurance and are going to Medicaid? Ben: This sample here, to be clear, is from one firm, like an employer, so this is not the individual market. This is the employer market, and we're taking the data moments from this market, and we're going to think about that in the context of the individual market. Your question about churn, I actually know this from another project I'm working on, not from this project, with this Utah database where we observe all insurers in the state and how people transition across them, including employer and individual. There's a pretty high degree of churn in these markets overall. I mean, I'm guessing here, but I'd say typical duration of a relationship is two years, one and a half years, something like that. Then the individual markets like the ACA exchanges, the duration is much shorter than that. I'd say there's a decent bit of churn in these markets, and I'll return to that at the end because that's actually both something that could be positive about these contracts, but also potentially make it difficult to implement in a policy sense. Audience Question: When you say one firm, is that one health insurance firm, or one firm that hires and this firm only hires from one health insurance? Ben: Neither. It's one employer, and they allow employees to choose between multiple insurance options that they carry as an intermediary. Thanks for asking. That's important. So the model, we're going to be thinking about someone who's living in our model from 25 to 65. We stop at 65 because in the U.S., when you hit 65 you go on to Medicare, so we don't care at that point about private insurance, though, in other papers, you may care about the impact of that market on spending later. People are going to have utility over these periods for consumption. Each person is going to be characterized by an income path where

12 income is known. We also have extensions to stochastic income and unobserved income from the perspective of the insurer. The health state, coming from the ACG score, is going to summarize expected health costs. You can think about that as the expectation of MT where MT is your health costs in a given year, and that's conditional on lambda T, where lambda T is your ACG score, your health status. The health expenses MT in the upcoming year, and your health status for the next year, your projected health status for the upcoming year are going to come from a joint distribution. It should be clear why those things come from a joint distribution. When you have an idiosyncratic outcome or medical event occur, you're both going to spend more money and it has implications for your future health status. Derived from this joint distribution, we're going to have an implied transition probability, F, which is going to say how your health status is transitioning over time, stochastically. For the given period T, you're going to have a density of health spending as a function of your health status. You're going to have an initial health state, which we'll talk about in the model. Some cases we allow that to vary, and some cases we think about it as just the ex-ante individual. For all of the paper, we're going to observe that the spending in a year and the health status are commonly observed by the individual and the insurer. I think that's a reasonable assumption for this paper. In general, you could think that could be wrong in two ways. You could think that an individual has some private information that they're not disclosing, and you could also think that the insurer, using these sophisticated algorithms, knows more than the individual about their future projected health spending. We're going to assume that in all of the analysis that we do, all of the different contingencies, we're going to assume that the market is competitive and firms are risk neutral. Audience Question: So the MT and lambda Ts are observed, not only by my current insurer, but also potential other insurers, so they are all observing me? Ben: Exactly. And the problem would be a lot more complicated, of course, if your insurer observes this and the others' insurers didn't, which is possibly the best way to characterize the market. Though there's data dissemination policies that can make something like this more like reality.

13 We have two assumptions, and these are very intuitive assumptions. I'll go over them quickly. The first one is that the expected spending is increasing in your health status score. This is just saying if you're predicted to have a worse health status, then your expected spending is going to be higher. This is almost, by definition, true, just to be clear. The second assumption is a first order stochastic dominance assumption, which is basically saying, as your health status gets worse, your distribution of health status for the next year, for that transition, first orders stochastically dominates that of a healthy person. Again, this is true in the data, so we can check that. Three benchmarks we're going to look at are first-best, spot contracting, and full insurance without inter-temporal smoothing. Here's the way to think about this. The first-best here is there's basically two sources of inefficiencies that consumers are facing in our model. One is their being exposed to risk from their health state over time and within a given year, and the second thing is that they may not have the ability to smooth consumption over time inter-temporally, which is a separate thing. Tthe first-best here allows consumers to do both of these things. They're smoothing consumption inter-temporally, and smoothing risk, so they have full insurance. Spot contracting here. They have neither of these things, so there's no inter-temporal consumption smoothing assumed or allowed in spot contracting. There's no notion of long-run insurance, so you arrive at the market and the insurer can risk rate you and charge you with your expected cost for the next year. Spot contracting means one year at a time contracts where you can be risk rated. Audience Question: And you're assuming that this kind of rate is the same as the interest rate? Ben: Yes. The third benchmark we're going to look at is full insurance without inter-temporal smoothing. Here, you should think about this saying, "Okay, because we're studying these two different things, we may want to have a notion of first-best, but first-best conditional on not having inter-temporal consumption smoothing," just to give us a sense of the traditional insurance value of the contracting design versus how the contract interacts with inter-temporal smoothing. Just to preview this, you may be wondering why I'm talking about this. The reason is that these contracts we're going to talk about, these long-run contracts with commitment, one property of these contractors is they're going to have front loading where the consumer pays in money upfront and that allows

14 reinsurance of their risk in the future. That, relative to the spot contracts, is going to change the path of spending in these contracts, and so that's why we're thinking about consumption smoothing, as well. Audience Question: When you have the income, do you think of income after taxes transfers and self-insurance incomes? How do you think about this income? Is it disposable income which you're going to just consume? Ben: Yes. That's the way I would think about it. Audience Question: But your employer doesn't see that. When you go to the data, what are you going to do? Ben: That's a good question. You know, that's a really good point. What we do is we take the income that we observed in the data, which is pretax, so not post tax. Audience Question: The employer has that? Ben: That the employer has, the information the employer has. I do not remember whether we made that post-tax based on people's marginal tax rates, which we observe, and I use that in a prior paper. I don't know if we use that here. This is either adjusted for federal and state taxes or not. I'm not sure from an empirical standpoint. Audience Question: Those transfers are happening on their own? Ben: Yeah, and that we don't observe. Audience Question: That's really hard. Ben: Yeah. That's potentially important, of course. One way to think about that issue because, I agree, is that you should think about this as we're going to show you this for an income path with dynamic contracts versus the other contracts. You should think of it in a theoretical sense as kind of what Chris was saying, a kind of after-tax, disposable income net of transfers, but that's not what we're putting in empirically, to be clear. When you see the path in the model it's acting like that, but the empirical moment is not quite the same thing. The problem, you can think of it as a recursive problem, and the design

15 problem is to set up a contract with one-sided commitment, where the insurer breaks even because this is a competitive market, and the consumer gets the best possible path of consumption for themselves going forward, given their health risk, given their income profile, etc. We actually conceptualize this contract in...you can think about it in a couple different ways. The way we have it written here is basically, where the insurer sets up a program for the consumer of consumption and subsidies, where the consumer is basically consuming an amount in one period, but via the contract, they're allowed to port over money to the next period, which is basically front-loading. What that means is that in a given period, the consumer is consuming something which is basically net of their premium and medical expenditures. But they can also consume less and send money over to the next period, which allows for reinsurance. Another way to say that, they can pay more money now into the contract. That's one way to characterize it, which is what's written here. Another way to characterize it, which I'll talk about in a minute, which I think is actually a little more intuitive, is the insurer is guaranteeing a premium path for people over time, and this premium path is a function of their health status. Audience Question: In a commercial sense, you would have kept track of promised utilities, right? And this S is just amount of time of promised utilities or something? Ben: Yes. Audience Question: And this S is just a transformation? Ben: That's exactly right. That's what I was saying with the premium. You could write this in terms of someone has consumption over time, and there's a promised path of premiums in the contract. Audience Question: So, in the model there's only one type of consumer, but in your data, employers and the consumers can select their PPOs or HMOs. And what if you have different type of consumers selecting different products? Do you only look at one type of contract to avoid dealing with these issues, or what do you do? Ben: Yes, we look at one type of contract. We look at the PPO contracts which have about 65% of the employees, and that's in part to deal with this issue. Exactly. The main reason to do that is if you're trying to

16 characterize risk aversion. If you have consumers really choosing between many different networks or having different networks of doctors, then it's multi-dimensional heterogeneity because people could have risk aversion, but also preferences for certain kinds of doctors. And so here, what we do in prior work is we show that the substitution between these massive plans is not meaningful. When people substitute, they substitute within one nest, PPOs and HMOs. We use that to say, "Well, we don't have to worry much about this much from an estimation sense." From an external validity sense, we're not capturing that. Audience Question: I'm probably just understanding the demographics structure here because there is this initial lambda, and that initial lambda be very different depending on how you model the demographic structure. If you have people I knew in this firm, some of which are 55 years old, some of which are 30 years old. And so sufficient generality here, where you see it's clear that this is conditional lambda, with insurance opportunities dynamic or otherwise, are a huge function of your status when you're nearing the contract at your age and your health status. Ben: Absolutely. The way to think about this here is we're making the assumption with this data that basically, we're assuming that problem away, is the way to think about it. We're saying, "Here's a population of consumers. They arrive at the firm for whatever reason. There's no model of selection into the firm," which is kind of what you're referring to. They have health statuses that transitions over time. We're going to characterize those transitions over time. What their selections within the space of contracts implies about their risk preferences for insurance, and that's it. When we move to the Utah data which we're working on right now, my answer here isn't very satisfying understanding why we're not doing it. In the Utah data we're not doing it yet either, but we could do it because we actually observe employers as well and how people transition between employers. Going forward, I think this actually is something that's a legitimate and interesting question, I would say. Audience Question: Let's get a graphic estimation? Ben: Okay, so there's contracts, optimal contracts, and let me talk about

17 them. What the contracts do is they guarantee a minimum consumption level for people conditional on their current health status. The idea here is basically the converse of downward rigid rate wages in Harris and Holmstrom, if you're familiar with that paper. The idea is there's symmetric uncertainty, there's symmetrical learning. The insurer is insuring you against bad realizations of that symmetric learning. In Harris and Holmstrom, that means that your wages are lower than they would otherwise be, but they don't go even lower if you have a bad productivity draw. What it means is, conditional on your health status, you're guaranteed better premium draws in the future if you have a negative health draw. If you have a positive health draw, so you get healthier to provide you with intuition, you can go to the outside market. What that means is the firm is required to offer you a better premium in the next year to keep you in the contract. These contracts are characterized by front-loading so there's subsidy up front. What the front-loading does is it allows consumers to reinsure themselves through the contract over time because it allows the insurer to offer them these guarantees in the bad health states in the future without earning a negative profit on these consumers in the future when those bad states happen. A couple other quick things and then I'll get to the estimates. You can think about this, as I said, as a guaranteed premium path, and for me, that's actually the way I prefer to think about it. We have thought about this with unobserved income paths risk aversion and discount factors, and we have a proof that actually shows that these optimal contracts we've kind of determined here are self-selective. What that means is even if there's unobserved income and risk aversion at the consumer side, they're going to select the optimal contract for themselves if the insurer offers a granular menu of these contracts. We also do an extension with switching costs and I'm going talk about how that's important. What are the elements of the data? The key things we take from the data are health status, health status transitions over time, costs as a function of health status, and risk preferences. So, I'm not going to show you the structural model and how we do those estimates here just in the interest of time. But basically, what we do is we take these consumers choosing from a set of PPO options, like we talked about. They have three options. The options are vertically differentiated by financial

18 coverage, so the more expensive options from a premium standpoint give you more generous risk protection. You can think about choosing between these three options conditional on your health status, that's kind of putting bounds on what your risk preferences are. If you choose the most generous plan, it means you're more risk averse than some level of threshold. If you choose the least generous plan, you're between zero and some level. And there's a middle plan for the middle range. Using these detailed data, we have estimates of your health states, the health state transitions, your risk preferences. And then we also have these transition matrices that are based on these estimates as age dependent. So we're going to estimate these transitions differently for 5-year age bins, 30 to 35, 31 to 35, 36 to 40, etc. And the idea there is to really characterize the health status transitions for people at different point in their lifetimes and use that to think about these long-run welfare calculations. So the estimate we get from the data is gamma equals 4 times 10 to the negative 4. That's fairly central if you look at the literature. But for those of you who have worked in this literature, you know that there are a wide range of estimates and they can actually be interpreted quite differently depending on how you use them. Audience Question: You implied disclosure benefit when you look at the median income? Private versus commercial and private? Ben: Oh, I don't remember. Audience Question: You used this to figure that. Ben: Yeah. Audience Question: So, contract to some baseline that perhaps, you know, is the current contract. If you can go over what that baseline looks like, what is your idea of what a contract means? Ben: Yeah, let me get to that as I get through the results. That will be there, and I'll try to pause and talk about that a little longer there if that hasn't come through yet. Audience Question: Suppose they change the health status in your model, can they switch plans? Ben: Yeah. Now, the contracts are designed so that they don't switch

19 plans because all the firms are symmetric, but they can switch plans. Audience Question: The assumption that you can identify risk fragments from plan choice on the assumption no part of information is correct. So really you're identifying a mix of risk preferences and people's believes about the future and expenses when they're in control. Ben: Right. And so in the choice model, which I haven't shown you, there's a rational expectations assumption that the ACG score and how that links to future health expenses which is comes from a detailed estimation model that that's what consumers perceive as their beliefs as well. Now, I have another paper which actually links survey data on people's projections of their health spending to these rational expectation measures, and show that they're centrally correct, but there's meaningful dispersion. So some people underestimate meaningfully and some overestimate, and so that's an assumption here. Now, in terms of the private information side, in an earlier paper, we do test on that in these data which is something akin to a correlation style test. Where we say, "Okay, conditional on the information we have about you, if you spend more ex-post, are you also likely to have chosen more generous coverage ex-ante." We don't find evidence that that's a strong effect, that that's a strong thing. One of the reasons that's not too surprising here is we have very detailed health information. In a case where we didn't have that, I would be very concerned about that. We take a discount factor of I'll just be honest. I have no idea why we chose that. So if you have advice on that let us know. Audience Question: I think because you chose the discount rate and the interest rate. Ben: Oh, okay. Thank you. I knew we chose it for a reason. But we do the analysis for different discount factors as we do in the last paper, and that's in some sense the way I think about it. You could have different discount factors depending on your underlying model and we allow for that in the simulations. We're going to investigate these different programs for four different income profiles, and these income profiles come from the data. And so like I mentioned earlier, I really want you to think about this as just investigating these different types of market set ups for different time profiles and income. Here, you should think about what we call managers, because that's

20 what we have data on, as the blue line. And so, you should think about that as higher income and a steeper income profile over time. We have non-managers who are flatter income profile and lower income over time, which is the purple line. We have flat net income which is basically saying let's take for the none managers the average income over the life cycle and make it flat instead of steep, when we want to do analyses that take out that steepness. And then we also have downscaled managers which means let's have the same expected income for the steeper profile and the flatter profile, but allow comparisons, apples to apples comparisons in that sense. Audience Question: So I didn't understand. But are these actually income profiles from your cross section? Ben: Yes. Audience Question: Except for the green bin? Ben: Except for the green bin, exactly. And except for the black one. The black one is it's related to what we observe, but it's downscaled. Audience Question: Actually, you're taking the blue and getting the black. And you're taking the purple and sort of flattening. Ben: Exactly, so the things which come from the data here are the blue line and the purple line. What are the contracts? First, what I'm going to do is I'm going to kind of go through and show you what the dynamic contracts with one-sided commitment look like for consumers in the data. And then I'm going to move on to the analysis that compares this kind of insurance market design to the other ones which was the question that was asked before. So here, I'm going to show you the results the first for flat net income. You should think about there's going to be no inter-temporal consumption smoothing that's needed because people have flat income profiles over time. This is talking just about what the insurance contracts looks like when there is a need for risk smoothing. Here, what I want you to take away is the degree of front-loading, row number three. On top, we have the same ACG septiles. One is healthy. This is the last time I'm going to remind you of this, so listen up. One is

21 healthy and seven is sick. Here, if you're a healthy person in the first year of the contract, so think of this as someone who's 25, you're going to pay a premium in the optimal contract of 2,750. You're expected costs are 1,131. Your degree of front-loading is going to be 1,619. The way to think about the optimal contract if you prefer the first characterization we had is just what are your costs and how much are you front-loading. And your cost here, because we have income flat up front, your premium is directly determining your consumption, is the way to think about this. One other thing to point out, which I didn't mention, in these optimal dynamic contracts within a given year you always have full insurance because we're assuming no moral hazard. So, in the second year, which is where you start to see some of the implications of the model with flat net income, here on the rows you have health status in period one, on the columns you have health status in period two. And what you see is how the contract leads to guaranteed consumption paths and so how it leads to reinsurance. So look at health status one in the row. What the row is saying is as you become sicker and sicker in the next period, you're still obtaining guaranteed consumption level in this optimal contract, and this is the front-loading capturing and the reinsurance. But if you're in the worse health status in period one, which is group seven there, you're getting no reinsurance because there's no worse state for you to move to in the second period. But if you become healthier, you can do better. You can go back up because you ll get a competing outside offer relative to this contract. This is showing the same thing, but for premiums. Here you can see that if you're in group one, but you become sick, your premiums are guaranteed at a certain level. And if you're in group seven and you become healthy, your premiums go down, but if you don't become healthy, they stay high because you had no ability to reinsure yourself. Then here what we do is we say, "Well, there are 40 million histories here if you look at all the possible combinations of health status or health status transitions that you can have over time. Based on these histories in these dynamic contracts what's the mean degree of front-loading we see and what's the maximum front-loading we see?" This is yearly front-loading. Over the first 10 years of the contract, the average front-loading ranges from about 1,700 to 2,500 over time, and the maximum front-loading is in the 3,000 range. If you take this mean and you say someone's acting like the mean person in the contract over 10 years, they're going to pay in $20,000 to this contract, and that's

22 going to be used in a sense to reinsure themselves. Both for the future, but also reinsure the other people who became sick in the pool. Audience Question: Other than are that, what's driving the fluctuations here? Ben: What's driving the fluctuations here is that we're starting at age 25, and so as you get older, you are moving towards older ages, which are leading to both different levels of health expenses and different profiles of health status and different profiles of health status transitions in the future, so there's a number of ingredients which are changing. Audience Question: But this goes to age 60 then it starts getting negative or something. Am I probably right? Ben: Right, well, there's never going to be negative front-loading, but it's getting closer to those spot contracts you're doing much less reinsurance. I've got five minutes so I'm going to keep plugging along here. So with the downscale manager, so here you should think about these as people same average income as flat net income, but steeper paths. Basically, the lesson I want you to take away from this is there's lower degree of front-loading in these contracts. If you have steeper income paths you want to front-load less because you have lower income and earlier periods when you would otherwise do this front-loading. For the results from the welfare standpoint now what are we going to do? This relates to the question I was asked before. We're going to compare the results empirically with these moments and with the data we've used for five scenarios. First-best, which is two sided commitment. Spot contracts, which means one year at a time, but contracts can be risk rated. Dynamic contracts which I've been talking about the case with full insurance, but no allowance of inter-temporal smoothing, which is basically like a modified first-best. And then ACA, and here ACA is going to operationally mean that people are going to have 60% coverage in each year because of adverse selection. But they're going to be insured against reclassification risk because of community rating. And the results from the ACA come from our last paper where we found that with community rating in these markets the potential for adverse selection is high. But the regulation specifies that the insurer has to offer at least 60% coverage, and so

23 that's why that's the level here. And we're going to do these five comparisons for these four different time profiles of income paths. I think this is the main results' table here, and here what this is doing is it's talking first about the implied certainty equivalent for someone from 25 to 65 of being in each of these different systems. And so this is basically saying, what's the certainty equivalent of putting someone, given their lifetime income profile, in a given insurance market design. And what's the certainty equivalent of that over these 40 years, and this is in thousands of dollars here means $53,000. And so here, the first column is first-best, second is spot, third is dynamic, fourth is first-best without smoothing, and fifth is the ACA contracts. There's a couple things to take away. First, is that the dynamic contract when you're looking at the case of flat net income paths, which is the top row, the dynamic contracts add value relative to the spot contracts because they're allowing you to reinsure. And they're actually getting you most of the way to the first-best without inter-temporal consumption smoothing. In the flat net income paths, the dynamic contracts are better than the ACA contracts. The reason is they're allowing you to reduce this degree of adverse selection in the market, but you're not getting this negative impact where you're having to front-load in a state where you have low income. When you go lower down here to the steeper income paths, which our manager the ACA market designs, were studying are better for people than the dynamic market design. And the reason is that even though the dynamic design is better able to reinsure you against risk, it's forcing you to move income to these states, to these years, where you actually want income. So think about a young person who's being forced to pay $2,000 more than they want to into the insurance company. Audience Question: Would you clarify what you mean the adverse section under the ACA. What is it? Ben: Oh, yes. So in ACA when you have community rating, there's more likely to be adverse selection, as you know, than if you're doing risk rating. The reason is you're pooling together sick and healthy people and having them face the same premium. There, we're doing an equilibrium analysis of this static market with community rating. The idea is that in this market there is substantial adverse selection because of the community rating regulation, and we do it with and without risk adjustment in different cases. And that adverse selection leads to lower degree of insurance coverage within a year, but great protection from