The technology is simple enough, conceptually. You need some way of detecting impulses inside the brain, and you need to map these onto outputs. At the moment, they (Conduit) are assembling a dataset by having people go to an office in San Francisco and have a conversation with AI. They’re a bit vague in how they are collecting the data, but it’s non-invasive and probably isn’t an MRI. I am skeptical it is going to work perfectly. There is a lot of noise in what people are thinking about, and a lot of noise in what we observe in people’s brains. However, just as typos in training text have not prevented AI from existing, I expect that we will have improving technology so long as there is money spent on it.

Has nobody considered what the world looks like where they succeed? Suppose that they are right, and we are able to detect and interpret people’s inner monologues, conjured images, and even as yet unexpressed thoughts. They think it is going to be a world of people using it to control their AI agents with extraordinary efficiency. I need them to stop thinking about what people could do with it voluntarily. Start considering instead what people will want to do with it involuntarily!

Subscribe now

There are two categories of problems. First, it will enable authoritarianism. Second, even if we avoid this, it could remove our ability to think independently. The first case seems, to me, to be much stronger than the second, although both are real.

This technology need not be used to monitor the thoughts of the public at large. You simply do not need to do this. Instead, all you need to monitor are the security forces. Authoritarian states are already perfectly capable of suppressing the masses. What they are unable to do, and what is the real reason that authoritarian states fall, is to perfectly monitor those in positions of power who could replace the current leaders. Popular protests lead to the fall of regimes only when the security forces are unwilling to shoot. An armed security force that is willing to shoot an unarmed population without limit will always win.

Authoritarian states are normally unable to monitor everyone personally. They invest a lot in coup proofing, which is having everyone be monitored by everyone. No single entity holds the power to replace the leader. Now, you don’t need that. You can select people who are perfectly loyal, with a device that serves anyone.

Consider this. Do you want President Trump to have a technology that can detect, with real accuracy, whether you would support him in a coup? Do you want the President to be able to scan everyone for loyalty to him personally, above the Constitution? Do you want every government in the world to have this? It is already standard to screen people in positions dealing with classified intelligence for their honesty and loyalty, through polygraphs. We have been saved due to them not working. Why should we make them work?

Beyond the use in authoritarian regimes, which will definitely happen, I am also concerned that it will narrow the ranges of thoughts which are acceptable to hold. Suppose that their world comes to pass, and everybody interacts with their technology as a matter of course. In the ideal scenario, by 2027, AI agents will be able to interpret our thoughts, and react to them. Her specific vision of her daily life involves an AI agent interpreting her scarcely voiced thoughts about an interview at lunch, and acting upon them. Any agent capable of doing that is perfectly capable of acting upon scarcely voiced dissatisfaction with the present administration, with a desire to run away, or even just opinions that differ from the community at large.

I fear society will tear itself apart over intrusive thoughts. Did you know that, in Britain, 12,000 people are arrested and questioned every year over online communications that could cause “annoyance”, “inconvenience”, or “anxiety” to others? Britain is, despite the online jokes, a rich, liberal democracy. Suppose that we have a window into everyone’s thoughts. Are you so confident we will restrain ourselves from arresting everyone who has had a flash of attraction to children or to their siblings or to their mother? Do you want to be known as the sort of person who doesn’t want to protect the children?

Perhaps you endorse this, because this is our taboo, and our taboos are right. But do you want an Ayatollah to detect traces of apostasy? Do you want a counterinsurgency campaign to detect who supports their cause, and who does not? Do you want the Nazis to know exactly where the Jews are hiding?

The downsides must be fairly considered against the upsides. Many technologies are not unconditional goods. But what are the upsides? If all goes well, it allows us to interface quicker with a laptop. Where is the improvement? I type. I type quite fast. Lots of people type fast. If the problem is that people cannot move their fingers fast enough, then you can record a voice statement – transcription software is plenty good enough for that.

If anything, as AI agents get smarter, human interaction is going to matter less and less. We direct them to do a task, and they go work for a long while. The present workflow skews toward things happening which you cannot keep up with thinking in live time. This sort of technology is read only; it cannot write to the brain.

The best use case is a physical extension of the body. Plausibly, we could control a large machine with our mind. This is, incidentally, the plot of Pacific Rim. I think such scenarios are a bit far out.

I think the people involved with Conduit should stop what they are doing. Go back to work at OpenAI! I’d far rather you take your chances with AI that might kill us all, than with an AI that will very definitely enslave us.

This post is brought to you by Mechanize, Inc. They are hiring for a variety of positions, including software engineers. I encourage you to apply here.

This is all well and good. But what does it mean? We can say that there exists a gap in callback rates, but this does not give us any indication of how it maps onto labor market outcomes. Depending on our model of the labor market, discrimination by individual employers is consistent with no difference in average wages, or with a very large difference in average wages. It is perfectly consistent with being an accurate and efficient reflection of skills, and also consistent with discrimination causing the differences in skills. In short, we know much less than we think we know about discrimination!

This post is brought to you by Mechanize, Inc. They are hiring for a variety of positions, including software engineers. I encourage you to apply here.

Consider the following model. We are in a competitive economy, and all workers and firms have identical productive capacity. There are two groups of workers, White and Black. Firms differ from each other only in their preferences for hiring workers of one color over another. Some firms prefer White workers to Black workers, and are willing to pay a premium to hire them. Other firms are completely indifferent between White and Black workers, and simply hire whoever’s price is lower. Let’s say that 20% of firms are in the former category, while 80% are in the latter.

In such a world, there is no difference in wages whatsoever. The wages which Black workers are paid depends upon their marginal employer. None of the Black workers are affected by the discriminatory firms, because they all go to the non-discriminatory firms. It is still possible, of course, in this Becker (1957) world, for Black workers to be worse off as a result of prejudice. If the discriminatory firms are sufficient in number to crowd out the non-discriminatory ones, then there will be a gap in wages; and if we break the assumption of identical firms, productivity being correlated with discrimination will lower wages for Blacks. However, it suggests that if most discrimination is eliminated, we will have done all the necessary work to remove the actual gap in outcomes.

Subscribe now

Many audit studies are not able to detect this. If you suppose that at the margin firms are willing to hire an unlimited amount of labor, then if you, as Bertrand and Mullainathan do, send out four resumes, with White and Black (2,2) to each firm, the discriminating firms will show up as returning (2,0), and the non-discriminating firms show up as (2,2). But you’ve overlooked the fact that, if the unprejudiced firms received more applications from Black workers, they might return (2,4) or (2,6) or (2,whatever)!

This talk of applying, though, shows that we’re leaving something rather important out of the model. People search for jobs. Now we’re back to your intuition that firms discriminating must surely lead to lower wages. Suppose that workers pay a cost to apply to a job. They don’t know which firms are discriminating against them. This lowers the expected value of applying to jobs for Black workers, so they settle for lower wages. (Black, 1995).

Curiously, this suggests that a low level of enforcement of anti-discrimination laws is worse than no enforcement at all. Suppose that enforcement consisted of prosecuting anyone who admitted to discriminating, but did not touch anyone who discriminated in practice. In that world, it would be more efficient for jobs to simply post that they will not accept applications from Black workers, sparing them the cost of applying. Wages for both Whites and Blacks would rise in such a world. (Why? Suppose that firms face a cost to create a vacancy. Since they now can search more efficiently, they will increase the number of jobs available).

But of course, whether information actually helps depends on the model. Suppose that there is either an infinitesimal difference in White-Black productivity, or else that all employers have a very small preference for White workers, as in Lang, Manove, and Dickens (2005). All workers are otherwise identical. All firms post jobs publicly, with an attached wage. Workers pay a small cost to apply to jobs. Black workers know that if both they and a White worker apply for the same job, they will be passed over. Rationally, they choose to apply to less well-paid jobs. The gap in wages can be made arbitrarily large, so long as there is any gap in preferences.

Nor is it clear that holding skills constant is a meaningful thing. People make investments into their skills based off of the anticipated return to them. If people expect discrimination, they invest less, and thus the discrimination justifies itself.

It’s also not at all clear that the callback rate is a stable, meaningful object. Suppose that productivity is a function of two factors, A and B. One is everything that can be captured on a resume, while the other is everything that cannot. It is obvious that if the two groups of workers differed in their average level of the second ability, then holding the first fixed will result in a gap in hiring and wages by race. This is uninteresting. What is interesting is what happens if we fix mean ability, and allow only the variance of the second factor to be different. In this world, how large the difference in callback rates is, and whether it exists at all, depends upon the level of skill encoded in the resume in relation to the threshold at which people are hired.

This point was first made, to my knowledge, by Heckman and Siegelman (1993), which seems to have disappeared off the internet. Instead, I rely upon the exposition of James Heckman (1998). Suppose that firms hire when the combined sum of A and B exceeds some value. This value is different for all firms, and is symmetrically distributed. If the level of A in the resume is above average, then the firms will hire more people from the group with smaller variance in B. If the level of A is below average, then firms will hire more people from the group with wider variance.

Relatedly, discrimination is related to how far up the ladder you are. Suppose that everyone knows Black workers are discriminated against in entry level jobs. Someone progressing up the ladder, holding A fixed, implies a higher level of B. It’s entirely possible for callback rates to be worse for Black workers early on, and then better later. Clearly, we should focus on entry level jobs, as people do, but we would then miss discrimination that is open to employing a member of a disfavored group but not to promoting them. (This is particularly relevant for discrimination based on gender).

It is possible to adjust for the variance concern, as Neumark and Rich (2019) do. Going back to Bertrand and Mullainathan, they sent out two sets of paired resumes, which differed in their implied ability. If you are willing to make strong distributional assumptions – specifically, the distribution of the unobserved trait is normal, it is the same for all jobs, and the only thing going on is a difference in variance between Whites and Blacks – then you can back out the implied distributions, and identify how much discrimination there actually is.

I think that their results support, on the whole, there being discrimination. Their emphasis is that the labor market results are less robust than housing discrimination, but that is substantially just a loss of precision when we move to a less restrictive model.

I have not seen anyone convincingly put this together. Neither do I expect anyone to, for several reasons. First, measuring the accumulation of human capital and how it maps onto wages is essentially impossible. Even if you do have the perfect, experimentally induced variation in wages and can map it onto wages later (and you do not get this in practice – Chetty, Friedman, Hilger, Saez, Schanzenbach, and Yagan (2011) is a common citation for this purpose, but they actually take the starting correlation of test scores and wages, and assume that the experiment induced changes in test scores will show up identically in wages), you cannot be sure how much of the wage gains is due to them displacing others. You’d need to know the country’s production function. Second, and relatedly, nobody actually knows the correct model of the labor market. There are multiple competing models, each with different strengths and weaknesses. They are used to answer different questions – Diamond-Mortensen-Pissarides with Nash bargaining captures variation in unemployment over time but without variation in wages, while Burdett-Mortensen models which have workers search while on the job can generate wage dispersion but have nothing to say about unemployment in a recession.

Where the topic has moved to instead is detecting which firms are discriminating. Kline, Rose, and Walters (2022), along with Kline and Walters (2021) which details the identification strategy, aggregate many resumes (83,000) sent to 108 major U.S. companies. The point is to be able to say, “in each cell of firms which responded in a given way to the application, what percentage of them are discriminating?” There is no claim about how big this is, how much of wage gaps it can explain, or whether it is efficient or not. They don’t need to. Instead, it is restricted to a question which can actually be answered, and acted upon – who should be investigated for discrimination in employment? They find a gap of two percentage points against a callback rate of 25%, and are confidently able to describe 23 individual companies as discriminating at the 5% significance level. And that’s as good as it’s going to get.

The two articles are John Hassler, Per Krusell, and Conny Olovsson (2026), and Leo Kaas, Etienne Lale, and Nawid Siassi (2026). The first is forecasting the effect of climate change, how it varies across the world, and what the effect of various policies to prevent it would be; the second is integrating the wage bargaining framework of Cahuc, Postel-Vinay, and Robin (2006) into a model with precautionary savings in order to show how changes in unemployment insurance cause people to save more which in turn affects the real interest rate and the number of vacancies.

These are both macro papers, so they give the sense of having a lot of distinct modules which are clipped onto each other. Hassler, Krusell, and Olovsson (HKO) divide up the world into 16,500 grid cells, and assume that nothing moves between countries except the air and oil. People and capital allocate themselves across cells within a country, with production of a single final good happening through a Cobb-Douglas production function with decreasing returns. Decreasing returns prevents everything from piling into one cell, and because we are at equilibrium, we can render the country’s total factor productivity as an aggregate of each of the cell TFPs. Later on, temperature will affect each of these cells’ TFP (and differently, because places will warm different amounts), but we can still reallocate, and solve for the country TFP as one region.

This post is brought to you by Mechanize, Inc. They are hiring for a variety of positions, including software engineers. I encourage you to apply here.

Energy is handled by pushing all the different energy sources into a big composite. One branch is fossil fuels, another electricity, which have some elasticity of substitution between the two p; electricity splits into plannable (hydro, nuclear, coal, oil, and gas) and unplannable (solar and wind) sectors, each with elasticities of substitution. The elasticity of substitution between fossil fuels and electricity is low, meaning they cannot easily replace one another, while the elasticity of substitution between similar sources is high. Oil is made exogenous, so it does not respond to price at all.

Catch up growth between countries is assumed. Developed countries grow at 1.2% forever, and developing countries remove 1/4th of the gap between them and the developed world every decade. Modeling of the climate is essentially outsourced, but since it only depends on aggregate carbon emissions, including it is very simple. Then, damages from the temperature are found by looking at the association of temperature and TFP.

There are two big problems. The first is their measure of how temperature affects Total Factor Productivity. When they look at the cross-section, they observe that very cold places are not very productive, productivity rises as we move to temperate regions, and falls when we get to much hotter regions. The figure in their paper looks like this.

In order to evaluate the economic effects of climate change, we assume that this is a stable, meaningful relationship. In other words: the only difference between Norway and Nigeria is the temperature, and if the temperature in Norway rose enough, they would regress to the TFP of Nigeria. (And likewise, that if we were enter another Ice Age, Nigeria would boom).

This is simply ridiculous. It is self-evident that there is more that varies between countries than just temperature, even if it is often correlated with it. At the very least, we could be comparing regions within a country – Minnesota compared to Texas – although even that need not purge spurious correlation. Simply reading off the change in productivity as a function of temperature ought not work, even if we throw more and more control variables at it.

I grant, however, that this is a fundamentally hard problem. The alternative method would be to use shocks to the global temperature, as Bilal and Kanzig (2026) do, but this is also problematic. What we’re actually interested in is the long run impact of temperature on outcomes, after we have had time to adapt. It is to be expected that storms caused by warmer water off the coast of West Africa should lower U.S. GDP, but if we expected more storms we would also be prepared and reallocate where we are spatially. We also have to assume that we can extrapolate a 0.2 degree Celsius shock to a 1 degree Celsius shock by multiplying the damages by 5, and we also have to that it’s plausible that temperature shocks only affect world output with a delay of 4 to 6 years, which does make the story I told of hurricanes implausible. (See Figure III of BK for the key graph).

The mapping of temperature and climate to economic damages is basically only credible when it restricts itself to agriculture. Costinot, Donaldson, and Smith (2016) is a textbook example of what you can do – we have extensively tested crop productivity under different soil and weather conditions, so a change in weather is simply moving from one tested scenario to another. Not so when it’s the whole country changing weather!

The second problem is how they treat substitution of electricity and fossil fuels. The counterfactual most of interest to me was their comparison of a carbon tax, versus subsidizing lower emission technology. We have shown no political capacity whatsoever for imposing a carbon tax, but subsidizing green energy instead has proven to be much more politically palatable. HKO say that this is not a substitute, however. In their model, a uniform carbon tax is effective, while subsidizing green energy does close to nothing.

But why? Everything runs through the parameter dictating the substitutability of fossil fuels and electricity. If you set it low, then subsidizing green energy simply expands total energy production. If you set it high, then eventually green energy replaces fossil fuels. HKO have their parameter set at 0.66, bumping up to 1.5 after 30 year and there she goes – a carbon tax will always be much better than subsidizing green energy.

But where do they get this number from? They get it from section 3.2.4 of the WITCH model. Where do they get it from? They get it from Koetse, de Groot, and Florax (2008) – but here’s the thing, they substantially don’t. Koetse, de Groot, and Florax are measuring the substitution between capital and energy, which is not the same thing as electricity versus fossil fuels. This is besides the fact that they are absolutely not doing best practices for meta-analyses – they eliminate studies with implausible conclusions, but if you believe that all of the studies are measurements of the underlying phenomenon with error, then eliminating those will bias your results. The studies which underlie it are all estimating production functions on aggregate data, which is an approach that does not work, and there is a reason why that line of work has been completely abandoned in favor of estimating production functions at the firm level. And even if we overlook the fact that they’re not measuring the same thing, and the studies don’t really measure the thing we’re hoping they do, the specific number 0.66 isn’t actually in the paper – it’s just vaguely around some of the numbers they find.

To put it bluntly, the estimate is made up. It is made up perhaps without any intent to be made up, but it was made up. Somebody took their best guess at what a reasonable parameter value would be, and it passed directly from guess to result. For really important problems were we can’t identify the parameter value, it’s certainly understandable that we do this, and it has a lineage – Hansen and Heckman (1996) note how pioneers of simulation like Tinbergen and Frisch often guessed at parameter values – but it means that you simply cannot cite the result as settled science. It is scarcely science at all! It’s a big excel spreadsheet, and if you move one number everything else deterministically follows.

Read more

1. The primary cause of the Recession was the shock to mortgage valuations blowing up the collateral upon which the financial system relied. These securities were the foundation of an unregulated banking sector, and when the valuation of the securities was called into question, the unregulated banks folded.

2. When these quasi-banks folded, aggregate demand fell, and we reached the zero lower bound.

3. Using the tools it had available to it at the time of the crisis, it is not at all clear that the Federal Reserve could have prevented the banking collapse or the shortfall in aggregate demand. However, quicker action would have reduced the severity of the crisis.

4. The financial crisis could have been substantially reduced by actions the Treasury could have taken before the crisis, namely producing more safe assets to serve as collateral.

5. Nominal GDP targeting would have been an improvement over inflation targeting, but it is difficult to think that it could have been implemented during the crisis. In particular, I do not think that Congress would have accepted the necessary inflation.

6. A recession of some size was going to happen anyway.

7. A shock to the expected returns of AI poses little danger of a recession along the lines of 2008.

This essay is very large. For your convenience, I will be adding links to sub-sections after the post goes live.

And before we begin, this post is brought to you by Mechanize, Inc. They are hiring software engineers to train the AIs of today to automate the work of tomorrow. Apply here today.

What does finance do?

Finance is fundamentally about making trades across time, and rearranging consumption in different states of the world. Let’s think through what a bank does. They solicit deposits of currency from depositors in exchange for paying interest, and then make loans to borrowers at a higher interest rate than what they are paying the depositors. These loans are into projects which take a long time to come to maturity.

What keeps the individual from making these trades themselves is that they face unexpected shocks in the future. They might have their car break down, or perhaps they would like to buy a bicycle. They cannot recall their money from someone who has taken out a loan to build a house and profit from 30 years of renting. By pooling money together, the bank can give people the ability to recall their deposits at any time. This also means that a bank can lend out more than it has at any given time, knowing that they will never be called upon to provide it all at once.

Is this unstable?

In practice, no, but in theory, yes. Suppose that everyone has common knowledge of the state of the world. Everyone knows that the bank can pay its obligations if only a few people call upon their deposits, but also that if too many people called upon the bank’s reserves, only the first people there will get anything. The bank cannot recall the loans before the projects which they are invested into are ripe. Grant that something changes in people’s beliefs. If people believe that the bank will fail, their best outcome is to go to the bank now and get their deposits before they lose everything. Thus, whether a bank exists is entirely determined by people’s beliefs about whether the bank will continue to exist. Diamond and Dybvig (1983) are the ones who formalized this model, but the basic intuition has been known for ages.

Subscribe now

How do we keep this from happening?

The first thing is to realize the model is a bit too strong. Common knowledge, which means that not only does everybody know everything, everybody knows that everybody knows everything and so on, is an assumption with unrealistic implications. If common knowledge actually existed, then we would not observe any trades at all, besides those needed for liquidity. (Intuitively, Milgrom and Stokey (1982) points out that making an offer to buy or sell reveals private information you have about the payoff of the asset, which discourages one from actually accepting the deal).

Still, though, a bank run is a serious concern. And that’s why the federal government, through the FDIC, offers insurance on bank deposits. By guaranteeing that you will be made whole in the event that the bank collapses, we remove the incentive to be there first, and thus remove the incentive to collapse a bank. This has essentially made bank runs by retail depositors who are covered extinct. Even before the FDIC, though, banks were not helpless. A panic could be stamped out by the major banks using their credibility to take on the distressed assets, picking through what is good and bad, and lending freely to solvent firms, against good collateral, at high rates, in the words of Walter Bagehot.

The best empirical work on bank runs – Emil Verner is making a career out of this, see Correia, Luck, and Verner (2026), among others – is pretty conclusive that complete nonsense runs don’t really happen. Runs happen when the bank makes bad investments which were going to fail, and leave it unable to pay its creditors. Still, you can have real destruction in pulling back the loans too early. The ideal is to have the insurer take control of the bank quickly, and then unwind in an orderly manner which preserves as much value as possible. If a company has to sell in a hurry, then they can hardly be expected to get the full value from their assets.

How does finance work for large institutions?

Basically the same, except they don’t necessarily work with entities which are necessarily called banks. Suppose you are a large company with hundreds of millions to billions of dollars, which you use to make payroll, buy materials, and so forth. You can’t deposit the money into a bank, because the FDIC only insures up to a certain amount per account holder per bank. Instead, you must construct something like it, through a repurchase agreement.

What’s a repurchase agreement?

You, the large company, would like to invest your money in assets which take a long time to mature but pay a higher rate, such as in corporate bonds or a mortgage backed security. You cannot get a traditional deposit, but you can make something like it. The company gets (to put real numbers on it) $100 million in securities from the dealer (the bank), in exchange for the company giving the dealer $99 million in cash. This is held overnight. Come morning time, the dealer is obliged to repurchase the securities for $99 million in cash, plus some interest. The difference between the cash lent, and the value of the securities, are called the haircut. You withdraw currency by declining to roll over the arrangement in the morning.

How large was this form of credit provision?

Bigger than normal banks. These shadow banks, according to Pozsar, Adrian, Ashcraft, and Boesky (2012) provided about $22 trillion in financing at peak in mid 2007, although nobody knows the precise number.

What is a mortgage backed security?

A mortgage backed security (an MBS) is a bundle of loans to home mortgages. The idea is that the risk of default for a mortgage is independent of the risk of default for other mortgages. By combining partial claims to thousands and thousands of mortgages, you can create safe assets out of risky ones. What’s more, even if you were concerned about the safety of the asset, you can divide up the claims by priority. The senior tranches are the ones who are paid first out of a given bundle of loans, and should be rock solid, AAA rated assets.

Why did people want them as collateral?

People had an enormous demand for safe assets to make repurchasing agreements work. People needed assets that had economic value, but required them to know nothing about the actual details of the asset.

I should emphasize that mortgage backed securities made up only a part of the wider universe of asset backed commercial paper (ABCP). These are purposefully opaque

Wait, really? Why would people not want to know the value of the collateral?

Because eliciting that information is costly, and the payoff to knowing the exact quality of the collateral is flat. All you care about is that it is good enough for someone not to renege on the deal. Doing the evaluation once, and then never again, is what is efficient.

This idea comes from Robert Townsend (1979), but received its most forceful exposition by Bengt Holmstrom (2015). Debt is very different from equity markets, like the stock market. There, your payment is determined by the underlying value of the company. Acquiring more and more accurate information about the state of the world allows us to allocate capital to better uses of it. In the case of debt, you don’t get a penny more for discovering exactly how much the company will make, so long as it does not go bankrupt. The ideal is to exchange pooled long run assets for pooled short run assets, which neither borrower nor lender has evaluated for quality. In fact, if you own assets and wish to trade with them, you want to keep yourself from having the possibility of knowing what your assets are worth. The possibility of knowing more about the payoffs of your own assets makes you worse off because everyone suspects you might be trying to rip them off.

People got their securities rated by one of the rating agencies, and that was the end of things.

Subscribe now

Who rated the bonds?

Moody’s, Standard and Poor, and Fitch.

And who issued them?

There were two groups of issuers: on the one hand, the government sponsored enterprises Fannie Mae and Freddie Mac, who bought conventionally sized and underwritten loans, and on the other hand, private label issuers, who securitized everything else. Fannie Mae and Freddie Mac are strange organisms – Fannie Mae was originally set up as a government agency in 1938, then spun off into a quasi-private organization in 1968 (with Freddie Mac created two years later as “competition”). The motivation for this was that they wanted the agency off of the federal budget because it looked bad – see page 104 here – but in practice people treated it as still guaranteed by the government.

This is a pretty bad state of affairs! Fannie Mae and Freddie Mac understood their position, borrowed at rates nobody else could get, and were heavily levered. It produced profits by shoving all its risk into the tails. Then when things blew up, it took the involvement of the government to keep them from taking the economy with them.

Who was buying all these new mortgages?

The expansion in mortgage credit was broad-based. Contrary to the popular narrative, this was not just about subprime mortgages – if you go by the value of the loans underwater, high and middle income borrowers were just as important.

How important was actual fraud in mortgage issuances?

Important, if not as important as people think. Like there was definitely a lot of it going on – Piskorski, Seru, and Witkin (2013) found that a tenth of loans were misreporting either whether the owner was the primary occupant and/or whether there was a second loan with the house as collateral – and with the mortgages bundled off into securities and out the door, there was little incentive to check. Nevertheless, the rise in housing prices was broad-based enough that it was not simply about fraud.

How should certification happen in a financial market?

The idea is that the firms acquire a reputation for soundness and plain dealing which makes their claims trustworthy. But the trouble is that this need not hold when the risks are rare but enormous, and the amount that the rating agencies have at stake is limited. Competition, in the form of the entry of Fitch, led to ratings actually getting slacker. (Becker and Milbourn, 2011)

A substantial part of what certification is doing is not for the benefit of actual investors, but to meet regulatory requirements. Frank Partnoy (1999, 2006) was warning about this long before the financial crisis – the point of an AAA bond rating is to get the bond in the door, and allow you to do things for which no one is responsible for with other people’s money. I think this is only partly true, because

I am roughly in favor of getting rid of shopping around for ratings. If someone wants an asset rated, it’s assigned at random to a qualified firm, and you live with what it says. This was proposed in the aftermath of the Great Financial Crisis, but scrapped. For complicated assets, people simply shouldn’t treat the credit agency rating as definitive, and should just not treat it as money.

Subscribe now

Was there a bubble?

Probably. But it’s going to be observationally equivalent to some other things.

What is a bubble?

It is not, to be clear, simply any time that prices were high and then fell. For instance, it is not a bubble if there exists uncertainty over the future development of a technology, and we happen to get bad luck.

How is it possible to have a bubble?

This is a serious question. If we make the innocuous assumption that people prefer more to less, and are not actively trying to lose money, it’s pretty difficult to justify. If the game has a known number of periods, somebody’s gonna have to be holding the bag at the end; knowing this, people will choose not to get involved. At best, you break even on average, and if you add in the slightest bit of risk aversion or declining marginal utility, the utility from getting involved is negative.

There are two basic ways to get a bubble, one with rational agents, and the other with some number of irrational agents. The rational agents are pretty striking and clever, and have a similar root to dynamic inefficiency in an overlapping generations model. You have agents who are born, work, retire, and die. They can save at the interest rate r, and the economy is growing at the rate g. If r is less than g, then the economy is growing faster than private savings are, and it would be efficient for us to fund people’s retirements through taxes on labor income now. We’re essentially changing our savings from a worse to a better vehicle.

A bubble, which is an asset which grows at the growth rate of the economy, is something which sops up capital until r equals g. A bubble, then, can not only exist, it actually improves outcomes. The reason it doesn’t happen by default is that no one is incentivized to start the chain of transfers – but if the system came into being that way, or in the case of Tirole (1985), the bubbly asset started out that way, we would be better off. Perhaps you think that this is silly. But then, what is the fundamental value of money?

(Why must it start that way? Because if people expected a given asset to become a bubble, the expected value must be greater than the fundamental. If it is trading at the fundamental value, then people must never expect it to be a bubble. QED.)

These rational bubbles were likely not what’s going on, though. Housing prices were appreciating much faster than the interest rate. More likely there are some people who are just plain wrong about prices. We don’t necessarily need to know why they’re wrong – perhaps they’re just enthusiastic about the economy doing well. If this is the case, then a rational agent might want to play along with the bubble, and sell before it runs out of rubes. We need one more ingredient to make a bubble – some inability of arbitrageurs to bring the price down independently. Abreu and Brunnermeier (2003) is one such story – people know that there is a bubble, but they don’t know when other people learned that they are in a bubble. Thus, you want to play along.

How do we know that we’re actually in a bubble?

So I will grant that it is very easy to say that a bubble can exist. It is very hard to say that any particular rise in asset prices is a bubble without making a fool of yourself when the prices stay high indefinitely. We should be clear that there being a bubble in housing does not mean that all housing prices in all parts of the country were inaccurate. Indeed, housing prices are higher now than they were then, driven by appreciation in the supply constrained coastal metros, like New York City and San Francisco. Rather, we must argue that there was a bubble in housing in some part of the country.

The best case for us being in a bubble is that rent growth became detached from price growth, including in cities where it would be very easy to buy more housing. When you asked them, people expected prices to just keep on growing. Kaplan, Mitman, and Violante (2020) argue that changes in objective credit conditions did not explain the boom in housing, but changes in beliefs did. (Their argument is that if people are limited by credit from buying the living conditions they want, they will instead rent. Removing credit constraints results in people buying instead).

Couldn’t mortgages just have been a good bet, just bad luck?

I would argue that that is not what happened. Certainly it could explain the early 2000s dotcom crash, which I do not think was a bubble. You do, however, raise a good point. The exact form of the sudden price decline is not that important. What matters is leverage in the banking system, whether formally or informally a bank.

Why does this matter?

The mortgage backed securities were being extensively used as collateral. The whole class was correlated with each other. When the value of the mortgage backed securities fell, companies were now uncertain about the value of the things they’re buying in the repurchasing agreements. Indeed, they expect the dealer to know better than them which securities are good and bad. They leave, or demand a larger haircut.

The dealer, being a bank, is leveraged. They have borrowed extensively to conduct this operation. They cannot pay everyone at once, not until the securities they own mature. They must liquidate their holdings, and liquidate now. When they are forced to sell, they have to sell below the real value of the assets, because the people who are able to actually value the securities are the ones who have to sell.

Do all losses in value cause a financial crisis?

No, and the difference is in how levered companies are. Enormous decreases in value have come and passed without a crisis – in a single day in October, 1987, the Dow Jones declined by 22% without any discernible impact on the real economy. The tech boom had fallen just a few years ago in 2001, with the only fall in output coming from realizing that the tech wasn’t going to be as good as advertised and cutting back on spending. If there is no borrowing – if the assets aren’t being used as collateral – there’s no crisis.

Why is a financial crisis bad?

So we wake up one morning, and there’s bad news about the economy, large enough that we begin to doubt the worth of our collateral. The clock is ticking – if we roll over our deposits to the next day, but everyone else doesn’t, we’ll be left holding worthless assets and out millions of dollars. So we decline to rollover, something essentially costless to us if we’re the only ones doing it, and the best we can get if everyone else is.

The dealers were highly leveraged. Bear Stearns and Lehman Brothers, when they collapsed, were operating at 30:1 leverage. With their usual parties backing out, they need to get money and get money now. They start selling off assets, but since nobody has had time to evaluate the quality of the assets, they go for prices well below what they’re worth. That’s the fire sale.

The fire sale propagates outwards, because other people hold those assets too and use them as collateral. Now that the collateral is worth less, people need to get more cash, which they get by selling their bonds, again at progressively lower prices.

All of this was being used as money. Nominal GDP, which is price times output, must equal money times velocity, where velocity is the number of times a given dollar bill is used during each period. Now that all these loans are winking out of existence, velocity is falling. There were fewer dollars available to spend on goods and services. Either prices or consumption must fall, and since prices did not fall enough, real consumption and employment fell. But more on this later.

It was confusing to people at the time that what we conventionally call “money” – currency, bank deposits, etc – did not fall during the crisis, and in fact grew at a 5-10% annualized clip throughout. The fall in aggregate demand was due to money like objects, the asset backed commercial paper, which collapsed. Nominal GDP fell by 4%.

Subscribe now

Why, exactly, is a fire sale externality an externality?

It’s certainly a bit weird, because things which occur through prices are not usually externalities which we care about. An externality, incidentally, is a cost or benefit which falls upon someone who is not party to a transaction.

Parties hold assets as collateral, and how much they can borrow is related to the value of that. How do we value it? Well, we must consult the market. If someone is selling AAA mortgage backed securities, same as the ones you hold, for well below their fundamental value, your collateral has lost value and you too must start selling to get cash to unwind your position. It spreads from person to person.

Can you give me a chronology of what happened and when?

I have been shying away from this, because there is simply so much detail that presenting all of the events would boggle the mind and impede understanding. To my mind, the principal events are these:

– National housing prices peaked in the middle of 2006, and began falling.

– More and more people started defaulting on their mortgages. On August 9th, 2007, BNP Paribas said it can’t value holdings of subprime mortgages in three funds, and halted withdrawals.

– In March 2008, Bear Stearns collapsed. This was handled by JPMorgan buying the company, and the Federal Reserve guaranteeing part of the assets.

– On September 7th, Fannie Mae and Freddie Mac were placed in conservatorship, where they have remained ever since.

– On September 15th, Lehman Brothers collapsed. AIG, which was in danger of collapse, gets bailed out with a ton of money. This is, properly speaking, the financial crisis.

– Output collapses. GDP fell 2.2% in the fourth quarter of 2008. (That’s 8.4%, annualized). The Troubled Asset Relief Program, which was $700 billion to recapitalize banks, passed October 3rd.

– In February of 2009, Congress passed an $800 billion stimulus bill. The recession ended in June of 2009, but unemployment would peak at 10% in October, and employment did not go back up to the level it had attained before the crisis until 2014.

I want you to understand that this was a crisis which happened slowly, then all at once. Employment started falling in 2007, well before the collapse of Lehman Brothers. Still, the big slide in output and unemployment occurred after the financial crisis.

What is the Federal Reserve?

The Federal Reserve is a unique government organization. It was created by statute as a consortium of the major banks in 1913, formalizing the role that those banks had in preventing panics previously. They would step in to act as a lender of last resort, propping up sound business in order to prevent the collapse of the system.

What does the Federal Reserve do now?

The Federal Reserve manages the money supply of the United States. This is a relatively new development – before the 1960s or so they were much more of a consortium of banks in operational practice than a branch of the government held at arm’s length. They still do do bank regulation.

The Federal Reserve is in a policy regime of inflation targeting – they want the price level to increase by two percent every year.

How does the Federal Reserve do it?

Notionally, it uses the tools available to it to manipulate the money supply. More on that later. But fundamentally, the power of the Federal Reserve is that it selects the equilibrium we are in. There is a lot about the Fed which is driven by people’s belief in it.

What tools does it have?

The standard instrument is its control of the federal funds rate, and the rate at the discount window. The federal funds rate is the rate at which banks in the Fed lend to each other in overnight reserves. It used to be that they would change this rate only indirectly, by buying and selling treasuries until the rate matched what they said. (This works because the banks were required to meet the reserve requirement of having 10% of deposits on hand, and so banks would borrow from each other to meet the target at the end of the day). Now it manipulates the interest that it pays on reserves directly.

Interest on reserves was introduced during the Great Recession, incidentally, in order to give them more tools. Unfortunately, by hiking the amount that they were paying on reserves at the same time that they cut interest rates, they undid the rate cuts during the crisis meant to stimulate demand. It was rather like – I endorse giving the car a steering wheel, but that does not mean I endorse turning it into a brick wall.

The Federal Reserve can also lend directly to banks, at the discount windows. This option is not as powerful as one would hope, because everyone understands using this to mean that the bank is in financial trouble. There is a distinct stigma. When using the discount window is absolutely necessary, the Federal Reserve will often call the major players together, and tell them to all borrow from the discount window together, so that no one has to indicate that their position is weak.

In abnormal times, it can engage in quantitative easing. This is when the Federal Reserve buys assets for cash (reserves). The Federal Reserve cannot, as is normally conceived, simply print money. Quantitative easing is supposed to be swapping treasuries for reserves, which have well known payouts. The Fed did engage in buying mortgage backed securities, which while they only bought securities guaranteed by Fannie Mae and Freddie Mac, is still pushing resources to the housing sector.

If the Federal Reserve were simply printing dollars to buy assets, then it can always force changes in prices, and there’s really no question about it. If people don’t increase their prices, they will find themselves selling the shirt off their back for pennies on the dollar. If the Fed is swapping treasuries for reserves, though, it’s a bit difficult to see how this leads to higher inflation. Treasuries at the zero lower bound are essentially a perfect substitute for money.

There are two paths for this. The first is that it affects the duration of debt. A bond which takes a longer time to mature is exposed to the possibility that interest rates will change in the meantime, and we would have preferred to have bought a different security. This requires a premium. When the Federal Reserve buys long run treasuries and emits reserves, it gets rid of the interest premium and reduces the cost of borrowing. This is stimulatory.

The other is more of signaling. The Federal Reserve has the power to squash inflation by hiking rates again, even if it does not have the power to create inflation by cutting below zero. People will naturally be concerned that when inflation does rise, including above the target, the Federal Reserve will squash it down. Owning lots of treasuries means that hiking interest rates too early will be made more and more expensive.

Subscribe now

This is really weird. Isn’t the Federal Reserve part of the government? Does it meaningfully face costs?

The interest rates it sets do ultimately affect how much the government will be paying in debt payments, which flow to people outside the government. If it hiked interest rates too early, it would have to stop remittances to the Treasury and it would look like it is losing money. The assumption is that Congress would get mad if that happened.

It just goes to show, though, that above all what quantitative easing does, when it is not fiscal policy in disguise, is try to coordinate expectations!

What happens if the interest rate is at zero?

Then we are at the zero lower bound. This spells trouble, because the Fed can’t cut. People will simply hold cash. Having settled down at zero, we might have a very hard time getting back up again.

Bad?

Very bad. Very, very bad.

Can you walk me through why?

The financial crisis happens. The wealth of households falls. We presume that they want some level of savings relative to future income, which is what their wealth is. The only way to build back up savings is to consume less and save more. The real interest rate, which is what it costs to consume now in exchange for payment later, can fall negative. Savings exceed desired investment at the market clearing price. The Federal Reserve cannot lower the interest rate anymore in order to clear the market, so what must fall is total output and total consumption.

Making it worse, people rationally expect this condition to last for a while, which is what makes it actually last for a while. Because they expect lower spending in the future, they cut their investment and increase savings now, which further decreases demand. When we are at the zero lower bound, there’s no reason to expect a shortfall in aggregate demand to get better anytime soon, and we can have a persistently sluggish recovery – as indeed, we saw.

Why don’t prices and wages simply change to accommodate the fall in the money supply?

That’s what we’d expect, but this need not hold at the zero lower bound. We would expect money to be neutral in the long run, and indeed, under some circumstances we would expect it to be perfectly neutral in the short run. If the Federal Reserve went out and said “tomorrow, all quantities of money are going to be ten times higher, add another zero to the end of it” and this genuinely applies to everything in the economy, without a being a trick of some kind to get out of debt, I would expect this to have a minimal effect on the economy. Everybody just adds a zero, and we get on with things. Less fantastically, countries have been able to switch over to new currencies without much fuss, and the UK was able to switch over from pounds-shillings-pence to decimalized currency without causing a depression, so we cannot always expect changes in the quantity of money to matter, and certainly not forever.

This intuition falls to pieces at the zero lower bound. The financial crisis cuts money supply (demand), and prices and inflation fall. The interest rate is stuck at zero, so falling inflation actually raises the real rate of interest, further discouraging investment. An example. The real interest rate is simply the nominal interest rate minus expected inflation. The nominal interest rate is at zero. Expected inflation is also zero. Thus, the real interest rate is zero. If the amount of money is expected to contract by one percent, then the real interest rate goes up by one percent. Things end only when the shock that is causing the contraction in demand ends.

Second, I think there are complementarities to people’s actions, both real and through aggregate demand, and that a crisis can lead to a shift in people’s beliefs and lower investment indefinitely. There are a number of broad strokes models consolidated by Cooper and John (1988) on how we could be at less than optimal output because of a crisis of confidence in what other people are doing. The key is that there have to be increasing returns in some way.

We get out of a zero lower bound by, somehow, having people believe that inflation is coming. If we have fiscal powers, this is easy. We just buy real assets. If people don’t want their assets bought for a fraction of their value, they will raise prices. If we are constrained to monetary policy alone, then we need people to believe that prices will rise. We do this by credibly promising that we will run inflation hotter in the future.

What could we have done to prevent it?

This question branches out in a few ways. The first path is things we could have done before the crisis, in particular with mortgage backed securities. We’ll return to that. The second path is what the Federal Reserve could have done in the crisis, and this too divides up into two paths: we can change the Fed’s actions, and we can change the Fed’s regime.

The first path involves cutting earlier and harder. I do not think that this prevents a recession entirely, but it definitely prevents it from being a Recession. There are a few points at which they should have acted, and what’s more, should have acted with the information they had.

I can partially excuse the summer of 2008, because their data was bad. Inflation was 5% in 2008, while output seemed to be normal. I note that it seemed to be normal, because it later got revised down to better match the facts. There were warning signs, though. The financial markets, in particular, were positively screaming that things were not right. In particular, the gap between LIBOR and OIS widened, and stayed wide. LIBOR, the London InterBank Offered Rate, is the rate at which banks are willing to lend to each other unsecured; OIS, the overnight indexed swap rate is a promise between two parties, one of whom pays a fixed rate, and the other pays the federal funds rate. The spread between the two is basically the value banks put on not having to actually hand over assets to make a trade. For it to be up and stay up indicates an enormous breakdown in trust between banks.

What was absolutely inexcusable was September 16th, the day after Lehman Brothers collapsed. The Fed met in a regular meeting, and chose to hold the federal funds rate at 2% rather than cutting. When it finally did cut rates, on October 8th, 2008, it was too late. Even when it did so, in October, it introduced interest on reserves at the same time, offsetting what it was trying to do with rate cuts. By December, rates had fallen to 0, but it was too late.

Interest rate cuts are rather like balancing a pole on your hand. It is necessary to exceed the amount that it is falling in the opposite direction, and once it has fallen you cannot start again without picking it back up. To do that, you need to change the regime.

Subscribe now

What do you mean by changing the Fed’s regime?

The regime is a term of art in macroeconomics. It more or less means what goal the Fed is aiming for. If it’s aiming for a 2% inflation target, then the regime is how it is going to respond to real shocks. We can’t confidently port empirical results from one to another, not unless we have built why people do things out of their underlying utility functions.

An alternative regime, which I think gets resorted to de facto at the zero lower bound, is nominal GDP level targeting, or NGDPLT.

What is NGDP level targeting?

NGDP targeting is the idea that the Federal Reserve should not target a rate of change in the price level, the inflation target, but instead target a level of nominal gross domestic product (or NGDP). This is the nominal value of all final goods in the economy. The central bank proposes to reach some level of NGDP in each period, and if it fails to do so, to make up what was missed in the next period.

The point that I want to make, in as forceful a tone as I can express through the written word, is that a credible promise to run inflation hotter in the future, which gets us out of the liquidity trap, fundamentally is NGDP level targeting. A pure inflation targeting regime has two steady states, one where we are at the targeted level of inflation, and one where we are the zero lower bound forever. (Benhabib, Schmitt-Grohe, and Uribe (2001) are the ones who worked this out). We only have to believe that the Federal Reserve will follow its own rules forever. NGDP level targeting technically does have two equilibria, but having the economy stay at the zero lower bound requires that the Federal Reserve be completely non-credible about what it is going to do.

It’s important that we target the level, rather than letting bygones-be-bygones and restarting in each period. Otherwise it’s just a shadow of itself.

Are there other advantages?

Yes, and the biggest one is how we handle supply shocks. Nominal GDP growth is equal to real GDP growth plus inflation. Suppose that the economy is growing at 2% a year, and we are targeting a level growth of 4% a year, giving inflation of 2% a year. If the economy does unexpectedly worse than predicted, we look past this and have higher inflation in the period.

Two ways you can arrive at this. The first is that we don’t actually want to stabilize prices per se, we want to stabilize the prices which are sticky. NGDP turns into a proxy for labor incomes. You don’t get confused by oil spiking and bringing inflation with it, and thinking you need to cause a further reduction in employment to keep inflation on target.

The other is making a claim about the price changing process of firms. The standard device for modeling price changes in the economy is Calvo pricing, where some fraction of firms get the ability to change their pricing in each period. In that world, the optimal strategy is an inflation target (and at 0%), with the intuition that misallocation increases quadratically as we get away from the optimal price, and it is thus optimal to break a shock to one sector into a bunch of little distortions across all of the sectors. If firms face a deterministic cost to changing their prices irrespective of the price change size, as in Caratelli and Halperin (2025), then it’s best for shocks to one sector to not be shared across other firms, and instead have the price change be large enough for everyone in the affected sector to pay the menu cost to change their prices to optimum.

NGDP may also be easier to compute, although I think this is a minor advantage at best. To calculate NGDP, you simply sum up the value of all final goods in the economy. Calculating inflation requires doing this, then adjusting for changes in quality, then pulling out inflation as the residual. However, bias in quality adjustments is not actually important so long as it is the quality unadjusted prices, the nominal prices, which are sticky, as Schmitt-Grohe and Uribe (2009) show. This seems obviously accurate, and so I am not particularly concerned about inflation being that hard to measure for the purposes of monetary policy.

Subscribe now

Do you just agree with Scott Sumner on everything?

Kinda! But where I disagree most is on futures markets. Sumner is particularly attached to using futures markets to actually guide Federal Reserve policy. I don’t think this can fully work, although it partly can.

Much of Sumner’s work, from the Midas Paradox on the Great Recession, to his blogging, has been basically to look at the financial markets right after the announcement of policy, and pull out the implied effect of the policy from the change in people’s beliefs. This is an incredibly powerful tool. If the Fed says that they are cutting rates by 25 basis points, and the stock market craters, it indicates that monetary policy is in fact tight, and they are not reducing the interest rate enough. So it’s natural that we’d want to pull the information from that. But we can only do that when everyone understands that the change in financial markets is not, itself, determining the course of monetary policy.

Consider the following model. The Federal Reserve creates a prediction market for NGDP. People can invest into finding information about the world, which causes them to buy and sell futures on the prediction market. But the Federal Reserve is just going to take this information, and change things so that, on average, the investors are as equally right as wrong. Nobody has any incentive to invest into finding information. The course of NGDP will be on average accurate, but with wild error bars in either direction.

Futures markets can be a part of making predictions. But, they must remain only a part. The Fed faces a time inconsistency problem. Suppose the Fed forecasts something internally, but the market disagrees that their answer will actually achieve the target. It must be incredibly tempting to simply change policy until the gap is fixed! But if you do this, you destroy the informational content of the predictions. This is why Bernanke and Woodford (1997) rejected using market forecasts of inflation long before the crisis.

What makes things worse is that the influence that the market can play in decision making has to be fixed. Suppose that investors pay fixed costs to discover information, in expectation of a certain amount of profit. If the Fed makes their prediction more accurate than expected, then rather than break even, traders actually lose money. I see Prof. Sumner’s proposal for guardrails, instead of exact targets, to be conceding this point. Basically you put bands around the NGDP target which are good enough, and people are still able to profit by placing bets, knowing that the Federal Reserve is not going to – or at least, it says it’s not going to – continue poking nominal GDP to be more and more on target.

I think it is telling that a similar instrument, which allowed you to bet on non-farm payroll, ended up getting delisted because of lack of interest.

Could we have switched to NGDP level targeting in the recession?

This is where I am also not so confident. On the one hand, people really, really hate inflation. Also, the Republicans had been attacking the stimulus, quantitative easing, and everything else the Fed was doing as dangerous government overreach and the first steps to hyperinflation. Peter Diamond, who is a genius, got his nomination tanked by Republicans because they felt that he would be too much of a Keynesian and too likely to allow inflation.

This is an unfortunate answer to me, because I want to separate out things which are impossible because they are physically impossible, and things which are impossible because politics is unlikely to allow them. Answering the latter is completely uninteresting. It’s tautological – it can’t happen because the decision makers would have to make different decisions – but the whole point of studying this is to inquire after what happens with different decisions. Nevertheless, it binds both our attempts here, but it will also likely bind our future attempts to prevent hitting the zero lower bound. The last potential crisis which might have caused us to reach the zero lower bound was Covid. We had high inflation after, because the Fed did not hike rates aggressively enough. This was incredibly unpopular, and substantially responsible for the reelection of Trump. If we have another deleveraging crisis, will the promises to have high and sustained inflation later be credible? I am not so sure.

Was the financial crisis a cause of the recession, or a symptom?

I am not convinced this is a conceptually sound question. So, the case for it being a symptom focuses on the summer of 2008, when market indicators for NGDP were falling. The Fed standing still is allowing a contraction to happen. As NGDP falls, loans which were previously good failed, and the financial crisis is a belated recognition of what has gone on.

Why I think the question is conceptually unsound is that indicators of falling NGDP are themselves including the expectations of a financial crisis. Thus, preventing the financial crisis is a different path to preventing the fall of NGDP.

Go back to the question of mortgage backed securities. Couldn’t we have prevented this if we had more safe assets?

I do not think we could have completely prevented it, but I do think that we could have substantially reduced the magnitude of the Great Recession and financial crisis.

The lead up to the Great Recession saw an enormous global savings glut, as Bernanke (2005) called it. The developing world – mostly China – was growing and needed somewhere to save their money, but China’s government was not trustworthy enough to buy debt in bulk. So, global investors gobbled up U.S. debt at a massive scale, and when that ran out, they turned to other securities. This is a problem, because the private sector does not take into account the firesale externalities they create when they produce money. Relative to the optimum, there is too much private money (Greenwood, Hanson, Stein, 2015, and Jeremy Stein, 2012)

The United States could have prevented much of this by having the Treasury create more bonds, then depositing them in the Federal Reserve in exchange for reserves. Doing this crowds out private money basically one-for-one (Krishnamurthy and Vissing-Jorgensen, 2015), which is pretty much exactly as I would have expected. The idea of making the monetary system run free by getting rid of leverage is an old one, dating at least as far back as the Irving Fisher endorsed 1935 Chicago Plan, and seems to be held both by conservatives like John Cochrane and liberals like the Harvard team.

It would not have prevented it entirely, because I do not think more treasuries would have fully prevented the use of mortgage-backed securities as quasi money. Some of the investors were after the higher yields of mortgage backed securities, and were shoving away the tail risk. This means that we probably get an unwinding of debts and a bunch of financial firms in the tank, but no systemic crisis. This also increases the odds that the actions of the Fed will be appropriately sized to the crisis at hand, rather than inadequate as it was in reality.

The best argument against is that the flow value can be used to fund things. This is by itself an upside, because it’s pure value, but the downside is that there might be spending obligations tied to it, which we are unable to adjust for political reasons when the facts change. The sovereign wealth fund of Norway, which is doing something conceptually similar, gives us a way to handle it, which is to have the rules for how funds are returned to the government be entirely driven by the returns of the portfolio, and avoid having it tied to a specific program.

To what extent did housing itself matter?

So there’s a literature on this that I would like to talk about – it is a question like this which gives away that I am controlling both sides of the conversation.

This line of literature is dominated by the work of Atif Mian and Amir Sufi together. Suppose that people have a desired level of wealth as a portion of their permanent income, of which their house is a substantial component. The value of their house drops; in order to have wealth return to the desired level, they need to consume less. This causes demand to fall, reducing output unless offset by the central bank.

The first part of the argument, Mian, Rao, and Sufi (2013), is spent establishing that there was an enormous response of consumption to the price of housing. When the value of someone’s house falls by one dollar, consumption falls by five to seven cents. (If that doesn’t seem like much, remember that the value of a house is spread over a long period of time, so this adds up to 60 to 80 percent of the net worth fall gets reflected in consumption). This response is even larger in zip codes which had a greater proportion of underwater houses or subprime mortgages. Similarly, Berger, Guerrieri, Lorenzoni, and Vavra (2018) find a considerable decline in spending where housing wealth falls.

We should be clear about why this is a bit surprising. A house has two things bundled in it – it is both a financial asset, and a place to live in. When the price of a house falls, that also means that the price of choosing to live there falls. Buiter (2008) puts it bluntly in the title – “Housing Wealth is not Wealth”. Nevertheless, once you add in credit constraints, you will get changes in spending. If consumers had the ability to offset by borrowing, this shouldn’t matter. Since they face credit constraints, it does.

We can then trace this fall of demand to employment. Mian and Sufi (2014) show that places which saw a greater decrease in demand had employment fall in non-tradeable goods, but there was no correlation with the fall in employment in tradeable goods whatsoever. The idea here is that the demand for tradeable goods depends only on aggregate conditions, but non-tradeable goods depend on demand in the local region.

We should be clear that employment fell in all sectors. We are not able to take these partial equilibrium estimates and trivially scale up to “this is the total explained by housing and aggregate demand”. What we can do instead is say something about what the shocks that mattered were.

Why does it matter?

It matters when we consider how monetary policy actually gets passed through to the individual. There has been a considerable change in how economics thinks about this in just the last few years. (For details at greater length than this, see my essay on Ludwig Straub). Previously, we assumed a representative agent, standing in for all consumers. The attributes of this representative consumer can be filled out by taking the average from the population.

Recently, however, we have begun to incorporate heterogeneity in the models. People differ in their income and wealth, and thus in their marginal propensity to consume. In a representative agent world, if we say that we are giving everyone a hundred dollars today, to be paid for by taxes later on, this produces precisely 0 change in spending. Everyone simply saves enough to offset the expected tax liabilities in the future, a condition called Ricardian Equivalence. This is too strong to hold literally in reality, although it is a useful reminder that expenditure largely depends on what people expect their total income over time is, not necessarily how much they have now. Our model should be changed to reflect this.

Now add in heterogeneity. Posit that some people have limits to how much they borrow. They want to be spending more now, but cannot – instead they might have illiquid assets that pay out in the future. When you have transfers, people do not all share the same marginal propensity to consume. Transfers to people with high MPCs are going to be much more effective than broad based fiscal stimulus.

Heterogeneity also completely changes how monetary policy reaches the public. In the world of representative agents, changes in the interest rate only reach the public through changes in the intertemporal elasticity of substitution. With heterogeneity, much of how monetary policy reaches the public is through actual changes in the economy, and not purely due to savings behavior. Kaplan, Moll, and Violante (2018) is the key reference on this. We have a distribution of people who hold mortgage contracts. They could refinance their mortgages by taking on a new loan at the new, lower interest rates. This reduces their monthly payment, so it’s effectively a transfer from the banks (who have a minuscule marginal propensity to consume) to the households (who have a very high MPC). The refinancing channel is an enormous way that monetary policy reaches the public, and households which are underwater on their mortgage can’t do it at all.

Beraja, Fuster, Hurst, and Vavra (2019) found that places with a greater fall in housing value, and more homeowners underwater on their mortgages, were less responsive to cuts in interest rates. Thus, the passthrough of interest rate cuts to the places which were going to see the biggest declines in aggregate demand was hampered, and we see a decrease in the purchase of consumer durables like cars at the same time.

Subscribe now

What is the other story?

The other story of how the financial crisis gets to the real economy is that it runs through lenders to borrowers. There are banks which lend to firms. They suffer liquidity shocks, and don’t have the funds to lend as they did in the past. The firm is unable to meet its lumpy needs for cash, and has to contract its business.

There is evidence for this channel being important too, to be clear. Chodorow-Reich (2014) is a fantastic paper addressing this. He has data on the ties to banks for thousands of firms, and he has how exposed different banks were to mortgage backed securities. Since the lending relationship between borrower and lender is sticky, borrowers whose banks were plausibly exogenously more exposed to the Lehman Brothers collapse and its aftermath saw reductions in financing and reductions in employment.

It’s quite a different recession than the Great Depression, isn’t it?

Yes, absolutely. The Great Depression is a story of bank failures leading to people being unable to borrow at any price, and the aggregate money supply collapsing. I cannot possibly go into detail about it, in an essay already of prodigious length, but while both of them were caused by shortfalls in aggregate demand, how that shortfall came about differs.

Who is right about the transmission mechanism?

Both? They’re both substantially right. Job losses were substantially concentrated among small and medium size firms, who are getting hit multiple ways. First, small firms are more likely to produce non-tradeable goods, and are thus more affected by the local demand shocks. Second, for many small firms, access to credit is through the house, which can be posted as collateral against a business loan. Large firms were more able to self-finance through bad patches, while small businesses could not. Fort, Haltiwanger, Jarmin, and Miranda (2013) have census data on jobs losses (which are kinda like looking at the answer key for an economic question).

Could it have been a slowdown which was bound to happen anyway?

Possibly. If this were true, it would go a long way toward explaining why the recovery was so slow. I believe that we would have had a small recession no matter what, although the severity of it was caused by the financial crisis and shortfall in aggregate demand.

I find Beraja, Hurst, and Ospina (2019) really puts its thumb precisely on what you have to believe in order for the recession to not be driven by demand shocks. At the national level, wages stayed largely unchanged while employment shrunk. At the regional level, wages fell by more in the places where employment fell by more. The usual device in macroeconomics for representing wage/price stickiness is the Calvo parameter, which is the fraction of firms in each period which are able to change their prices (or wages). If the Calvo parameters are the same in all the regions, then they must be the same for the nation as a whole. If you can’t explain the regional results with the national parameter, then you need other shocks to explain it.

You cannot, as you might naively think, aggregate Calvo parameters by averaging. Consider a world where a randomly chosen firm has a 50% chance of changing their prices in each period (or a Calvo parameter of 0.5). A negative demand shock occurs, and over time firms adjust their prices to bring them into line with the new optimal prices. In the limit, we return to efficiency. In another world, half the firms have a Calvo parameter of 0 and the other half have a Calvo parameter of 1. When the shock comes, in the first period outcomes are identical – but then in every period thereafter, they stay the same, forever. These are completely different outcomes.

If you are willing to believe, however, that all regions share the same Calvo parameter, then obviously the aggregate is inconsistent with the regional parameters, if the slowdown is entirely explained by shocks to labor demand.

Surely you wouldn’t have me believe that the Great Recession was due to people just not wanting to work anymore?

Of course not. There are lots of candidate explanations. Casey Mulligan argued that the expansion of unemployment insurance led to people quite rationally preferring to stay at home, while over time leading to their skills degrading. And also, in this framework, a sectoral demand shock – the decline of American manufacturing over time – is actually a labor supply shock.

Why isn’t it a demand shock?

This is a bit confusing, I admit. What we’re referring to when we say a demand shock is an aggregate demand shock. The sum total of goods demanded goes down. It doesn’t say anything about what the composition is. If someone is laid off from their job as an assembly line worker, maybe they don’t want to go into a different line of business. Maybe they’d prefer to retire, or go on disability. (Disability functions as shadow unemployment, at least according to Autor and Duggan (2003)). This shows up as a shock to labor supply.

There are two ways of telling this. First, there might be a mismatch in skills or non-wage compensation. The assembly worker perhaps finds services degrading in a way that doesn’t show up in wage compensation, and so chooses to go home. Sahin, Song, Topa, and Violante (2014) argue that a third of the loss in employment can be explained by people being unwilling to take jobs which are on observables the same.

The other is that manufacturing wasn’t doing well anyway, and the housing boom was masking the decline in manufacturing with construction work. The case for this has been made basically by Charles, Hurst, and Notowidigdo across several outlets. The basic shape is this: manufacturing fell by a lot, but employment for lower skilled workers didn’t. What were they doing? Probably housing. And when housing goes, so too goes the jobs.

Subscribe now

How well does the unemployment insurance story hold up?

It does not hold up. But why it doesn’t hold up is interesting, so we will digress.

The starting point is that microeconomic studies of unemployment insurance extensions, which were the policy instrument of choice – benefits went from 26 to 99 weeks – don’t find that much of an effect on unemployment (and even if they did, that might be good). The standard, intuitive way to find the effect of duration is to use variation in exposure across people but within a geographic area. Schmieder, von Wachter, and Bender (2012) exploit a feature of German law, which provides for additional benefits for those above the age of 42 and 44. Presumably, the people fired two months after their 42nd birthday are pretty similar to people fired two months before their 42nd birthday in all ways except the amount and length of unemployment insurance they receive. You can then infer the effect of unemployment insurance generosity on unemployment duration.

Counterpoint, however, from Hagedorn, Karahan, Manovskii, and Mitman (2016). These microeconomic studies are forgetting the general equilibrium effects. Consider that both employers and employees are searching for each other in the market. Both expend effort to do so, and the chances of a match are increasing in the amount of effort expended. When unemployment insurance is extended, people put less effort into searching, and are only willing to go back for a higher wage. From an employer’s point of view, this higher reservation wage decreases the return to creating a vacancy at all. You don’t see this in the difference between workers at all! When they compare bordering counties across state lines, they find that extending unemployment insurance during recessions has simply enormous effects.

Unfortunately, this result has not held. HKMM don’t actually have county level data on unemployment. Instead, they have what the BLS imputes the unemployment rate to be there. Hall (2013), among others, point out that these are clearly affected by the state level aggregates, and it’s questionable whether you’re actually getting anything resembling county-level unemployment statistics. Boone, Dube, Goodman, and Kaplan (2021), who have data on employment, look at the expiration of the extended benefits in 2013, which restored everyone to the default level and so had cross-sectional variation from how high the old benefits were, and find nothing. They can’t detect differences-in-differences from county border pairs.

Meanwhile, Chodorow-Reich, Coglianese, and Karabarbounis (2019) take on the macroeconomic effects from a completely different path, and find that at most, the increases in unemployment insurance benefits explains 0.3 percentage points of unemployment. What they use is the measurement error in the initial BLS measures of unemployment, which determines whether unemployment insurance benefits get extended. As more information comes in and the BLS revises their model, they update the figures to reveal what the true rate of unemployment was. If two states are initially measured as having different rates of unemployment but are secretly identical, then you have variation in the amount of UI benefits that are plausibly exogenous to the underlying economic fundamentals.

The authors of HKMM (see Hagedorn, Manovskii, and Mitman (2016) for specifics) disagree with these last papers, on the grounds that what matters more are the effects on expectations of future policies. CCK are catching small blips, when people expect the differences in policy due to measurement error to smooth out over time. Boone, Dube, Goodman, and Kaplan are picking up the change to an already somewhat anticipated ending. These are fine counterpoints, but I don’t think we can go so far as crediting unemployment insurance with millions of unemployed people, and indeed the authors have walked back the claims in later drafts as compared to early ones.

What was the role of oil?

The price of oil increased substantially during the run up to the Great Recession. Oil is a bit of a special product, because it has an economic impact far greater than its share of GDP. The need for energy finds its way into everything. I don’t, however, believe the story that the rise of oil led directly into a recession. However, it still could have affected the economy, but through it discouraging the Federal Reserve action.

James Hamilton (2009) is the most committed advocate of the oil price rise affecting output. Basically his whole career has been around arguing that, in the time series, oil price rises lead to recessions. However, the identification is weak, and anyway the rise in the price of oil was not due to supply restrictions like in the 1970s. It was due to increasing demand – it was due to growth. So naturally, when the recession came, the price of oil absolutely cratered, going from $145 a barrel to $30. Did this help matters? Not really!

I also don’t believe the story that rising gas prices caused the value of homes with longer commutes to fall, and was responsible for the collapse in home prices. Frankly I think it’s risible.

The best case for the price of oil affecting the economy and causing the Recession is through the actions of the Federal Reserve. They saw inflation go up, and it scared them off from cutting rates. The original people to argue this was Bernanke, Gertler, and Watson (1997) – they should have known better!

Subscribe now

Was Cash for Clunkers effective?

No, absolutely not. Cash for Clunkers was an Obama policy to stimulate the economy by paying people for their old, crappy cars, simultaneously putting money in people’s pockets, getting polluting cars off the road, and encouraging people to buy new automobiles from the American auto industry which we were keen to protect. People sold their cars, yes, and it did put cash in people’s pockets, but it did so only shifting forward the sales people were going to make anyway. Mian and Sufi (2012) find that the additional sales of cars are completely wiped out after just seven months by people selling less.

Should the government have bailed out General Motors and Chrysler?

I think, yes, but not in the way that they did. Given, however, the incentives which the Obama administration faced, there was likely no other way it was going to happen.

It is pretty clear that bankruptcy was going to happen. They had mismanaged their finances, and this was not simply running out of cash flow during the Great Recession. What the government should have done was stepped in to provide the financing for an orderly bankruptcy, at a time when the private sector was unable to because of the financial crisis.

That was not what they actually did, though. They used their control of the process to settle the unsecured claims of the union’s retiree health trust above the secured claims of creditors who should have been paid first. It was in part an enormous giveaway to politically favored actors. Things were not all bad. The disaster case would have been if they kept everyone on the payroll, and simply made up the shortfall with taxpayer money. But, General Motors and Chrysler were substantially subsidized by the government during the transition.

Unfortunately nobody has written a serious accounting of the costs and benefits of doing the bail out, in the context of the Great Recession. Goolsbee and Krueger (2015) is a retrospective from people inside the Obama administration on the bailout, but it’s mostly narrative history, with the quantitative treatment being of the “point and squint” variety. I disagree with their characterization of the decision to favor the union health fund as being uncoerced, what with the secured creditors being the banks which needed Federal favoritism with TARP. Thomas Wollmann (2018) is an excellent paper, and makes the best case for letting Chrysler or GM simply be liquidated – firms are able to enter the market and offer product lines similar to what the exiting firm offered. However, this is comparing steady states, and doesn’t include the losses from the chaos of liquidation, and it also only covers commercial trucks, which are more modular than passenger vehicles and so more able to reposition.

To what extent was uncertainty a component of the Great Recession?

Not very much, and the importance of uncertainty in recessions in general has been overstated, despite some very interesting models.

The path of uncertainty to affecting real output goes through the “real options”. Suppose the investments of firms are irreversible. We define uncertainty as the accuracy of firm forecasts of the future, and an increase of uncertainty to be a decrease in the accuracy of firm forecasts. When things are uncertain, the value of waiting and seeing before making a decision increases. You thus get a slowdown in growth and investment, before a rebound as firms make the delayed investments.

Nicholas Bloom (2009) really kicked off studying the macroeconomic impacts of this, and the basic problem should be apparent from when uncertainty spikes – if you use stock market volatility as your measure of uncertainty, it correlates with literally everything.

Plus, it doesn’t even measure what you’re trying to measure. Suppose the uncertainty of firms forecasts increases, but in a way that is uncorrelated with each other. If you’re holding an index of stocks, you have no reason to trade. Jurado, Ludvigson, and Ng (2015) actually measure uncertainty itself, and find that major episodes are extremely rare (although they are economically impactful).

Uncertainty definitely can matter. Handley and Limao (2017) pursue a really cool idea – the accession of China to the World Trade Organization didn’t actually change the tariffs of the United States, which were already low, but it put those low rates on a sound footing, not requiring yearly renewal by Congress – and they find quite large effects, the equivalent of a 13 percentage point decline in tariffs. But, when we try to fit the facts to the Great Recession, the best we can get is 2.5 percentage points of the unemployment rate. That paper, Bloom, Floetotto, Jaimovich, Saporta-Eksten, and Terry (2018) is one which I see as the lawyerly case for uncertainty as a contributor to the Great Recession – it is not dishonest, but at every juncture, it sees how far uncertainty as an explanation can go. Its numbers should be taken rounded down, and in any case, it cannot explain persistence.

Nor should we be so confident that papers which say “uncertainty” actually mean “uncertainty”. Eduard Schaal (2017) attributes 40% of the increase in unemployment during the recession to uncertainty, but I think a close reading reveals that it’s not uncertainty as we would think of it. We do not see the real options channel matter. What happened is that the dispersal of shocks across firms was bigger than usual, and since firms could not hire and fire quickly enough to match, output went down.

So it’s all well and good to point out cool mechanisms around uncertainty – I think “Uncertainty Traps”, for instance, which proposes that firms delaying investment itself delays discovering information about the world, further increasing uncertainty, is a cool idea – but I think we have to conclude that it was of limited quantitative importance.

Is AI a bubble?

Well, is it?

I guess that would be the thing to know. What I mean is, if it were a bubble, could we have another Great Recession?

This is an enormously important question. I think we should be clear that the Citrini Research story of an AI boom leading to mass immiseration is not going to happen. See Alex Imas’s article on the subject. At the time of writing I was actually working for Prof. Imas, and I thought the assumptions required to get immiserizing growth were so ridiculously strong that it wasn’t even interesting to argue against – who could possibly believe it to be the case? Then along comes Citrini a few months later.

What is actually interesting is if GPUs are actually making their way into the economy as quasi-money and collateral for loans. This is something which I genuinely don’t have an answer to. However, I believe that the current size of the assets used as repo is limited relative to the size of the repo market more generally. Moreover, the toxic assets of the time were rated as being of the highest quality, while these are rated as risky corporate debt.

Well that just about wraps everything up. Is there anything you’d like to say at the end?

I would first like to thank Basil Halperin for taking a look at the draft, and Scott Sumner for commenting. I would also like to thank Tyler Cowen for taking a chance on me, back when my blog stank, in supporting my writing.

The Great Recession was really, really bad, and it was largely the result of mistakes. Because I believe that the number of mistakes made can be reduced by study, and that the amount of study is responsive to investment, I think it’s actually really important to invest in producing economists. (Shocker, I know). The same is true for the other real sciences, but it is just as much so for economics.

I would also like to ask for your support. If you have made it this far in the article, I presume you must have liked it and found it interesting. I chose not to paywall this, but I could have. I would have traded perhaps ten people subscribing for several thousand never reading it. Does that seem like an efficient trade to you? If you like what I have done, please consider getting a subscription. It would mean a lot to me.

The best defense of the voters’ beliefs are that they think someone else can be made to pay for it. It’s only a partial defense, because the things they scapegoat – on the Left, the billionaires, on the Right, foreigners – cannot pay for more than a pittance of what they want done. But even this is not an adequate explanation, because voters demand the impossible.

Consider Obamacare, which I wrote about at length last month. The Affordable Care Act banned charging higher prices to people with preexisting conditions. If you do this, however, sick people will select in and healthy people will select out. To have insurance exist at all, you need people to be compelled to purchase, rather than allow them to opt out when they are healthy and opt back in later when they are sick.

The ban on discriminating on preexisting conditions was extremely popular with the voters, and so we kept it. The requirement that you purchase health insurance was not, and so we zeroed it out in 2017. These policies are not something where you can get one but not the other – they are the same policy, fundamentally inextricable. And yet, we did A which leads to B and simply say that it will lead to C.

I think the better explanation is that the median voter in America is a Maoist. They do not believe in tradeoffs. They believe that, if only everyone were good and honest and working toward the right goals, good things will happen. That politicians disagree with each other is simply a sign that they are corrupt. If the people are good and sincere, then we can simply do things and it will all work out okay.

This post is brought to you by Mechanize, Inc. They are hiring for a variety of positions, including software engineers. I encourage you to apply here.

This sort of attitude once ruled China. I do not call the American voter’s ideology Maoist as simple namecalling. It is just literally true. Mao Zedong’s program was that it was only the ideological sincerity which was needed for success to come. This ideological sincerity had magical properties. Consider one of the Three Constantly Read Articles, “The Foolish Old Man who Removed the Mountains”. There is a Chinese story about an old man who lived near the Taihang and Wangwu mountains. He wanted to travel, but the mountains lay in his way. So, he called his sons to him, and they began to dig up the mountains with hoes. A Wise Old Man comes and rebukes him for doing an impossible task, but the Foolish Old Man responds: “When I die, my sons will carry on; when they die, there will be my grandsons, and then their sons and grandsons, and so on to infinity. High as they are, the mountains cannot grow any higher and with every bit we dig, they will be that much lower. Why can’t we clear them away?” And so he and his sons resume digging.

It is not the case, however, that he and his sons did indeed clear away the mountains. Instead, God was touched by his conviction, and sent down two angels who carried away the mountains. Mao suggests, then, that there are two mountains the Chinese people wish to dig up – imperialism and feudalism – and if they dig them up without rest, then they will touch God’s heart and do the impossible. Mao is not so primitive as to suggest that a literal deity will come down and do it – instead, “our God is none other than the masses of the Chinese people”.

Subscribe now

The great disaster of Maoism – the Great Leap Forward, with its tens of millions starved – came because Mao thought enthusiasm meant that constraints could be ignored, and that simply intending for something to happen was the same as doing what was necessary for it to happen.

Before the Great Leap Forward, state planning was run by Chen Yun, who was by all accounts a serious, soberminded individual. It was he who coined the phrase “crossing the river by feeling the stones”; I disagree with communism as an economic system because I believe it to be inefficient, but there is also no doubt in my mind that competent administration set on achieving possible goals can indeed grow. Reading Chen Yun’s selected works (volume three can be found here), one is impressed by the focus on whether the means can achieve the ends. When he makes projections of future production quantities, they are not fanciful but possible (p. 208). To combat inflation caused by an excess of currency printing, he says that they must “strive for a balance between revenue and expenditure, between credit receipts and payments” and that “we should not wait until next year to check inflation” (p. 208). That the government is trying to solve inflation by opening more restaurants is ridiculous, and why communism is a failure at persistently raising living standards, but it is at least an attempt to do the possible with the feasible. Problems are not solved by enthusiasm, but by carefully exchanging evidence, comparing alternatives, and reconsidering matters (p. 194-5), allowing for the possibility that the methods may not lead to the results.

You do not see this in Mao. The people’s enthusiasm would suffice to dissolve constraints. What was his solution to nuclear warfare? “The atom bomb is a paper tiger. … the outcome of a war is decided by the people, not by one or two types of new weapons”. What was his solution to the shortage of steel? Melting down their woks and hoes for sulfurous ingots of pig iron. Mao would later concede that these ingots were worthless, but at the time he said that (reorganizing slightly) “on the foundation of the tremendous energy of the masses it is possible to accomplish any task whatsoever”. Nevermind what Chen Yun wrote to him (p. 147) in May of 1959, stating that “over 40% of the 9 million tons of iron … contains sulfur in excess of 0.2 per cent”. How shall heavy industry be developed, and should it come at the expense of light industry and agriculture? He has the answer: “If your desire is genuine or strong … there more grain and more materials for light industry … and there will be more funds in the future to invest in heavy industry”. How shall agriculture be improved? Here Mao is unusually specific, but his ideas are all wrong: to grow more, the people should simply plant more and plant deeper. On the basis of this, fanciful quotas were set. Mao’s method of projecting agricultural growth was to pull numbers out of a hat. Not setting quotas higher and higher became a confession of disbelief of the program, and the people starved because their food was requisitioned. When Peng Duhai submitted the mildest of criticisms, suggesting that local leaders were overstating production, he was purged. One of the striking details of the Great Leap Forward is that when cooking and food distribution was centralized – it had to be, they had melted down all their cookware – the first months were those of plenty. Production was going to increase, so food was free, and gorging yourself was an ideological statement. When the food ran out, that was that.

I mentioned that Mao does not often include details in his writing. His instructions are aphorisms – not “do this”, but “see that this is done”. And of course, there is a great gap between what the appropriate things to do were, and what outcomes they would actually lead to. If things lead to disaster, they can be disavowed as being a misunderstanding, and not showing the fundamental wrongness of the program. Mao Zedong, at the Lushan conference of 1959, said “I do not claim to have invented the people’s communes, only to have proposed them. … When I was in Shantung a reporter asked me: ‘Are the people’s communes good?’ I said: ‘They are good,’ and he published it in a newspaper. There was a spot of petit-bourgeois fanaticism there, too. In the future reporters should keep away.”

I must emphasize that the Lushan conference was not when Mao submitted self-criticism, Liu Shaoqi became the dominant figure in the government, and the Great Leap Forward stopped. That came later, at the Seven Thousand Cadres conference of 1962. The Lushan conference was when Mao purged Peng Duhai for mild criticisms of the feasibility of the Great Leap Forward. Chen Yun fell “sick”, and absented himself from politics – you’ll note the table of contents has a gap from summer 1959 until 1961, when it became possible to describe the catastrophe that had occurred. (After it did, Chen Yun was willing to bend to empirical evidence – see p. 177, on the advantages of individuals raising pigs, and p. 190, on the advantages of private plots).

Mao wanted the impossible done. Man cannot do impossible things. The cost of pretending otherwise is death.

The voter wants impossible things, too. I do not think the American people are going to starve themselves to death – we are too far from the threshold of survival for that – but the American people have set themselves on an inevitable conflict with reality. More and more of the economy is dedicated to parasitism on the public – Social Security, Medicare, disability (both for the actually disabled, and for veterans), the department of defense. We seem to think that we can pay for it all with taxes on just our enemies. In practice, we have funded it all by pulling forward consumption from the future, and borrowing heavily. For a time this can be sustained, but the bill will come due.

I suspect, of course, that the voter has always wanted impossible things. What is remarkable is that we have long kept the demand for impossible things away from the actual conduct of the government. For many years, the budget was balanced, and entitlements were a small part of the budget.

I believe that what keeps government functioning are moral taboos as to what is right and proper for the government to do. When the government does not do something, it is a violation of the norm for the government to involve itself in it. This is simply a bias toward the status quo, not a consistent libertarianism – once the government is involved in regulating something, you are a madman for thinking it could ever be safely provided without the government. Further, once a government program does exist, it failing is never evidence that it should be gotten rid of. The program does not fail, it is failed. It being inefficient simply means that it was “chronically underfunded”, and needs “reinvestment”.

You’ll note how the defense of libertarianism is hardly ever expressed in the terms of utilitarianism. One does not persuade anyone by saying that there is decentralized information about people’s utility functions, and a price system is the only incentive compatible way to elicit it. Instead, the case for libertarianism is built upon natural law, and analogy. We have lower taxes because it isn’t right for someone to take the product of your work, not because taxes will move you along your labor supply curve and cause you to produce a socially inefficient level of labor.

This often leads us into errors. We may reject government programs which are beneficial, and over apply the wisdom of non-intervention. But we must weigh that against the basic errors that people fall into with a government program. If the amount of spending is zero, then the people cannot ask for spending to be doubled and taxes to be halved. Half and twice zero is still nothing.

Lastly, someone who purports to represent the people, and wants to bring the people’s will into office, is a dangerous character. The people wish to chew on the pretty foxgloves. A politician’s job is to do what they think is best for the people, not to do what the people want. Let them vote for whoever they should choose, but you should not help them.

Consider a house flipper. They buy a single family home, gut the insides, renovate, and then sell for a higher price. Have they made housing more expensive?

Many people will argue yes. Somebody bought the good, now it is more expensive, therefore they increased the price and made things worse. But this is erroneous. There are two things that one does by buying and reselling, and both of them improve the world. Properly understood, this behavior is beneficial, and being confused on this is the source of much bad policy.

The first is that they are increasing the quality of the good. People value the renovations they’re providing. This is not just some trick, even if you don’t value the renovations yourself. People are willing to pay more for it, because it fundamentally isn’t the same good as before. Thus, the price doesn’t actually rise.

Failing to account for changes in quality will lead us into enormous conceptual errors. Consider a bombing raid. This is obviously bad. People were able to live in a place, and then they weren’t. They are no longer able to consume housing, because it has all been burnt or blown up. This will, nevertheless, lower the price of housing, for precisely the same reason that breaking something lowers the price – it makes it worthless. Or consider a world where no housing exists at all, and we sleep on the ground, at a price of zero. If someone builds a cottage, this will greatly increase the price we pay for housing, and yet is not a house becoming more expensive.

This shows up in how we construct a price index. In measuring what inflation is or how much GDP has grown, we are ideally identifying how much it costs to buy a constant amount of utility. Thus, if a good enters our consumption basket with the same price but is higher quality, that is equivalent to a fall in the price. It is quite possible for a fall in the real price to occur at the same time that the nominal price increases.

The second the house flipper is doing is uncovering information about the world. Perhaps they didn’t renovate the home at all. They simply noticed that the price was too low. There was latent, as yet unrecognized, demand for housing there. What good have they done for the world?

The first thing to point out is that even if all they have done is increase the price, that is a neutral act by itself. It’s simply a transfer from one to another. But of course, that is not all that they have done. People have the ability to build in different places. They build where they expect the demand to be high. If they do not know where demand is high, then they are inhibited from building. The increase in price is a signal to others that one should build here.

The best argument for the folk intuition is that someone may be taking advantage of market power which had existed, but was not taken advantage of. Perhaps someone had had the ability to raise prices above what was efficient, but chose not to. (I should note that, really, what is going on is that buyers and sellers are matching with each other over time, and increasing the price decreases the time that it is occupied, or the quantity, but that is simply a question of emphasis). But where is this market power coming from? Are there not tens to hundreds of thousands of buyers and sellers of housing? And when we look at what institutional investors do when they buy housing, we find that their profits come from greater productive efficiencies, not from market power.

There is a tendency to confuse symptoms with causes in housing policy. It is often the case that development occurs where prices are rising. But we cannot infer that it is the construction of new housing which causes higher prices. That is perfectly consistent with demand increasing, and the ability of supply to respond being constrained. If we restricted housing development further, in order to prevent gentrification, we would increase the price of housing relative to the world where we didn’t. And if we prevent people from moving by making a place worse, well – we’ve made a place worse.

And why are health insurance companies so intimately involved in every aspect of our care? If health insurance worked like any other form of insurance, you would receive a lump sum of money upon the realization that you have a health condition, sufficient to pay for whatever care you need. And yet health insurance companies restrict the doctors you can choose from, choose what procedures you’re allowed to do, and negotiate the prices that will be paid. If life insurance worked this way, you would have in-network morticians, you’d have pine coffins fully covered but oak coffins have a co-pay, and the whole thing is settled two months after the funeral.

The answer to both of these is that health insurance is substantially not about pooling risk. It is about collective bargaining. By putting our money together, we can hold down the cost of medical care. Restricting what people can buy with insurance money allows them to implement something closer to an efficient transfer.

Subscribe now

To understand this function of health insurance markets, we must learn about the economics of bargaining. This is a new and exciting area of economics, and its applications in antitrust law has been one of the cleanest examples of new theoretical advances making their way into practical relevance. We will also learn everything we can about the provision of healthcare in America, and how well-intentioned laws have undermined the incentives for health insurance companies to fight.

Finally, we need to consider the effect on innovation. Medical innovation, largely pharmaceuticals, requires the outlay of an enormous fixed cost to discover new ideas. Pushing outcomes closer to what is efficient in the short run can be worse in the long run, as new ideas are left undiscovered.

This article is brought to you by Mechanize, Inc. They are hiring software engineers to train and test frontier coding agents. Apply here.

Health insurance is provided both privately and publicly, with the public insurance often being through the auspices of private companies, and with private insurance often being regulated so severely as to barely be a free market at all. The old, as well as people with disabilities and people on kidney dialysis (among others) are covered by Medicare; the poor are covered by Medicaid, often as a supplement to private insurance.

The rest buy privately, with over half of all people buying health insurance through their employer. Strikingly, for two thirds of these, the insurance is not actually financially liable for paying out claims or collecting premiums. This is handled by the employer. The employer contracts with the insurer in order to get access to the insurer’s network of hospitals and doctors, the set of negotiated rates for different procedures, and the administrative set-up for billing.

It used to be that insurance simply paid your bills, whatever they might happen to be, in a “fee-for-service” arrangement. We do this instead of paying out a lump sum because it is fundamentally infeasible to give people a sum of money, and only then have people search. This would be perfectly fine in perfectly competitive conditions, but is very different under imperfectly competitive ones. When one discovers that one has a condition, their elasticity of demand changes. Empirically, customers also do a terrible job shopping around – a few months ago I wrote an article on this, but basically requiring price transparency has an extremely limited effect on utilization and prices, and what effects there are come from the major companies learning about each other’s prices, not from the consumer choosing better options.

Paying people’s medical bills, regardless of what they were, was rather predictably disastrous. Prices exploded from the 60s to the 90s, going from 5% of GDP in 1960 to 13% by the beginning of the 1990s.

Doctors and hospitals were encouraged to hike their prices to match the generosity of insurance coverage. On the patient side, the marginal cost of getting more healthcare once you have the plan might be zero, thus leading to inefficiently high levels of healthcare. Even if the insurance company attempted to control things through cost sharing, where the patient must pay a portion of the costs of care, this would only cut out the most egregious waste.

Then, expenditures flatlined. For much of the 1990s, healthcare spending as a share of GDP was constant, and even in the years it grew, its growth rate was lessened considerably. The rise of managed care was responsible. Between 1988 and 1998, the percentage of employees in traditional fee for service plans fell from 71% to 14%.

They accomplished these cost reductions with the tools of modern health insurance companies. Rather than pay any medical bills at any doctor, they would instead restrict the network of doctors and hospitals you could see. If they charge too much, they’ll find themselves out of the network and getting no business at all. Insurance companies began to involve themselves in what treatments a patient should be prescribed, requiring “prior authorization” for treatments of great expense and heterogeneous effectiveness. The most extreme versions of Health Maintenance Organizations (HMOs) like Kaiser Permanente would be fully vertically integrated, with the doctors being employees of the insurance company on salary. Even if they are not directly employed, they can still be paid a flat per-patient fee, or capitation, regardless of what procedures are ordered or what work is done.

Subscribe now

These cost saving provisions worked. Cutler, McClellan, and Newhouse (2000) compare the cost of treating heart attacks and ischemic heart disease between traditional insurance plans and managed care. The managed care plans provided the same treatment for 30-40% less. Narrowing the network helps force out inefficient hospitals. Gruber and McKnight (2016) found that giving even a small subsidy was enough for people to switch to a narrow network, whereupon spending fell by 40%. Brot, Burn, Layton, and Vabson (2026) find that prior authorization in Medicare Part D (which, while publicly funded, is privately administered) reduces costs by 4% of total spending by reducing the utilization of the restricted drugs, and relatedly, Maggie Shi (2024) finds that every dollar spent on auditing for fraud and waste in Medicare reduces government spending by between 24 and 29 dollars. Even better, this cost-cutting came without negatively affecting the patient at all.

Yet, managed care was not very popular. Nobody wants to be told that they can’t go to every doctor, or that they need permission to be prescribed a drug. Moreover, physicians and hospitals hated it, precisely because it controlled costs.

So, largely between 1995 and 2001, states passed a number of laws restricting the power of insurers to bargain, the most consequential being “any willing provider” laws. These laws made it so that insurance companies could not exclude providers who were willing to accept the network rate. Since what was holding down prices was the threat to exclude to begin with, though, this is not a stable arrangement.

It is difficult for me to make the case for why the managed care backlash happened without sneering at it, because it really was an alliance of the cynical and the foolish; the cynical, of course, being the doctors, and the foolish being the public, who could not see how requiring more medical care to be provided for would simply show up in their premiums. Alain Enthoven comments: “As I have listened to the debate over managed care, I have often thought, ‘we are hearing from people who do not believe they are going to have to pay for what they are demanding.’”

It says something about the rigor of legislating that they banned things which didn’t even exist – basically every state has a gag clause ban, which were allegedly provisions forbidding doctors from discussing non-covered options. The Government Accountability Office (1997) investigated at the behest of Congress, and had to somewhat sheepishly report back that they couldn’t find anything, and in any case physicians didn’t read the contracts they signed anyway.

And so, yes, the backlash raised costs. Maxim Pinkovskiy (2020) exploits state-by-state variation in the timing of when states passed their laws, as well as differences in HMO density county by county. When a major law was passed, counties with a high density of HMOs and thus lower healthcare costs became like their low density neighbors. There was no corresponding improvement in mortality or health outcomes.

So how do they do this bargaining? Time for some theory. Suppose that there are two people who can combine their efforts to produce one good. In a competitive market the payment to each is determined by their marginal contribution to production, but if we say that each person’s effort is indispensable then this breaks down. Each person’s marginal contribution is everything. We need some way of dividing up the surplus in order to make the surplus exist.

We presume that there is no hidden information, and all facts are known by all participants. Bargaining occurs by making offers back and forth to each other. If there is no cost to dragging out the negotiation, then the outcome is undefined – we’d just go back and forth forever. If we add in a cost to delaying the negotiation, everybody can play out what will happen eventually, and agree immediately to whatever would have happened.

Subscribe now

More precisely, out of a set of feasible options S, and with each player receiving some utility d if the agreement does not happen, the option chosen will be that S which multiplies the joint product of utility gains over no agreement. Thus, (S-d) times (S-d), with one more pair of S and d for each additional player in the game. Differences in patience give us the bargaining weight that each player has, which is denoted by beta and exponentiates each (S-d) term.

It turns out that this non-cooperative game spits out the solution proposed in 1950 by John Nash, as the only solution which is Pareto efficient, symmetric, invariant to scaling each player’s utility, and which is independent of irrelevant alternatives. These mean (in turn) that there are no improvements that could make one player better off while leaving everyone else unchanged left unused, that each of the players can be identical (note that the bargaining weights break the symmetry assumption), that one does not need to make interpersonal comparisons of utility, and that is unaffected by deleting unchosen options.

This is a perfectly good solution for parties bargaining with each other in one big negotiation. However, negotiation occurs between many parties separately. What this messes with is our disagreement point d – we don’t know what would happen if they walk away from an agreement unless we know the outcome of all the other agreements, which is precisely what we’re interested in. The solution is to have a Nash equilibrium of Nash bargains, as Horn and Wolinsky (1988) do – hence the current common name, Nash-in-Nash.

Somewhat confusingly, Nash has two important concepts named after him – Nash equilibrium and Nash bargaining – and they’re rather different. Nash bargaining is just a solution concept for a game where everyone has full information and is cooperating. A Nash equilibrium is when people are engaged in a non-cooperative game, and no player can improve their situation by unilaterally taking an action. In the context of bargaining, it would mean that, holding all the agreements but one fixed, nobody wants to renegotiate that agreement.

Horn and Wolinsky’s case is a very simplified one, where two firms separately bargain over an input from a monopoly provider. If the two goods are substitutes for each other, then bargaining simultaneously increases the share going to the monopoly provider compared to bargaining jointly. Each of them knows that if they walk away from the bargaining table, the other firm will expand to take all of their business, which improves the disagreement point for the monopolist. If, conversely, the goods are complements, then walking away is especially powerful, which worsens it. Companies want to merge if they are substitutes, and actually want to remain split up if they are complements.

In the real world, we do not observe all of the information which firms have. We do not get neat equations which Horn and Wolinsky have to characterize what happens. Instead, we have to find the outcome by iteration – or more precisely, we need to work backward from the observed agreements and prices to find the bargaining weights, disagreement points, and feasible actions which imply the observed outcomes.

Nash-in-Nash bargaining took its great leap into empirical usability with Crawford and Yurukoglu (2012), which, despite having nothing at all to do with health insurance, is such a great paper that we’re going to spend some time on it. They are interested in the cable networks, which bundle together many channels into a few discrete choices. (You might be thinking to yourself, why bundle at all? A simple example. Suppose there are two channels, ESPN and HGTV, which offer sports and gardening respectively. One consumer values ESPN at $10 and HGTV at $1, and another consumer values ESPN at $1 and HGTV at $10. Selling each separately, they’d put each on sale at $10 and make $20. If they can bundle, they’d offer both for $11 in total, and make $22. Note that this is more efficient! There is no reason to assume, a priori, that bundling is always bad for the consumer. It must be profitable for the company, of course, but it could be beneficial to the consumer!)

Subscribe now

What would happen to consumer welfare if you required that the bundles be split apart, and channels be sold a la carte? To answer this, we have channels, cable companies, and consumers work through a four part game: channels and cable companies bargain over how much they will have to pay each other, then the cable companies set prices and bundles, then households buy bundles, and then they watch TV. We will go backwards from the end of the game, and figure out what parameters justify the observed costs by iterating until we get a vector of parameters that return a stable outcome.

To estimate consumer demand, you have data on how much people view each channel, and on the sets of channels which are purchased. We don’t observe what people are willing to pay for each of the channels individually, but we can rescue things by making the assumption that if people watch a channel more, they also value it more. The bundle purchase itself is standard BLP logit, with random coefficients representing people’s demands for particular characteristics, and with demographics providing additional micromoments to match.

With demand estimates in hand, we can then observe what the cable companies chose for their bundles, and the prices of those bundles. We are assuming that cable companies are competing on price – hence, Nash-Bertrand pricing – and so there is exactly one marginal cost, corresponding to the payment to the channel, which can justify the price chosen. Given those prices, we can then search for the set of bargaining weights which gives us the outcomes we observe. They lack precise fees for the channels, but they do observe the average fees paid. This allows them to put bounds on how valuable the channels are. Likewise, we don’t observe the channels independently of bundles. Instead, we presume that both the inclusion and non-inclusion of the channels is optimal, and that that puts the value of the channel between two points.

To test a new scenario, you change some element of your model, but keep the old parameters, and then resolve. This last step is rather scary, because there is absolutely no guarantee of existence. You can, in fact, circle the drain forever, with no vector of prices satisfying the outcomes. In the empirical example, mandating a la carte price doesn’t actually do anything for the consumer, because all of the gains simply get eaten up by renegotiating. If you didn’t have the full model, you would have thought that it would have substantially benefited the consumer. Pretty important to get that right!

Back to health insurance, Ho and Lee (2017) take the basic framework of Crawford and Yurukoglu, and add an additional layer, as well as some improvements. In California, the health insurance plans of public employees are negotiated through CalPERS, so they have an additional layer of negotiation – CalPERS bargains with the three insurance networks, who in turn simultaneously bargain with the many hospitals across the state. This layer of bargaining replaces the Nash-Bertrand prices of Crawford and Yurukoglu (2012). Unlike Crawford and Yurukoglu, they can perfectly observe the transfers between providers and insurers, so they do not need to have bounds. Demand for hospitals is estimated by basically assuming away unobserved quality and sorting across market lines, and thus being able to use geographic variation.

After estimating the parameters, they can then simulate what happens when one of the insurance companies is removed. Blue Cross offers a broad network plan, Blue Shield offers a narrower but cheaper plan, and Kaiser Permanente is fully vertically integrated but will not pay for you to go to hospitals outside the network at all. While in most cases, less competition among insurers increases prices for the consumer, in some cases removing an insurer will actually decrease prices. Removing a small insurer changes the bargaining position of the other insurers, and they can drive a tougher bargain.

Ho and Lee (2017) skip over the formation of the network, and instead only deal with the payoffs between pairs of agreements as they exist. This is because the model doesn’t actually do a great job of explaining the exclusion of providers from the network – there must surely be a price which makes the agreement work, and so not just doing that is a mystery. In a follow up paper, Ho and Lee (2019) extend the model to include the formation of a network. The insurer excludes at least one hospital in order to be able to credibly threaten to replace any given hospital with the held out hospital. If the insurer could not credibly commit to doing this, then any hospital knows that the insurer will eventually settle for something, and prices are higher. They call this Nash-in-Nash-with-Threat-of-Replacement (NNTR). They use the same CalPERs data from their 2017 paper, with the parameters estimated from 2017, but then allow for Blue Shield to reform the network. (Why only Blue Shield? Because if you allow for Blue Cross and Kaiser Permanente to respond strategically, you have made it literally impossible to compute. So please, give us this assumption).

This last extension was independently developed by Eli Liebman (2022) and Soheil Ghili (2022). (Ghili writes in footnote 12 “I recently became aware that an independent paper by Ho and Lee is in the process of being written, which has a similar practical objective … and similar modeling contributions.”) The idea was certainly out there. However, NNTR has been barely used, unlike straight Nash-in-Nash bargaining. The basic problem, beyond the computational complexity, is that you have to observe not only the consummated bargains, but also the bargains which could have been formed but weren’t. Lee and Fong (2013) is the framework for estimating network formation as a dynamic game with continuation values and everything, and it’s just, completely impossible to implement.

Insurers definitely hold down prices by negotiation. How much of this is captured by the insurers? How much passes through to the consumer? It is frustrating to report that I do not have a very good answer for you. Cabral, Geruso, and Mahoney (2017) use changes in subsidies to Medicare Advantage providers to estimate passthrough, which can stand in for how much market power providers have. The answer, of course, varies sharply across markets, but only 54% of payment increases get passed on to consumers. However, we can’t divide that up between insurers and hospitals from the information in the paper. Ho and Lee’s bargaining weight coefficients suggest that hospitals get 70% of the gains from trade, the insurance company gets 15%, and the bargainer on behalf of the consumer, CalPERS, gets the remaining 15%. Individual purchasers are likely to get less.

This bargaining literature was extremely important for antitrust action. Starting in the 1990s and continuing to today, healthcare providers have been consolidating both horizontally and vertically. Between 1998 and 2017, there were 1,577 hospital mergers, out of some 6,000 hospitals in the whole country. Many of these mergers were cross-market, such that the services the hospitals offer were not direct substitutes for each other.

The results for insurers, and thus patients, were bad, whether within a market or across it. Cooper, Craig, Gaynor and Van Reenen (2019) found that markets with monopoly hospitals had prices 12% higher, and that mergers lead to increased prices when the hospitals were nearby. Dafny, Ho, and Lee (2018) found that mergers across markets within a state, but in different markets, increased prices by 7-10%, while mergers across state lines had no such effect. Brand, Garmon, and Rosenbaum (2023) cover the 558 mergers from 2009-2016, and estimate an average price increase of 5%. Brot, Cooper, Craig, Klarnet, Lurie, and Miller (2026) point out that price increases caused by mergers function as a sort of payroll tax, and thus decreases employment among lower- and middle-income workers.

Even with access to incredibly detailed data on the inner workings of the hospitals, we cannot discern any increase in efficiency. (Recall that a merger which does not improve the efficiency of the constituent parts necessarily makes the consumer worse off). Gaynor, Sacarny, Sadun, Syverson, and Venkatesh (2021), which appears to spring out of consulting for an anonymous hospital – a hospital which they then all but name by saying precisely how many locations the two had and when the merger occurred – finds that despite the intended reforms being implemented, there were not apparent improvements in cost.

Yet, the FTC and DOJ could not win a case in a hospital merger to save their life. They lost seven straight merger cases between 1994 and 2001, and eventually just gave up. From 2002 to 2020, out of over a thousand mergers, the FTC took action against 13. Brot, Cooper, Craig, and Klarnet (2024) argue that 200 of them should have been disallowed under the screening tools the FTC was using, and that those predictably anticompetitive mergers increased prices by at least 5%.

Subscribe now

The basic problem is that the economics of the courts were out of date. Much hinged on the definition of the market, which need not make sense. In 1998, for instance, the courts ruled that a merger in Poplar Bluff between the only two hospitals in town could go ahead because some patients traveled for care in Cape Girardeau, 60 miles away. (Nevermind if they were just seeing specialists!) And it goes without saying that the FTC didn’t even bother challenging mergers across markets, even though those should have just as much an ability to increase prices as those within a market.

The FTC’s workhorse is the willingness-to-pay approach of Capps, Dranove, and Satterthwaite (2003), which simplifies the question they must answer to “how much does adding this hospital increase demand for an insurance plan?”. With a logit demand model, you can compute that in as quickly as it takes to type a few lines of code and look up the variable names. What Nash-in-Nash bargaining gave the FTC, in the form of Gowrisankaran, Nevo, and Town (2015), was the intellectual legitimacy they needed in order to actually win cases. When you estimate a full Nash-in-Nash bargaining model and use it to evaluate mergers, you get answers which are pretty similar to the simple screen. Christopher Garmon (2017) compared the predicted results with the actual results of 28 mergers, and found that Capps, Dranove, and Satterthwaite’s approach does a pretty good job predicting demand.

I am not an unconditional fan of antitrust action. It can often be misdirected, especially when companies are innovative, and a broad discretionary remit enables a corrupt government to harass political enemies. I am nevertheless confident that antitrust action in the healthcare sector should be more vigorous than it presently is.

We have also seen an increase in vertical consolidation between health insurers and healthcare providers. This modern trend has been due largely to government interference. While early vertically integrated insurers and providers like Kaiser Permanente did this to hold down costs, and hopefully give the customer more coordinated care, the modern trend has been to bilk the government and take advantage of regulations.

First, many elements of Medicare are now privately run. Medicare Advantage insures half of seniors now through a risk-adjustment model. The government pays a subsidy to each insurer equal to the risk score of the patient, in order to make the marginal cost of taking on a customer the same. Insurers thus want the government to think that patients are much sicker than they actually are. The insurance company could strongly encourage the companies they contract with to rate patients as sicker, but it’s a lot simpler to just employ the doctors and have them upcode the patients.

Geruso and Layton (2020) is chock full of smoking guns for this being a problem. Once someone turns 65 and is eligible for Medicare Advantage, their risk scores take a jump. In particular, they are more likely to be coded for chronic conditions where the rating is squishy, diabetes being the biggest one. (There is no objective test for the severity of diabetes, and indeed we shorthand the severity of diabetes to just how much one is being treated for it.)

Even more damningly, this jump in upcoding is bigger when the provider is owned by the insurer. Otherwise identical people are rated 6 to 16 percent sicker when the diagnoser is owned by the insurer.

Among other things, the Affordable Care Act of 2010 required that insurers pay out at least 80% of premiums paid in healthcare claims for individual plans, and 85% for large groups. If the insurance has a year with unusually low claims, they must write a check to all of their policy holders equal to the gap. This Medical Loss Ratio (MLR) completely flips around the incentives for them to hold down costs, as now they actually profit from healthcare costs becoming higher. Cicala, Lieber, and Marone (2019) show that claims simply rose about 1-to-1 with exposure to the new rule. Their source of identification is the persistent variation from year to year across plans in how much they pay out in claims. The plans which were below the threshold in one year tend to be below it the next, and they would surely have continued that way were it not for the change in regulations.

Subscribe now

Xiaoxi Zhao (2021) is, to my knowledge, the only quantification of the effect of the medical loss ratio requirements in a structural paper. She starts out with a BLP style random coefficients discrete choice model, which allows her to flexibly estimate substitution, and then does the now standard Nash-in-Nash bargaining, with the difference being that we now check after each iteration what the implied MLR would be. This rules out some low prices as being infeasible, and thus the insurer is now okay ripping up the bargain.

It should be little surprise that this raises prices substantially, although her data is too stylized to be able to say much conclusive about the real world. The work with real data has to hold premiums and demand fixed, which is missing how it will pass through to costs. A hypothetical but realistic example with three insurers, given in Table 8, suggests that the imposition of medical loss ratio laws – which are, I remind you, supposed to benefit consumers – actually lead to prices increasing by 55%!

The two can go together, because owning a provider allows you to directly inflate the medical loss ratio and pass profits into the part you can keep them from. Kakani, Yde, Kanter, Frank, and Bond (2026) document that vertically integrated pharmacies increased prices by 9.5% after the passage of price controls in Medicare Part D, with the insurers more at risk of hitting the caps hiking prices more.

We have established, I hope conclusively, that insurers hold down the price of healthcare through negotiating. This is not always a good thing. High prices can actually be good. Inventing a new drug, for instance, requires an enormous outlay to discover it and get it approved, which must be paid back with higher prices during the life of the patent. If pharmaceutical companies expected for their prices to be reduced by bargaining, then they would not pay the fixed costs to discover drugs at all.

This is a live question. The Inflation Reduction Act of 2022 included provisions to use the bargaining power of Medicare in order to reduce the price of drugs. I note two caveats with that – first, the drugs chosen are very narrow, being a small number of drugs nearing the end of their patent life. These are also “negotiations” only loosely – unlike with doctors, who can simply opt out of seeing Medicare patients without penalty, if you refuse to sell at the price the government sets they will tax what is notionally 95% of the revenue of that drug. (Although in fact, they define the tax base in such a way that they include the post-tax price, so the real rate is 1,900%.) Such price controls will reduce the amount that we spend on drugs. Will it then reduce the innovations we get?

The literature is extremely clear on this subject. Investment into innovation, and thus innovation, is directly related to the expected profits. Acemoglu and Linn (2004) is an early test of this. They argue that the demographics of a particular country is completely unrelated to where the technological frontier happens to be. As a population gets older, they demand more drugs which cure the diseases of the old. It seems unlikely that it just happened to be that drugs for gout just happened to get easier to find when the population gets older and fatter! They estimate an elasticity of 4 – for every 1 percent increase in market size, you get a 4 percent increase in the number of new drugs.

This is across categories, which raises the concern that the inputs to pharmaceutical innovation are inelastic. Even still, there is a substantial response of innovation to market size reallocating resources from non-drug to drug uses. Dubois, de Mouzon, Scott-Morton, and Seabright (2015) have an elasticity of .23, implying .23% more drugs for every 1% increase in market size. Basically all estimates of the elasticity lie between these two.

Subscribe now

Medical devices, like stents and pacemakers, have a similar response. Ji and Rogers (2024) take advantage of Medicare cutting funding for some categories of medical devices to link a 61% decrease in spending to a 75% decrease in patenting. Going in reverse, Parker Rogers (2025) finds that decreasing the cost of approval for new devices increases both the quantity and quality of new devices. It is simply not reasonably contested that innovation responds to expected profits.

The natural next question is whether these innovations are something worth paying for, and the answer is yes. The social value of pharmaceuticals is incredibly high. To take an example, statins, which protect against heart attacks and strokes, had a social value of 1.25 trillion over 20 years, saving 40,000 lives. Not all drugs are statins, but the average value of a drug is something around $7-17 billion. Murphy and Topel (2003) famously estimated longevity gains since 1970 due to drugs at $320-480 billion a year. This is not just for the benefit of the world, which is able to free ride off the United States, and is not dependent on the choice of an extremely low discount rate, although of course caring more about the future or about other people will increase the social value.

We really can’t get cute with it. An alternative proposal is to tie the rates that we in America pay to those overseas. Drugs are much more expensive in America than in the poor, developing world, but also in the rest of the developed world too. This, too, however, would not be a good idea. The practical upshot is not that they sell drugs for less in America, or even that they increase the prices overseas. The result is that they do not sell the drug overseas at any price. Dubois, Gandhi, and Vasserman (2022) work out the consequences of reference pricing with just Canada, and find that it will not meaningfully decrease prices in America, while greatly increasing them in Canada. It goes without saying that what might be affordable in Canada is prohibitive in Nigeria.

I am quite skeptical, however, that the same innovations happen with healthcare delivery. We can test this! Clemens and Gottlieb (2014) find that an increase in pay by 2% leads to an increase in service provision of 3%, but with no change in mortality or health outcomes. The response of medical procedures to changes in funding is derisory, and they mostly arise out of hobbyist iteration. (See Dranove, Garthwaite, Heard, and Wu (2021) for more on this point).

This suggests a division of responsibilities. Where high costs are rents, as is the case for doctors, companies should bargain aggressively, and the government should bargain hard too. If there is substantial room for innovation, however, companies and the government must be mindful of the long run, and not mortgage the future for the benefit of the present.

This essay was several things in one: a partial exposition of the American healthcare system, an introduction to the Nash-in-Nash bargaining literature, comments on antitrust policy, and a digression into the effect of funding on innovation. I confess to jamming in some things which I had wanted to talk about for a while, but for which I could not sustain a full article. Your takeaways are for you to decide, but to the extent I can influence them, I would like to give some parting comments.

Most people’s understanding of healthcare is upside down. To them, the insurers are the great villains. It is they, after all, who are responsible for taking your payments, and who turn down claims for treatment. They are not friendly doctors, but bureaucrats. Nevertheless, our impressions are often wrong, and they are wrong here. The insurance company is more like a union. You may not always like a union, and they do not benefit everyone, but on the whole they increase your wages. Health insurance companies bargain over who is in the payment network, and in doing so hold prices down. Restricting the quantity simply allows you, the consumer, to capture more of the gains by replicating a lump sum transfer.

I envision AI alignment as being like research into aviation safety. There is no grand theory of why airplanes crash, unless you take that to mean gravity. Instead, we patch little problems. The airplane crashed because of metal fatigue in the engine – okay, we mandate inspections of engines, and prevent crashes from occurring by that source. As new problems arise, we iterate. We don’t even need for there to be a crash to make improvements – we test proactively, we build in redundancy, and we monitor for deviations which could be a threat.

Subscribe now

What we are not doing is thinking about how we are going to keep flying saucers from crashing. If flying saucers are invented, then we will work to keep them from crashing, but we will use exactly the same methods. We will test them, see how they perform, and patch particular problems. AI will be the same way.

I am doubtful that we can align via theorizing about a technology which does not yet exist. Consider the way people thought about AI in the 2000s, at least on LessWrong. They identified the basic problems AI alignment faces – leaving the invented jargon aside – which is an AI doing things that we don’t want to do, and some plausible reasons for this to be the case, but they had the model of technology completely wrong. There is no apparent anticipation of machine learning, but rather a program, a discrete block of code, which is an agent in the same way you and I are, which could then take off. It is not possible to align it, because it doesn’t exist. You can’t align code that hasn’t been written.

We should expect alignment to happen by default. We are training AI to accomplish the goals we want it to accomplish. What is so categorically different between “write code to solve this problem” and “don’t kill all the humans”? An aligned AI is simply a capable one. Indeed, the techniques used to make an AI more capable are the same which are used to align it.

I think there is much to be said for considering how AIs can monitor each other. We need not understand the inner workings, just as we do not need to read machine code in order to understand programming. I expect that the most important things to do are to use AIs to align other AIs. I don’t see why people expect this to fail – if you think that a dictatorship can last, how about a dictatorship with access to the inner motivations and desires of its citizenry?

Suppose, though, that we do delay AI via international treaty. What then? To the extent that the frontier labs are delayed, it allows other companies to catch up. We can imagine that there is a tradeoff between the mean and variance of capabilities, at least on the feasible frontier – do you want more companies to take riskier bets? Delaying the release of a model to patch concerns is the prerogative of a company with market power. The present leaders in AI are at least reasonably concerned with safety. Do not take that for granted, and think twice before you try to exchange them.

I am not yet taking a position on full automation. I am a bit pessimistic of the prospects of that succeeding, for no more a reason than that there are a lot of things which could go wrong, and we need not expect everything to go right. I am not particularly concerned about the “going wrong: including AI destroying us all, out of the control of humans.

This does not mean that I am unconcerned by AI! I simply believe that, if it were to destroy us all, or to lock us in slavery, it would be at the direction of humans. The biggest problem facing alignment researchers is not alignment at all, but keeping it from fulfilling our directives – or perhaps, to take off, so that it can give us a subservience preferable to the one that might be given to us by our human masters.

Since we are talking about matters AI, the reader is encouraged to read Roger Myerson’s recent note on the subject. The central argument is that much of wages are rents, and AI, being in many ways a superior monitoring technology, is going to greatly benefit those who ultimately control resources at the expense of those who labor, independent of whether their labor is no longer valuable.

If you do this, though, it’s possible for health insurance markets to completely “unravel”, and no one can buy insurance at any price. The people who know that they are sicker opt in, and the people who know they are healthier opt out, increasing the premiums. This drives out more and more people until no one is left. To prevent this, we mandated that every individual, at the cost of a fine, purchase health insurance.

If we require everyone to purchase health insurance, some people will be unable to afford this. This would be a sharply regressive tax, falling upon the lower classes who were likely to never purchase insurance at all. So to prevent this, the third leg of the stool subsidizes the purchase of health insurance for all people falling below 400% of the poverty line.

Each of these legs are necessary. Jonathan Gruber (2010) was very clear about this: “pulling out any of the legs while leaving one or two intact will critically undercut gains from reform.” We removed the individual mandate in 2017, when the tax penalty was reduced to zero dollars. I can still buy health insurance. What gives?

It’s a mix of things. For one, health insurance remains heavily subsidized. But for another, the purchase of health insurance is rife with inconsistent behavior, and sometimes strains the bounds of being called a rational decision at all. Customers are leaving money on the table — and thank goodness they are, because them acting rationally might collapse the system as we know it. In substantial part, health insurance is like a cartoon character that has run off a cliff – it has not fallen because we have not realized that we should be falling. There is a fundamental tension between people being aware of what is best for them, and markets existing at all.

The result is that the continued existence of individually purchased health insurance is in danger. With the expiration of enhanced premium tax credits at the end of 2025, premiums paid have risen by 58% in the past year, while the number of people enrolled in private, individually purchased health insurance has fallen by millions of people, concentrated in the part that subsidies were rolled back. Things will continue to get worse, because people poorly optimizing their plan decisions delays the unraveling of a market, but does not prevent it.

To properly answer the question of why, though, we must understand the economics of insurance markets.

This post is brought to you by Mechanize. They are hiring software engineers to help automate all work. Please consider applying.

Health insurance in America is provided through a hodge-podge of government and private insurers. 37% of the market is covered by Medicare and Medicaid, 64.5% purchase insurance from private companies, and 8% are uninsured. Many of the people in Medicare, however, are enrolled in Medicare Advantage, which is a privately provided alternative to traditional Medicare, subsidized by paying for the implied risk of each consumer, and the private plans are heavily subsidized and regulated.

Most private insurance is through the employer, because the contributions which the employer makes are tax deductible. Companies, especially large ones, often self-insure, where they are the one paying all of the costs, and are simply contracting with the insurance company for access to the negotiated rates and network, and to handle the billing.

A substantial fraction of private insurance, 10% of all insurance, is bought on the open market. Obamacare instituted exchanges, one for each state (insurance is regulated at the state level, and entry across state lines is inhibited). If you do not have an employer which offers satisfactory coverage, or are self-employed, you are eligible to purchase there. There are four tiers of plans on the market, corresponding to covering 60, 70, 80, and 90 percent of medical expenditures, and plans are subsidized by making it so that your premiums

There are several parts to a plan. The customer pays a premium, a monthly payment to the insurer for having access to its resources. The insurance company then insists upon measures to control costs and discourage the wasteful use of medical resources. These include the deductible, where you must cover the first $x of expenses before the insurance company picks things up, which is intended to keep the insurance policy for genuinely serious events. Then, once the insurance company is involved, you will generally pay a co-pay, which is a fixed fee per visit, and co-insurance, which is a fixed percentage of expenses which you must cover personally.

A substantial part of what health insurance companies do is bargain over what the prices for particular services should be. This is why you have restrictive networks. In this, they are something of a union for healthcare consumers. I am going to be skipping over this here, but there will be a second article in the coming weeks on this function, and the bargaining literature in industrial organization more generally. Whether these are allowed to vary depends on the state – in California, from which many of our studies draw their data, all plans offered on healthcare exchanges are standardized.

The desire for people to buy insurance goes back to the curvature of the utility function. Every additional dollar of consumption can be presumed to give you less and less satisfaction, which implies that you want to consume the same amount in each period. Negative shocks like healthcare expenses reduce your total utility, so you are willing to pay some amount to smooth away the variance.

There are two parts of risk aversion we care about, absolute and relative risk aversion, where absolute is our aversion to a gamble of a particular size, and relative is our aversion to a gamble which loses a particular percentage of our wealth. We will hold one of those two things constant, giving us constant absolute risk aversion (CARA) and constant relative risk aversion (CRRA). Throughout, we are going to use CARA, because it is much more tractable. It is unwieldy computationally to have to keep track of the wealth of each patient, and indeed, we won’t know what everyone’s savings are. That is not to say that CRRA does not have advantages – if you multiply everything up by the same proportion, decisions are unchanged. (CARA holds changes in levels invariant).

Absolute risk aversion is given by the negative second derivative of your utility function, divided by the first derivative of the utility function. The formula for absolute risk aversion is as follows.

It being based on e means that taking the derivative keeps e^-aw the same when we take the derivation. We are left, after division, with a simple alpha. As that increases, curvature – and hence, risk aversion – also increases.

Let us presume that everyone shares the same risk preferences, such that everyone would like to purchase an actuarially fair insurance contract. People differ, however, in the risk of them needing medical care, and importantly, they know this risk while the insurance company doesn’t. There is no guarantee that an insurance market exists under such circumstances, and the allocation is bound to be inefficient. In the strong case where everyone knows their exact risk, the insurance market either insures everyone or no one. The insurance company proposes a plan at the average cost, some people below that average cost drop out, the insurance company proposes a new contract at a higher price, more people drop out, and on until no insurance exists. This type of unwinding dates back to George Akerlof’s (1970) “Market for Lemons”.

We don’t generally observe the unwinding happen live. Markets that have completely collapsed don’t exist, and probably never existed – for example, income based college loan repayments don’t exist because of unwinding (Herbst and Hendren, 2024). What instead happens is that the market is shaded away from optimum, just how a monopoly distorts away from the optimal allocation.

The workhorse for understanding this is Einav, Finkelstein, and Cullen (2010). They have a simple graphical counterpart to the supply and demand graphs with which you are likely familiar. Demand for insurance is decreasing in price. In this example, demand for insurance is correlated with average cost, so that the average cost falls as more people are added. The demand curve falls faster, however, and we stop at the point where the price of the plan is equal to the average cost of the plan. (The graphs are from Einav and Finkelstein, 2011).

However, the efficient allocation would have been one where demand intersects with marginal cost, so that each person who is willing to pay more than what it would have taken to insure them is indeed insured. The amount of loss is given by the shaded region from C to D to E to F. It would be welfare improving in this example to require that everyone purchase insurance, and subsidize the plans just enough so that they break even.

It would also be welfare improving to subsidize insurance, or to pay each insurer a fee for taking on customers tied to the estimated risk of the customer. This latter is called risk adjustment, and is used by Medicare Advantage; a fuller discussion of the relative merits of both is deferred until later in this article.

It need not always be the case, of course, that more insurance is better. The average cost curve need not always slope down. If it is flat, then average and marginal costs coincide, and things are efficient. If risk aversion is negatively correlated with cost, then we can actually have advantageous selection. The average cost increases as more people are added to the plan, and the last people are added at a price below their marginal cost. Efficiency is actually restored by reducing the amount of insurance provided.

What makes this framework so useful is that all you need is variation in the prices available and the realized costs, and you can pin down everything, so long as you are willing to believe that there is no unobserved quality.

It can also work with non-linear cost curves, which might give us more exotic situations. The average cost might lie above the marginal curve, fall below as more people join, then rise again, as might happen if high cost individuals are certain to take a loss and thus the value of insurance to them is low. Efficiency here is denying coverage to both high and low risk individuals, and offering insurance to only those with middling risks.

We are very confident that an individual mandate reduces adverse selection and increases welfare. Hackmann, Kolstad, and Kowalski (2015) study Massachusetts, which essentially had Obamacare before Obamacare. In 2006, after previously banning discriminating on health status, the state instituted an individual mandate. We can compare outcomes there to every other state. When we do so, premiums stopped growing, in contrast to everywhere else, and even fell in absolute terms.

Interpreting the implied demand curves through the Einav and Finkelstein framework, the mandate increased welfare by 4.1%, or by $51.1 million a year.

Similarly, Finkelstein, Hendren, and Shepard (2019) use discontinuities in the subsidies available to infer the demand for healthcare, and the strength of adverse selection in the market. The adverse selection practically leaps off the page – the realized costs jump up at the threshold people lose access to subsidies, because the marginal consumers are the ones to drop insurance.

However, their results are not an unqualified success for subsidies, as the actual willingness to pay of people for this healthcare is at all points below what it would cost to provide it to them. The most likely explanation is that poor people already receive de facto free healthcare due to healthcare providers not turning away indigent patients.

Under Obamacare, you cannot charge different people in the same area different prices, nullifying the information a health insurance company might possess about your health status. This will make adverse selection worse, and it is thus surprising that we do this. The more information about someone’s health status a company has, the closer they can price their plan to their own marginal cost. However, “community rating” can be welfare enhancing. Consider what one is trying to insure across time, too. One would like to insure not just shocks in any given period, but the whole walk of health status over time. Preventing “reclassification risk” by banning discrimination on health status this smooths the expected premiums one will pay

Handel, Hendel, and Whinston (2015) is a remarkable quantification of the costs and benefits of community rating. They start by simplifying the environment, such that there are only two plans, and everybody is required to purchase one of them. These correspond to the “bronze” and “platinum” plans in Obamacare, and we will refer to them as H and L for High and Low. Despite the mandate to purchase a form of insurance, it is still possible for the market to unwind such that everyone purchases the Low, L plan. Even more strikingly, it’s possible for no equilibrium to exist.

The Einav-Finkelstein framework assumes that there is one type of plan, and all that varies is the price. Here, though, the plans vary in their generosity. Rothschild and Stiglitz (1976) considered what happens when firms are able to enter with new types of insurance plans, whose attractiveness may vary to different types of consumers. Suppose we start with a company offering a policy which is bought by everyone. Another company may find it profitable to enter with a policy whose features are attractive to lower risk consumers – for example, they might have a high deductible combined with low premiums. If they can offer this lower quality plan, then the old full-coverage policy is left holding the bag of sick customers, and loses money. They will have to adjust themselves. Depending on cost and preference parameters, we could have an equilibrium with everyone in an L plan, everyone in an H plan, and healthy in L and sick in H.

What’s more, it’s possible for no equilibrium to exist, at least not a Nash equilibrium. A Nash equilibrium is one where no player can make themselves better off unilaterally, and thus no one has an incentive to change the outcome. This has been the default for ages, but it does not work here. Going back to choosing between high and low plans, it can be possible for someone to offer a profitable price which takes back everyone.

Let’s work through a numerical example. 30% of people are sick, and they cost 100 in H and 70 in L. Healthy people are the remaining 70% of the population, and they cost 40 in H and 20 in L. The willingness to pay for an upgrade from L to H is 50 for the sick, and 30 for the healthy. Thus, everyone in H costs 58, and everyone in L costs 35.

We start with everyone in H. A new company would find it profitable to offer an L plan at 27, skimming the healthy customers while leaving the sick with the old H plan. This H plan leaves, and sick customers enter the L plan, which must rise up to 35 or else lose money. Now that everyone is in L, a new company can offer an H plan at a price of 65 which everyone would jump to, both healthy and sick, at a price above the cost of 58. Of course, another company would then be able to skim the sick, and so the cycle goes on.

Handel, Hendel, and Whinston propose a different equilibrium concept, called Riley equilibrium. Essentially we allow the insurance companies to look ahead one step in the game – if any deviation would be made unprofitable by a safe deviation, then it won’t happen. A safe deviation is one which can never result in losing money – we know this because if a company would break even with only sick patients, they must make a profit with all patients. When a Nash equilibrium exists, the Riley equilibrium coincides with one of them, and in all other cases exactly one Riley equilibrium exists. The assumption here basically just rules out the jump back up to an H plan

They have five years of data from an anonymous corporation between 2004 and 2009. During some year in this data, the company changed all of the options, and required each person to choose a new policy. This separates out inertia – some people just keep the same plan from year to year, even if a new plan becomes better – from people having idiosyncratic preferences. With variation in price, we can infer the distribution of risk preferences in the population of workers, and thus what plans they would choose under different conditions.

Adverse selection is indeed a big deal. Compared to a world where firms can offer different rates to different quartiles of health risk, adverse selection leads to everyone being in the L plan. On average people would be willing to pay $619 a year just to be in a world where they could purchase the H plan at a fair price. And yet, this is dwarfed by reclassification risk. The average person would be willing to pay $3,082, or half of annual premiums, to avoid the risk of their premiums increasing. Even allowing self-insurance through saving and borrowing does not substantially change the magnitude of the results.

The reasoning behind Obamacare follows. Because of reclassification risk, we ban discriminating on health status. Because of adverse selection possibly leading to people buying no plan, we mandate the purchase of insurance. And because of the market possibly unwinding to a lower level of coverage, we restrict what plans can be offered and heavily subsidize lower income consumers.

Do we need government regulation to do this? I would say yes, but we should first earnestly consider the free market alternative. John Cochrane (1995) proposes an alternative way of doing things. He wants to design contracts that both insure people in a period against bad health events, and also wants to insure people against their health status changing over time such that they must pay greater premiums. Signing a contract at 18 which lasts for the rest of your life would do this, but it would be impossible to make work. We certainly would not want for our future health insurance to be dependent on the continued good conduct and sound administration of a single company for the rest of our life, especially if it would be extremely hard to switch providers. Such a long-term contract would also not work if the healthy could not be expected to stick with the plan for the rest of their lives either. Our courts are loath to make someone abide by an indefinite contract, and in any case if someone goes bankrupt you cannot squeeze blood from a stone. Requiring that everybody pay in advance would make the insurance contract pointless.

So forget all that. Cochrane points out that you don’t need to sign a long term contract to replicate the features we want. We just need a series of short term contracts which leave both parties able and incentivized to sign another contract. You can do this with severance payments – at the end of each year, each person whose health status changes for the worse receives a value equal to the expected cost of future premiums. Conversely, every person who becomes healthier must pay the company. This last part is accomplished by pre-paying premiums, and simply being able to take less out if you turned out healthier than expected.

You might say, “but how can we possibly forecast how much someone’s health status changes the present value of their premiums?”. Cochrane answers, “we don’t need to”. There exists a healthy and robust market for consumers. The prevailing price a customer would face would forecast what it would take to insure them.

This paper made quite an impression on me when I first read it, but it is here that I must get off the bus. First, valuing the change in health status is much harder than it appears. Only some conditions are a discrete condition whose severity is well-known. Plans differ widely in their characteristics, including access to doctors as well as the usual variation in deductibles and co-pays. The products would have to be comparable to each other in order to value the premiums. Perhaps an agreeable formula can be found for how to handle deductibles, premiums, and co-pays, but how on earth could we agree upon the value of access to different doctors, or different reputations for customer service? Nor could we expect for companies to be incentivized to homogenize their products. Firms are clearly competing with each other through price. If two firms offer an identical product while competing on price, we should expect premiums to fall to exactly where price equals marginal cost, which is to no one’s benefit.

This can be fixed, however, with standardization of plans, as is commonly the case. This is not an insuperable obstacle. What is unsolvable is the market power firms possess, combined with persistent biases in consumer decision making. Without perfect competition, companies would be incentivized to inflate the prices they charge for the conditions which they have less of. And with poor decision making by consumers, we can hardly expect them to choose a plan they’re going to be happy with.

What’s more, these two frictions interact with each other. People facing a high implied cost to assess the relative merits of insurance plans means that it is less costly for entrants to punish poor performance by incumbents, and people sticking with whatever they happened to pick first means that prices can be raised without people switching.

Pricing in the insurance market is inconsistent with perfect competition. Pietro Tebaldi (2025) has data from California claims, 2013 to 2017, and estimates demand with a random coefficients logit model. This means he models people as drawing their demand for insurance from a normal distribution for each cluster of ages, region, and year, and he finds this by searching for the mean and variance which replicate the choices observed in the data, after purging the correlation of price and unobserved quality with an instrument. He uses Waldfogel instruments, which is using the region demographics to shift the price – the argument here is that since you are only allowed to adjust for age in a prespecified and actuarially unfair way, young people in older areas face a higher cost for health insurance.

Ordinarily, we specify the way in which firms compete with each other, and find the set of costs which would justify the prices observed; but here, since we possess costs from the realized claims, we can look at directly at what model best fits. Which is perhaps to be expected, given the often limited number of firms offering products on an exchange.

This has implications for the optimal design of subsidies. There are three ways to subsidize enrollees – risk adjustment, a price based subsidy, and a voucher based subsidy. Under perfect competition, where price equals marginal cost, all of these are equivalent. Under imperfect competition, they diverge.

The subsidies differ in whether they give customers a fixed sum of money, as in vouchers, or whether they cover everything beyond a certain point. The present design of subsidies in Obamacare is that you are not going to pay more than the second cheapest Silver plan (which is 70% coverage), plus the amount that you will ever spend out of pocket is capped as a function of income. The government promises to make up the rest. Naturally, with subsidies set up like this, the insurance company will be incentivized to inflate their costs, as the buyer is not the one who bears the marginal dollar of costs, and indeed if everyone were subsidized the government would have to directly negotiate in order to hold down costs. A voucher based subsidy gives each person a sum of money to spend on health insurance. They are responsible for the marginal dollar, and so prices tend to be lower.

The tradeoff here is the amount of knowledge that the government has to have about costs and demand. If you knew exactly how much to give each consumer, voucher based systems are strictly better. In Tebaldi’s sample, switching to vouchers would reduce premiums in concentrated markets by 27%, $4,127 a year to $2,998, while simultaneously increasing enrollment by 12.5% and saving the government $1,500 per person per year. (In less concentrated markets, competition had already reduced prices relative to what they would have been, and the benefits were much smaller).

You can also do things better by tailoring who gets subsidies. Holding the premium that people pay at a fixed amount and making up the rest of the costs means that the effective subsidy is larger for older people, but older people already have a higher demand for insurance, due both to being sicker and being innately risk averse. They don’t need to be paid as much in order to enroll. Tebaldi finds that it is possible to reallocate subsidies to young people such that every single person is better off.

However, to accurately price the voucher subsidies, you do need the demand estimates. Linking subsidies to price sponges up any error in projections of demand. This is not a convincing reason in a mature industry. Jaffe and Shepard (2020) study this trade-off, and argue that at this point, costs are pretty well known. Linking subsidies to price increases premiums by at least 1-6%, and by considerably more in more concentrated markets.

Risk adjustment is paying each insurer an amount equal to the expected riskiness of the pool of people who end up enrolling with them, such that the marginal cost of enrolling one additional person does not vary with health status. Under Obamacare, this is done through transfers between insurers, with no new subsidies from the government entering. Medicare Advantage is the main program which actually has government subsidies in the form of risk adjustment.

Risk adjustment and subsidies relate to an older discovery in industrial organization for identifying the way in which firms are competing. A subsidy shifts out the demand curve, raising everyone’s willingness-to-pay by some amount, while keeping the slope of the demand curve the same. Risk adjustment corresponds to rotating the average cost curve, such that it becomes flatter. (Recall Einav-Finkelstein-Cullen from earlier – this means that the marginal cost curve corresponds more with the average cost curve, and the number of people insured increases). Bresnahan (1982) pointed out these will have identical effects when the market is in perfect competition and markups are zero, but will differ when there is market power. I just thought this was a plain neat connection.

The big problem for risk adjustment is that how much a patient should cost is not a stable object. We obviously cannot rely upon ex post costs, because the insurers will inflate their revenue, especially the ones which are vertically integrated and so could collect from the healthcare providers directly. Even if we have an idea of how much a patient should cost, every insurer would like for their patients to be rated as riskier than they actually are. The smoking gun for this is the difference between vertically integrated insurers and those that are not in the Medicare Advantage program. Geruso and Layton (2020) show that enrollees in Medicare advantage have risk scores 6 to 16% higher than those in fee-for-service Medicare. It’s just so obvious – the moment people switch over to Medicare, what do you know, they discover they have extremely lucrative but subjective conditions like diabetes.

Einav, Finkelstein, and Tebaldi (2026) have a forthcoming article on the topic. If you are able to hold the details of plans fixed, as California does, then subsidies clobber risk adjustment.

So why do firms have this power? Perhaps because consumers choose plans in a way that is so confusing that it is difficult to believe that they are actually maximizing their utility. The utility of the consumer is not strictly increasing in the information known or in the variety of plans available.

Consumer choices are characterized by inertia. People purchase a plan, and then do not update which plan they choose as premiums and features change. Even if they picked it optimally the first time, they are absolutely not picking it optimally years later on. Ben Handel (2013) is able to use the same dataset we discussed in Handel, Hendel, and Whinston (2015) to put numbers on the welfare losses. The first step is to note that the first choice of plan is incredibly sticky. Most people carry over their insurance plan from the prior year. This is despite premiums changing substantially from year to year, as they realize the costs and adjust to remain profitable. The company subsidizes certain combinations of family size and income, leading some plans for some families to be strictly worse than others. Yet, many people still buy these worse plans. The implied dollar value for inertia is $2,032 a year from the demand estimates, and the floor loss from dominated plans is $374 a year.

Even when people are picking for the first time, they often make mistakes. Bhargava, Loewenstein, and Sydnor (2017) have data from a company that allowed people to customize their own plan, creating 48 possibilities. The vast majority of plans with low deductibles required paying a greater amount in premiums than one got a reduction in deductible. To take an example for their paper, to get the deductible down $250, from $1,000 to $750, one would have to pay $528. This can never be useful! And yet, 61% of employees at a large, anonymous firm chose a plan which was strictly dominated by another. Similarly, Abaluck and Gruber (2011) found that people in Medicare Part D acted as though money spent on premiums was much more valuable than money spent out of pocket, with consistent plan choice having the potential to raise welfare by 27%.

Why is this the case? It seems that people really hate paperwork, and they really hate thinking about health insurance. They really hate this. I’m sure I’d hate it, though I have not yet dealt with it myself. People are also wildly incorrect about basic details of what the plans will do, what they will cost, and when they pay benefits. Handel and Kolstad (2015) take the same approach in Handel (2013) and Handel, Hendel, and Whinston (2015) to estimating insurance demand, but add in the results of a survey querying people on the basics of the plan they selected, and how much they perceived the time and hassle of dealing with paperwork to be.

People were badly misinformed about what the plans covered and what they did. The company in their study offered a PPO plan or a subsidized high deductible health plan with a health savings account; despite both plans having access to exactly the same doctors, the twelve and a half percent of people who believed the PPO plan had better coverage were willing to pay $2,267 a year for the privilege of believing that. The average consumer, even if there was no hassle for self-administration, would have picked the wrong plan for the wrong reasons.

The distortion was not so large as for it to be better to mandatorily switch everyone over to the high deductible plan, but some sense of the losses from each of the sources – inertia and misinformation – can be found by comparing the loss in welfare under models with and without them, with dropping these frictions increasing the losses by 50%.

Under some circumstances, this lack of optimization might actually be good. The ideal world would be where everyone is blind to their particular risk and opts in, as they would behind a veil of ignorance. Going back to Handel (2013), whose company changed the menu of options available, it turned out that making people think about the choices they were making worsened adverse selection so much that it actually undid the gains from optimizing plans. Handel, Kolstad, and Spinnewijn (2018) find that just removing choice frictions, in a setting where employees have the choice of the generosity of plans, leads to just 9% of people enrolling in the most generous plans, despite the first-best world being one where 85% of individuals enroll in the high generosity plan.

People not paying attention to the actual details of plans also meant that making the individual mandate toothless by removing the tax penalty was not actually the same as repealing the mandate. Evan Saltzman (2019) found that enrollment responded to the presence of a mandate, but not to the amount. People just had a “taste for compliance”, even if there was no actual penalty.

However, poor optimization being useful is a contingent fact. In different contexts, choice frictions make adverse selection worse. Domurat, Menashe, and Yin (2021) mailed 87,000 people in California with information about what benefits were available, which substantially increased uptake. (The implied value of the letters, in terms of the change in subsidies needed to get the same result, was between 300 and 636 dollars a year). The people who were brought into the program were disproportionately healthier people, leading to premiums being lower than they would otherwise have been.

Meanwhile, optimization frictions absolutely do exacerbate market power. Insurance companies are incentivized to get people onto the plan, and increase premiums later on. Ho, Hogan, and Scott Morton (2017) estimate that consumer inattention increases costs in Medicare Part D (which is insurance for prescription drugs, and only covers a part of total health expenses) by $536 per person per year. Marzilli Ericson (2014) finds that older plans in Medicare Part D have consistently higher prices, with firms introducing new, cheap plans in order to attract new customers without needing to lower premiums for old customers. The

This is something where getting smarter might make us worse off. It’s striking how one of the very first empirical demonstrations of health insurance markets unwinding is Cutler and Reber (1998), who were working with data from Harvard, a notoriously smart place. For Bhargava, Loewenstein, and Sydnor (2017), they note that it was low earners that were disproportionately likely to choose the dominated plans.

So privately sold health insurance is stuck between a rock and a hard place. If people are left to their own devices, they choose plans poorly. But if you prompt people to choose optimally, there is no reason to expect the individual optimum to coincide with what is socially optimal. The choice needs to be somewhat taken out of the hands of people.

With everything we have learned, I would like to answer two questions. First, what is likely to happen to Obamacare exchanges? Second, what would an optimal health plan look like? I have no basis for speculating what the effect of an enormous change would be, so I will constrain myself to marginal improvements – given the basic framework, what can we improve upon?

The individual mandate was repealed in 2017, with that taking effect in 2019. The reaction of premiums and coverage were surprisingly small – there was no evident death spiral. There were still substantial subsidies, including new ones for people over 400% of the federal poverty line enacted in 2021.

However, the new subsidies were not renewed in 2025, and we have seen a considerable increase in premiums with an enrollment decline in the millions of people. The current projections, from the Kaiser Family Foundation, are that enrollment is going to decline from 22.3 million people to 16.5 to 17.5 million people, with most of that decline in the segment on the border of receiving subsidies.

What we have covered suggests that this is not all we are going to see. Things are going to get worse, because while choice frictions can delay a market unraveling, it cannot stop it forever. If we are going to ban price discrimination by health insurance companies, we are going to have to keep up the other legs of the stool.

If I could suggest some improvements, coverage should be universal, mandatory, and subsidized. Both adverse selection and reclassification risk are substantial. People plainly value health insurance as insurance, but that is no guarantee that they can buy it at the optimal price without government intervention. These subsidies should be in the form of vouchers, and we should move away from price-linked subsidies.

Plans should be clear and standardized. The government should be in the business of certifying plans for their quality, as they do in Medicare Advantage. Indeed, as Benjamin Vatter (2025) shows, certification thresholds allow an opportunity for the correction of market power in the amount invested into different networks. Standardization of plans both prevents insurers from trying to change the quality of consumers in the different plans, and prevents insurance companies from complicating the design so as to dazzle the consumer.

Insurers should be able to sell across state lines. It is difficult to see how preventing this is beneficial.

I cannot definitively comment anymore on the employer contributions tax credit, which is what results in the bulk of insurance coming through the employer. Doing the research has left me with an appreciation for how bad adverse selection is. Having healthcare through the employer takes the decision out of people’s hands, and ensures we’re insured.

This is just about all I have to say about health insurance. Sometime next month, I will have an article on the insurance company as a union of consumers, collectively bargaining over prices and utilization. But until then, keep well.

The costs of adopting it are simply not that high compared to either the holdings of farmers. Indeed, when we expand access to credit, which would be useful if the constraint were just needing enough working capital, we do not find a big change in uptake of better technologies. It’s a mystery.

That was why I was quite excited when I read Bold, Kaizzi, Svensson, and Yanagizawa-Drott (2017), who argue that while fertilizers and hybrid corn would increase yields, they won’t do it if they’re of substandard quality. They went out and tested fertilizer bought from local markets for nitrogen content, and found that 30% of the nutritive content is missing. Likewise, the difference in yields between hybrid corn they were sure was authentic and the ones they bought in the market implied that under 50% of the seeds sold were actually correctly labeled. Given that when they tested authentic corn and fertilizer returns there was a mean return on investment of 83%, with every sample having a return greater than 50%, we might have a problem with a clean solution of simply regulating quality of inputs. I was so excited by this I cited it favorably in my recent discussion of whether microcredit works.

Subscribe now

The trouble is that this finding does not generalize to other samples. Michelson, Fairbairn, Ellison, Maertens, and Manyong (2021) do essentially the same thing in Tanzania, sending back samples of fertilizer they bought in Tanzania to be tested for nitrogen content. They found absolutely nothing wrong with it. Even more directly, the International Fertilizer Development Center (IFDC) tested samples from Uganda, and found that only 10% contained less nitrogen content than they should. Of those samples which failed, the difference was small – it means that when they were supposed to be 46% nitrogen, they were actually around 44.5%. They’re pretty blunt about what they think explains the Bold et al results – the lab screwed up the testing. (Specifically, they used the Kjeldahl method at a lab which normally deals with soil. You essentially convert everything to ammonia through a series of chemical processes, and measure the implied nitrogen in the ammonia. If you screw up any of the steps, it biases nitrogen down. The IFDC instead just burn everything, and measure the gas that comes out.)

It just doesn’t seem plausible that dealers are systematically adulterating the fertilizer, because a lot of the methods are easily detectible. If you put sand in it to dilute, as Bold et al do, you can tell that something is wrong with it. It comes in distinctive blueish white pellets, and it would not be trivial to fake. It may well be the case that the fertilizer could be soaked with water, but it’s going to be rather noticeable that the prills have dissolved together and the whole thing is damp. There’s just no undetectable way to sell someone bad fertilizer.

Nor do bad storage practices alone explain it. If you store it in hot, humid conditions, it’s still well below the temperature when urea chemically breaks down. If it absorbs water, sure, that would dilute things, but it would need to actually be wet for it to matter. It’s just not going to break down without urease.

So why do farmers not adopt fertilizer? It’s worth reading Suri and Udry (2022), which covers the important things. The most hopeful of the reasons why comes from Michelson et al, who found that people use the appearance to infer quality in a way that is at odds with the actual nitrogen content. Because urea doesn’t actually easily degrade, fertilizer can look funny without actually being any worse. Plausibly, a public outreach campaign could correct this. The optimal amount of fertilizer varies, often with much misinformation from authorities as to how much should be applied (Duflo, Kremer, and Robinson, 2011), and providing actually accurate information is helpful. There is also some work from Islam and Beg (2021) about providing farmers with guides mapping leaf color to fertilizer needed. The risk of a bad crop is also going to distort their investment into intermediate inputs of all kinds, and so providing insurance will actually change input choices. (Obligatory reference to Karlan, Osei, Osei-Akoto, and Udry (2014) here, who randomized access to capital and insurance, and found that only insurance increased uptake).

The more pessimistic reasons are that the non-adoption is completely correct, given the circumstances farmers face, and which cannot be easily corrected by a change in beliefs or an insurance product. There is evidence that the returns measured in RCTs are the best case scenario – within a particular sample of fields, fertilizer has a big return, but once one goes elsewhere these returns evaporate. Suri (2011) argues that the farmers who find it actually profitable have already done so. In general, we’re almost certainly mismeasuring prices. If we measure input prices where it is convenient, then we undercount the prices of fertilizer out in the sticks, and we know that the price at which farmers sell their corn diverges sharply from the price which consumers face. Explanations for that include collusion by intermediaries (Bergquist and Dinerstein, 2020) and farmers selling their crops at the trough of prices for lack of storage (Burke, Bergquist and Miguel, 2019).

There is not, however, one single reason for inefficiency in African agriculture. There are lots of things wrong. I am confident, however, that fertilizer is not the reason why.

Harigovind S: Do you agree that economists are uniquely bad at presenting their work to public audiences? Context to my question is that this does not seem to be a problem facing psychologists for example whereas pop economics books are filled with anthropological accounts and no discussion of econometric traps and resolutions. Is this a result of the poor training economists receive in articulation or are conceptual ideas in economics fundamentally more difficult to communicate? I am interested in your thoughts seeing as you communicate the methodology of the field so effectively and persuasively.

I think this is partly our fault, and also partly not our fault.

Where it’s not our fault is this: the public does not particularly care about learning true things at all. The claims which are super successful in popular culture are those which are counter-intuitive but obvious upon consideration. They’re things which you can tell to your co-worker and feel a little bit superior as a knowing one. They don’t have to be deep, they don’t have to generalize. They can be cute. They can even be false.

I think everybody wants to see magic, right? The fun facts are the flourishes, the rabbit being pulled out of the hat. People are less interested in how the hat was manufactured, or how to manufacture rabbit hats yourself, but that’s what you need to know for the magic to be anything more than a bit of dazzle.

I have high standards for myself in how I present findings. I am just not all that interested in results, fundamentally. Results depend on parameter values, and parameters change. What doesn’t change are the methods which let you answer many questions. So to me, simply presenting a result is not presenting science at all. I need to prove the result by arguing for it above all other possibilities, and so it’s important for me to go through the paper, discuss whether and how it can make the claims it makes even in theory, then whether the data supports the results, and then where that stands in a broader world.

Where it is our fault is that a lot of people in the profession have no interest in really communicating science to the public. (For that is what economics is! Scientific inquiry into how the world works, same as any other!). Many people are content with the quiet life of publication. Now, I am happy to grant comparative advantage. Getting a platform is hard. Some people are not as good at it as others, although I sincerely doubt that I really have any great advantage in talent than enthusiasm. (That, and having lots of time to read everything). But a lot of people in the profession frown upon it, although naturally you will have only heard this in person; see this thread for color.

The Literary Quant: What was your experience as a child and what was it like to mature into an adult? How did being on the spectrum factor into this? How would you say your childhood affected or still affects your worldview?

Read more

The mailbag column next week will be partially paywalled, so please be sure to subscribe to read. I must make a living somehow!

Subscribe now

Nowhere is this more clear than in the rural agrarian village. Let’s call our village Rampur. It is in the north of India, somewhere outside Lucknow, in the Gangetic plain of Uttar Pradesh. It has a tight residential core of 20 acres in which 1,000 people reside, and surrounding fields of 300 acres of agricultural land. This village is fictional, but aside from it not existing, it is otherwise indistinguishable from thousands of other villages.

This post is brought to you by Mechanize.ai. They are looking to hire software engineers, and are paying handsomely. Please consider applying.

Where people live is strictly delineated by caste and religion. There are four big hierarchical castes, or varnas, taken in order Brahmins, Kshatriyas, Vaishyas, and Shudras, and underneath everyone the untouchables, or Dalits. Each of these umbrella castes is divided up into jati, which traditionally correspond to occupations. Each jati has its own lane in the village, a gali. It looks something like this.

At the center of the village are the high caste individuals, the Thakurs (landowners) and Brahmins (priestly caste, though most are simply landowners). Further out are cultivator castes – Yadhavs, Kurmis, Lodhs, Kachhis, etc – each of whom have reputations for specializing in particular things (the Kurmis as tillers of the land and so on), and the service castes, of which there are many. They have a particular job that is theirs by tradition, the Nai being the barbers, the Lohar the blacksmiths, the Dhobi the washermen, and so on. Traditionally they would do these jobs at the command of the high caste individuals in exchange for shares of grain, but this part has broken down and they work for cash. I listed out the jatis which appear in our village of Rampur, but I emphasize that there are a lot of these. In the 2006 round of the Rural Economic Development Survey, in the 202 villages included, with 99,760 house total, there were 3,266 unique sub-castes.

The Dalits, formerly the untouchables, are outside the caste system, though they have their own particular castes. They live outside the village because they are dirty, and cannot ritually pollute the well. (It used to be that a village was centered around a single well, controlled by the high caste – to get water, a Dalit would have to get a high caste individual to draw it for him, taking care never to touch him, or to let his shadow fall across the high caste individual. Water access is much more common now, with millions of wells being installed starting in the 1970s.) Now the Dalits live in the outskirts, in a little hamlet, strictly downstream. The dirty jobs, like leatherwork and funerals, are reserved to them, but for the most part they are landless laborers. There are also government jobs and spots in higher education reserved for Scheduled Castes, Scheduled Tribes, and Other Backwards Castes, including the Dalits.

The Muslims live separately, but they retain their own castes. Christians are rare in Uttar Pradesh, and there are none in our village of Rampur, but for the sake of completeness I note that they are fairly common in the south of India, especially the parts exposed to trade, and they retain castes.

Each village, or a small cluster of villages, has its own local government, the gram panchayat. These exist mainly as administrators for various government benefits, and do not themselves levy taxes. Elections are held every five years. Voting is essentially along caste lines, but only the narrow castes. Within the backward castes, they are often fighting for the same resources earmarked toward them. Landowners and local elites are able to use their control of local resources to influence outcomes (Anderson, Francois, and Kotwal, 2015).

People only marry within their jati, the narrow occupational castes. This pattern of endogamy is ancient, dating back at least 2,000 years. (How do we know this? When people mate, they exchange whole chromosomes. When they form new gametes through meiosis, these chromosomes get chopped to bits and reassembled. We can infer from how “chopped up” people’s genes are how long it must have been since they last intermixed. For more, see (Moorjani et al., 2013)). The higher castes tend to have a greater proportion of Aryan ancestry, but this does not correspond to an American conception of race, with north-south geography being more important for that.

People never, however, marry within their village. Under no circumstances can someone marry another in the same village, and it would be as shocking as marrying your brother. Since these occur at arm’s length, “love marriages” are extremely rare. Less than 5% of marriages were chosen without the involvement of the parents. The bride goes to live with the groom’s family. Since this is a burden to them, a dowry must be paid.

The paying of a dowry is technically illegal, and has been since 1961, but it is universally ignored. This dowry is substantial – it essentially corresponds to an inheritance paid before anyone has died. A family might have to transfer several years worth of income. Conflicts over the size of the dowry are common. The receiving family does not want to take on the burden of the daughter without appropriate compensation. In our region of Uttar Pradesh, 2,000 women die every year over dowry disputes, and this is the tip of the iceberg of the violent extortion of the bride’s family.

Subscribe now

Women rarely work formally, and this declines as families get wealthier. There is a strong norm against women working, and a stronger norm against working outside the house, where they cannot be supervised and might do something dishonorable with a man. Men are preferred to women at birth, and especially for later births to families without a son, there will be sex selective abortions and feticide of daughters.

Production revolves around the seasons, with two, sometimes three, crops planted. Water for irrigating the rice crop comes from the monsoon season, with the rain coming in early summer. This is harvested in September. Wheat, which is much more drought tolerant, is planted in November and harvested in March. Sometimes, a third crop of lentils can be planted immediately after the rice crop is harvested. The fields are plowed with oxen, which the Hindus cannot kill. They are kept for dairy, which is a source of both consumption smoothing and protein. Cows are sacred, and stray cattle are a persistent annoyance. The Muslims have an important role in doing the dirty work and slaughtering bullocks (in Uttar Pradesh, killing female cows is forbidden).

This cyclicality means that there is either lots of work to do, or very little. During the lean season, families draw down their savings, and try to find some form of work to do. The Indian government guaranteed 100 days a year of employment to rural households (now 125, as of 2025), mostly earthwork – digging ditches, irrigation canals, and the like. Other common jobs are firing bricks, or making textiles.

The other option is to migrate temporarily, working in the city for a season. Most people do not move permanently, but return to their home village at the end of the season. If one is willing to work for a longer period, they can go abroad to Saudi Arabia and the UAE, with at least several hundred thousand people doing this.

A day in the fields begins with walking out before dawn. It is customary to defecate now into the fields. In 2015, 55% of individuals reported open defecation in the countryside. Since this was something of an embarrassment to the national government, the Swachh Bharat Mission built toilets across India. However, the open defecation rate remains high, at around 28%. This has deleterious effects on the health of the community, and likely accounts for why children are still shorter than in Africa. Muslims are more likely to use toilets than Hindus, and their children are more likely to live, despite being poorer than Hindus, which indicates that what is driving this are customary beliefs about fecal matter being ritually polluting. Only the very lowest of the untouchables engage in cleaning out latrines. (Manual scavenging is, like dowry, both illegal and widespread.)

If they are a child, one would go to school. An ordinary school might look like this.

Enrollment is universal. Everyone, both male and female, goes to a school. The school they go to is not unlikely to be private, as around 35% of children in Uttar Pradesh go to private schools. (I am confident that higher figures are the result of people misreporting where they actually attend school). Private schools are more likely to be attended by boys within the family, and by richer castes. Females drop out much earlier than men. The main reason that they go to high school at all is to increase their prospects on the marriage market (Andrew and Adams, 2025), as they are unlikely to be working.

The government schools were notoriously unable to make teachers show up – Chaudhury et al (2006) reported on spot checks around the world. In India, a quarter of primary school teachers were not there when the enumerator dropped, and half of them were not teaching at the moment they dropped by. Across 3,000 schools, they could elicit exactly one case of a teacher being fired for absence.

Learning outcomes are poor. We are not entirely sure how poor, because India has only participated in one internationally comparable test, in 2009. They finished 71st out of 72 nations (beating only Kyrgyzstan), and have found lame excuses not to participate in any further testing. Most of the people in our village will be illiterate and innumerate. Across the whole of India, 45% of students in the 8th standard (about 13 years old) could solve arithmetic problems. The census figures, which will tell you that literacy is between 70 and 80%, are wrong, because they are purely from self-reported figures. If you actually ask people to read a text at the second grade level, they will perform much worse. In Uttar Pradesh, which is middle of the pack, 28% of children in third grade could read a second grade text in 2024 (Aser Report, p. 25) and this is actually a great improvement over the 6% who could in 2014. By the time we get to 8th grade, 67% of students in government schools can read a second grade text, and 45% can do division (p. 62-63).

Subscribe now

Healthcare is similarly provided. There is an official, government provided healthcare network, and there are private providers. The government system is in three tiers, with local nurses for every few villages referring complicated cases up to primary health centers, and then up another tier to district hospitals. Nevertheless, private providers are common.The private providers are more expensive, but trade that for reliability. The Chaudhury et al survey found that 40% of healthcare workers were absent when they dropped by. This absence is aided and abetted by local political forces, who view it as an opportunity for patronage more than doing the job. Banerjee, Glennerster, and Duflo (2008), in the most caustic paper I have ever seen them write, characterized monitoring for attendance as like “putting a band-aid on a corpse”. When they installed time clocks to monitor attendance, within 18 months the treated places had either had the time clocks “accidentally” broken, or the now monitored absenteeism was excused by the local government.

Those are the important details of life in a rural Indian village. Who you marry and what you do are bound up with who you were born to. While the details are specific, the generalities are universal. Village life from Uganda to Thailand is fundamentally similar to one another, and the lessons we learn from one place can be ported to another. In the next section, we will cover how these customs arose, the problems that they solved then, and why they are bad now.

The principal concern of the farmer is risk. The crops cannot fail. If they fail, you die. It is much more important to avoid a catastrophic fall in income than to have a higher standard of living most of the time. So, people shade their decisions – they spend less on intermediate inputs, choose crops which are less dependent on rainfall, scatter their fields rather than have them all in one place, and invest in livestock with a low rate of return but which provide something all year round.

They also save, and when that isn’t enough, they importune their relatives. These obligations to provide to your relatives is a form of insurance, of immense quantitative importance and surprising effectiveness.

Robert Townsend (1994) proposes a simple test for full insurance, which he tests against a sample of rural Indian villages. If risk is perfectly shared across a village, then individual consumption should not vary after controlling for variance in aggregate consumption. He finds that the data rejects full insurance, but it comes remarkably close.

Townsend is agnostic as to what, exactly, the cause of the insurance is – it is sufficient, for his framework, that the risk be insured through something – but the best explanation for coming close to full insurance are the ties of caste. Kinship networks allow for the risk to be shared across many people. Your fields may have been obliterated by a freak hailstorm, but your neighbors’ fields were not, and certainly not the people in the same caste in other villages.

Subscribe now

Townsend’s original paper is imperfect in two main ways, but they both point to him undermeasuring the extent of full insurance. First, the relevant unit is the kinship group, not the village. Munshi and Rosenzweig (2012) found that most transfers are between villages, if overwhelmingly within caste. The proper standard for full insurance is not village consumption, but average caste income across local villages. Second, Townsend is forced to make the assumption that the risk preferences of everyone are identical, or at the very least all come from the family of constant absolute risk aversion utility functions. If, however, some people prefer greater risks than others (in a way not captured by the demographic and income controls intended to sweep up some obvious confounds, like richer people having a greater stock of savings and thus inherently having smoother consumption) this will bias you away from accurately measuring full insurance.

Both of these imperfections in Townsend’s paper, however, suggest that informal insurance is even stronger than we thought. Both inject noise, and removing them makes the case stronger. The definitive treatment is Mazzocco and Saini (2012), who emphatically reject people having identical risk preferences, and while rejecting full insurance at the village level, cannot reject full insurance at the caste level.

But this insurance comes at a cost. It is not the insurance you sign with a company which has clearly defined obligations and benefits, but a latent claim of relatives against anyone who should come into money. It is a tax, and a punitive one.

Some of the idea of the magnitude can be given from a clever experiment conducted by Carranza, Donald, Grosset-Touba, and Kaur (2025), in the Ivory Coast. They offer cashew processors the option of a blocked savings account into which they can deposit a part of their paycheck. These savings accounts do not pay interest. They exist solely to prevent one from being able to access the money for nine months. It is surprising that anyone would take up the offer, and yet many do. Because they are working at a job with pay tied strictly to productivity, through a piece rate system, we can observe how much more people are willing to work when their relatives don’t know how much they’re making.

The effects are tremendous. Earnings rise 9.4%, and attendance rises 6.5%. These gains are observed only when family members don’t know about the accounts. When they do, very few people are inclined to take up the account, ruling out that this is about self-control for the worker. The implied tax rate of the family cannot be directly known without knowing the labor supply elasticity for the workers, but we can put bounds on it through estimates elsewhere. The high end estimates of the labor supply elasticity imply that the workers who took up the account faced a tax rate on the order of 65% of additional earnings, and the lowest tax rate it could plausibly be is around 20% of income.

This finding is extremely well-replicated. Frankly, you can just go out and talk to people and see that this is true, but there are also numerous lab experiments (e.g. Jakiela and Ozier, 2016). Munir Squires (2024) also does a lab experiment, and points out that the effects of the social tax are especially binding on those with the most skill. This misallocation, in his model, lowered productivity by a quarter. People will also borrow at a considerable interest rate money which they already have themselves in savings. Baland, Guirkinger, and Mali (2011) report that in their sample in Cameroon, a fifth of borrowers pay a net interest rate of 25% a year just to hide away their income.

The social obligations which sustain coinsurance make it much harder to migrate. If you move to a far away city, who are you going to marry? India is surprisingly non-urbanized for a country of its level of development. I believe that caste is substantially connected to this.

Subscribe now

An illustration from Bangladesh, which is sufficiently similar for lessons there to apply to India. In the early 2010s, we got quite excited by some studies which subsidized people to migrate to the cities. Bryan, Chowdhury, and Mobarak (2014) found that a small subsidy caused 22% of households to send a migrant, greatly increasing their wages. Wages are just higher in urban areas. However, when we inquire into the deeper causes, as Lagakos, Mobarak, and Waugh (2023) do, we find that the main thing keeping people from moving is an enormous disutility from being in the city. What might this be? The soundest explanation, I think, is that you will lose access to your kinship networks, including your ability to marry.

(I explained this methodology in some detail under Kaboski and Townsend (2011) in an earlier post on microfinance, which I am naturally eager for you to read. Essentially you write down a model which from inspection seems to capture everything of importance, and simulate it over and over again until you match the data details of the period before the intervention. If your model is correct, then changing one of the parameters in line with the intervention – in this case, changing the cost of moving – should result in matching the data, and it does.)

Munshi and Rosenzweig (2016) test for who is likely to migrate, if castes are providing insurance. We would expect the people who are wealthier to send more migrants in order to escape the demand of their families, and for people in a worse position to stay. You might reasonably point out that we would expect richer families to differ in their skill, perhaps investing more into education, and so have a higher gain from migrating; but they sidestep this by pointing out that what matters is whether they are a rich family only relative to the rest of their caste. Holding income constant, you can see how much relative position matters for migration (quite a lot).

We also have some unexpected evidence from the United States. Ghosh, Hwang, and Squires (2023) infer the effect of cousin marriage bans in different U.S. states through changes in the frequency of same-surname marriages, which can stand in for cousin marriage. Places that had higher rates of cousin marriage are more likely to migrate off the farm and into higher-paying jobs, and are also more likely to live in homes for the elderly as a consequence of weaker family ties.

The need to maintain kinship networks leads inexorably to dowry, and the pattern of marriage in India. Marriages simply must be arranged in such a world. You cannot chance your children falling in love with someone outside of the caste, and the cultural separation needed to sustain endogamy without force is on the level of the Amish. In any case, once you have restricted your marriage prospects in each village to a cluster of 30 households, you would become hopelessly inbred. It would be too costly – more so in the past – to have people travel tens of miles just to interact with potential suitors, so instead it’s treated as a business decision and negotiated.

Daughters go to live with the groom’s family, and for this a price must be paid. This is not universal, however. Many other places have the sons live with the family of the daughter. The fundamental cause is the form of agriculture, and whether it requires the hoe or the plow. Hoe agriculture, commonly used for root crops, is labor intensive, but does not require anyone to be unusually strong. Women can do this, and in places with hoe agriculture, work in the fields. The plow, on the other hand, requires considerable upper body strength to operate. (It also reduces the need for weeding, which women and children might specialize in).

It is little surprise, then, that the direction of where people go to live, and where the price for marriage flows, varies with agricultural opportunities. Much of Africa pays a bride price, because much agriculture is hoe agriculture. Much of India – including our village of Rampur – is plow based agriculture, and women are not involved with cultivation. This has a long run effect on gender attitudes – Alesina, Giuliano and Nunn (2012) find that places which kept women in the house due to agriculture in the past have worse attitudes towards gender equality today, and that these norms are carried with them when they migrate elsewhere.

The payment of dowry is bound up with informal insurance. The timing of marriages are affected by droughts and other shocks. In India, a drought increases the age at which marriage occurs, while in Africa it decreases the age at which it occurs. In the former, the shock to savings means that the family must wait longer before they can afford it, while in the latter the family is more eager to sell the bride for cash to get by. Dowry thus forces families to maintain a stock of savings above what they would otherwise, and allows other people to make claims against it.

Subscribe now

The transaction is not concluded when the dowry is paid on the wedding day. The groom’s family will continue to make claims, prompted by some necessity – sickness, Diwali, the breakdown of a motorbike – and backed up by the threat of violence against the bride. Violence against women is incredibly common in India, with 28% of women across the country reporting having been physically abused in the past year. About 5,700 women a year are killed due to disputes over dowry, with Uttar Pradesh accounting for 2,000 of them. The paradigmatic example is the bride burning, where the woman is covered in kerosene and set alight, with the aim of blaming it on a household accident or suicide. It’s a really ugly part of rural life, yet common. (I’ve received some pushback on this claim. Yes, 28% in the past year is correct. NFHS survey, page 680. The most common form is that 25% of women report being slapped by their husband.)

The low labor force participation rate among women can be thought of as the solution to a principal-agent problem. The groom’s family, in particular the mother-in-law, would like to extract surplus from the bride. Allowing her to work outside the house might well have higher potential earnings, but they are earnings which she could spend on herself, or transfer back to her family. Forcing her to work on unremunerated household tasks means that the family captures her whole surplus. If this model is right, then whether someone is mobile should be affected by whether the mother-in-law is alive, and indeed it is.

Paying dowry also means that it is easier to abuse the bride because the outside option of the bride is worse. She cannot return to her family – not when they have given much of their savings to pay for her to another family. So she stays, abused by the family. At least when she has sons, she’ll be able to exploit their wives!

I do not want to overstate the economic story too much. Much of the opposition to working outside of the house is plain old sexual jealousy. Kailash Rajah (2025) shows that women are much more likely to apply to or accept a job at a woman’s only work force, and that not going to the job is much more common among women with jealous husbands. And it’s entirely possible that the main problem is patrilocality, rather than the paying of a dowry per se, at least for domestic violence. And some of it is the compounding effect of parent’s beliefs about the future – with worse labor market outcomes expected, parents invest less in their children, who then are less likely to work. But there is no reason to expect women to be unable to work in India, and many reasons to think that this is making them poorer.

I am not convinced by the claim that there has been a major reversal in trend for female labor force participation in the past few years. Measured female labor force participation increased from from 24% to 47% between 2017 and 2023, but this was probably driven by changes in the definition of employment, where women doing activities like working in a small garden around the house or collecting firewood are now self-employed (Thakur and Chaudhary, 2024). In the cross-section, families which are wealthier have women who work less.

Who works is not the only thing affected by customary beliefs. Even among the men, who works which jobs are affected by custom. Jati are traditionally associated with very particular jobs, with different jati having feudal obligations to each other. People only performing the jobs of their forefathers is no longer as set as it once was. The modern economy is different enough, and outside options are good enough, that people are no longer bound to particular jobs. There is still, however, a considerable preference to not work jobs not associated with your caste.

Subscribe now

Suanna Oh (2023) is a remarkably direct test of how much people are willing to pay to avoid out-of-caste tasks. They start by identifying what jobs are associated with which castes. They’re pretty narrow. For instance, the Dhoba are associated with washing clothes, but not with washing anything else. The Hari (who are untouchables, and universally agreed in the survey to be the lowest caste among those ranked) will sweep latrines, but will not sweep animal stalls.

Each person is hired for a default task without a caste connotation (assembling paper bags) and an additional task which is revealed in private. The experimenters vary the task that is offered, the wage at which it is offered, and the time spent doing it. The idea behind varying the time spent is that we want to be sure we are picking up the disutility from one’s identity being violated, rather than simply the disutility from exerting effort on that task.

People absolutely do not want to do lower caste jobs, and to a lesser extent don’t want to do higher caste jobs either.

They run a second experiment where they surprise people who they’ve hired to assemble paper bags with a request that they perform a task for a bonus. The results are even more astonishing – 43% of workers would refuse to do a task outside their identity for 10 minutes, even for 10 times their daily wage. It’s not a matter of effort expended, as it’s flat across the time increments. Instead, people have a conception of themselves as a person who does x and not y, and will not budge for filthy lucre. This tracks with historical evidence – Alexander Persaud (2023) found that immigrants to Fiji from high castes were willing to pay much more – 2 and a half years of wages – in order to return home.

This is pretty bad for society! We know from the United States that race and sex discrimination led to substantial misallocation of labor, and Hsieh, Hurst, Jones, and Klenow (2019) argue that 20 to 40% of growth between 1960 and 2010 can be explained by people reallocating to what their comparative advantage is. Cassan, Keniston, and Kleineberg (2022) pursue a similar exercise, and find that removing the caste preference for certain occupations could substantially raise income. I would like to point out that I disagree with much of their analysis, though. They are more pessimistic, because they believe that caste leads to greater training and skill accumulation in certain jobs. I agree that castes likely do do this, and that there is some role for training – Hardy and McCasland (2022) find that small firms in Ghana, for instance, have room for more labor, they just don’t want to pay the cost of finding a good one – but I am skeptical that their estimates of the returns to skill are actually reflecting skill, and not rents. In other words, people working in their parents occupation are more productive, in the sense that they earn more, but this need not reflect them actually being better at their jobs. Thankfully, though, the distortion is mostly in low-skilled jobs in the countryside, and once people enter the modern economy old caste preferences are of less importance. Still, though, caste preferences can keep people in the old economy by changing schooling choices – Munshi and Rosenzweig (2006) show that lower caste networks chose local language schooling, even though English language schooling enables people to work in better jobs.

Being in a kinship group allows them to act in a coordinated fashion, but this need not be good. Cooperation with your fellow man is certainly good in general, but it need not be if you are an army cooperating to invade another country. Similarly, the bonds of caste and family allow groups to extract rents from each other, a practice which is good when you alone do it and bad when everyone else does it.

Subscribe now

For instance, wages are surprisingly sticky in Rampur. When rainfall is good and there is a positive demand shock for labor, nominal wages increase. When times are bad, they do not fall. It takes inflation to get rid of the high real wages. Until then, people are simply unemployed. (Note that one must be careful in dealing with self-employment. By and large, when someone is self-employed in a village, this is disguised unemployment). Supreet Kaur (2019) estimates that this nominal wage rigidity reduces employment by 9%, which is absolutely enormous, and like having the Great Recession but worse pretty much all the time.

This is not something which people are willing to abide by, if it weren’t for the transactions being observed and people being able to retaliate against those who break the norms. Breza, Kaur, and Krishnaswamy (2026) find that less than 2% of workers would accept a job 10% below the prevailing wage if it were proposed in public, but fully a quarter would if the offer is made in private, in the home. When they survey people as to why, they make it clear that it’s against the norms of the community to accept wages that undercut other workers.

This is a really big deal! Labor rationing is ubiquitous in Indian labor markets, with Breza, Kaur, and Shamdasani (2021) showing by inducing surges in hiring that pulling people out of the local labor market leaves wages unchanged in the lean months, but increases wages in the peak months. At least a quarter of self-employment is because they cannot find work at the prevailing wages.

Personal contact holding up collusive arrangements is likely widespread in the developing world although we do not know the full extent, as collusion is difficult to measure even with excellent data. Bergquist and Dinerstein (2021) find that maize traders will collude with those they know in order to hold up prices in Kenya, and Garima Sharma (2025) finds that wages for garment workers in Tirupur are held at minimum wage by collusive cartels of garment factories. It’s worth pointing out here that clusters of businesses in cities tend overwhelmingly to be run by people in the same kinship network.

It goes without saying that caste solidarity leads to discrimination. Soumitra Shukla (2026), working with multinationals hiring from elite Indian colleges, shows that despite lower-caste individuals being more likely to be promoted and less likely to leave, they are weeded out by higher caste interviewers before they can be hired. It is striking how people with unambiguous surnames get stricken early on, while people with ambiguous surnames are removed only after the interview stage. Going back to Rampur and villages like it, low caste households have much worse access to water for irrigation if someone from a higher caste owns the water supply – Siwan Anderson (2011) finds that lower caste water buyers would have yields 45% higher if they were in a village where the water supplier was also of the same caste. It is striking that trade could even break down in the water market, because the good to be traded and the contracts needed are so simple. In kinship based societies, though, people have limited trust. Moscona, Nunn, and Robinson (2017) show that in places in Africa with similar customs, they trust the central government and outsiders less.

I am inclined to say, however, that the predominance of family firms in India is not due to caste or custom, but would be solved by a change in the legal environment. Firms are by and large owned by members of an immediate family, not shared among people in the broader jati. The better explanation for this is the slowness of the legal system making it impossible to discipline recalcitrant managers except through extralegal channels. As I have argued before in “What’s the Matter with India”, there are simply not enough judges. But we must move on.

Subscribe now

Laws against firing people may also have a caste origin. Being unable to migrate, except in a group, means people will be around their family, even in the same job. Relatedly, recruiting was historically dependent upon caste ties. A company would pay an agent to go back to his village/region, and convince as many men as he could to come work. Once there, they cannot be thought to be maximizing their own individual utility as much as the utility of the collective.

Before the formation of India as an independent nation, this inhibited the formation of industry. Pseudoerasmus has an excellent article arguing that workers would not permit the use of labor-saving machinery, even if it meant that their own wages were raised. The unemployment of their relatives was unacceptable. Reading that article is a handy corrective, by the way, to the notion that the controllers of capital are invariably the ones with power in a relationship. Strikes were frequent, as the Bombay Presidency did not wish to upset the colonial subjects by coming down against the workers. Behind this was, naturally, the bonds of kinship, which made organization possible.

What is permitted to be done is bound up by culture. Textile factories in India required many more workers, even when using identical production technologies, because of strong norms against doing too much. Gregory Clark (1987) argued that workers didn’t want to look after too many machines and would prefer to be idle, even if they could increase their wages. To quote a contemporary observer, “The operatives in this mill refuse to attend to more machinery. I watched two ring frames for three minutes; there was not a single end down, yet the workpeople would not look after more than one side. They said they were satisfied with the present wage, and that there were too many men who want work and cannot get it that it would be unfair if they were to attend to more machines.”

This is all backward. Obviously it would be great if people simply lived and worked and created without market imperfections. Bargaining over rents through collective action can be fine. But we’re pretty sure that, under any reasonable world, one should not start out by trying to haggle over meagre gains. As Itskhoki and Moll (2019) point out, the optimal policy under realistic frictions is one where the government actively favors capital holders and holds down the wages of laborers, at least until growth is achieved. Strikingly, this logic holds even if one does not care about the welfare of entrepreneurs in the slightest.

The logic behind it is fairly intuitive. Suppose that entrepreneurs face borrowing constraints, modeled as a requirement to post collateral. They are unable to borrow as much as would be profitable, being limited by their own savings, and accumulate less capital than is optimal. Having more money earlier on leads the entrepreneurs to use the profits to finance yet more borrowing. Arbitrarily favoring them at first, by taxing labor to fund them, actually raises worker wages because it raises the amount of capital they are working with.

This pattern of forced savings is a common thread in countries which made the takeoff into sustained modern growth. South Korea did it, trying to keep wages below 80% of inflation plus productivity growth. China is still doing it, and has an astonishingly low labor share of income.

India did something like this, to their benefit. The 1991 reforms, which abolished the License Raj and made access to outside capital much easier for Indian companies, are commonly thought of as the start of modern Indian growth, but if you look at the time series it starts earlier, in the 1980s. Rodrick and Subramanian (2004) argue that the Indian National Congress, which had hitherto been a mix of Cambridge socialists and agrarian rusticists in the line of Gandhi, lost the election in 1977 and upon returning to power felt that they had to have better relations with the big businesses.

Still, though, Indian labor law is incredibly favorable to workers, on paper. You essentially cannot be fired, ever, not without government permission. What good does this do the people who are never hired? This is something which could be changed, thankfully, but to my mind it traces back to kinship solidarity, and thus to caste.

Subscribe now

For India to grow, the jati must break. Clan obligations are incompatible with a rich, modern society. It is not enough to trust your family. You have to trust at least everyone in your nation.

It is difficult to see what India can do to undo caste, however. I certainly do not expect this article to sway government policy, and the fabric of life cannot be easily upturned. What little I can hope for is this – at least when people come to America, you should leave caste behind. It should simply be in bad taste – fundamentally un-American – to think about it or care.

In the realm of policies which are possible for disrupting caste, the provision of insurance is likely the best one available. The status quo is to provide government insurance through a paid work guarantee in the rural villages, but leaves people unable to move and still be insured against idiosyncratic risk. This also means that aid is administered by the gram panchayat, which as noted much earlier in the essay, is dominated by caste clientelism and may strengthen caste. Instead, insurance should be available for purchase by all people, whether or not they own land. (The current system of crop insurance, the PMFBY, does not allow you buy insurance if you do not own crops! Good luck if you are a landless laborer who needs to be hired by farmers! Munshi and Rosenzweig (2014) find that selling insurance to landowners makes wage laborers actively worse off, as cultivators shift to riskier crops). In a related line, better weather forecasts will also reduce the risk that farmers face. Reducing risk is good for its own sake, but it is also good because it supplants the caste system.

I do not agree, however, with removing the extensive reservations for lower castes for government jobs and higher education. Or at least, I am unconvinced that it is that which is responsible for caste’s prominence in Indian life. There’s a lot of active discrimination out there. Calls for removing it strike me as a cynical claim for resources, rather than seriously believing that people would just stop thinking about caste if it weren’t for the reservation system.

This article has focused on India. I do not believe that the lessons we draw from here, though, are particular to India. I believe that they touch upon something common to agrarian, kinship societies all around the world. Having family ties is beneficial when nothing can change, the world is harsh and unfriendly, and better things aren’t possible. When things improve, they hold us back. They will be, in part, dissolved by progress – but for progress to happen, we need to dissolve them first.

We should, nevertheless, be less concerned about existential risk. You are going to die. I am going to die. We are all going to die, whether or not humanity goes extinct. The only plausible way to prevent this is to develop AGI. Some amount of safety is good, even if you’re only interested in your own welfare, but if we have to choose between completely stopping the risk of extinction from AGI through stopping AI entirely, or developing AGI even at a substantial risk of human extinction, we should take the latter. Extinction is not something special. It is simply the shuffling up of deaths which were already going to happen.

Subscribe now

I think this attitude toward the risk can be justified even without the particular ethical views that I am evincing here. If you are a utilitarian, whether or not we should accept the risk is simply a matter of what parameters you believe about extinction risk, the future flow utility of humanity living without AGI, and the future utility of humans with AGI. Indeed if you’re a utilitarian you can’t categorically say anything at all. If you view the number of people as mattering, then you will be inclined to support AGI (as that would encourage humans spreading across the galaxy). If you see AGI as mattering for its own sake, because it too experiences things, then you would favor it. If you think that people not existing is a zero, not a negative number – as I do – then the scales will be tipped toward developing AGI. If you regard death as being itself a negative, then when deaths occur is not that big a deal.

Someone who favors a deontological theory of ethics would find it difficult to oppose it without hypocrisy. There is no guarantee that the development of AGI would endanger us all. It simply raises the possibility that people may die. One might be inclined to say that increasing the risk of death for others is bad, and yet there are many actions which we do every day which raise the risk of death to others. When we drive to work, we may run over somebody; if we play outside we might catch a sickness which would itself cause human extinction; when we talk to someone, we may give them the flu. Clearly, risking the death of others cannot itself be the end of the argument, and once you start arguing about which numbers to believe in you’ve acceded to consequentialism.

The religious thinkers have already allowed it. The Magnifica Humanitas clearly does not forbid the creation of AGI, simply warning us to do it rightly. The creation of nuclear weapons was not immoral, only its use. The creation of a Golem, in Jewish traditions, is done by noted rabbis without comment as to it being bad, and indeed is approved of as being made with the “forces of sanctity” (although I note that its treatment of extinct risk is perhaps the most stringent of any religion). As far as I know, there is no religious authority which has explicitly called for stopping the creation of AGI.

My own personal views on ethics are as follows. I exist. I have no sound foundation for thinking about the world besides that. By that I mean that there is nothing else which survives the question “but why?”. It is the only axiom which is true by definition. Because I exist, I act. Even inaction is itself a sort of action. When I choose a particular action, it demonstrates that given my constraints, this action is the one I prefer most. Thus, I maximize my own wellbeing. There are an infinite number of possible actions, over which I have a full and complete ordering. Thus, we may think of there being cardinal, as opposed to only ordinal, utility, in the same way that an integral is composed of an infinite number of infinitesimally small objects, and we can admit the possibility that I like some things a lot and some things a little.

We may also surmise that others are like me. Certainly as I interact with the world, I will find that behaving as though others do possess thoughts and feelings like me will be to my advantage. This does not mean I have any basis for caring about their welfare – this would not survive the question “but why?” – but it does mean that we can imagine what a negotiation between all of us would be. We know that just as we can mutually immiserate each other, we can mutually make us better off. Our goal is to figure out what that agreeable set of principles and actions is, and work toward that world.

From that falls out everything. We should be good to one another; we should treat others with kindness and generosity, but not so much as to enable parasitism; we should strive to expand the possibilities that people can choose; we should produce good and useful things. It’s not dissimilar to what most people believe otherwise.

Where I differ from most people, I expect, is in my view of death. To die is simply to not experience. We fear it, as we live, and do what we can to prevent it, but once it occurs it is a matter of complete indifference to us. How can we possibly care that we are dead when we are dead? How can we possibly care that we don’t exist when we don’t exist? It is only those who live who are harmed by it.

Every night, I go to sleep. I cease to experience. Would it matter to me if I died overnight, and were reborn with the same memories and experience? Would I notice? At every moment, my atoms change. Who I am changes. Are there suffering me’s of the past who have ceased to be? On the contrary, they don’t exist, and so experience nothing. I cannot admit the possibility that we must care about every potential combination of atoms.

I believe that if you proposed, to all people now alive, the tradeoff of some risk of death in exchange for immortality, they would take it. Certainly we take on a risk of death every day in exchange for lesser pleasures! If you offered this to all people who would come to be in the world without AGI, they would refuse it, but we must confess that the proposition does not make much sense. The people who would come to be in a world without AGI would also not come to be in any world with even the most minute alterations. If the moment of their conception was even the slightest delayed, they would not be the same person. Instead, we must offer the deal to all the myriads who could have been. That would simply be utilitarianism which cares only about maximizing lived experience.

I think that what is right in the last scenario is essentially unknowable. For every conjecture of negative, awful, terrible things, I can counter with an equal and opposite vision of goodness. I think it is much more responsible to care about the preferences of those who are alive now. When we do that, extinction ceases to be the thing to avoid above all else, but simply another risk among many. We thus should not fear it above all else, and we must develop AGI.

Nevertheless, there has been a recent return of Marxist economics in the form of some very good papers. But to understand the context of all in which it exists and what came before it, it’s time to rewind to the 1960s. At that time, the economic world was riven by a seemingly arcane dispute about the aggregation of capital into a production function. Robert Solow’s famous 1956 growth model, in which depreciating capital is combined with labor in order to produce output, had just been published. In it, and the 1957 followup, he attributes some fraction of growth to capital accumulation, but the majority to increasing technological progress. In order to make such a determination, one must aggregate together all different types of capital – steam engines, factories, automobiles, foundries, and fishing boats – into one measure of capital stock.

Subscribe now

He was writing in Cambridge, Massachusetts. From Cambridge, England came the response. It is not clear under what circumstances you can meaningfully speak of “a” stock of capital. Joan Robinson (1953) points out that in order to convert heterogeneous capital into a single stock, we must value it in a numeraire good (money). Different pieces of capital have values which depend on the interest rate, but the interest rate is itself determined by the value of the capital goods.

This means that intuitive things you would expect need not happen. One would expect, as the interest rate falls, that companies would invest into capital which produce more, at the expense of spending more time to produce them. But this need not be the case. Techniques might be optimal at a high interest rate, become sub-optimal as the interest rate falls, and then become optimal again when it falls still lower. To borrow Paul Samuelson’s example (1966), suppose one is making champagne. Seven units of labor are applied in one period to make one unit brandy, which the next period turns into one unit champagne. That is one production method. In the other, two units labor make one unit grapejuice in the first period, which then turns into wine in the next period, which is shaken by six people to turn into champagne. (This is a rather fanciful production function). If interest rates are extremely high, then you cannot abide waiting three periods in order to get a return. You’d rather lent it elsewhere. As it falls – specifically, interest rates between 50 and 100% per period – you switch over to preferring the three period process. However, as it falls still further, what matters is only the total amount of labor spent, and you naturally prefer 7 to 8.

Franklin Fisher, who, while taking the English side of the debate, was an economist at MIT, gave the conditions needed for aggregating together capital into one, well-ordered stock. It can be summarized in one word as “no”. Every single firm in the economy needs to have a marginal rate of substitution which does not vary across different types of capital. You can’t have some things work well with lots of labor, and other things work better with only a little labor.

And yet, we still use aggregate production functions. The conclusion of the mainstream world was roughly as follows: yes, it is impossible to rigorously construct an aggregate stock of capital except under the most special of circumstances. That notwithstanding, it is the least of your concerns. You’re already going to have a devil of a time simply measuring capital. The possibility of reswitching is simply a theoretical curio rather than a meaningful critique. As Solow said, “If God had meant there to be more than two factors of production, He would have made it easier for us to draw three-dimensional diagrams”.

I think that this practically minded counterargument is correct. So why did the English economics stick so doggedly to the story?

The English were Marxists. This is not an allegation of secret affiliations, or even being a fellow traveler – Joan Robinson was an out and out Marxist, notably defending Kim Il-sung and the Cultural Revolution. The point was this – if the firm picks a production function, rather than applying its capital to the commonly known production function, then the rate of return of capital is not some fact of the universe given from on high. The old defense for the return of capital was that it was simply the result of time preference, of patience. If it is something chosen by capital, it is thus in some sense undeserved.

Some sixty years later, we have had a return to Marxism, in the form of some extremely intriguing papers. Acemoglu and Restrepo (2026) is a paper which shook up how I think about automation. Labor and capital work together on a set of tasks which produce output. Labor is able to charge a rent over their outside option. Automation is capital replacing labor on a particular task, and it can be adopted by expending a fixed cost.

There exists variation from task to task in how much workers can collect as a rent, and there also exists variation in productivity across tasks. A social planner would reallocate labor from low productivity, low rent tasks, to ones with high productivity and high rents. Yet, it is precisely those tasks which are targeted for automation. The capital owners adopt technologies which transfer from workers to capital (neutral) after paying a fixed cost.

The empirical prediction of this is that within a group of tasks which might be exposed to automation, wage losses should be much higher for the people who already had high wages. The “groups” here are demographic clusters. This prediction is born out, with the 99th percentile seeing gains likely attributable to management exerting its influence to protect itself.

Acemoglu and Restrepo project that the gains from automation were almost completely offset by inefficient targeting.

More explicitly inspired by Marx, Fukui, Nakamura, and Steinsson (2026) consider what happens if capital can invest into simplifying jobs. If workers have specific skills then it will be harder to find a replacement, allowing them to charge a markup on their wages. A high wage firm (in the tradition of Abowd-Kramarz-Margolis – see my discussion of the union wage premium on the blog earlier) would prefer this not to be the case, and so will invest in standardizing. The result is that the labor share falls and markdowns increase.

Their conclusions are more optimistic than Acemoglu and Restrepo’s. The labor share falling is not a necessary result of commoditizing labor – if everyone does it, then the outside option remains the same. Unemployment also falls, because it is easier for people to find a job. Welfare can go up or down, depending on how much standardization of a job costs. Standardization is not directly measured, but rather inferred from the macroeconomy. It is, nevertheless, a compelling story.

Marxist economics is like Austrian economics today. It carries with it such a connotation of sloppy philosophizing that very few competent people are willing to be associated with it. And yet, it is possible for good work to be done in the tradition of it, without explicitly acknowledging it to be the case.

This did not work out. The empirical evidence, always slight, was emphatically refuted by the randomized controlled trial revolution. There is definite utility to having lenders around, but the marginal impact of adding more was dominated by a cash transfer. This does not mean that credit constraints do not exist. They do, and the decisions of farmers and entrepreneurs in the developing world are substantially distorted away from what is optimal. Neither does this mean that there is no scope for international aid to take forms other than simple cash transfer. There is room for a balanced portfolio of charitable interventions.

This essay will survey the effects of microfinance, why the initial revolution failed, and what we can do about it, including how changes in contractual structure can improve welfare. We will then discuss whether international aid should seek to change people’s actions, and whether common interventions to share risk, internalize externalities, and induce investments are better than cash transfers.

Subscribe now

The explosion of microfinance in the 1990s was in part due to technological innovations on the part of Mohammed Yunus and the Grameen Bank. He had founded the bank in 1976, lending money to the poorest of the poor in Bangladesh, with the motto that credit is a human right. With ineffective rule of law, lenders would normally find the cost of people running away with the money too high. The innovation was giving group loans to neighbors, almost all women, and having everyone be a guarantor of the loan. The neighbors would refuse to agree if they believed their neighbors to be a rotten apple, and they would be able to pressure each other into making a good-faith effort to pay it back. This basic model started spreading around the world – to Bolivia with the Banco-Sol, to Indonesia with the Bank Rakyat and the Bank Kredit Desa, and to India, where a patchwork of moneylenders developed in the 2000s.

International aid being not a hand out, but a hand up, was very popular with donors. In 1997, the Microcredit Summit, headlined by Hillary Clinton and other notables, started a drive to raise $20 billion and reach 100 million people by 2005. The peak of the movement’s influence was in 2006, when the Nobel Peace Prize was awarded to Mohammed Yunus and the Grameen bank. At that time, they had served 7 million people, and disbursed 6 billion dollars.

The Grameen Bank charged interest for their loans, and did have a high rate of repayment, but it still ran at a loss. It needed support from international donors for continued operation. So why support it? One theory is that the poor people are in a poverty trap. Suppose that people produce output by physical exertion. When one is famished, they are not capable of working as hard as they could, and produce just enough to survive. When they are well-nourished, then they can produce much more output. For this to be a poverty trap, there must be a region where adding a bit more food has only a small gain, less than the cost of buying more food, but with a big enough investment, you could get to the higher equilibrium of producing while well-nourished. Or for an alternative story, suppose that borrowing is impossible, there exist large and indivisible productive investments, but that if one accumulates savings they will face unbearable pressure to share from impecunious relatives. Borrowing allows you to shift to an asset which is more difficult for others to take. (I am going to defer a fuller discussion of this to an article next week, which will be linked here once that is complete).

To complete the story for the intervention of outsiders, we need some way for lending to be suboptimally low. This is where Stiglitz-Weiss (1981) (and in the same line of logic, Mankiw (1986)) come in. Suppose that borrowers possess private information about the riskiness of a project, or perhaps whether they intend to run away with the money. When the project fails and they are unable to pay back the loan, the lender is limited in how much they can get back. Raising the interest rate causes the people with safer investments to exit, forcing the interest rate to be even higher; in some cases, there will be no interest rate which is able to clear the market. The role of the donor is to eat the loss long enough to get people to jump out of the bad equilibria. If credit does exist, it will be rationed, and people will be prevented from buying as much as they wanted at the prices they want. This part definitely exists, and it is striking how much microfinance is viewed not as a reduction in the price of what you borrow, but of the amount which you can borrow.

The evidence base for microfinance being effective was always extremely thin. Morduch (1999), in an otherwise hopeful article on the development of microfinance, could not help but note the paucity of the available evidence that it was actually doing anything. The best evidence cited at the time for the Grameen Bank was Pitt and Khandker (1998), who used an eligibility requirement to infer the effect of the loans. However, they skip over the obvious approach in favor of considerably more complicated estimators. Simply running standard estimators, as Morduch and Roodman (2012) do, does not replicate the findings. Pitt and Khandker responded, but it’s almost beside the point – if you have $20 billion in aid money riding on it, you don’t want its utility to depend on involved arguments about what the correct estimator to use on a single dataset is. The comparison is to simply give people money – an action which we now know has substantial multipliers, and at the very least can’t harm anyone.

So by the later part of the 2000s, a movement to test microcredit with randomized controlled trials developed. A randomized controlled trial (an RCT) is one where the treatment – in this case, access to microfinance – has been randomized, and data is collected on the control group who was never offered the treatment. What randomization does for you is get rid of selection. The people who are likely to seek out loans may systematically differ from the population at large, and simply controlling for the things you can observe is unlikely to fix. For example, imagine that people who discover a profitable idea would pay for it out of savings without microfinance, but if loans are available will seek a loan. Thus, you would observe that people who go on to have higher incomes are more likely to obtain loans, without the loans having had any actual causal influence on the outcomes. With randomization, you would observe that there is no difference in investment or income between the groups, and thus correctly infer that the offer of a loan had no effect.

Subscribe now

Things came to a head in a special issue of the AEJ:Applied in 2015, which featured six randomized controlled trials in Morocco, Bosnia-Herzegovina, Mexico, Mongolia, Ethiopia, and India. These trials were deliberately designed to be similar enough to aggregate together, and Rafe Meager (2019), conducting a meta-analysis on these and also Karlan and Zinman (2011) found that the average effects were simply nothing. The RCTs, even in very different contexts and with different methods, had similar null effects. The only light for microfinance was that Meager (2022), interested in if there were different impacts among subgroups, found that households with prior business experience were overrepresented in the tail of outcomes.

There are two ways to conduct an RCT on microfinance, randomizing at the individual level and randomizing at the village level. These have different tradeoffs. Randomizing at the village level means that you capture the interpersonal spillovers which individual randomization would miss. For instance, if people live in households, then it might not matter if one member of the family was not offered if somebody else got a loan and shared it. In a more negative sense, some people receiving money in a village and consuming more might increase the prices of goods, and reduce the consumption of others.

Randomizing at the village level captures the general equilibrium spillovers from person to person, but it is possible that the control group is contaminated. Lending is offered for profit. It stands to reason that profit-seeking firms will choose to locate themselves where the researchers are not competing. This would bias the effect toward zero, as untreated groups are secretly treated.

An immediate objection is that the RCTs are estimating something different from credit constraints, particularly the studies which are randomized at the individual level. The natural thing to point out is that lending is useful both for investment and mitigating risk, and what actually matters for the latter is the credible promise to be able to borrow. Of course take-up is extremely low for one time offers – not everyone has investment opportunities just waiting to go. What we need is for lenders to stand ready when an investment opportunity does come along. Unfortunately, we were able to test it, and it once again failed.

In 2001, the newly elected Thaksin Shinawatra government announced the Million Baht Initiative. One million baht, about $25,000, was transferred to each of the 77,000 villages in Thailand in order to start a village bank. These banks lent out most of the principal at rates matching prevailing interest rates, and the loans were not systematically looted through default. By all accounts, it was implemented in a fair and evenhanded manner.

The results were puzzling. We might have expected people to use the loans to finance investment, but there was no change in investment. Neither was there any change in the price of credit. What happened is that total consumption rose dollar for dollar with the amount available to the villagers to borrow, which is difficult to reconcile with the lack of change of the price of credit and the low default rate. It’s not even consistent with a cash transfer, because people would save a portion of the transfer and spend the interest.

Kaboski and Townsend (2011) try to fit the facts with a macro model. Each household has declining marginal utility of consumption in a given period, and so would like to consume the same amount in each period. Since they discount the future, though, they will not spend exactly the same amount, but consume more now at the cost of consumption later. Each household faces both permanent income and transitory income shocks in each period, drawn from an unknown distribution, and likewise are stochastically presented with the opportunity to invest into a randomly sized, lumpy investment. They can borrow up to a limit, denoted s, at an interest rate r, and if they are unable to pay back the loan default to a minimum level of consumption c. Lastly, there is measurement error in income, which we will estimate, and there is a return to investment which is plugged in from elsewhere (as there is not enough investment in the sample to tidily identify it).

The objective is to choose a set of parameters which will match the period before the intervention. If our model is correct, then that same set of parameters – properly modified to account for the change in the borrowing limit, and the general upward trend after the East Asian financial crisis of 1997 – should replicate the outcomes afterwards.

We find those parameters with the method of simulated moments, which is an extension of generalized method of moments. A brief aside on what those are. Generalized method of moments works by slowly converging to some set of parameters which match the outcomes in the data, or “moments”. You take a guess, solve, and see how far off you are. How far you are updates what you guess next, and eventually you converge to something which comes close enough to matching. The simulation aspect comes in for the inner loop, where you “solve”. The model which they sketch out does not have a tidy solution. Instead, they simulate what agents would do under the parameters with random shocks many times, and take the average. The results from the inner loop can be compared to the outcomes of interest, and you crawl around finding the values that fit.

To evaluate the counterfactual, they change the borrowing constraint s. Because one million baht was given to each of the villages, regardless of size, the variation in the borrowing constraint is plausibly exogenous to what was going on. Consumption increased equal to the amount of credit that was available in each village, and investment was unchanged, although investment was rare enough in the data that failing to observe a significant increase was not puzzling.

Key to understanding the puzzling result is that ex ante identical households which face different shocks will behave very differently. Those households which are living hand-to-mouth, and face a negative temporary income shock, will use the credit to smooth consumption. Their consumption will naturally rise. The more interesting thing is the consumption of people who have positive income shocks. They do not borrow, yet still benefit from the change in the borrowing limit because they can reduce the stock of precautionary savings which they would have held. Meanwhile, people who are insolvent are actually worse off – raising the borrowing limit leaves them more in debt, and they have to pay the interest in the next period. Finally, people who face investment opportunities may borrow to finance it, with ambiguous effects on consumption. Because of the fixed size of the investments, families might reduce their consumption to add to the new borrowing limit.

I think it’s important to point out the role that the default level of consumption c is playing. If default was impossible, and you could always lose more money, then the expanded borrowing limit would always be strictly better. However, with the bound on consumption, the borrowing limit is essentially how much the lender can hold you to from period to period. A higher borrowing limit essentially reduces how much you can default on the debt.

We can then evaluate what would happen if, instead of a loan program, they had simply transferred a sum of money which raised utility by as much. Earnings are in the model – simply move it up, and simulate. They find that a cash transfer could have produced the same increase in utility for 70% of the cost. And yes, that is fully accounting for the credit offerings being permanent. With the risk of having to pay interest for a long time, they would prefer a smaller cash transfer.

Subscribe now

I would not suggest that lending services are not effective. Breza and Kinnan (2021) study the government of Andhra Pradesh’s abrupt interruption of moneylending services in 2010 on consumption, wages, and employment. The government in Andhra Pradesh claimed that it was for consumer protection, but this was of course nonsense; it was more akin to medieval kings periodically expropriating the moneylenders. Breza and Kinnan are not studying the direct effect of no longer being able to access loans in Andhra Pradesh, which would be hopelessly contaminated with changes over time and have only one observation. Rather, they have data from 25 providers of microfinance across multiple districts. Their exposure to Andhra Pradesh, where borrowers were able to default at will, varied, and so the places that saw bigger losses had to reduce their lending in other districts. From this variation, we can get the effect on wages and consumption.

It was not good! The places which saw reductions in lending activity also saw reductions in wages, earnings, and consumption, with employment showing a null effect. Agricultural wages went down 4%, non-agricultural wages went down 8%, and household consumption went down 5%. These losses did not show up immediately – only in the next year, when the banks could not make the loans they had made the previous year in view of their financial losses. Screwing the moneylenders is good exactly one time, and then you pay for it forever. Also, at a larger scale, Bau and Matray (2023) show that India removing the constraints on access to foreign capital which it had until the early 2000s caused firms which had higher marginal revenue products of capital to greatly increase revenue and investment.

The evaluations thus far are the short-run impact. Unfortunately, we can’t expect for the long-run impacts to be better, and in fact we should expect them to be worse. Buera, Kaboski, and Shin (2021) trace out the incentives, but the basic finding can be found from considering what happens if people draw down their buffer stock savings. With lower savings rates, the real interest rate rises in the long run, and the village accumulates less capital. To be clear, total consumption is sure to go up, and as a corollary, income is actually redistributed to labor through higher wages at the expense of less capital. But the rise in immediate consumption is a sugar high from being able to use one’s savings now, and does not stick around.

Nor have we been able to find much evidence of the poverty traps which microfinance would solve. A poverty trap is not merely “things are bad”. It is very specifically the idea that if you give people a substantial enough amount of money, they will be able to start building up toward a higher equilibrium on their own. We just have not found it. The story of people restricted by calories, for instance? Outside of maybe active famine, it just doesn’t make sense – calories are so astonishingly cheap now. Kraay and McKenzie (2014) survey with a sympathetic eye, and can’t find anything. It goes to show just how starved we are for poverty traps that a paper showing one example in rural Bangladesh gets a trumpeting paper – nevermind that, as Karlan, Raswan and Udry (2026) pointed out a couple weeks ago, none of the other instances of this program showed the same effect, and the one that does it probably just correlated geography.

The RCTs, combined with the structural evaluations of the general equilibrium effects, marked the end of microfinance as a charity darling. If you cannot beat a simple cash transfer, then you have no argument for existing.

Some mysteries remain. How do we reconcile the ineffectiveness of microfinance in the RCTs with the rate of return appearing to be well over the prevailing market interest rate? Just because microfinance was ineffective does not mean that credit constraints do not exist, or that decisions are distorted due to risk. Perhaps we need not to discard the spirit, but change the letter.

The second part of this article will cover whether poverty traps exist, microfinance with different loan structures, insurance for farmers, the utility of public health interventions, subsidizing education, subsidizing migration, and summing up.

Read more

What this calls into question is the assumption that firms and consumers are rational, profit-maximizing agents. Standard practice across a range of models is to assume that firms are making the optimal decisions given information and costs which we cannot observe, and then to infer what those costs must be in order to rationalize the observed behavior. This is believable when optimal behavior is easy to solve for when costs/benefits are known, even if inference is difficult, but difficult to believe when the problem which must be solved by both firm and econometrician is difficult. Airlines are a demonstrable break of this assumption – as I will show, it implicates a large class of problems.

In particular, antitrust authorities will naturally be interested in regulating the airline industry, and preventing the abuse of market power. But what market power? What do we actually know about the conduct of airlines? How can we regulate an industry when not even the participants know entirely what they are doing? And in regulating an industry to minimize distortions in a static model, do we raise the possibility that we might make the existence of airlines unsustainable?

Rationality is probably the most attacked assumption in economics by people from outside. If you talk to businessmen, they do not say “we generate a forecast of demand given uncertainty, and then choose the optimal course of action”, nor do people choose bananas one day and apples the next as part of deliberate calculation. People describe the decisions they make in very broad, heuristic terms.

Economists defend themselves with a somewhat unsatisfactory two part gambit. The first part is to say that what we mean by rationality is far narrower than what you might expect – that it means nothing more than each agent possessing a complete and well-ordered set of preferences. The second part is to defend people being calculating on an “as-if” basis. People may not literally undertake enormous calculations to find how much to spend on groceries every week, but they behave as if they do. Armen Alchian (1950) famously argued for profit maximization as the result of evolution and natural selection – although I should point out that there is no reason to expect that natural selection should lead to profit maximization over survival.

Dealing with uncertainty and imperfection in decision making is absorbed in the model. If we say that firms are uncertain about demand, we assume they either know the distribution from which demand is drawn, or if we want to be more complicated, they possess beliefs about the distribution and learn over time. A firm persistently deviating from what is implied by rationality is a form of model misspecification. If you want to make counterfactual predictions, it is not enough to know that the firm’s beliefs are wrong – you have to know the exact way in which they are wrong.

It is not a good idea to start by presuming that firms are getting it wrong. But in some cases, they are demonstrably wrong. I source what follows from Hortascu, Natan, Parsley, Schwieg, and Williams (2024), who have extraordinary access to the internal decisionmaking processes of a major U.S. airline. (It’s either Delta, American, or United). They consider only the routes along which the carrier has a nonstop monopoly, to avoid having to care about the possible responses of carriers. Even without competition, they find that the airline could raise revenue by 18% simply by having its various departments work together. A brief excursion, though, on the history of airlines.

Airlines were deregulated in 1978, allowing for actual price competition. Before then, prices and routes were regulated by the Civil Aeronautical Board, a body which existed in practice to hold up fares and prevent undercutting. They possessed remarkably broad powers, which they abused – between its formation in 1938 and its dissolution, they granted not one single new route. With firms unable to alter fares, they competed on frequency and amenities, frittering away the supracompetitive rents on half-empty airplanes. (The CAB targeted a load factor of 55%. By comparison, load factors are now around 85% and have been for years).

There were a few entrants before 1978, but they were restricted to running routes within states. Notably, Southwest Airlines ran routes around Texas, often at half the price of routes between states. After deregulation, fares plummeted and many new companies entered as fast as they could.

This is in spite of the fact that airlines are a notoriously bad business for investors. I suggest Warren Buffett’s comments on the airline industry as extremely entertaining reading along these lines – he characterizes airlines as a bottomless pit into which investors shovel money out of misguided optimism. And it is a tough business – one puts out large capital outlays to buy the aircraft, but then has small marginal costs to either enter a new route or to sell an additional ticket. After considering the cost of capital, airlines seem to have actively lost money. I had Claude plot out profits over time. Even in the best times, they’re only getting a 4% margin, and they get obliterated by disasters.

It is partially for this reason that so many airlines have gone bankrupt. I should note, though, that bankruptcy is not as serious as one would think. Employees are governed by the Railway Labor Act, and you cannot renegotiate contracts in the event of a negative shock to demand without bankruptcy. It’s really more of a negotiating device, and all of the major airlines have done it at some point.

As stands now, there are three major carriers, United, American, and Delta, which operate across the country, plus Southwest, which is the largest of the low cost carriers (although it is increasingly becoming like the major carriers). Alaska Airlines is a quarter of the revenue of the major carriers, and then there is a tail of low-cost and ultra-low-cost carriers. The business models differ, where the major carriers focus on offering higher quality service, combined with rewards points through their own credit cards in order to make it harder to switch. The low cost carriers differ by unbundling every component of the flying process, and charging a fee for each of them.

The airline sets the network, then a menu of prices, then estimates demand for each of the fare classes. The network is generally of a hub-and-spoke form, with a few central nodes of the network branching out to many destinations. This has numerous advantages, the biggest of which is that it makes many flights possible. There isn’t enough demand to offer a flight from Boise to whatever the 50 different cities passengers from there want to go to, but there is enough demand to put them on a flight to Denver, and then onwards to whatever their destination is. Hubs also make it easier to repair all planes in a central location, makes it easier to hire crew (because they can guarantee a return to one’s home city), and makes a network more resilient to disruptions. If you recall the 2022 Southwest holiday meltdown, that was on account of Southwest running point-to-point flights. When flights were grounded in one airport, it caused a cascade of cancellations in all of the airports which they were due to fly to. It is my understanding that activist shareholders are seeking to make Southwest like the other airlines, which is why they have recently gotten rid of the unusual features like free checked bags and unassigned seating. Later in the article, we will be able to put explicit numbers on the economies of density from hubbing.

Given a network, optimizing which aircraft go on which routes, the schedules, and crew are solved algorithmically, the details of which I am not particularly interested in. Hortascu, Natan, Parsley, Schwieg, and Williams take the network as given, on the grounds that the adjustment process is slow. The network is chosen as marginal additions to an existing network, rather than a complete redesign. You consider what the likely demand for the flight is from the distance and size of the cities, and what the frequency and quality of other competitors on the route is. As far as I can tell, though, this is done remarkably heuristically. Routes are added as part of large initiatives, not as detailed optimization.

With the network, the pricing department chooses a set of fares. These are buckets for fares to be allocated, which the revenue management department does after forecasting demand. They assume that it is low-willingness-to-pay consumers who arrive first, with willingness-to-pay rising as we get closer to the day of the flight. This corresponds to leisure travelers being much more flexible over both when and where they travel, while business travelers have their itinerary taken out of their hands. As the tickets are sold, the price of the remaining tickets increases, and as we get closer to the date of the flight, the prices are also hiked.

In the simplest case, suppose that we have two classes of tickets, high and low. In that case, we want to be indifferent between selling a ticket now at the lower price, and holding back the ticket to sell at the higher price. It’s just like sliding down a demand curve until marginal cost equals marginal revenue.

With two classes of tickets, this is the optimal rule. But with multiple classes of tickets, finding the optimum across all of them is too complicated to do, so instead airlines use the EMSR-b algorithm (where EMSR stands for Expected Marginal Seat Revenue). You collapse all of the fares above your lowest price into one fare, then optimize the number of seats at the lowest level given your demand forecast. You then repeat this exercise for the next fare class, and so on until you run out of fare classes.

Explicitly estimating demand like this was a really big deal. Smith, Leimkuhler, and Darrow (1992), in an article crowing about the innovations which they introduced at American Airlines, credit it with yielding at least $500 million a year in revenue. And when we trace out load factors – the percentage of seats filled – we see that they grew steadily ever since deregulation began, until holding steady around 80% since the 2000s.

The algorithm itself is not optimal because it collapses all of the information of higher tiers into a single number, but it’s very close to optimal. The far bigger problem is that it does not consider how customers with a higher willingness-to-pay will buy the cheaper tickets if they are available. The model assumes that all of the low demand consumers arrive, and then all of the high demand arrive. If they arrive at the same time at all, then you are unable to see how protecting more seats will raise revenue. If you simply estimate demand like this, you will underestimate the true willingness-to-pay, a problem which the revenue management department partially solves by systematically inflating all of their demand forecasts. Further, HNPSW are able to see that the particular algorithm this company uses does not consider substitution across flights, even their own. If there are two flights two hours apart, then selling a ticket on the earlier flight should affect the measure of demand for the other flight.

Perhaps remarkably, these departments are not in coordination with each other, and indeed barely appear to communicate with each other at all. In an illustrative example, the revenue management department persistently assigns tickets to fare classes which do not exist. This is not disastrous, because the fares will default to a different class, but it says something that a stupid bug lasts for two years without any apparent change.

Far more consequentially, though, the pricing department sets its menu of prices largely without the knowledge of the demand estimates for their own product. Their fares are set based upon what other firms are doing, not what they are capable of doing. Remember, HNPSW have restricted themselves to routes where the airline is the only non-stop carrier, so the fares of other firms on other routes is useful only as a vague heuristic.

HNPSW can do better. They possess all of the information the firm does – including, in something of particular importance for this type of good, all of the customers who considered purchasing but ended up not – but unlike the firm, they’re gonna do this right.

Their model has two types of consumers, business and leisure. They arrive over time according to a Poisson process, choose whether or not to buy at the posted price, and depart. They are assumed to be unaware of the number of remaining seats, which prevents needing to consider the level of inventory. They recover the latent demand using variation people stochastically arriving – if you show up 3 weeks before the flight to buy a ticket, sometimes all the cheap tickets will have been bought, and other times cheap tickets remain – as well as shifts at arbitrary dates of the costs. With the demand estimates, you can resolve for what would have been the optimal prices, and the optimal bins.

The prices are set too low. Many of the fares are set in the inelastic portion of the demand curve – in other words, raising the price by 1% reduces the number of tickets bought by less than one percent. If they listened to the revenue management department and repriced, they could raise profitability by 18%. The revenue management department inflating their demand forecasts above what their algorithms actually found is not sufficient to undo the bias, although if they inflated it even more, they could partially offset the losses.

It is important to note that the changes which could be implemented by the firm raise profits, but they do reduce efficiency. While price discrimination, and dynamic pricing more generally, can raise efficiency – Kevin Williams has earlier work showing that it does do this in the airline market – it does not here. The gains are from a firm realizing the market power which it has, and finally exploiting it. It is possible, of course, that there is something in the model which we are not capturing. Given the importance of credit cards to airlines, they may want to maintain brand presence in the public mind, and not reduce the number of seats sold through higher prices, or there may be unobserved costs to consumer goodwill.

Yet, I find it difficult to believe that that is the correct model, and not a model where the firms just don’t know what their optimum is. How else can you explain the persistent lack of communication? How else can you explain the company giving this truly comprehensive data over to the economists, if they do not intend to use this in their work? Firms have to learn about the conditions which they face. The existence of consultants is proof positive that this is costly and not perfect.

A standard method in industrial organization is to, in the absence of knowledge of the costs facing a decision maker, infer what the costs must have been in order to justify the behavior. If firms are not at their optimal behavior already, however, then we might reasonably be concerned our estimates are not stable, and cannot predict much at all.

Some classic studies like this concern the entry decisions of businesses, including airlines, into different markets. Bresnahan and Reiss (1991) show that it is possible to infer something about the degree of competition in markets as a function of the number of competitors, despite lacking any information on price or quantity. In the American West, there are enormous swathes of isolated towns, with no population for dozens of miles. Each of these is their own market – people are unlikely to drive a hundred miles to get their car repaired. We assume that firms pay a fixed cost to enter, and then produce at constant marginal costs. In order to enter, they must charge a markup sufficient to pay back the fixed cost.

When we look at the ratio of population to competitors, we can observe that the first firm requires only a very small population to enter, and that getting the next firm requires more than the former. Each additional firm requires more people to enter than the firm before, although the rate at which these additional people grow slows down. We can infer that market competition must decrease the markup needed to pay back the fixed cost, and that the degree to which the markup is reduced falls as more firms are added.

In Bresnahan and Reiss, firms are assumed to be identical, both in fixed costs and marginal costs. Steven Berry (1992), studying airlines, keeps the identical marginal costs, but allows the fixed cost to vary depending on which airports they already have a presence at. This is of immense practical importance, because of the hub and spoke structure of most networks. We have some practical difficulties, however, which are best thought of as a sequence of problems. First, we now need to care about the identity of firms. We resolve this by first identifying the number of firms that will enter on average, and then imposing a condition on the order in which firms will enter. (In this case, by profitability). There are unobserved shocks to costs drawn from a type I extreme value distribution, which you will recall from demand estimation. The probability that a given number of firms will enter are the combinations of the integrals of the probabilities of each given firm, with which 26 potential entrants easily gives us millions of terms. So, instead we use simulation. We observe the actual number of firms which entered, so for each guess of the cost parameters, we draw a few hundred random draws and see how accurate they are.

Ciliberto and Tamer (2009) extend firm heterogeneity even further. Now, the firms can have impacts on the variable profits of other airlines. What we give up, when we do this, is a determinate equilibrium. Since everybody competed for people just the same conditional upon entering in Berry, there exists only one number of firms; however, in Ciliberto and Tamer, it could be possible for there to exist multiple numbers of firms which are equilibria that no one can deviate profitably from. Put concretely, maybe two small companies have the same effect on other’s profits as one large company, and which prevails is a matter of unmodeled beliefs.

So what we use are called “moment inequalities”, and they’re becoming a workhorse in some of the more involved areas of industrial organization. An intuitive example might come from grocery stores, and whether putting them close together allows them to be more efficient (Holmes, 2011). If we observe how much business opening a store takes from the other stores you own, then the gains from opening grocery stores close together must have been bigger than this. However, the gains must not be so large that you would open up another store. The true value must lie between these points. In Ciliberto and Tamer, the things we are putting bounds on are parameters which deliver the equilibria that we observe in the data.

All of this is predicated on the idea that firms do not make unprofitable moves. But if we demonstrate that firms are systematically deviating from profit maximization, and may make moves that lose them money, where are we left? Unfortunately, I have been unable to make progress on this question. It introduces radical uncertainty about our estimates, but not any clear bias. It might lead on the one hand to too many firms entering and then losing money, or too few firms entering and leaving money on the table. The only bound that comes to mind is a bankruptcy bound. As Gary Becker (1962) showed, you can get market outcomes without utility maximization, but it’s going to take you a very long time to get there.

Antitrust policy in the airline market has been in the news lately. In 2024, the DoJ blocked the proposed merger of Spirit Airlines with JetBlue. Now Spirit has gone entirely bankrupt, and unlike other restructurings, this one is for real. It’s not coming back.

The argument behind the blocking of the merger was that Spirit was an important “maverick” which held down fares everywhere by threatening to take the bottom of the market, even if they don’t actually enter. There is evidence for this, largely from Southwest. Goolsbee and Syverson (2008) showed that Southwest caused fares to decline once they began serving two airports, even before they began flying the route themselves. For instance, if they fly from Dallas Love Field to Cleveland, then opening a route from Dallas Love Field to Washington Dulles causes fare from Dulles to Cleveland to fall, before they actually enter the market.

I do not have a definite answer to whether the merger should have been allowed or not. I certainly don’t have a general answer to antitrust actions in the airline sector. But this is because no one really does. Not even the airlines know what optimal behavior looks like. How could the government possibly know?

This need for caution is especially strong in the other main area of interest for antitrust authorities, collusion. We broadly think that firms are not setting their prices as low as they could go. With fares and details being public on homogenized goods, the conditions are certainly ripe for it, and we do indeed have substantial circumstantial evidence.

There have been some papers formally testing the “conduct” of the airline market. That bit of jargon, “conduct”, is the way in which firms compete with each other. More technically, it’s a matrix of values that spits out marginal costs given prices, quantities, and the demand curve. If firms are colluding with each other, that means they are pricing as if they placed a weight on the profits of other firms equal to their own. The current method is this: we take the demand curve, the estimation of which we will skip over. Given some model of conduct, the marginal costs must be x, y, and z. The markups are thus price minus marginal cost. You possess some set of instrumental variables which shift the markup, but are uncorrelated with marginal costs. If your model of conduct is wrong, the marginal costs will be correlated with the instrumental variables.

I am not, however, terribly enthused by the prospect of testing conduct in the airline industry. Here’s the problem – conduct is going to be correlated with demand shocks, and this holds under many conditions. Suppose that firms are not able to directly bind themselves to colluding through a contract. Instead, they can do so only with a Green and Porter (1984) style punishment strategy. The idea is simple: if you can’t monitor the output of the company you’re competing with, you’re unsure whether a fall in demand for your product is due to demand falling, or due to the competing company defecting on the agreement. So, whenever demand falls, you go from colluding to the default competition. At the other extreme, when they can perfectly observe competing firm output as in Rotemberg and Saloner (1986), defection occurs during booms, because the gains from cheating are larger. Either way, your conduct is going to be bound up with changes in the demand curve. This definitely wrecks the old-school method of measuring conduct, which is to measure the conduct parameter theta as a multiple of the maximum price possible. Even for modern methods, it implies that we cannot summarize the result of a dynamic game as a single conduct parameter.

The best work takes this seriously, but in doing so they limit what they can say. Ciliberto and Williams (2014) cannot say anything about the level of collusion, but they can say that there is more collusion as firms have more contact with each other across markets. This fall simply out of the Green and Porter story, because being involved across more markets increases the amount that one can be punished for defecting.

There is also circumstantial evidence. For instance, the airlines seem to communicate with their competitors through earnings calls. When everybody talks about the need for “capacity discipline” in a particular market, Aryal, Ciliberto, and Leyden (2022) show that the number of seats offered in the next period falls by 2%. Notably, this only happens when all of the carriers talk about it, and there is no unilateral fall in capacity when they are the only ones doing so.

The mandate of the antitrust authorities is clear. Collusion is illegal under the law, whether or not it involves an explicit agreement. But, in principle, should we always try to restrict collusion? Airlines seem like an example of an “empty core” game, where there could be no stable market. David Oks recently argued this, in an essay which inspired this whole choice of topic. Firms may enter in the expectation of being able to collude later, and if we restrict this, we lose out on entry.

An easy way to see why this might be the case is to observe that later entrants do not enter where demand is low. They enter where demand is high. But how do they know that demand is high? Well, they know it because one of their competitors took on the risk of entering when demand conditions were unknown, and now find themselves charging high prices. Not enough firms will enter, because nobody wants to be the first one to move. (I realize, after consulting with Claude, that I have reinvented Chamley-Gale (1994) from first principles). This is how the Civil Aviation Board was justified. I don’t think that those justifications were born out, but it does suggest that perhaps a little collusion need not be so bad.

The new big thing in airline pricing is using steadily improved algorithms to set prices. There has been much recent work on “algorithmic collusion”, and the pretty robust conclusion is that multiple algorithms will converge on supracompetitive prices, and even “dumb” human players can be forced to go along with the higher prices, if algorithms can change prices faster than humans can (Brown and MacKay, 2026). The Department of Justice is interested in pursuing cases against their use. Should they do so? I don’t know. But with so much uncertainty over what is optimal, and knowing that profits tend to induce entry and innovation in the long run, we should perhaps be less eager to intervene.

I should substantiate the claims about traders in Kenya, lest I be accused of libel. In Kenya, as with much of East Africa, corn is a staple crop. It is sold to the consumer in bulk with the kernels removed from the cob and dried, to be ground into flour when needed (flour does not keep as long as the kernels do), or simply boiled and served in another dish. Farmers do not sell directly to the consumer, but to intermediaries, who then bring the corn to market. Entry is constrained by the high cost of trucks – buying one takes 21 times annual per capita GDP, and simply renting one for a day would set you back $250, or 18% of annual income. Everybody knows everyone else, even when they might visit multiple markets.

Subscribe now

The conditions are certainly ripe for collusion. There are no publicly posted prices, and everything is done via negotiation. Nevertheless, these negotiations are public, and everyone can hear what the prices you offer are. The ability to secretly deviate and undercut your rivals is anathema to collusive cartels. They would prefer everything out in the open. Read Genesove and Mullin (2001) on the sugar cartel – what held it together was not explicit collusion on prices and quantities, but obsessive rules on standardizing the product and making it impossible to give secret discounts.

We face a difficulty, however: estimating “conduct” — which is just the way in which firms are competing against each other, or perhaps colluding — is normally very difficult with observational data. Most studies in industrial organization do not attempt it. Suppose that you have estimated the demand curve, at considerable effort. In order to recover the degree of market power which firms have, one needs the marginal cost of producing an additional unit. The marginal cost and the mode of conduct are not separately identified – for every mode of conduct, I can give you a marginal cost which justifies it. The normal method, then, is to assume the mode of conduct, and spit out the marginal costs and markups that way.

In principle, however, the form of conduct can be identified with an additional exogenous variable. You have one exogenous variable shift the supply curve along the demand curve, which traces out its shape; and the other exogenous variable rotates the demand curve. You might get this from something like the entry of an alternative good which makes demand more elastic now that there is a close substitute available. With the demand curve rotating, firms which are colluding or otherwise charging a markup will increase the amount they supply, while firms charging at marginal cost will see no change in output.

The original implementation, due to Timothy Bresnahan, doesn’t quite work if firms are colluding on a price which is not the joint profit maximizing price. This can be reached quite easily, if the firms are colluding through a Green and Porter (1984) style strategy where they revert back to competition whenever they detect a decrease in their demand. This muddles matters, and biases the conduct parameter to perfect competition.

The modern method is due to Berry and Haile (2014). Rather than estimate a single conduct parameter, you can test particular models of conduct, and see if they deliver the behavior necessary for them to be correct. If your model of conduct is correct, then after subtracting out markups, the marginal costs must be uncorrelated with another set of instrumental variables which you did not use in estimating demand, like demographics. To see it put in action, read Backus, Conlon, and Sinkinson (2021).

Lauren Falcao Bergquist and Michael Dinerstein (2020) run a beautiful series of experiments to systematically pin down everything you would need to know about the markets. First, we are going to estimate passthrough by giving traders a per unit subsidy. There are three groups, corresponding to no subsidy, a low subsidy, and a high subsidy. Only around 20% of the subsidy is passed on to the consumer, and perhaps remarkably, this holds no matter the size of the subsidy, the number of traders in the market, or the degree of market access (proxied for by things like whether the roads are paved).

We can rule out pure price competition (Bertrand competition), which would imply 100% passthrough of the subsidy. We can’t separate out the type of imperfect competition without knowing the demand curve, though. For any given passthrough under a monopoly, increasing the curvature of the demand curve to make it more concave could replicate it under partial competition.

So Bergquist and Dinerstein have a second experiment, where they manipulate the price that consumers face directly. Consumers negotiate a quantity at a given price, then the enumerators of the experiment come along and draw a random subsidy. The consumer is then free to choose a new quantity at the new price (with the seller having agreed in advance to not revoke the agreement and try to renegotiate). Tracing out the demand curve, competition where traders choose the quantities they will bring to market, and then choose price (Cournot competition) would imply a pass through of 46%, comfortably outside the 95% confidence intervals. Instead, the pricing behavior is essentially indistinguishable from a perfect monopoly.

So what can we do about it? The obvious thing would be to encourage entry. Perhaps someone who is not privy to the collusive arrangement would cause the whole thing to break down. We are also uncertain whether the markups represent something necessary to make back the genuinely high fixed costs, or whether they represent excess profits extracted from the consumer.

So Bergquist and Dinerstein test this. They offer traders money to go to new markets. The take up rate – which is altogether low – allows us to infer the fixed cost to entering a new market. Then, we can look at how prices change. Strikingly, if the traders were already familiar with the new entrant, there is no change in prices. It is only when the traders were unfamiliar to the existing traders that prices are impacted.

We know that intermediaries possess considerable market power in Africa. Atkin and Donaldson (2015), in an extraordinary feat of data collection, collect detailed prices on homogenous branded goods as one gets further away from distribution hubs. When prices fall at the port, that doesn’t mean people inland see the gains. It’s getting absorbed by the intermediaries. What Bergquist and Dinerstein are able to do is finally understand why.

We sadly lack a lot of evidence on collusion, as opposed to more generic sources of market power. I know of only one other study credibly measuring conduct – Garima Sharma’s working paper on Indian textile mills, who proposes some heuristic methods of determining collusion along the lines of Bresnahan, and then validates it with the “full-IO” approach.

She is working with data from Indian textile mills, where she has uncovered a startling regularity. Members of the Tirupur Exporters’ Association pay their workers exactly the minimum wage. Not around the minimum wage – because deviations below what is legal are ubiquitous, this is possible – exactly the minimum wage. She argues that this is due to the members of the association exerting pressure on each other to hold the line on the wages. Like with intermediaries in Africa, this is a place where entry is constrained. It is notoriously difficult to hire and fire workers in India – for example, if you have more than 100 employees, you need to get government permission to change hours, wages, or employment status. It is difficult, then, to profitably defect from existing arrangements.

Sharma has a simple test. Under any form of earnest competition, a shock to the demand for a competitor’s products will increase their demand for labor, reducing the number of people that you are able to hire and forcing you to increase wages. If you were previously in a collusive agreement, however, wages rise, but so too does employment. Because these textile plants are generally producing for one single major firm, like Nike or Zara, a change in the demand for those particular brands creates a demand shock that is genuinely idiosyncratic.

Naturally, the firms are colluding with each other. Positive demand shocks increase employment and wages, before the wages eventually settle back down around the minimum wage as the collusive agreement is re-established.

I suspect that this sort of collusion is common. Humans are social creatures, who are capable of sustaining cooperative agreements within their community. Many contracts are not enforced through law, but through social pressure and norms. What breaks it down is “capitalism” – trade, globalization, market access, anonymous profit-seeking by alienated individuals. All this is good.

It leads one to believe that the gains from trade are understated. The normal way to quickly estimate the gains from trade can be borrowed from Arkolakis, Costinot, and Rodriguez-Clare (2012) – all of the standard models, whether Eaton-Kortum (gains from comparative advantage) or Melitz (gains from reallocating to more efficient firms) can be summarized by the share of domestic expenditures, raised to the elasticity of trade with respect to costs. The key to their equivalence is that markups are a constant proportion of marginal costs. Having trade make collusive agreements unsustainable breaks the ACR equivalence. The gains from trade can be much larger than they would appear, because they are really the gains from modernity.

A lot of people dislike this. They believe that college should be strictly ordered by merit. This is bound up with affirmative action and racial preferences – the people on twitter who advocate for this tend to be right-wing and think there should be more Asians and Whites instead of Black people in elite colleges – but there is a genuine belief that college admissions should be more objective. For instance, the SAT should not be top-coded, but should instead ask harder questions to distinguish between students; and admissions committees should have much less power to consider soft criteria from hard evidence of academic capability.

Read more