Jeff Horwich serves as interlocutor in “Interview with Ariel Pakes: To understand the broader economy, unlock the inner workings of product markets” (Federal Reserve Bank of Minneapolis, September 25, 2026). Pakes is an industrial organization economist, so he is congenitally suspicious of macroeconomics aggregates and broad statements about “how markets work.” Instead, he wants to examine actual markets.
One of his most prominent papers, written with Steven Berry and James Levinsohn, was called “Automobile Prices in Market Equilibrium,” and published in Econometrica in 1995 (63(4), 841–890). The problem the authors faced was that if you want to think about how demand for one kind of car will be affected by a change in price, you also need to think about how demand for all other cars will be affected: for example, are there lots of close substitutes for the car where the price is rising, or not very many? But with 200 models of cars, this problem quickly becomes unwieldy. Here’s Pakes:
As an IO [industrial organization] economist, I have to analyze a market. In the auto market, there were 200 models. If I was to do a demand system—let’s say, how many people want to buy an Oldsmobile Ciera—you’d need the prices of all the goods. Just for the demand for the Olds Ciera, you’d need 200 price coefficients at least, even if you just did it linearly. That’s what we call, in IO, “product space.” The only thing you know is the products and their prices. And that’s how demand systems used to be analyzed. If you did that, I’d have Olds Ciera quantity on one side, 200 prices on the other, and a similar equation for every one of these 200 products. That’s 40,000 coefficients. There’s no data that could effectively estimate 40,000 coefficients.
What we did in the demand systems—the BLP [Berry, Levinsohn, Pakes] group—is we said, If we know the characteristics of the products, and the distribution of preferences over those characteristics—like you have a preference for car size that depends on your family size and price that depends on your income—then all I really need is the distribution of people’s preferences, some of which I can get from age and things like that. If that distribution is, say, normal, and there are 10 characteristics, there’s something like 50 covariance terms. And from that, I can give you 40,000 cross-price elasticities. I just determine what everybody would do if a price changed. I need to sum over households. But with modern computers, that’s easy even for a hundred million people. And then I say, What would happen if I changed the price? I see what you would buy if the prices changed, or what everybody else would buy, and I find out the price elasticity.
Working in “characteristic space” like this has two advantages. One, it allows you to estimate demand in a market with many products. Two, it allows you to do the second question in IO, which is, What products do I want to develop? If all I have is past price and quantity, I can’t tell you what would happen if I put out a new product. But if I know the characteristics of a new product, then I get some idea of what the demand for a new product would be. That’s the incentive for the development of new products.
Pakes had a follow-up story about this paper, when General Motors heard about it, that I had not previously heard:
There was a guy named Mustafa Mohatarem in charge of research who had a Ph.D. in economics and was a very smart guy. He asked us to come to General Motors headquarters and present on the basic stuff. It didn’t have any micro data, just aggregate characteristics and quantities demanded. Afterwards he said to us, “I know exactly the markups on every product, up to the options that are on the product. You guys got it almost right. What would happen if I gave you real data?” So, he gave us their micro data. They had, at the time, an annual 64,000-person sample from new purchases, new registrations across the U.S. And then they had another 64,000, asking the people what they liked and what they didn’t like and what their second choice was. So, we got an old copy of that.
Why I brought up the Olds was because one of the questions he asked us is, What would happen if we killed the Olds department of GM? This is because they had Pontiac, they had Chevy, they had many midsized cars or family cars, and they thought they were just cannibalizing each other. They were not expanding. They killed Olds before we got back to them because it took us a while to figure out how to use the micro data. But after we did this, we got a phone call from them asking for the programs. We put the programs together in a way they could use them, sent them to GM, and they never talked to us again about this. They just wanted the programs. I hope they ended up using them.
This idea of thinking about the product market in terms of characteristics of products, not just the products themselves, turns out to have wide applications, and Pakes touches on a number of them in this survey. For example, when calculating inflation, there is a problem with looking at the prices of products where technology is evolving rapidly: for example, the specific laptop computer you bought last year may no be on the shelf, because this year’s model has different characteristics. When calculating price change, one has to make a “hedonic” adjustment for the change in characteristics. Pakes was instrumental in getting the US government to use hedonic measurement for certain areas in the Consumer Price Index measure of inflation.
Or think about the official poverty line, which is a certain dollar amount, adjusted for family size, that rises by the rate of inflation each year. But the characteristics of what a low-income family can actually purchase have changed over time, so that a poverty line based on characteristics of goods and services might look quite a bit different over a few decades from one based on adjusting a dollar amount.
Or think about antitrust. If two large companies want to merge, how should government regulators evaluate the threat to competition? Rather than that looking at meaures like share of total industry sales, perhaps the more relevant measure is about the characteristics of goods that would be produced by the merged firm and the characteristics of goods produced by other firms in the market. Indeed, this approach is now being used by many antitrust regulators.
At the end of the interview, Pakes offers this thought on the interplay of data and theory:
I think it’s the philosophy of science, that you have to put together data and theory to understand the world. Data by itself cannot prove anything. You need the theory that generated the data to have an answer. The theory is not perfect either, but you’ve got to start somewhere. The issue isn’t that we get the answers right. The issue is that we get the answers better than the next best person could get them. The world is too complicated to get them exactly.
