The China Shock

When can trade make people worse off? In short, whenever there are frictions. If people and resources are able to frictionlessly transfer from use to use, then trade always makes us better off. When there are frictions, then trade can create winners and losers, and even create more losses than gains.

The first paper to seriously consider what restrictions need to be made in order for free trade — or really any increase in market size to be worse than otherwise is Brander and Krugman (1983). They consider two firms manufacturing the same product, each with a home country. The firms are engaged in Cournot competition, where they choose the quantity they wish to produce, and then decide where and at what price to sell it. Goods can be sold costlessly within a country, but shipping from one country to another destroys a portion of the good. Under autarky, which we can define as the share destroyed equalling 1, each firm behaves as a monopoly and restricts output to earn profits. If trade barriers fall, each company finds the first units shipped to the other country profitable, and so both companies end up producing less in total. The consumer is better off, but the companies are even worse off for it.

The key is that other firms are not allowed to enter. With free entry, falling trade barriers necessarily improves welfare. (The result is also sensitive to the form of competition – the results flip if you say they pick their price first, and then produce a quantity). Krugman (1981) does not create a scenario where total welfare can decrease, but does show how trade can create winners or losers when there is trade based on factor endowments. Imagine there are several factors of production – say, land labor and capital – and there are also increasing returns within an industry. If two countries are identical in every way, then trade will only be due to increasing returns, and everyone is better off. If trade is due to different resources, then it is possible for some to win and lose – the owner of the scarce factor of production in each country’s income could fall.

In this case, the creation of winners and losers is driven by the inability to frictionlessly transform one factor into another. In models where, to simplify things, there is only one factor of production (labor) like Eaton-Kortum (2002) it is impossible for trade to create winners and losers. In heterogenous firm models, like Melitz (2003), then allowing for labor market frictions can create winners and losers (because the least productive firms go out of business). We can think of this generally – people invest in particular skills and in particular areas, and cannot change what they do costlessly. It is absolutely possible for some people to be made worse off, and not merely in relative terms, but in absolute terms.

But did it? I am going to cover the work of David Autor, David Dorn, and Gordon Hanson, its followup literature, and its reevaluation. Autor, Dorn, and Hanson have several papers on “the China Shock” – the opening of trade with China, largely in manufactured goods. Between 1990 and 2011, China’s share of world manufacturing exports increased from 2 to 16%, much of which was shipped to the United States. At the same time, the level of employment in manufacturing in the United States fell. Was this good? Who was it good for? Was this due to Chinese import competition? ADH find that greater exposure to import competition led to a rise in unemployment, lower labor participation, and reduced wages, while increasing welfare payments; and they attribute one quarter of the decline in US manufacturing employment to the China Shock.

Autor-Dorn-Hanson is the canonical shift-share paper – so canonical that when Borusyak, Hull, and Jaravel gave the now definitive formal mathematical treatment of what shift-share instruments really are, they illustrated all their points with examples taken from it. In ADH, they isolate exogenous variation from the change in relative exposure to imports, so that if China gets really good at manufacturing a particular item, a place where 10% of employment is in that exact product is much more exposed than a place with 0% employment in that product. They use commuting zones as their unit of analysis, which are defined as areas which share one labor market as seen in commuting data, and use the shares of employment ten years prior (to account for the possibility that companies might begin winding down in anticipation of increased competition).

You cannot just naively use the growth in exports, however. Suppose that there is a positive demand shock for baby carriages in the United States, on account of an increase in the birth rate. What China produces is influenced by these demand shock, which also benefit the workers in the United States. You could conceivably get the sign completely reversed! What they do instead is use a panel of China’s exports to other countries to get a sense of what they are getting better at producing.

We’re not quite done identifying the effects. It is conceivable that there are demand shocks which are common to all of the countries in the panel. To get around this, they use a gravity equation to estimate what trade flows “should” be. Trade flows increase the closer two countries are together, and the larger they are. This gives them some sense of what trade flows “should” be, and they can use that to eliminate the possibility of correlated demand shocks. Now, it is possible that China wasn’t actually getting more efficient, but that firms in the richer nations simply had coincidental declines in productivity. This is obviously unrealistic; and with that, we’ve identified the paper.

ADH, with Jae Song, also look at the effects from the point of view of specific workers in a separate paper. Here they find effects which persist to the present day. However, the regional reduction in employment goes away over time, as workers change out of manufacturing into service employment. The older people who lost their jobs simply age in place, and don’t get new jobs. Strikingly, people migrate less when a negative shock comes, which is the opposite of what you might intuitively expect.

The main criticism of Autor, Dorn, and Hanson is that there’s something of a bait and switch – perhaps not intentionally by them, but certainly in how it has been interpreted by the public. ADH are, in spite of what you may think, fundamentally not asking whether workers were made worse off overall. Rather, they are asking if they were made worse off relative to other places in the United States. It is possible for everyone to be made better off on the whole, including the people who benefited less.

Acemoglu, Autor, Dorn, Hanson, and Price (2016) take the work of ADH (2013), and try to find the general equilibrium effects. Firms are linked together in complicated ways, so the actual gains and losses can be bigger or smaller than apparent. For example, an input becoming cheaper might lead to job losses in the American company which formerly made that input, but increase employment by even more in the sector which uses that input. AADHP use input-output tables to measure linkages between companies, and find that there were, in fact, regional winners and losers.

Adao, Arkolakis, and Esposito (2025) focus on the gap between reduced form estimates like ADH, and general equilibrium trade models like we’ve covered before on the blog. Why does this arise? GE trade models don’t take into account intermediate goods, thus ruling out linkages. They try to get around this by essentially estimating it twice – first a reduced form estimate of the regions, and then a second reduced form estimate of the effects which regions have on each other. They come to large and substantial estimates on employment.

Wang, Wei, Yu, and Zhu (2018) take AADHP, and change two things. First, they separate out intermediate goods from final goods to see how exposed places are to trade. Second, they use exporter specific information to tie imports to sectors. This second one is the really big one – the approach of AADHP is simply to assume that the allocation of inputs matches the allocation of inputs from other countries, like Germany. While this is error – the direction of the bias is unclear from theory – it turned out that adjusting for this flips the sign of ADH. Even in places which lost manufacturing employment, the increase in service employment outweighed that of manufacturing. There were still winners and losers when it came to changes in wages, but there were no severe unemployment shocks in commuting zones. They make bare what causes the divergence from ADH by using the same specifications, except including two terms to represent the influence of the supply chain. The key assumption needed for this to work is that there was not much migration from region to region, which holds up.

Bloom, Handley, Kurmann, and Luck (2024) fleshes out the story of changing from manufacturing to services, and shows that there was substantial reallocation by firms from manufacturing to services. 40% of the manufacturing job losses were simply changes in what the firm was producing, without an actual loss of employment as a result. They are able to do this since they have better data access than ADH, using the Longitudinal Business Database that the Census bureau maintains. Here though, there was considerable regional heterogeneity, with more of the switching taking place in higher skilled areas of the coasts, and not so much in the heartland. They agree with AADHP that there were actual job losses, especially in manufacturing, but that these were systematically overstated due to the firms changing their business.

I came away unimpressed by some of the criticism of it. Take, for example, Jonathan Rothwell’s “Cutting the Losses”. The thrust of the criticism is that there were different macroeconomic environments playing out differently in different areas, so you should include controls for time. In addition, he argues that some results change with different specifications. However, the referee report from ADH, I think, deals with this pretty well – the core claim, that import competition reduced manufacturing employment, holds up no matter what you throw at it. Rothwell’s paper is quibbling over magnitude, but not really disputing the sign. Adao, Arkolakis, and Esposito from earlier tried out the ADH approach with a ton of specifications changes and found that the result was robust to whatever they threw at it. There is better work.

One naturally wonders how much a role America’s land use restrictions played in worsening the distributional consequences of change, and indeed how much they worsen the distributional consequences of any economic change. Ganong and Shoag (2017) showed how the rising cost of housing stifled regional income convergence. No longer could people move from the poorer countryside to the cities, where even in low-skilled jobs they could earn more. As everyone agrees upon, migration was very low, and it is quite plausible that there would have been a lot more if housing supply were more elastic.

My takeaway from all of this is that trade is good, but it may have distributional consequences. Whether those distributional consequences actually led to some people being made substantially worse off is simply not clear – I do not know enough to confidently say whether WWYZ or AADHP is right. I am inclined to believe that WWYZ is the last word on the complete effects, however. They have the most detailed approach. I am mildly concerned about its non-publication, after seven years – I do not have the time to deeply examine the papers I read, and so it may have some fatal flaw I am not aware of.

The greatest effect of “the China Shock” may not have been on economic outcomes at all, but on the 2016 election. Autor, Dorn, Hanson, and Kaveh Majlesi take their method, and apply it to measures of political polarization, in “Importing Political Polarization?”. Exposed districts which were majority white were more likely to elect a Republican, and have greater rightward political polarization along various measures. Non-white districts were more likely to elect a progressive. In the 2016 election, this swung districts toward Trump.