The China Shock’s Wage and Living-Standards Findings Did Not Survive a Revised Design

The “China shock” became a widely used explanation for a painful change in the American economy: the loss of manufacturing jobs and the decline of communities that depended on them. Economic historian Joseph Francis argues that the evidence behind that explanation is weaker than its influence suggests. A 2022 correction to the best-known research design cut its estimated effect on manufacturing employment roughly in half. More consequentially, Francis says, applying the correction to the paper’s other results largely erased its reported effects on wages and living standards.

The research by David Autor, David Dorn, and Gordon Hanson—often called ADH—asked whether rising Chinese imports caused losses in American manufacturing, rather than simply occurring at the same time. That is a difficult distinction. Imports into the United States respond to Chinese production and exports, but also to American demand. A rise in imports and a decline in factory employment, by themselves, cannot show which caused what.

ADH aimed to isolate a supply shock: the effects of China’s growing capacity to produce and export, apart from changes in US demand. Instead of relying only on imports into the United States, they used Chinese imports into seven or eight other developed economies, including countries in Europe and Japan, as an instrumental variable. In Francis’s explanation, imports to those countries served as an external measure of China’s growing supply.

The question mattered because several large changes overlapped. Russ Roberts put the scale of the employment decline this way: from 2000 to 2010, the United States lost about six million manufacturing jobs, roughly a third of the total. During that decade, imports from China rose from about $100 billion to $365 billion. The increase was substantial, though Roberts noted that the figures represented about 1 percent and 2.5 percent of GDP, respectively. At the same time, manufacturing was changing through technological progress and productivity growth, which made it possible to produce more with fewer workers. The statistical challenge was to separate China’s contribution from those other forces.

ADH’s 2013 paper reported that Chinese imports reduced manufacturing’s share of employment. It also reported adverse effects on unemployment, wages, household income, and government transfers. Later papers associated Chinese import exposure with outcomes including mortality, marriage, and political behavior. Francis focused his review on the manufacturing-employment result, which he said had received most of the attention, and then asked what happened to the broader results when the research design was corrected.

The correction came in a 2022 paper by Kirill Borusyak, Peter Hull, and Xavier Jaravel. Francis says they found that ADH’s design did not adequately distinguish manufacturing’s long-running decline from the decline specifically caused by Chinese competition. In effect, the original comparisons partly set manufacturing areas against places with little or no manufacturing. But manufacturing employment had been declining in the United States since at least the 1970s. Those comparisons could therefore capture the longer trend, not just differences in exposure to Chinese imports.

Borusyak and coauthors introduced controls that changed the comparison. They compared US commuting zones with similar levels of manufacturing but different exposure to Chinese imports. A commuting zone is a group of places linked by work and travel around an urban center. The revised comparison asked whether places with comparable manufacturing levels experienced different outcomes as their exposure to imports varied.

With those controls, the estimated effect on manufacturing’s share of employment fell from about 0.6 to about 0.3. Francis described the reduction as roughly half. The coefficient measures exposure to Chinese imports, though Francis could not recall its precise units.

That halving did not, by itself, produce a revised number of jobs. Roberts recalled that when he interviewed Autor in 2016, the authors’ claim was that China accounted for about 25 percent of the decline in manufacturing employment from 1990 to 2007, and about 45 percent of the decline from 2000 to 2007. For the latter period, Roberts said, that could correspond to roughly 1.5 million to two million jobs. If the estimate were cut in half, he suggested, the equivalent might be about 700,000 to one million.

Francis cautioned that the percentages Roberts cited were not directly produced by the regressions in the way that translation might imply. The 25 percent and 45 percent figures appeared in ADH’s abstract; Francis said the regressions could not produce them. The 2022 correction changed a raw regression coefficient, not those abstract percentages directly. Roberts’s job figures were therefore an illustration of what a halved estimate might mean, not a revised count established by the correction.

The correction did not make localized losses or their human costs disappear. Roberts noted that disruption concentrated in particular places could be severe even if the nationwide number were smaller than previously reported. Affected workers might not quickly find equally attractive jobs, and disruption in a limited number of communities could have political consequences.

Francis’s larger finding concerns what the 2022 corrections did to ADH’s results on living standards. After examining the online statistical appendix to Borusyak and coauthors’ paper, he says he found that the corrected design largely eliminated the reported effects on wages and other living-standard measures. The table appeared on page 45 of the appendix. In Francis’s account, the main article referred readers to the appendix table but did not draw attention to the result.

The revised design left an estimated effect on manufacturing employment, smaller than ADH’s original estimate. But Francis says the associated findings on wages and living standards did not survive the correction. A finding that import exposure reduced manufacturing employment in some places does not, by itself, establish that Chinese imports lowered living standards more broadly.

As Roberts put it, the consequences of imports cannot be evaluated only by looking at industries and areas facing competition. Cheaper goods may leave resources available for spending elsewhere, and demand in other sectors may support other work. That does not settle how quickly workers adjust or whether particular communities bear lasting costs. It does mean that a local employment estimate and a claim about economy-wide living standards answer different questions.

Francis says he can only speculate about why the appendix result was not more prominent. He points to the incentives of academic publishing, especially pressure on younger researchers to publish in leading journals and the possibility that they may avoid antagonizing senior economists who influence publication. He does not know whether ADH noticed the appendix result. He says they incorporated some of Borusyak and coauthors’ corrections in a later handbook chapter, but it was not clear to him how many they incorporated, and they did not discuss the broader implications for living-standard results.

The revised employment finding does not settle what caused the longer decline in manufacturing employment or what governments should do about the people affected. Roberts and Francis distinguish the possibility of local harm from the broader claim that Chinese imports explain the decline of American manufacturing.

Roberts emphasized that the share of US employment in manufacturing had been falling well before the period associated with the China shock. He argued that falling manufacturing employment should not be confused with falling manufacturing output or proof that the US economy was being hollowed out. Factories may produce more while employing fewer people, as technology and productivity improve. That change can be painful for displaced workers without showing that the economy as a whole is deteriorating.

The causes matter for policy. If a loss is driven by foreign competition, one response might be to protect domestic producers. If it is driven by technology or rising productivity, a policy aimed at imports may not address the cause. Roberts’s point was not that Chinese imports had no effect, but that attributing a broad employment trend to China requires separating it from domestic changes in production and other developments occurring at the same time.

Francis offered a historical interpretation, while marking it as a prior rather than a researched conclusion about this specific case. Countries, he said, tend to industrialize and then deindustrialize as they become wealthier, partly because people no longer want to work in manufacturing. Globalization facilitates that process. If China had not supplied the American market, he suggested, another country—perhaps Japan, Malaysia, or India—might have done so. China may have accelerated the shift, but Francis said he had not investigated that proposition enough to present it as an established finding.

That interpretation does not make the distributional consequences trivial. Francis noted that international trade has regional effects and that economists have long recognized the existence of winners and losers. If a country embraces globalization, he argued, it should consider how to support people and places that lose out. Regional programs are one possibility; the appropriate response is a matter for debate. Treating “China caused it” as a complete explanation, he said, can bypass that debate.

Roberts added that workers who lose manufacturing jobs may face constraints beyond the availability of jobs elsewhere. Some have less education; moving to a place with better prospects can be expensive, particularly when housing makes relocation difficult. Such obstacles help explain why a national adjustment that looks manageable in aggregate can be difficult in particular communities. The question is not only whether the economy can adjust, but who bears the costs and how long adjustment takes.

Francis agreed that the political anger associated with these changes deserves attention, but objected to treating China as the settled explanation. He sees the China shock story as politically useful because it assigns responsibility to a foreign country. Recalling a New York Times column by Nick Kristof from the early 2000s, he said Kristof had argued that American elites used China as a scapegoat for their own shortcomings. Francis’s view is that this pattern continued, with politicians from both parties blaming China.

Roberts warned that a convenient cause can lead to poor policy when other changes are ignored. The issue is not whether trade can hurt particular people; both speakers accepted that it can. It is whether the China-specific evidence supports the scale and scope of the claims made in its name.

The disagreement over the China shock also raises a question about how empirical findings become credible. Roberts described the choices researchers can make when running regressions: they can change controls, try a different instrument, or test another specification. He recalled asking an economist how many regressions he had run. The question was partly rhetorical, he said, but intended to point to a practical problem: researchers may keep testing variations until they find a result they believe they can defend or publish.

Roberts argued that when clear, dramatic causal results are rewarded, while null or less striking results are less likely to appear, the published record may show only part of the work. He referred to the “file drawer” or “garbage can” problem: results that do not look compelling may be set aside, leaving readers with a more decisive-looking literature than the full set of attempts would justify.

Francis connects the issue to what economists call the credibility revolution: a shift toward empirical work and econometric techniques intended to establish causal effects. He recalls an earlier generation of researchers treating the methods as powerful new tools—“shiny toys”—that could demonstrate that one thing caused another. He believes younger econometricians are now asking harder questions about whether results remain robust when the design is examined closely. The Borusyak, Hull, and Jaravel paper is an example: it changed the comparison and showed how the estimate changed, rather than treating the original result as settled.

Francis’s paper title, Whimsy and Fragility in the China Shock Narrative, refers to economist Edward Leamer’s critique of econometrics. Francis describes fragility as what happens when changing a few choices makes a result disappear or shift. He uses “whimsy” for assumptions built into research that may not withstand scrutiny. His analogy is to an old house: remove a small patch of plaster and a much larger hole may appear.

His criticism is not that econometrics has no value. He thinks it can answer focused questions, including questions about Chinese imports, imports from other countries, or automation. But a broad account of deindustrialization requires more than one estimate of “X causes Y.” Those smaller empirical questions need to be brought together in a wider analysis that takes account of multiple causes.

Francis said he would use long-run time-series analysis, revisit debates from the 1980s and 1990s, and describe the historical process more fully. He also wants economic history to be informed by theory. In the case of deindustrialization, he said, researchers should consider whether it is an expected consequence of a country becoming wealthier. He acknowledged that he had not himself studied US deindustrialization beyond the China shock paper. His proposed approach was a description of how he would investigate it, not a claim that he had already established its causes.

Roberts agreed that complex changes rarely yield a single, tidy explanation. The incentives, he said, favor simple answers, even when social and economic change has many possible causes. Francis hopes that critiques like his can make it more rewarding to test influential findings after publication and to show where apparently robust results depend on choices that were not fully examined.

For Francis, scrutiny after publication is necessary because peer review alone does not guarantee that a paper’s weaknesses will be found before publication. He calls peer review necessary, but says economics often operates as a “black box”: readers see a working paper and, years later, a published article, without knowing what happened in between. Reviewers may object and an editor may publish the paper anyway. Francis says he has heard from people who reviewed articles they considered poor but saw them published regardless.

Francis contrasted this opacity with journals such as Nature, which sometimes publish peer-review reports and sometimes identify reviewers. He also described scientists publicly scrutinizing a questionable paper, finding irregularities, and prompting a journal to issue an expression of concern within a week or two. He does not see a comparable post-publication culture in economics, where published work can be treated as settled.

Economists do make replication packages available, Francis said, but his impression is that few people examine them. When he has raised concerns with an economics journal, he has been directed to Econ Journal Watch, which provides a venue for post-publication criticism. He sees that outlet as both a sign of a weakness in routine scrutiny and a strength: economics has a place where such challenges can be aired.

Francis said he sent a working-paper version of his critique to Autor, Dorn, and Hanson, and that David Autor thanked him for sharing it. He described exchanges with Hull and Borusyak, while distinguishing those interactions from the substance of his article. He recalled that Borusyak criticized a separate statistical package Francis had released; Francis withdrew that package. Francis said he was not aware of an econometric error in his article.

The speakers did not present economics as uniquely unreliable. Francis said different disciplines have different problems. He had heard from a sociologist that sociology may be more cautious about accepting results simply because they pass statistical tests, while facing resistance to making data available. His broader point was that each field’s standards and incentives shape what can be checked and how challenges are received.