The China Syndrome: Autor, Dorn and Hanson (2013), replicated and extended to 2024

Between 1992 and 2007, China's share of US goods imports rose from under 5 to over 16 percent. Autor, Dorn and Hanson split the country into 722 local labour markets and asked what the surge did to the places making the same goods, using other rich countries' imports from China to strip out anything driven by US demand. The more exposed a place was, the more manufacturing jobs it lost: about 0.6 percentage points for every $1,000 of import growth per worker. Our extension through 2023 traces the paper's own design forward, one import shock fixed at 2000–2007, the paper's instrument and controls, one regression per year, and finds that the damage to manufacturing employment kept deepening for a decade after the shock window closed, reaching the paper's own headline estimate by 2018 (−0.61 against −0.596, in the paper's units) and still standing at about ninety percent of that peak in 2023.

The paper's main claims

1. A $1,000 per worker rise in decadal import exposure cuts manufacturing employment per working-age adult by 0.596 percentage points.
2. That explains about one quarter of the US manufacturing decline over 1990–2007.
3. It also raises unemployment by 0.221 points and non-participation by 0.553 points.
4. Average weekly wages fall by 0.759 log points.
5. Transfer payments rise by about $58 per capita.

What we were able to do

We used the paper's analysis dataset, workfile_china.dta, containing 722 commuting zones observed over two periods, along with its trade and figure data. The original-period regressions started from the authors' prepared data.

We reimplemented the authors' Stata analysis in Python using pandas and linearmodels, including weighted two-stage least squares and standard errors clustered by state. We checked the estimates against the authors' saved Stata output.

For the extension, we took the paper's 2000–2007 import shock and instrument directly from its workfile and assembled manufacturing employment, unemployment and participation, wages, and transfers for its 722 commuting zones from public Census, BLS, and BEA sources. We harmonized county identifiers across datasets before aggregating them to commuting zones. We also rebuilt the paper's import series from CEPII BACI trade data through 2024 for the national picture in Figure 1.

What we got when we re-ran it

Our re-run reproduces every reported coefficient in Table 3 below at the paper's published precision. Our clustered standard errors come out very slightly larger, agreeing with the published ones to within 0.002, a small-sample degrees-of-freedom convention in the clustering. The central finding holds: regions more exposed to Chinese import competition experienced larger declines in manufacturing employment. Some appendix results were only partly reproduced or were not examined; the Table 3 comparison does not establish that every result in the paper was reproduced. Italicised rows show our estimates, printed at the precision our pipeline reports.

Table 3. Imports from China and Change of Manufacturing Employment in CZs, 1990–2007: 2SLS Estimates

Dependent variable: 10 × annual change in manufacturing emp/working-age pop (in % pts)

Panel I. 1990–2007 stacked first differences

Panel II. 2SLS first stage estimates

Notes: N = 1,444 (722 commuting zones × 2 time periods). All regressions include a constant and a dummy for the 2000–2007 period. First stage estimates in panel II also include the control variables that are indicated in the corresponding columns of panel I. Robust standard errors in parentheses are clustered on state. Models are weighted by start of period CZ share of national population. *** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level. Replication: italicised rows print our re-run's output at the precision our pipeline reports. Every coefficient agrees with the published value at the paper's precision, while our clustered standard errors are larger by up to 0.002, for example 0.1004 against the published 0.099 in column 6, a degrees-of-freedom convention in the clustering. Our pipeline emits the main coefficient and the first-stage coefficient for each column; the control-variable rows, the first-stage standard errors and the R² are shown from the paper only, with no italic counterpart. Our replication output prints coefficients and standard errors but no significance stars, so the stars shown are the paper's.

The paper's coefficient, period by period:

China's WTO entry as the shock

The policy event behind the import surge was a single decision: China's accession to the WTO in December 2001. To ask what entry did to US local labor markets in the long run, we use the paper's design in a fixed-shock form, also used in the follow-up chapter "Places versus People". The shock is the paper's own measure of Chinese import growth per worker over 2000–2007, taken directly from its workfile, instrumented by the paper's eight-country measure over the same window. For each year T from 2001 to 2023 we then regress the change in an outcome since 2000 on that same fixed shock, with the paper's column 6 controls held at their 2000 values, the paper's population weights, and standard errors clustered by state:

Because the regressor never changes, movement in θ over the years is the unfolding effect of one well-defined shock, not a moving exposure measure. The shock is measured in the paper's own Table 3 units, decade-equivalent thousands of dollars of import growth per worker (the seven-year 2000–2007 change scaled by 10/7, averaging about $2,600 per worker), so every θ below reads on the same scale as the paper's headline −0.596; per raw accumulated $1,000 of 2000–2007 import growth, the same numbers are 10/7 larger, a trough of −0.87 and −0.77 in 2023. The first-stage F-statistic is 29.3; the same first-stage regression applies at every horizon.

The long-run effect of the fixed 2000–2007 shock, outcome by outcome

Each cell: 2SLS coefficient on the fixed 2000–2007 import shock, per $1,000 per worker in the paper's units; the outcome is its change since 2000; standard errors in parentheses, clustered by state.

Notes. 697–722 commuting zones per cell: disclosure suppression in the county business data and a Connecticut county-definition change trim the manufacturing sample to 718/717/699/697 across the four columns; the other rows keep all 722. Population-share weights from 2000; the transfers row ends with the 2022 BEA vintage; the unemployment, participation, and wage series are observed in snapshot years only, while manufacturing and transfers are annual. The non-participation measure is a proxy built from labor-force and population counts, and wages and transfers are deflated to 2000 dollars.

The manufacturing row is the signature result, and its message is that the effect of one fixed shock was not static. It starts near zero in 2001, is significant by 2003, and reaches −0.26 by 2007, the end of the shock window. Then it keeps deepening for a full decade after the window closed: −0.36 in 2008, −0.45 by 2012, and a trough of −0.61 in 2018, where the year-by-year effect of the 2000–2007 shock alone crosses the paper's own stacked 1990–2007 headline of −0.596. By 2023 it has eased only to −0.54, about 88 percent of the peak. One well-defined shock did roughly twice its 2007 damage by 2018, and most of that damage is still there two decades after accession. The other rows say how places absorbed it: the adjustment shows up in labor-force withdrawal (+0.33 in 2023, though at a standard error of 0.18 this margin is not statistically distinguishable from zero) rather than in measured unemployment, the real-wage penalty is significant early and fades to nothing by 2023, and transfer receipts rise modestly.

How this relates to later research

Autor, Dorn and Hanson (2021, Brookings Papers on Economic Activity) pushed the paper's realized-import design to 2019 and found the employment gap between more and less exposed places still open, and the authors' follow-up chapter "Places versus People" traces outcomes year by year against a single import shock fixed at 2000–2007, a structure our extension shares, with differences in data and measurement detailed below. Our series agrees with the persistence finding and carries it to 2023: the manufacturing gap kept widening for a decade after the shock window closed, peaked around 2018–2019, and about 88 percent of the peak remains in 2023. A separate literature beginning with Pierce and Schott (2016, American Economic Review) measures the same episode through the removal of tariff uncertainty at accession rather than through realized imports. That design estimates a different effect, so its magnitudes are not directly comparable to our realized-import estimates.

Notes

These are the judgment calls and data limits behind our numbers, so you can decide how much weight each result deserves.

• The significance stars are the paper's, not ours: our pipeline reports coefficients and standard errors and we reprint the paper's stars alongside them.
• The exact numerical agreement reported here applies to Table 3's coefficients; our clustered standard errors differ from the published ones in the third decimal, and some appendix results were only partly reproduced or were not examined.
• The extension measures jobs, wages, and transfers from public series (county business data, local unemployment statistics, wage records, transfer accounts) rather than the census microdata the paper used, while the shock and instrument are the paper's own workfile variables. So the 2007 point, −0.263 on the county-business-data outcome, is not directly comparable to the paper's census-based 2000–2007 estimate of −0.72; the comparison the design supports is the long-run path against the paper's stacked −0.596, which is in the same units.
• County identifiers are harmonized to the original commuting-zone boundaries. Oglala Lakota County, SD, is absent from the labor-force series, affecting one rural commuting zone.
• Since 2017 the county business data are protected with infused noise, which adds year-to-year wobble to manufacturing counts in small commuting zones.
• The extension holds the paper's column 6 controls at their 2000 values at every horizon, where the paper's own stacked design refreshes controls at the start of each sub-period. These baseline controls account for differences in local labor-market characteristics at the start of the extension.
• In “Places versus People,” Autor, Dorn, Hanson, Jones and Setzler (2025) use linked employer–employee records through 2019 to follow workers across jobs and locations and separate changes due to migration, aging, and sector shifts. Our public aggregate series extend local outcomes through 2023 but cannot trace individual workers. A fading local average wage penalty therefore does not establish that the workers initially exposed recovered their earnings. Both studies hold the 2000–2007 shock fixed, but the chapter measures import growth relative to 1991 domestic industry absorption and reports effects per standard deviation of exposure. We retain the 2013 paper's imports-per-worker measure and report effects per $1,000 in its decade-equivalent units. The chapter examines both employment-count changes divided by initial working-age population and changes in employment-to-population ratios, separating the contributions of employment and population changes. Our extension reports changes in the ratios without that decomposition. The shared timing supports comparison of the adjustment paths, but the coefficients are not directly interchangeable.
• The fixed shock comes directly from the paper's workfile; the BACI series through 2024 supplies the national picture in Figure 1.

Data sources for the extension studies

The extension studies pulled fresh public data on August 23 and September 22, 2026. In full:

• CEPII, BACI International Trade Database at the Product Level, HS92 revision, version 202601 (1995 to 2024), downloaded from cepii.fr.
• David Dorn's data archive for Autor, Dorn and Hanson (2013): the HS6-to-SIC87dd crosswalk, SIC87dd bilateral trade 1991 to 2014, County Business Patterns commuting-zone employment, and the county-to-commuting-zone crosswalk, downloaded from ddorn.net.
• Autor, Dorn and Hanson, data and code for "On the Persistence of the China Shock" (Brookings Papers on Economic Activity, 2021), ddorn.net, used for validation only.
• U.S. Census Bureau, County Business Patterns, county files for 2000, 2007, 2012, 2019 and 2023, plus the remaining annual county files 1998 to 2022 for the year-by-year manufacturing series.
• U.S. Census Bureau, Population Estimates Program, county population by age and sex (intercensal 2000 to 2010, Vintage 2020, and Vintage 2024).
• U.S. Bureau of Labor Statistics, Local Area Unemployment Statistics, county unemployment and labor force for 2000, 2007, 2012, 2019 and 2023, retrieved via FRED, Federal Reserve Bank of St. Louis.
• U.S. Bureau of Labor Statistics, Local Area Unemployment Statistics, Miami-Dade County unemployment and labor force (series LAUCN12086000000004 and LAUCN12086000000006), retrieved directly from the public BLS API on September 22, 2026.
• U.S. Bureau of Labor Statistics, Quarterly Census of Employment and Wages, annual county files for 2000, 2007, 2012, 2019 and 2023.
• U.S. Bureau of Economic Analysis, CAINC35: Personal Current Transfer Receipts by County (1969 to 2022 vintage).
• Federal Reserve Bank of St. Louis, FRED series PCEPI, MANEMP and CNP16OV (deflator and national context).
• The paper's analysis dataset (workfile_china.dta) and derived tables, reused for the original-period anchors and controls.

The full replication working directory, including the extension studies, is public at github.com/living-sciences/living-china-syndrome.