The China Syndrome

In 1991 the United States imported $26.3 billion of goods from China; by 2007, $330.0 billion, in 2007 dollars. Over the same years the share of working-age Americans employed in manufacturing kept falling. Autor, Dorn and Hanson asked how much of that decline the import surge caused, by comparing local labor markets that were exposed to China very differently. This lab reruns every regression table of the paper from the authors' file archive in OpenEconometrics and compares 867 printed numbers with the run, under two conventions for clustered standard errors: the SDK's default and the one of the authors' 2013 Stata run.

A surge of imports and 722 local labor markets

Manufacturing employment / working-age populationImports from China / US spending on goods

Chinese import penetration rose tenfold between 1990 and 2007, from 0.47% to 4.68%, while manufacturing employment fell from 13.2% to 8.4% of the working-age population, faster after 2000. A national time series cannot separate China from automation, recessions and everything else that happened in those years.

The paper's unit is the commuting zone: 722 groups of counties with strong commuting ties that together cover the mainland United States. Each is a local labor market, where people live and work and where laid-off workers look for new jobs. Because industries are spread unevenly, the same national surge hit some zones much harder than others.

Import exposure per worker turns that into one number. For each industry, the rise in US imports from China is allocated to commuting zones in proportion to their share of the industry's national employment, summed over industries and divided by the zone's total employment. Equivalently, it is the zone's industry mix times import growth per US worker in each industry. It is measured in $1,000 per worker (2007 dollars) per decade; the seven years 2000–2007 are scaled to ten.

San Jose7.32

Providence4.99

Los Angeles3.59

San Diego3.08

Portland2.96

Washington0.86

New Orleans0.70

Orlando0.59

Ten-year-equivalent change, $1,000 per worker (2007 dollars). The population-weighted mean over all 722 commuting zones is 2.63.

San Jose leads with 7.32 ($1,000 per worker per decade), followed by Providence (4.99) and Los Angeles (3.59); Washington (0.86), New Orleans (0.70) and Orlando (0.59) are at the bottom. Across all zones, weighted by population, exposure rose by 1.14 in 1990–2000 and by 2.63 per decade in 2000–2007.

Why imports alone can mislead

Part of the rise in US imports from China is a supply shock: China's productivity growth, falling trade costs and its opening to world markets. That is the effect the paper wants. But imports also rise when US demand rises. If demand for a product booms, Americans import more of it, and the US plants that make it hire more workers too. Ordinary least squares mixes that positive link with the negative effect of competition and is pulled toward zero; measurement error in exposure pulls the same way.

An instrument from other rich countries

The instrument builds the same exposure from Chinese imports into eight other high-income countries (Australia, Denmark, Finland, Germany, Japan, New Zealand, Spain and Switzerland), with employment shares from ten years earlier, so that they cannot already reflect the imports that were coming. Their imports from China rose from $28.2 billion in 1991 to $262.8 billion in 2007, much like US imports. If the surge was driven by China itself, it shows up in all of these countries at once, and a US demand boom should not raise German or Japanese imports industry by industry. That exclusion restriction cannot be tested: a demand or technology shock shared by all rich countries would break it.

In this long difference the first-stage slope is 0.815 (state-clustered SE 0.092, t = 8.88), the "$815" of the paper's text: each dollar of predicted exposure goes with about 82 cents of actual US exposure. The reduced form, the change in the manufacturing employment share on the instrument, is −0.340 (0.071). In the stacked model with all controls (Table 3, column 6) the first-stage coefficient is 0.631, with a state-clustered first-stage F of 47.64 (97.54 with only the period dummy).

Stacked first differences

The outcome is the ten-year-equivalent change in manufacturing employment as a percentage of the working-age population, so every fixed difference between zones drops out. The two periods, 1990–2000 and 2000–2007, are stacked (722 × 2 = 1,444 observations) with a period dummy. Controls, measured at the start of each period, are the manufacturing share of employment, census division dummies, the shares of college-educated and foreign-born residents, women's employment, the share of employment in routine occupations and an index of offshorability. Each zone is weighted by its start-of-period share of the national population, and standard errors are clustered by state (48 clusters).

The main result

(1) Period dummy only−0.746 (0.068)

(2) + manufacturing share−0.610 (0.094)

(3) + census divisions−0.538 (0.091)

(4) + demographics−0.508 (0.081)

(5) + job tasks−0.562 (0.096)

(6) All controls−0.596 (0.099)

Standard errors in the convention of the 2013 Stata run (oe.ivreghdfe, small=False).

With only the period dummy the coefficient is −0.746 (SE 0.068). Adding the initial manufacturing share gives −0.610; census divisions, demographics and job tasks keep it between −0.51 and −0.56; with all controls, the preferred column 6 gives −0.596 (0.099). A $1,000 rise in a zone's import exposure per worker over a decade lowers its manufacturing employment by about 0.6 percentage points of the working-age population.

Controls	OLS	2SLS	First stage	First-stage F
Period dummy only (column 1)	−0.397 (0.058)	−0.746 (0.068)	0.792	97.5
All controls (column 6)	−0.171 (0.028)	−0.596 (0.099)	0.631	47.6

OLS of the column 6 model gives −0.171 (0.028), so 2SLS is 3.5 times as large; with only the period dummy, OLS gives −0.397 against −0.746. That is the pattern the demand story predicts: US demand shocks raise imports and employment together and pull OLS toward zero.

The paper's own example (footnote 27): from 2000 to 2007 a zone at the 75th percentile of exposure (3.11) saw 1.06 ($1,000) more import growth per worker over the seven years than a zone at the 25th percentile (1.60). With the column 2 coefficient, that predicts a manufacturing share 0.65 points lower.

A placebo and other outcomes

1990–2000−0.89 (0.18)

2000–2007−0.72 (0.06)

1990–2007, stacked−0.75 (0.07)

1970–1980, future exposure0.43 (0.15)

1980–1990, future exposure−0.13 (0.13)

1970–1990, stacked, future exposure0.15 (0.09)

The last three rows regress past changes on future exposure (the placebo).

Estimated separately, the effect is −0.888 in 1990–2000 and −0.718 in 2000–2007. The placebo asks whether the exposure that zones would face after 1990 predicts their manufacturing changes before China's rise. It does not predict a decline: −0.130 for the 1980s, not significant, and +0.431 for the 1970s, when the later-exposed zones if anything gained manufacturing. Stacked, the two decades give 0.148 (SE 0.095, p = 0.12).

Manufacturing employment−0.596 (0.099)

Other employment−0.178 (0.137)

Unemployment0.221 (0.058)

Not in the labor force0.553 (0.150)

Disability insurance (SSDI) receipt0.076 (0.028)

Percentage points of the working-age population per $1,000 of exposure (Table 5, panel B, all controls).

Per $1,000 of exposure, manufacturing employment falls by 0.596 points of the working-age population and other employment by 0.178 (not significant); unemployment rises by 0.221 points, labor-force exit by 0.553 and disability insurance (SSDI) receipt by 0.076. Average weekly wages fall by 0.759 log points, driven by wages outside manufacturing (−0.761). Transfer payments rise by $57.73 per person, and household wage and salary income falls by $549.3 per adult. The working-age population barely moves (−0.050, SE 0.746): people stay, and the local labor market absorbs the shock.

What the paper says it implies

Period	Exposure rise	Predicted fall	Observed fall	Share	× 0.48
1990–2000	$1,140	0.68	2.07	33%	16%
2000–2007	$1,839	1.10	2.00	55%	26%
1990–2007		1.78	4.06	44%	21%

Times the mean rise in exposure, $1,140 per worker in 1990–2000 and $1,839 over 2000–2007, column 6 predicts falls of 0.68 and 1.10 points in the manufacturing share, against observed falls of 2.07 and 2.00: 33%, 55% and, over the whole period, 44% of the decline. Counting only the supply-driven part of import growth, which the paper puts at 0.48, gives 16%, 26% and 21%, the abstract's "one-quarter": about 548,000 manufacturing jobs in the 1990s and 982,000 after 2000, 1.53 million in all (from unrounded inputs the first figure is 549,000).

Two cautions. The 0.48 cannot be rebuilt from the archive: with the paper's own printed inputs its formula (B2) gives 0.513, and the reproduced decomposition gives 0.443. And the benchmark reads relative effects across commuting zones as national totals; effects that hit every zone alike, through prices or national demand, are invisible to this design.

The scorecard

Target	Paper	This lab	Result
Table 3 · col. 6 · 2SLS coefficient	−0.596	−0.5964	✓ Match
Table 3 · col. 6 · SE, 2013 convention (small=False)	(0.099)	0.0988	✓ Match
Table 3 · col. 6 · SE, oe.ivregress (G/(G−1))	(0.099)	0.0998	Difference
Table 3 · col. 1 · 2SLS, 2013 convention	−0.746 (0.068)	−0.7460 (0.0680)	✓ Match
Table 3 · col. 6 · first stage, HC1 SE	0.631 (0.087)	0.6310 (0.0870)	✓ Match
Table 3 · col. 6 · first stage, state-clustered SE	(0.087)	0.0914	Difference
Appendix Table 4 · OLS, all controls	−0.171 (0.028)	−0.1711 (0.0278)	✓ Match
Table 2 · col. 6 · placebo, 2013 convention	0.15 (0.09)	0.1484 (0.0947)	✓ Match
All 867 printed numbers · oe.ivregress	867	677 match · 5 rounding · 185 differ	Difference
All 867 printed numbers · 2013 convention	867	818 match · 8 rounding · 41 differ	Difference
Authors' log · coefficients	214	214 of 214 equal	✓ Match
Authors' log · 2SLS SEs, 2013 convention	183	183 of 183 equal	✓ Match

Group	Numbers	oe.ivregress: match · rounding · differ	2013 convention: match · rounding · differ
Table 1 · trade flows	40	32 · 0 · 8	32 · 0 · 8
Table 2 · periods and placebo	12	11 · 0 · 1	12 · 0 · 0
Table 3 · main results	60	43 · 0 · 17	60 · 0 · 0
Tables 4–9 · other outcomes	237	154 · 3 · 80	231 · 4 · 2
Table 10 · other exposure measures	98	72 · 1 · 25	78 · 3 · 17
Appendix Tables 1–2 · descriptives	274	267 · 0 · 7	271 · 0 · 3
Appendix Tables 3–5 · robustness	117	72 · 0 · 45	107 · 1 · 9
Figure 2 and the text	29	26 · 1 · 2	27 · 0 · 2
All printed numbers	867	677 · 5 · 185	818 · 8 · 41
Authors' logs (values printed)	818	612 · – · 206	791 · – · 27

The run is compared with 867 printed numbers: every coefficient, standard error and printed R² of Tables 2–10 and Appendix Tables 3–5, the descriptive tables, Figure 2 and 23 numbers in the text. With the SDK default, oe.ivregress, 677 match at the printed precision, 5 differ only in rounding and 185 differ. With the covariance convention of the authors' 2013 Stata run, 818 match, 8 differ only in rounding and 41 differ. Against the authors' logs, which print about seven significant digits, all 214 compared coefficients are equal.

Two conventions for clustered standard errors. With few clusters, a cluster-robust variance is often multiplied by G/(G−1). oe.ivregress does so by default, 48/47 with 48 states; the 2013 Stata run applied no finite-sample factor to its 2SLS standard errors. Every oe.ivregress standard error is therefore larger by exactly √(48/47) = 1.01058, checked for all 2,263 coefficients: 0.0998 instead of 0.0988 in column 6. oe.ivreghdfe(..., small=False) uses the 2013 convention and reproduces all 183 2SLS standard errors in the log to every printed digit. Of the 185 differences under the SDK default, 141 are standard errors and t statistics diagnosed as covariance conventions, including the first-stage ones below; the convention also changes two significance stars. Neither convention is wrong, but a replication has to say which one it uses.

The first-stage standard errors. The notes of Table 3 say that standard errors are clustered by state. For the column 6 first stage the paper prints 0.631 (0.087); clustered by state, as the notes say, the standard error is 0.0914 (t = 6.90). The printed value comes from Stata's ivregress ..., first display, which reports 1,398 "clusters" and prints .0870084 in the log; no variable in the data has that many groups. The heteroskedasticity-robust (HC1) standard error, 0.0870, reproduces every printed first-stage standard error of Table 3, so the paper's first stages are not clustered by state. The coefficient is unaffected, and the instrument is strong either way.

What still differs under the 2013 convention, each kept with a diagnosis:

• 31 from data vintage. The archive's rest-of-world import series reproduces the authors' log, $905.8 billion of US imports in 1991, but not the paper's Table 1, column 5, which prints $322.4 billion; Appendix Table 4, column 5 shows the same gap (exposure mean 9.04 against a printed 2.73). In Table 10, panel B, the archive and the log give −0.415 (SE 0.053) where the paper prints −0.51 (0.08). The published versions were built from data that is not in the archive.
• 9 slips or inconsistencies in the paper: "Forth Worth, TX", a standard deviation copied from the next row, three last digits (7.8, 0.150 and 0.260), three statistics in the notes of Table 10, and the 0.48 supply share.
• 1 unexplained: βₑ = −0.029 in the theory appendix has no code in the archive; the closest regression gives −0.120.

None of these touches Table 3, which matches in full.

Where this replication stops

The archive ships finished commuting-zone files. Exposure combines industry trade data with County Business Patterns employment by industry and commuting zone; the raw County Business Patterns files are not in the archive, so the lab starts from the authors' workfiles and cannot rebuild exposure itself. The same missing employment shares rule out modern shift-share inference (Adão, Kolesár and Morales; Borusyak, Hull and Jaravel), which allows for zones with similar industries sharing shocks wherever they are; the standard errors here are the paper's state-clustered ones. The SDK's effective F and Anderson–Rubin set take no analytic weights, so they are not run on these weighted models.

The paper's OLS with lagged-employment exposure (−0.273), the regression behind βₑ and the industry regressions of footnote 24 have no code or data in the archive, and Figure 1 plots published series, whose data are exported but not re-estimated. Finally, the national benchmark rests on reading relative effects as aggregate ones, which no regression across places can test.

Run it yourself

Watch the companion video above. The lab package will be published here. You will need the authors' file archive, which we never redistribute:

1. Run prepare_sources.py from the lab package. It downloads the 2.95 MB file archive from David Dorn's data page and checks the archive and the 33 files the lab uses against their recorded SHA-256 fingerprints.
2. Open a Python document in OpenEconometrics, point the launcher at the lab, the authors' files and a new results folder, and press Run.

from lab import run_lab

result = run_lab(
    source_root=source_dir,
    output_dir=results_dir,
    display_callback=display,
)

Each 2SLS model is one call, here Table 3, column 6; the lab runs it again with oe.ivreghdfe(..., absorb=["_const"], small=False) for the 2013 convention:

fit = oe.ivregress(data=cz, y="d_sh_empl_mfg", endog=["d_tradeusch_pw"], instruments=["d_tradeotch_pw_lag"], x=FULL, weights="timepwt48", weight_type="aweight", cluster="statefip")

The run writes reconciliation.csv, with every printed number, the authors' log value and the lab's result under both conventions.

The paper

David H. Autor, David Dorn and Gordon H. Hanson, "The China Syndrome: Local Labor Market Effects of Import Competition in the United States", American Economic Review 103(6), 2121–2168, 2013. doi:10.1257/aer.103.6.2121. Data and code: file archive [P3] on David Dorn's data page; no licence is attached, and the authors ask users to cite the paper as the data source. The paper, the authors' data and code keep their own rights; this page and the lab describe and recompute them.