To Find Relative Earnings Gains After the China Shock, Look Upstream and Outside Manufacturing

Board of Governors of the Federal Reserve System International Finance Discussion Papers ISSN 1073-2500 (Print) ISSN 2767-4509 (Online) Number 1431 January 2026 To Find Relative Earnings Gains After the China Shock, Look Upstream and Outside Manufacturing Justin R. Pierce, Peter K. Schott, and Cristina J. Tello-Trillo Please cite this paper as: Pierce, Justin R., Peter K. Schott, and Cristina J. Tello-Trillo (2026). “To Find Relative EarningsGainsAftertheChinaShock,LookUpstreamandOutsideManufacturing,”International Finance Discussion Papers 1431. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/IFDP.2026.1431. NOTE: International Finance Discussion Papers (IFDPs) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

To Find Relative Earnings Gains After the China Shock, Look Upstream and Outside Manufacturing∗ Justin R. Pierce † Peter K. Schott ‡ Cristina Tello-Trillo § January 7, 2026 Abstract We find that US workers outside manufacturing exhibit relative earnings increases afterUStradeliberalizationwithChina. Theserelativegainscumulateovertimeasthe beneficialeffectofaworker’supstreamexposure—increasedcompetitionfromChinain input markets—more than offsets the detrimental impact of her own and downstream (customer) exposures. These relative gains are smaller for non-manufacturing workers with less ex ante firm tenure and lower initial earnings, and are absent among manufacturing workers due to a lack of upstream gains and stronger downstream losses. Keywords: Trade, Worker Earnings, Uncertainty JEL Codes: F13, F14, F15, F16 ∗We thank Rafael Dix-Carneiro, Gordon Hanson, Brian Kovak, John McLaren, and Mine Senses for detailed comments. Any opinions and conclusions expressed herein are those of the authors and do not necessarily represent the views of the US Census Bureau, the Board of Governors, or its research staff. The Census Bureau’s Disclosure Review Board and Disclosure Avoidance Officers have reviewed this data product for unauthorized disclosure of confidential information and have approved the disclosure avoidance practices applied to this release. This research was conducted under DMS project 7517031. DRB Approval Numbers CBDRB-FY21-CES006-002, CBDRB-FY21- 327, CBDRB-FY22-CES020-001, CBRDB-FY23-CES020-004 and CBDRB FY25-CES020-06. This paper was previously circulated under the title “Trade Liberalization and Labor-Market Gains: Evidence from US Matched Employer-Employee Data.” †Board of Governors of the Federal Reserve System; justin.r.pierce@frb.gov ‡Yale School of Management & CEPR & NBER; peter.schott@yale.edu §U.S. Census Bureau; cristina.j.tello.trillo@census.gov 1

1 Introduction Large literatures in labor economics and international trade investigate the impact of labor demand shocks on worker outcomes across a wide range of economies, including the United States (Jacobson etal.,1993),India(Topalova,2007),Brazil(Kovak,2013;Dix-CarneiroandKovak,2017),andCanada (Kovak and Morrow, 2022). A major area of recent inquiry is the negative reaction of US manufacturing workers to the “China Shock,” driven by US trade liberalization with China in 2000 (Autor et al., 2013; Pierce and Schott, 2016). In this paper, we study how US workers outside manufacturing respond to this change in policy via the exposure of the counties in which they work and the position of their industries in the supply chain. We show that a substantial share of non-manufacturing workers exhibit relative earnings gains as the beneficial effect of upstream exposure (i.e. competition in input markets) more than offsets the detrimental impact of direct and downstream (i.e., customer) exposure. We thus find long-suspected but previously missing empirical evidence of relative labor market benefits of this US trade liberalization.1 Ourapproachtostudyingthisshockrequiresdetailedinformationonworkers’industryand county of employment tracked over a prolonged period of time, which is not typically available, even in other confidential datasets. Toward that end, we make use of the matched employer-employee data from the US Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) program. The LEHD is well-suited to our inquiry for two reasons. First, it tracks the earnings of nearly all workers – manufacturing and non-manufacturing – among US states participating in the program, permitting investigation of outcomes across sectors and counties. Second, workers in the LEHD can be matched to a rich set of personal and professional characteristics via links to other Census datasets, e.g., worker traits in the Decennial Census (DC), plant and firm attributes in the Longitudinal Business Database (LBD), and direct exposure to international trade via the Longitudinal Foreign Trade Transactions Database (LFTTD). Controlling for these attributes allows for cleaner comparisons of worker outcomes than can be achieved at higher levels of aggregation, such as across industries or regions, or with other individual-level datasets. US trade liberalization with China can affect non-manufacturing workers through both industry and spatial channels of exposure and via input-output linkages, and our empirical work is the first to examine these potential channels simultaneously, a contribution that matters crucially to estimated results. While non-manufacturing workers’ industries typically don’t face tariffs on their output, they may benefit if the liberalization leads to a reduction in input prices or productivity growth among their suppliers (Amiti and Konings, 2007; Goldberg et al., 2010; Topalova and Khandelwal, 2011), or lose if it induces difficult-to-replace customers to shrink or exit. Non-manufacturing workers may also lose if displaced manufacturing workers reduce demand for local services, or increase competition for service jobs. Regarding the latter, we find via simple decompositions of worker movement between 2000 and 2007 that about half of US non-manufacturing sectors (representing 41 percent of 2000 US private employment) receive net inflows of former manufacturing workers that account for more than 2 percent of their ex ante employment. 1Jaravel and Sager (2019) highlight benefits to consumers arising from lower prices. 2

WeevaluatetheoverallimpactofamajorUStradeliberalizationvis-a-vis China–thegrantingof Permanent Normal Trade Relations (PNTR) – on non-manufacturing (NM) and manufacturing (M) worker outcomes using a series of worker-level difference-in-differences (DID) regressions. As mentioned, a key innovation of this analysis is that it is the first to focus on workers’ industry and county supply chain exposure to the China Shock. “Own-industry” exposure is derived directly from the US tariff schedule and therefore defined only for workers in goods-producing industries. “Own-county” exposure is a Bartik-style average of the exposures of the industries produced in the worker’s county, using employment shares as weights. For NM workers, own-county exposure captures the spillover effects of being located near directly affected manufacturing workers. To assess workers’ exposure via their supply chain, we use total requirements data from the US input-output table to compute industry and county up- and downstream exposures. Industry up- and downstream exposures measure workers’ sensitivity to their industries’ suppliers and customers. County up- and downstream county exposure, by contrast, measure workers’ more general susceptibility to the exposure of all suppliers and customers of the industries in their counties, again weighted by employment shares. Our analysis yields several novel insights. First, consistent with binding geographic frictions to labor mobility assumed in many spatial general equilibrium trade models (e.g., Caliendo et al., 2019; Ad˜ao et al., 2019), we find that county exposure is more influential than industry exposure in determining worker outcomes for both NM and M workers. Second, we find that exposure along the supply chain has a substantial impact on worker earnings and employment, and that it is asymmetric across NM and M workers. For NM workers, we find large and precisely estimated DID coefficients for county upstream exposure – indicating relative increases in earnings as competition in input markets increases – but that these coefficients are small and not statistically significant for M workers. On the other hand, while estimates for downstream county exposure are negative for workers in both sectors, they are larger in magnitude for M workers. Taking all forms of exposure into account, these estimates indicate that NM workers in the vast majority of county-industry pairs exhibit relative earnings gains of 3 to 29 log points depending on worker tenure at their employing firm. By contrast, M workers experience substantial relative earnings losses. These results suggest that the potential negative spillovers to NM workers from the shock to manufacturing are offset by the positive impact of greater import competition in supplier industries. Our findings provide comprehensive reduced-form empirical evidence of relative earnings benefits arising from increased Chinese import competition in input markets, and reveal that adopting a broaderinput-outputperspectiveiscriticalforunderstandingworkeroutcomesoutsidemanufacturing. While Pierce and Schott (2016) and Acemoglu, Autor, Dorn, Hanson, and Price (2016) include upanddownstreamexposureintheirindustry-levelstudiesoftheimpactofChineseimportcompetition on US manufacturing employment, neither finds evidence of any positive effect.2 Here, we find that a failure to account for input-output linkages both spatially and by industry leads to the underestimation of NM workers’ relative gains as well as under-estimation of M workers’ relative losses. In this way, our findings are more consistent with recent research examining the 2018-19 US-China 2Feenstra and Sasahara (2018) use world input-output tables to calculate the amount of domestic employment supported by US export growth. 3

tariffs, which find that increases in protection negatively affect downstream producers (Flaaen and Pierce, 2019; Bown et al., 2020; Handley et al., 2020; Javorcik et al., 2025).3 Relative to these papers that focused on short-term responses to recently imposed US-China tariffs, our results characterize the implications of supply-chain exposure to trade shocks over a much longer time horizon. They also provide reduced-form support for the input-output linkages highlighted in the theory and calibrated modelcounterfactualsofCaliendoetal.(2019);Ad˜aoetal.(2019). Andlastly,theyhighlightpotential harms to workers in NM sectors that could arise from reviving trade restrictions on goods inputs. One potential explanation for why NM workers receive more help from upstream exposure and arelessharmedbydownstreamexposurethanM workersisanasymmetryinthesesectors’sensitivity to supply-chain disruption. If multiple links of a manufacturing supply chain tend to move offshore together due to correlated shocks or the benefits of remaining co-located, as posited in the theoretical literature (Baldwin and Venables, 2013; Antr`as and Chor, 2013), downstream links may not be able to benefit from greater upstream exposure, and upstream links may be particularly susceptible to higher competition downstream. For NM sectors, such co-offshoring may not be possible, e.g., a hospital must stay near its patients, and a hotel near its guests. Consideration of an “annual” (event-study) version of our baseline specification circumstantially supportsthisinterpretationwhileprovidingnovelestimatesofthedynamiceffectsofexposuretotrade shocks via the supply chain. For workers outside manufacturing, we find that the impact of owncounty exposure is negative and significant immediately after PNTR, suggesting a rapid and negative impact of the shock. On the other hand, this own-county coefficient is relatively small in magnitude and becomes insignificant, while the up- and downstream coefficients become significant several years later. This timing is consistent with NM industries benefiting from lower costs as they find new, potentially foreign sources of inputs in the years after PNTR, as well as their being harmed by the loss of customers, with the former dominating. The estimated coefficients for M workers exhibit a different pattern. There, the immediate negative and significant impact of own-county exposure gives way to negative, statistically significant and large-in-magnitude downstream coefficients around the time of the Great Recession. This pattern suggests that M industries that initially hang on during theirownnegativeexposuremaynotsurvivethelossoftheircustomers,andthatthissurvivalbecame more difficult with the subsequent downturn in manufacturing sparked by the 2007 financial crisis. We also examine how the impact of PNTR varies according to several forms of worker heterogeneity. First, following the mass-layoff literature described below, we investigate the potential effect of firm-specific human capital on worker outcomes by reporting all regression results for two sets of workers: thosewhoareemployedfortheentiretyof1993to1999pre-PNTRperiodbutnotnecessarily by the same firm, and those those employed by the same firm during the entire pre-period. We refer to these groups as “mixed-tenure” and “high-tenure” respectively. We find that relative income gains followingPNTRaregenerallyhigherforhigh-tenureNM workersalongboththeextensiveandintensive margins: they are relatively more likely to remain employed, and they exhibit relatively higher 3Waugh(2019)findsthatChineseretaliatorytariffshadnegativeeffectsonUSmotorvehiclesales. Greenland, Ion, Lopresti, and Schott (2020) show that firms’ reactions to PNTR vary widely within narrow industries, in part due to theiraccesstocheaperinputsfromChina. Aghionetal.(2021)findsimilarheterogeneityamongFrenchfirms’reactions to the China shock. 4

earnings growth conditional on being employed. Such relative gains are smaller for NM workers with lower tenure. Second, we investigate how responses to PNTR vary by workers’ initial characteristics or their firms’ attributes using triple interactions of these traits with the county and industry exposure terms. We find that relative earnings gains are higher among both NM and M workers with ex ante higher earnings. For NM workers, this relationship suggests they possess human capital that boosts their competitiveness ina morecrowded labormarketvis a` vis ex ante lower-paid workers. ForM workers, itmayindicatetheseworkerspossessskillsthataremoreeasilytransferabletootherindustries, areas, or firms. An alternative explanation worth exploring in future research is that such workers may have savings allowing them to be more selective in accepting a new position after the shock. Two other results also stand out. First, we find that manufacturing workers at smaller and nondiversifiedfirms–i.e. thoseengagedsolelyinM orNM activitiesbutnotboth–haverelativelybetter earnings outcomes than workers at firms that are larger or diversified. The former result provides worker-level evidence consistent with Holmes and Stevens (2014)’s hypothesis that small firms may be more likely to produce customized output less substitutable with Chinese imports, while the latter suggests that a focus on manufacturing may contribute to this ability.4 Second, outside M, we find that relative earnings gains are higher among women, whites and younger workers.5 This paper is most closely related to Autor et al. (2014), who use US Social Security Administration earnings data to examine the outcomes of M workers before and after the growth of US imports from China. Here, we use the LEHD to perform a conceptually similar analysis, but one that is focused on workers outside manufacturing. Our analysis differs in two other important respects. First, we consider workers’ exposure along the supply chain, which we find to be a key determinant of both NM and M worker outcomes. Second, we find geographic exposure to be a more important determinant of subsequent earnings than industry exposure, a result that may arise, in part, because our data contain more complete information on a worker’s location of employment. Another paper related to ours is Carballo and Mansfield (2023), who also exploit data from the LEHD to examine how workers are affected by firms’ exposure to PNTR via their direct firmlevel import and export participation. However, while Carballo and Mansfield (2023) find that the negative effects of import competition in manufacturing spill over to other sectors, here we show that workers outside manufacturing generally experience relative earnings gains via the supply chain linkages noted above. We note that our approach – which uses input-output tables to identify greater access to cheaper imported inputs – allows for this channel to affect firms that source inputs from 4ThisresultisconsistentwithevidenceforCanadianfirmsfollowingtheCanadian-USFreeTradeAgreement(Head and Ries, 1999; Kovak and Morrow, 2022) and US firms facing increased import penetration from China Autor, Dorn, and Hanson (2013). 5Kahn, Oldenski, and Park (2022) examine the heterogeneous effects of Chinese import competition and find that Hispanic workers exhibit greater manufacturing employment loss during the China shock, while Autor et al. (2019) focus on differing effects by gender. Conlisk et al. (2022) use data from the Current Population Survey and find differences across gender in terms of labor market outcomes, the college-attendance income premium, and educational attainment decisions. Kamal, Sundaran, and Tello-Trillo (2020) illustrate how import competition results in a decline intheproportionoffemaleemployees,promotions,andearningsatfirmssubjecttotheFamilyandMedicalLeaveAct, compared to firms not subject to this policy. A more general discussion of labor-market adjustment to trade shocks is surveyed in McLaren (2017), McLaren (2022), and Caliendo and Parro (2022). 5

domestic suppliers and wholesalers in addition to direct importing (i.e. being the importer of record themselves). Furthermore, we allow these input-output linkages to have an impact on firms via both industry and geographic dimensions. Our ability to control for worker, firm, and geographic controls, along with both industry- and geographic-level exposure to import competition also sets our work apart from Wang, Wei, Yu, and Zhu (2018)’s commuting-zone-level analysis of supply chain effects of the China Shock, which uses an augmented version of the approach in Autor, Dorn, and Hanson (2013). And while they also find beneficial effects of increased competition in input markets for services industries, our use of annual worker-level data allows us to consider individual-level earnings, track effects over time, and examine the relevance of worker heterogeneity. Our focus on US workers outside manufacturing also relates to the commuting-zone level study of the China Shock in Autor, Dorn, and Hanson (2013) and the worker-level analysis of NAFTA in Hakobyan and McLaren (2016). The former finds that greater own-region exposure to imports from China has no impact on non-manufacturing employment but exerts a negative impact on wages. Hakobyan and McLaren (2016) document a decline in wages of 8 percentage points among lesseducated NM and M workers in US industries and regions with greater exposure to NAFTA. Here, considering both spatial and industry exposure, as well as supply chain linkages, we find relative earnings gains among NM workers after PNTR, but that these relative gains are less robust for workers with less firm tenure and who have lower ex ante earnings.6 Our characterization of worker earnings and employment before and after PNTR relates to the broader literature investigating the short- and long-run consequences of “mass layoffs,” typically defined as separation by workers with three to six years tenure from an establishment shedding 30 percent or more of its labor force within a year. Papers in this line of research – e.g., Podgursky and Swaim (1987); Jacobson, LaLonde, and Sullivan (1993); Stevens (1997); Sullivan and Wachter (2009) – have documented earnings drops of 30 to 40 percent upon displacement before staging a modest but often incomplete recovery in the subsequent decade. Here, we provide context for such large declines in earnings among displaced manufacturing workers using a plausibly exogenous shock to US trade policy as an alternative approach to identifying “mass layoffs.” Finally, our results offer insight into recent research suggesting regional responses to import competition vary according to relative endowments (Bloom et al., 2019; Eriksson et al., 2019). Bloom et al. (2019) find that higher import competition boosts services and total employment in highhuman-capital commuting zones, with no gains to overall employment in manufacturing-intensive low-human-capital regions. While we also find that workers – particularly NM workers – in some geographic areas benefit from increased import competition, we identify a mechanism for this observation that operates through input-output linkages. We also find that for the workers that are 6Several papers examine the effect of trade liberalization on manufacturing workers outside the United States, e.g., in Brazil (Dix-Carneiro, 2014; Krishna et al., 2014), Denmark (Utar, 2018) and Canada (Kovak and Morrow, 2022). KellerandUtar(2023)andKellerandUtar(2022)investigatetheimpactofimportcompetitioninDenmarkonworker polarization and gender inequality, andDeng, Krishna, Senses, and Stegmaier (2021) study differences inhow industry versus occupational exposure to import competition affect German workers’ income risk. Focusing on a major trade de-liberalization – the collapse of the Finnish-Soviet bilateral trade agreement – Costinot, Sarvima¨ki, and Vogel (2022) find scarring effects on both employment and wages, while also considering industry- and geography-level exposure to the trade shock. 6

(relatively) harmed by PNTR, the negative effects on earnings are long-lived, persisting through the end of our sample period in 2014, consistent with findings in Autor, Dorn, and Hanson (2021). The remainder of the paper proceeds as follows. Section 2 summarizes the matched employeremployee data we use, and Section 3 describes the trade liberalization we study. Section 4 describes our empirical strategy and presents our main results. Section 5 describes heterogeneous outcomes by worker attributes, and Section 6 concludes. 2 US Employer-Employee Data We examine the relationship between US worker outcomes and exposure to PNTR using longitudinally linked employer-employee data from the US Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) program, created as part of the Local Employment Dynamics federal-state partnership. The earnings and employment data are derived from state unemployment insurance (UI) records and the Quarterly Census of Employment and Wages (QCEW). In each quarter in each state, firms subject to state UI laws submit the earnings of their employees to their UI program, where earnings are defined as the sum of gross wages, salaries, bonuses and tips.7 States match the firm identifiers in these records to the QCEW, which contains information about where the firms are located and their industries of activity, and pass these data to the US Census Bureau. Census adds information about workers’ age, gender, race, birth country and educational attainmentderivedfromseveralsources,includingtheDecennialCensus. Thisinformationiscollected in the LEHD’s Individual Characteristics File (ICF).8 Birth country is either US or foreign. Racial categories are White, Black, Asian and Other. Education attainment levels are less than high school, high school or the equivalent, some college, and bachelors degree or higher.9 Censususesseverallevelsoffirmandestablishmentidentifiersacrossvariousdatasets. Firmsinthe LEHDareidentifiedbystateemployeridentificationnumbers(SEINs). ConcordancesbetweenSEINs and Census’ other identifiers allow us to match workers in the LEHD to a firm in the Longitudinal Business Database (LBD), which tracks employment and other attributes of virtually all privately owned businesses in the United States. Via the LBD, we are able to measure the size, age, and multi-unit status of a worker’s firm. In any given year a worker may be employed by more than one firm. We adopt the convention among LEHD researchers of assigning each worker in each year to the firm at which the worker’s earnings are highest. Firms can have multiple establishments, and these establishments can have different six-digit NAICS industry codes and be located in different counties within the state.10 We 7AsdiscussedingreaterdetailinAbowd,Stephens,Vilhuber,Andersson,McKinney,Roemer,andWoodcock(2009) and Vilhuber and McKinney (2014), state UI records cover approximately 96 percent of all private sector employees as well as the employees of state and local governments. Prime exceptions are agriculture, self-employed individuals and some parts of the public sector, in particular federal, military, and postal workers. 8Workers in the LEHD are identified via anonymous longitudinal person identifiers (PIKs) which have a one-to-one correspondencewiththeirsocialsecuritynumbersandwhichareusedtoidentifyworkersinarangeofCensusdatasets. Except for Minnesota, UI records do not contain any information about firms except their identifier. 9Notethateducationalattainmentisimputedforthevastmajority(92percent)ofPIKsintheLEHD.SeeVilhuber and McKinney (2014) for more details. 10We use the updated “FK” NAICS industry identifiers provided by Fort and Klimek (2016). 7

assign workers to establishments within the firm (and, thereby industries and counties) using the firm-establishment imputation in the LEHD’s Unit-to-Worker (U2W) file. As illustrated in Appendix Figure A.1, the number of states for which data are available in the LEHD varies over time. For the descriptive results on workers’ industry switching, in Section 4.3 (Table 3, Figure 5 and 6), we use information from the 46 states whose data are in the LEHD starting in 2000.11 In all of our regression analysis, we rely on data from the 19 states for which information is available over our full pre- and post-PNTR sample period (1993–2014).12 To track workers across state lines, we also use information from the 46 states that become available after 2000. This allows us to keep in the regression sample individuals who were observed in one of the 19 states during the pre-period but subsequently moved to one of the other 46 states after 2000. Table 1: 19-State Sample Worker Attributes in 1999 Non-Manufacturing Manufacturing MixedTenure HighTenure MixedTenure HighTenure Female 0.48 0.46 0.30 0.28 (0.50) (0.50) (0.46) (0.45) White 0.84 0.87 0.85 0.87 (0.37) (0.34) (0.36) (0.34) Black 0.10 0.08 0.08 0.07 (0.30) (0.26) (0.27) (0.26) AmericanBorn 0.89 0.89 0.85 0.85 (0.32) (0.32) (0.36) (0.36) LessthanHS 0.11 0.09 0.14 0.13 (0.31) (0.28) (0.35) (0.33) HS 0.28 0.27 0.34 0.34 (0.45) (0.44) (0.47) (0.47) SomeCollege 0.34 0.34 0.33 0.32 (0.47) (0.47) (0.47) (0.47) CollegeorMore 0.28 0.31 0.20 0.21 (0.45) (0.46) (0.40) (0.41) Age 34 37 35 38 (7.8) (6.5) (7.4) (6.2) Earnings($000) 34 47 39 47 (113) (210) (89) (230) Source: LEHD, LBD, and authors’ calculations. Table reports the mean and standard deviationofnotedgroupsofworkersin1999. Samplesare5percentstratifieddrawsfromthe 19stateswhoseinformationisavailableintheLEHDduringthepre-PNTRperiod,1993to 1999. Workersabovetheageof50in2000areomitted. Ageandearningsareinyearsand dollars;allotherattributesaretheshareofobservationsforwhichthenotedattributeistrue. Thenumberofworkersinthousandsineach5percentsampleis789(Mixed-TenureNM),209 (High-TenureNM),194(Mixed-TenureM),and69(High-TenureM). Our regression analysis compares outcomes across workers with “mixed” and “high” firm tenure. High-tenure workers are employed by the same firm during the entire 1993 to 1999 pre-PNTR period. The mixed-tenure sample includes workers that are employed for the entirety of the pre-period, but 11The46-statesamplerepresents96percentofUSoverallandmanufacturingemploymentin2000. Missingfromthe 46-state sample are Alabama, Arkansas, New Hampshire, Mississippi, and the District of Columbia. 12The19statesareAlaska,Arizona,California,Colorado,Florida,Idaho,Illinois,Indiana,Kansas,Louisiana,Maryland, Missouri, Montana, North Carolina, Oregon, Pennsylvania, Washington, Wisconsin, and Wyoming. Together, they represent 47 percent of total US employment and manufacturing employment in 2000. 8

not necessarily by the same firm.13 For computational convenience, we draw representative 5 percent samples from the population of both groups of workers for our regressions. These draws include all workers from “small” counties (i.e., those in the first size decile, with population at or below 5,327 according to the 2000 census), plus a 5 percent random sample of workers from all other, i.e. “large,” counties, stratified according to worker attributes (age, gender, race, ethnicity and educational attainment). All regressions are weighted by the inverse probability of selection into the sample. Finally, we exclude workers who would be older than 64 in 2014 to abstract from normal-age retirement. Within each sample, workers are classified as initially in manufacturing (M) if they are employed inanestablishmentwhosemajoractivityin1999isinNAICSindustriesbeginningwith“3.” Allother workersareclassifiedasinitiallynon-manufacturing(NM). Workersnotpresentinthesampleduring some or all of the post period are classified as not employed (NE) in those years. The predominant reason for NE status is lack of employment – unemployment or labor force exit – but it may also be the result of death, movement to a state (or country) outside the sample of states, for which we don’t have data, or movement to a job that is out of scope of the UI system. We note, however, that workers in our regression sample that move to one of the 46 states available in the LEHD after 2000 remain in the sample and are not classified as NE as a result of such moves. Table 1 summarizes the attributes of the NM and M workers in our high- and mixed-tenure samples as of 1999. Across both groups, high-tenure workers are less likely to be female, more likely to be white, and are more likely to be college educated. They are also older and have higher average earnings. Vis `a vis their M counterparts, NM workers are more likely to be female and college educated. This disproportionate presence of women in NM employment will be relevant when we examine the heterogeneous effects of exposure to import competition by gender. 3 Defining Industry and County Exposure to PNTR The US granting of PNTR to China in October 2000 was unique in that it left assessed tariff rates unchanged, but altered the way US imports from China were considered under the two sets of tariffs that comprise the US Tariff Schedule. The first set of US tariffs, known as NTR tariffs, are applied to goods imported from fellow members of the World Trade Organization (WTO) and are generally, but not uniformly, low due to repeated rounds of trade negotiations during the post-war period. The second set of tariffs, known as non-NTR tariffs, were set by the Smoot-Hawley Tariff Act of 1930 and areoftensubstantiallyhigherthanthecorrespondingNTRrates. Importsfromnon-marketeconomies such as China are by default subject to the higher non-NTR rates, but US law allows the President to grant such countries access to NTR rates on a year-by-year basis subject to annual approval by Congress. US Presidents granted China such a waiver every year starting in 1980, but, as documented in Pierce and Schott (2016), Congressional votes over annual renewal became politically contentious and less certain of passage following various flash points in US -China relations, in particular the 13Because the high-tenure and mixed-tenure samples are drawn independently, in principle, some of the workers in the high-tenure sample could be in the mixed-tenure sample as well. 9

Chinesegovernment’scrackdownonTiananmenSquareprotestsin1989. Asaresult,firmsconsidering engaging in US-China trade prior to PNTR faced the possibility of substantial tariff increases, raising the option value of waiting for a more permanent change in policy (Pierce and Schott, 2016; Handley and Limao, 2017). This uncertainty ended with passage of PNTR, which “locked in” China’s access to NTR tariff rates, eliminating the disincentive to US-China trade caused by the annual renewal process, and effectively liberalizing trade between the two countries. Following Pierce and Schott (2016), we measure industry i’s exposure to PNTR as the rise in US tariffs on Chinese goods that would have occurred in the event of a failed annual renewal of China’s NTR status prior to PNTR’s extension, Gap = NonNTR −NTR . (1) i i i WecomputeNTR andNonNTR for6-digitNAICSindustriesusingsimpleaveragesoftherespective i i Harmonized System (HS) level ad valorem equivalent tariff rates provided by Feenstra, Romalis, and Schott (2002), mapping HS to NAICS using the concordance developed by Pierce and Schott (2012). Wecomputethisgapusingtariffsasof1999,theyearbeforePNTR.AsdiscussedinPierceandSchott (2016), an attractive feature of this measure is its plausible exogeneity to employment outcomes after 2000,as79percentofthevariationintheNTRgapacrossindustriesarisesfromvariationinnon-NTR rates, set 70 years before. This feature of non-NTR rates rules out reverse causality that would arise if NTR rates were set to protect industries experiencing surging imports: To the extent such activity occurred, the higher NTR rates would result in a lower Gap , biasing results away from finding an i effect of the change in policy. We follow Topalova (2007) and Pierce and Schott (2020) in computing a Bartik-style county exposure to PNTR as the employment-weighted average Gap of the industries it produces. Measures i such as these are useful for gauging effects on workers that arise from local labor market shocks. For each US county c, (cid:88) L1990 Gap = ic Gap , (2) c L1990 i i c where the employment shares for 1990 are based on county-industry employment recorded in the US Census Bureau’s Longitudinal Business Database (LBD), which tracks the employment of virtually all US firms and establishments from 1977 to the present.14 In this computation, Gap is positive i only for industries whose outputs are subject to US import tariffs, primarily in the manufacturing sector. For industries whose output is not subject to tariffs, such as service industries, the industry gap is set to zero. The measure of geographic exposure to trade liberalization could also be calculated at a higher level of aggregation. Pierce and Schott (2020) show that measures based on Public Use Microdata Areas—which contain a minimum population of 100,000 and are larger even than 14AnadvantageoftheLBDversusthemorecommonlyusedandpubliclyavailableCountyBusinessPatterns(CBP) for computing county-industry labor shares, e.g., as in Autor, Dorn, and Hanson (2013) and Pierce and Schott (2020), isthatitcontainsemploymentcountsforallindustriesandcounties,therebyavoidingissuesofsuppressiontomaintain confidentialityinthepublicversionoftheCBP(Eckertetal.,2020). Bloom,Handley,Kurmann,andLuck(2019)make use of the LBD for the same reason. 10

Commuting Zones—yield similar effects to those based on counties. Figure 1 displays the kernel densities of Gap and Gap where, for ease of exposition, the former i c is restricted to industries that appear in the US tariff schedule. As a result, the industry-level distributionomitsalargemassatzerorepresentingnon-goodsindustriesthatarenotsubjecttotariffs. Gap has a mean and standard deviation of 33 and 14 percent, while Gap has a mean and standard i c deviation of 7 and 6 percent. Intuitively, the distribution of Gap lies to the left of the distribution c of Gap , reflecting the presence of service industries with NTR gaps of zero. The correlation between i Gap and Gap across workers in our 19-state regression sample is 0.26.15 Interquartile shifts in i c exposure are 20.5 and 7.7 percent for industry and county, respectively. Figure 1: Distribution of Gap and Gap i c Source: LBD, Feenstra et al. (2002), and authors’ calculations. Figure displays the distributions of the 1999 NTR gap across 6digit NAICS industries (Gapi) and US counties (Gapc). The former is restricted to the 473 industries that appear in the US tariffschedule. Trade liberalization episodes such as PNTR may also affect US workers’ earnings via their supply chains, i.e., the upstream industries from which their firms purchase inputs or the downstream industries to which they sell their outputs. Measuring supply-chain exposure to PNTR is especially important for determining effects on NM workers who can face effects of import competition via supply chains, even if their industries are not directly affected by tariffs.16 We compute up- and downstream NTR gaps using total requirements information from the 1997 BEA input-output tables. Gapup is the weighted average of all 6-digit NAICS industries k used by industry i and not sharing i the same 3-digit root as i, using total requirements input-output coefficients (ωup) as weights, ik Gapup = (cid:88) ωupGap . (3) i ik k k 15Autor et al. (2014) report a correlation of 0.12 across workers’ industry (four-digit SIC) and region (commuting zone) exposure to Chinese import penetration. 16A number of recent papers emphasize the importance of examining input-output linkages when estimating the impact of import competition, e.g., Amiti and Konings (2007); Goldberg, Khandelwal, Pavcnik, and Topalova (2010); Pierce and Schott (2016); Acemoglu, Autor, Dorn, Hanson, and Price (2016); Flaaen and Pierce (2019). 11

Gapdown is the analogous weighted average for all the downstream industries outside i’s 3-digit root i that use industry i.17 WecomputeGapup andGapdown bytakingemployment-weightedaveragesofGapup andGapdown, c c i i e.g., Gapup = (cid:88) L1 ic 990 Gapup. (4) c L1990 i i c Upstream exposure is therefore higher when the county has more employment in industries whose upstream industries have higher exposure to PNTR.18 Industries vary intuitively in terms of their up- and downstream gaps.19 Hospitals (NAICS 622), for example, has above-median upstream exposure (0.08) as a result of sourcing from Chemicals (NAICS 325), which includes pharmaceuticals, Plastics and Rubber (NAICS 326), and Miscellaneous Manufactures (NAICS 339), which includes medical devices and scientific equipment. As its sales are mostly to final consumers, it has negligible downstream exposure. General Warehousing and Storage (493110), by contrast, has below-median upstream exposure (0.04) but above-median downstream exposure (0.11), as Chemicals (NAICS 325), Electronics (NAICS 334), and Transport Equipment (NAICS 336) are among its most important customers. Software Publishing (NAICS 511210) is an interesting case in that its up- and downstream exposure are both high (0.08 and 0.26) because it has substantial purchases and sales to Computer and Electronics (NAICS 334). We provide examples of counties with relatively high and low up- and downstream exposure in Appendix Figure A.2. 4 DID Analysis of Workers’ Earnings Response to PNTR We examine the link between PNTR and worker earnings using a generalized OLS difference-indifferences (DID) specification. This approach compares the impact of county versus industry exposure to the change in policy while controlling for initial worker (j), firm (f), industry (i), and county (c) characteristics—a much richer set of controls than possible with more aggregate data—along with worker and time (t) fixed effects, α and α . Our baseline specification interacts our six measures of j t exposure with a dummy variable, Post, which takes a value of one in the years following PNTR, (cid:88) (cid:88) {E > 0 ,LN ,CR } = ϕzPost×Gapz + ϕzPost×Gapz + jfcit jfcit jfcit i i c c z∈own,up,down z∈own,up,down Post×X β +Post×X β +Post×X β +X β + j,1999 1 f,1999 2 i 3 it 4 δ Post×MSH +α +α +(cid:15) . (5) 3 c,1999 j t jfcit As worker earnings may be zero in some years, we consider for our dependent variable three transformations of earnings recommended by Chen and Roth (2023). The primary transformation used 17Weomitup-anddownstreamindustrieswithinthesame3-digitrootgiventheirhighcorrelationwithownexposure. 18The means of Industry Gapup, and Industry Gapdown, County Gapup, and County Gapdown are 11.3, 11.0, 7.5 i i c c and 6.5 percent. Their standard deviations are are 4.3, 8.3, 0.8 and 1.5 percent. Their interquartile ranges are 5.1, 6.6, 1.7 and 1.9 percent. 19AppendixFigureA.2plotsup-versusdownstreamgapsbyindustryandcounty,revealingtheirpositivecorrelation. 12

throughout the paper is referred to as “CR” (for Chen-Roth), which captures the combined impact of the extensive and intensive margins of employment. As described in Chen and Roth (2023), this approach first replaces any zero earnings with the minimum observed earnings in the sample, then dividesearningsbytheminimum, and, finally, takeslogs. Asaresult, CRwillbezeroforobservations whose true value is zero and for those whose true observation is the minimum.20 The second transformation is a dummy for earnings greater than zero, “E>0,” which captures the extensive margin of employment. The third transformation is log earnings, or “LN,” which measures the intensive margin of earnings conditional on employment by dropping zeros. These latter two transformations are used in certain portions of the analysis as a complement to CR and to highlight the role of particular margins of adjustment. As discussed in Section 2, the sample period is 1993 to 2014, and our regression sample is a 5 percent stratified random draw of workers aged 64 or younger from the 19 states for which employeremployee data are available in the 1993 to 1999 pre-period. Also as noted in Section 2, we are able to follow these workers across the 46 states available in the LEHD throughout the post-policy-change period. As is standard in LEHD research, regression observations are weighted by the inverse of the probability of being in the sample. Worker, firm, industry, and county attributes are as of 1999. The first two terms on the right-hand side of the equation are the county and industry DID terms of interest. Industry exposures catch the impact of the policy on the industry in which the worker is employed, which is based on their establishment of employment. County exposures capture the spillover effect of exposure by nearby workers through factors such as labor market competition or changes in aggregate demand. We expect own exposure to have a negative relationship with workerlevelearningsasitcapturesthedirecteffectofincreasedimportcompetitionintheoutputofaworker’s industryorcounty. Weexpectupstreamexposuretohaveapositiveimpactonworkerearningstothe extent that greater openness with China results in lower input prices or greater productivity among input suppliers (Amiti et al., 2014). Downstream exposure, by contrast, is expected to dampen earnings to the extent that it disrupts sales to customers. Our regression specification thus represents a reduced-form approach to assess the kinds of spatial and industry mobility frictions assumed in a range of spatial general equilibrium trade models, e.g., Caliendo et al. (2019); Ad˜ao et al. (2019). Similar specifications are employed in a number of empirical analyses, e.g., Hakobyan and McLaren (2016) and Autor et al. (2014). The next four terms on the right-hand-side of equation 5 are controls for 1999 worker, firm, and industry characteristics interacted with Post as well as controls for time-varying industry characteristics. We multiply the 1999 worker, firm, and industry characteristics – which do not change over time and would be completely absorbed by the worker fixed effects – by the Post dummy to allow for the relationships between these attributes and earnings to change at the same time as PNTR was granted, assisting us in isolating the impact of the policy change.21 20We find results similar to CR(Earnings) using two alternate approaches: the arcsin of earnings and the log of earnings plus 1. 21Initial worker attributes are age, gender, race, foreign-born status and education. Initial firm characteristics are firm-size categories, trading status, and diversification. Trading status is import only, export only, both or neither. Diversification is an indicator for whether or not the firm operates both manufacturing and non-manufacturing establishments. Initial industry characteristics include exposure to reductions in Chinese import tariffs and production 13

The remaining terms on the right-hand side of equation 5 represent the worker and year fixed effects needed to identify the DID terms, as well as an interaction of Post with county c’s 1999 manufacturing share (MSH ). This term, recommended by Borusyak et al. (2021), addresses the c,1999 issueof“incompleteshares”thatariseswhenindustryweightsforcalculatingcounty-levelexposureare based on total employment, rather than only manufacturing industries, as in equation 2. In addition, it controls for the extent to which a county is manufacturing-intensive, allowing us to separate the implications of variation in exposure to the policy change. 4.1 Baseline Estimates of Workers’ Overall Earnings Response to PNTR We start by analyzing the impact of PNTR on workers’ overall earnings using the CR transformation describedintheprevioussection. ResultsarereportedinTable2,wherethefirst4columnsreportthe DIDcoefficientsofinterestforNM workersandthesubsequent4columnsfocusonM workers.22 Even columns report results for the full specification described in equation 5. To highlight the relevance of controllingforexposureviathesupplychain,oddcolumnsdisplayestimatesforasimplerspecification that excludes up- and downstream exposures. Hereafter, we refer to the full equation 5 specification as “Own+IO” and the simpler specification as “Own.” The former is our preferred specification, while the latter serves as a comparator to highlight the importance of controlling for input-output linkages. For the Own+IO specifications, we include a bottom panel that reports two F-tests: one testing the joint significance of the own, upstream, and downstream industry exposure terms, and the other testing the joint significance of the corresponding county exposure terms. Table 2 contains several novel results. First, we find that PNTR affects both NM and M workers primarily through spatial rather than industry exposure. That is, industry exposure terms are jointly statistically insignificant in the even columns for both groups of workers, while the county exposure terms are jointly significant across the board. While this primacy of county over industry exposure among manufacturing workers might seem surprising given the strong negative impact of PNTR on manufacturing industry employment reported in Pierce and Schott (2016), the two findings are consistent. ThecriticaldistinctionisthatPierceandSchott(2016)focusonindustry employment, while our attention here is on worker earnings. As we show below, there is a net outflow of workers from manufacturing to services after PNTR, lowering employment in highly exposed industries. The estimates in Table 2 demonstrate that what matters for workers’ subsequent earnings are the conditions in their county: If the county is highly exposed, workers experience relative earnings declines as labor market competition rises and aggregate demand for local goods and services falls, even if they are initially employed in low-exposure industries.23 By contrast, workers in less-exposed counties – even if initially employed in high-exposure industries – can transition from manufacturing to services with less of an earnings penalty.24 subsidies. Time-varying industry characteristics (X ) capture the elimination of US quotas on textile and clothing it productsaspartofthephasingoutoftheglobalMultifiberArrangement(MFA).ThesevariablesaretakenfromPierce and Schott (2016) and Pierce and Schott (2020); their construction is described in Appendix Section B. 22Appendix Section E reports coefficient estimates and standard errors for other earnings transformations that are reported visually in this Section. 23This relationship is also present in raw data as shown in Figure 6, below. 24We note that it continues to be the case that higher exposure to PNTR is associated with relative declines in 14

Table 2: Response of Overall Earnings (CR Transformation) to PNTR Mixed Tenure NM High Tenure NM Mixed Tenure M High Tenure M (1) (2) (3) (4) (5) (6) (7) (8) Dep Var: CR CR CR CR CR CR CR CR Post x GapOwn . . . . -.06 -.05 .11 .12 Industry . . . . .11 .12 .19 .19 Post x GapDown . -1.72* . -1.04 . -.31 . -.4 Industry . .96 . .98 . .28 . .38 Post x GapUp . 2.24 . 2.47* . .09 . .23 Industry . 1.39 . 1.4 . .75 . 1.21 Post x GapOwn -4.9*** -5.12*** -3.36*** -3.92*** -3.01*** -1.83 -3.06*** -1.36 County .87 1.17 .88 1.11 .63 1.17 .81 1.41 Post x GapDown . -3.21** . -3.92*** . -3.92** . -6.34*** County . 1.49 . 1.47 . 1.52 . 2.1 Post x GapUp . 6.2* . 9.41*** . 1.85 . 1.81 County . 3.55 . 3.62 . 4.4 . 4.9 R-sq .45 .45 .44 .44 .45 .45 .44 .44 Obs (000) 17,360 17,360 4,605 4,605 4,274 4,274 1,520 1,520 Worker FE Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Specification Own Own+IO Own Own+IO Own Own+IO Own Own+IO Cluster N4, Cty N4, Cty N4, Cty N4, Cty N4, Cty N4, Cty N4, Cty N4, Cty Industry Gaps F 1.633 1.657 .472 .399 Industry Gaps p .198 .193 .703 .754 County Gaps F 14.4 8.462 11.7 5.811 County Gaps p 0 0 0 .001 Source: LEHD,LBD,Feenstraetal.(2002),andauthors’calculations. TabledisplaysDIDtermsofinterestfromworker-yearlevelOLS regressionsofequation5forworkersinitiallyemployedinnon-manufacturing(NM)andmanufacturing(M). Thedependentvariable istheChenandRoth(2023)earningstransformation,whichisthelogofearningsafterreplacinganyzeroswiththeminimumobserved earnings. The sample period is 1993 to 2014. Mixed-tenure workers are employed throughout the 1993 to 1999 pre-period, but not necessarilybythesamefirm. High-tenureworkersareemployedbythesamefirmduringthepre-period. Postisadummyvariablefor yearsafter2000. Standarderrorstwo-wayclusteredby4-digitNAICSandcountyarenotedbelowcoefficients. Finaltwopanelsreport F-statistics for the joint significance of the industry and county exposure DID terms, respectively. ***, **, and * represent statistical significanceatthe1,5and10percentlevels. Observing both industry- and geographic-level exposure to trade or policy changes is rare, as it requiresaccesstodetailedworker-leveldata. Amongthissmallsetofresearch,Autoretal.(2014)and Hakobyan and McLaren (2016) find that both dimensions of exposure are associated with negative US labor market outcomes when studying import competition from China and NAFTA, respectively. Costinot et al. (2022) find that the negative effects of worker-level exposure to a reduction in export demand caused by the breakup of a Finland-Soviet bilateral trade agreement are amplified by being employment for more-exposed manufacturing industries. Indeed, we find results in line with Pierce and Schott (2016) when we aggregate our M samples to the industry-year-level and implement a regression as in that paper. 15

located in a more-exposed area. None of these studies, however, considers the type of industry- and geography-level input-output linkages, which we find can contribute to relative earnings gains. The second key message of Table 2 is that accounting for supply chain linkages is important in evaluatingworkers’responsetoalabormarketshock. ForNM workers,wefindpositivecoefficientsfor county upstream exposure and negative estimates for county downstream exposure that are large and statistically significant at conventional levels. These estimates indicate that NM workers’ earnings risewithhigherexposureininputmarketsandfallwithgreaterexposureamongcustomers. Including these channels of exposure is critically important, as they can flip overall estimates of the relationship between exposure to PNTR and NM workers’ earnings from negative to positive. In particular, when only including an NM worker’s own county exposure (the first column under the headings for mixed-tenure NM and high-tenure NM ), we find a negative relationship between NM earnings and exposure to PNTR. Inclusion of input-output terms – specifically accounting for more competition in input markets – uncovers offsetting effects, and as discussed in Section 4.2, the richer specification reveals that PNTR is associated with relative earnings gains for most NM workers. Two aspects of therelevanceofupstreamanddownstreamexposurefoundherearenovelintheliterature. First, they provide detailed empirical evidence supporting the input-output linkages highlighted in the modelbased GE effects of the China Shock in Caliendo et al. (2019); Ad˜ao et al. (2019). Second, while the benefitsofupstreamexposuretotheChinaShockhavebeenlong-suspected, previousresearch(Pierce and Schott, 2016; Acemoglu et al., 2016) has not found such effects because it did not consider the geographic input-output linkages included in our analysis.25 For M workers, county upstream (and own) exposure coefficients are smaller in magnitude and imprecisely estimated relative to those for NM workers, while downstream estimates are large and statisticallysignificantatconventionallevels, particularlyforHigh-TenureM workers, indicatingthat M workers are relatively more susceptible to customer exposure than NM workers.26 One potential explanationforM workers’anemicresponsivenesstoupstreamexposurerelativetoNM workersisan asymmetry in their sensitivity to supply chain disruption. In manufacturing, several links of a supply chain with varying levels of exposure might move offshore together if productivity depends heavily on proximity, as posited in Baldwin and Venables (2013), i.e., less-exposed downstream links may co-offshore with highly exposed upstream links, or vice versa.27 In that case, the former’s upstream exposure affords no benefit, and the latter’s downstream exposure is particularly disruptive.28 For nontradable consumer-facing services like health care and tourism, such co-migration of stages of the 25An exception is Wang et al. (2018) who find benefits of higher import penetration in input markets using the identification strategy from Autor et al. (2013). 26AsdiscussedbelowandillustratedinFigure7,theimpactofeachchannelofexposurevariesovertime. Forexample, the coefficient on own-county exposure is negative and statistically significant in the years immediately after PNTR. 27JohnsonandMoxnes(2023)provideamodelinwhichtradebetweenupstreamanddownstreamindustriesbecomes more spatially concentrated as trade costs fall. 28SupportforthisexplanationcanbefoundintheeconomicgeographyandexistingChinaShockliteratures. Ellison, Glaeser,andKerr(2010)findthatIO-linkedmanufacturingindustriestendtoco-agglomeratewithintheUnitedStates. Pierce and Schott (2016) and Acemoglu et al. (2016) show that US manufacturing plant and industry employment fall with downstream exposure to China but do not rise with upstream exposure, consistent with up- and downstream industries moving offshore in groups, potentially to China. Finally, Long and Zhang (2012) find that manufacturing industries within China become more spatially concentrated, and its regions increasingly specialized, after the China Shock. 16

supply chain may not be feasible to the extent that they must remain near their customers.29 Below, wediscussfurtherevidenceconsistentwiththisco-offshoringhypothesisthatisrevealedbyexamining the timing of effects of the various exposure measures. 4.2 Economic Significance Assessing the economic significance of the DID coefficients via the standard approach—-an interquartile shift in exposure—-is not applicable in our setting: workers are simultaneously exposed along six different dimensions, so shifting exposure one coefficient at a time does capture their potential joint distributions. As a result, we illustrate the economic significance of our estimates by computing predicted relative earnings for all county–industry pairs in our samples using our estimated coefficients and actual county and industry exposures. Specifically, for each county–industry pair, we use the DID estimates in Table 2 to compute (cid:88) (cid:88) PredictedRelativeEarnings Growth = ϕˆzGapz, (6) ic y y y∈{i,c}z∈{o,u,d} where the first summation is over county and industry and the second is over own (o), up- (u) and down- (d) stream exposures.30 Figure 2 reports the distribution of predicted relative earnings growth across NM and M countyindustrypairsundertheOwnandOwn+IO specificationsfortheCRdependentvariable, bysample. The substantial mass of the solid orange curves to the right of zero in the first two panels reveals that the majority of NM county-industry pairs – 64 percent for mixed-tenure and 95 percent for high-tenure – exhibit relative earnings gains when accounting for exposure along the supply chain. The magnitude of these effects is larger for high-tenure sample than the mixed-tenure sample, with their respective averages being 3 and 29 log points (3 and 34 percent). By contrast, the third and fourth panels indicate that M workers in nearly all county-industry pairs experience relative earnings losses. These first two panels also highlight the relevance for NM workers of exposure via input-output links mentioned above. The orientation of the blue curves for the Own specification to the left of zero show that failing to account for input-output relationships – in particular the benefits of exposure via input markets – would erroneously indicate that all NM workers experience relative earnings losses associated with PNTR. For manufacturing workers (the third and fourth panels), by contrast, failing to account for input-output linkages would lead to underestimation of the relative earnings losses 29PNTRmayalsobenefitNM workersbyinducingentryof“factorylessgoodsproducers”likeFitbitandRokuthat take advantage of a greater ability to outsource and offshore the physical transformation stages of goods production (Fort, 2023). While difficult to identify using existing BEA input-output tables, this activity may be reflected in M workerflowsintoWholesale(NAICS42)andProfessionalServices(NAICS54). Wehopetoaddressthischannelofjob creation more directly in future research. 30In principle, one could use the DID coefficients presented in the last section to compute worker-level conditional predicted relative earnings by multiplying each worker’s exposure by the estimated coefficients and taking their sum. However, in addition to it not being possible under Census disclosure guidelines, performing this more disaggregated calculation would not add useful information since each worker within an industry-county cell has the same exposure to PNTR. 17

associated with PNTR, as indicated by the orange curves being shifted left relative to the blue. For these workers, accounting for the large negative effects of downstream exposure reveals additional relative earnings losses. Figure 2: County-Industry Predicted Relative CR Earnings Growth, by Sample Source: LEHD,LBD,Feenstraetal.(2002),andauthors’calculations. Figuredisplayspredictedrelativeearningsgrowthacross county-industry pairs for mixed-tenure and high-tenure NM and M workers implied by the estimated Own versus Own+IO DID coefficients. In each case, predictions are for the CR earnings transformation reported in Table 2. Predictions for each county-industrypairaretheproductofthereportedcoefficientsandactualexposures. Figure 3: Own+IO County-Industry Predicted Relative CR Earnings Growth, by Sample Source: LEHD,LBD,Feenstraetal.(2002),andauthors’calculations. Figuredisplayspredictedrelativeearningsgrowthacross county-industrypairsformixed-tenureandhigh-tenureNM andM workersimpliedbytheestimatedOwn+IODIDcoefficients fortheextensiveandintensiveearningsmarginsreportedinAppendixTablesA.3andA.4. Predictionsforeachcounty-industry pairaretheproductofthereportedcoefficientsandactualexposures. Notethataxesscalesdifferacrossplots. Next,inFigure3,weexaminethemarginsresponsiblefortheestimatedrelativeeffectsonearnings. For M workers, the negative relationship between exposure to PNTR and earnings arises both from a lower probability of remaining employed (top row) and lower earnings conditional on employment (bottom row), as indicated by most of the curves’ mass appearing to the left of zero. The converse 18

is true for high-tenure NM workers; their relative earnings gains are due to both higher probability of remaining employed and higher earnings. For mixed-tenure NM workers, however, their modest gainsareprimarilyfromahigherprobabilityofremainingemployed,whilemostcounty-industrypairs experience relative earnings losses conditional on employment (left panel of bottom row). We explore differences in reactions for mixed-tenure and high-tenure workers in detail in the next sub-section.31 4.3 High-Tenure versus Mixed-Tenure Outcomes and Worker Flows Figure4providesanalternativeviewofthepredictedrelativeearningscomputedinequation6witha directcomparison ofoutcomesforhigh-tenure andmixed-tenureworkers, byearnings transformation. Asthetoprowofthisfiguremakesclear,therelativeearningsgainsofhigh-tenureNM workersexceed those of mixed-tenure NM workers, overall (left column, CR), and along both intensive and extensive margins (center and right columns), as indicated by the solid teal curves being shifted to the right of the dashed ochre curves. Figure 4: Own+IO County-Industry Predicted Relative Earnings Growth, by Sample Source: LEHD, LBD, and authors’ calculations. Figure displays predicted relative earnings growth across county-industrypairsformixed-tenureandhigh-tenureNMandMworkersimpliedbytheestimatedOwn+IO DID coefficients. The first panel in each row reports results for overall earnings using the results for the CR earnings transformation reported in Table 2. Subsequent panels in each row present analogous predictions using the coefficient estimates for the extensive and intensive earnings margins reported in Appendix Tables A.3andA.4. Predictionsforeachcounty-industrypairaretheproductofthereportedcoefficientsandactual exposures. 31GivenUSCensusBureaudisclosureconstraints,wereporttheeconomicsignificanceofourresultsinFigure4using county-industry pairs as a unit of analysis, and treat each county-industry equally. In Figure A.4 of Appendix Section E.1,weshowthatestimatesofeconomicsignificancearequalitativelysimilarwhenthekerneldensitiesareweightedby the number of workers in each county-industry pair. 19

One potential explanation for this difference among mixed-tenure versus high-tenure NM workers istheformer’sgreatersusceptibilitytolabor-marketcompetitionfromdisplacedmanufacturingworkers. Indeed, Table 3, which documents US workers’ transitions among non-manufacturing (NM), manufacturing (M) and non-employment (NE) between 2000 and 2007, reveals a net flow of 1.9 (=5.8-3.9) million workers from M to NM in the years after the US change in trade policy. This flow represents 1.6 percent (=1.9/118.6) of the number of NM workers in 2000. Table 3: Gross Flows to and from Manufacturing, 2000-7 Millions PercentofInitialLevel Sectorin2007 Sectorin2007 Sectorin2000 NM M NE 2000Total NM M NE 2000Total Non-Manufacturing(NM) 85.0 3.9 29.6 118.6 72 3 25 100 Manufacturing(M) 5.8 8.3 4.3 18.3 32 45 23 100 NotEmployed(NE) 42.3 3.2 . 45.6 93 7 . 100 2007Total 133.1 15.4 33.9 182.4 73 8 19 100 Source: LEHD, LBD and authors’ calculations. Table reports the transition paths of employed and not-employed workersfrom2000(row)to2007(column)forthe46stateswhoseinformationisavailableintheLEHDstartingin 2000. Alabama, Arkansas, New Hampshire and Mississippi as well as the District of Columbia are excluded. Left panelreportslevelsinmillionsofworkers. Rightpanelreportssharesofinitiallevels. Figure5providesinformationaboutthetransitionsofformermanufacturingworkersacrosssectors that has not been previously reported and can only be computed with longitudinal worker-level data such as the LEHD. It also speaks to potential changes in labor market competition as workers flow at varying rates into different destination sectors. The left panel of this figure decomposes the 1.9 million M worker outflow according to 2-digit NAICS destination sector. As indicated in the panel, the largest net outflow in terms of number of workers is towards healthcare (NAICS 62), with sizable net flows of approximately 200 thousand workers to several other sectors. In the right panel of Figure 5, these net flows are divided by the ex ante (year 2000) employment of their destination sector, yielding measures of the extent of net inflows relative to sector size, thus providing a more meaningful measure of the extent of potential labor market competition across non-manufacturing sectors. As indicated in this panel, half of the 2-digit non-manufacturing NAICS sectors exhibit net inflows of manufacturing workers of 2 or more percent of their initial level, including Mining (NAICS 21), Transportation (NAICS 48-9), Wholesale (NAICS 32), Utilities (NAICS 22), Construction (NAICS 23) and Administration, Support, and Waste Management (NAICS 56).32 Such increased labormarket competition for NM workers in these destination sectors could hold down earnings growth for agivenleveloflabordemand, particularlyforlow-tenureworkerswhocycleinandoutofemployment more frequently and are thus more likely to be engaged in the job search process. M worker outflows also provide intuition and context for the sharp earnings declines among M workers estimated in Table 2 and displayed in Figures 2 and 3. In Figure 6, we report the “quasi” medianchangeinnominalearningsfrom2000to2007exhibitedbyM workersalongallpossible2-digit 32Transitionstosomeofthesesectors,e.g.,wholesale(NAICS42)andprofessionalservices(NAICS54),areconsistent with workers switching industry but not necessarily occupation (Traiberman, 2019), e.g., an R&D scientist formerly located in a manufacturing plant might move to a research lab. 20

Figure 5: Net Manufacturing Employment Outflow by Transition Path, 2000-7 (46 States) Source: LEHD, LBD and authors’ calculations. Left panel reports 2000 to 2007 net transitions (outflow less inflow) out of manufacturingbyworkers’initial2-digitNAICSsectorinthe46statesforwhichinformationisavailableintheLEHDforthese years (Alabama, Arkansas, New Hampshire, Mississippi and the District of Columbia are excluded). Right panel reports these netflowsdividedbydestinationsectors’initiallevelofemployment. NAICS transition paths.33 The left panel of the figure reveals outcomes for all workers, with earnings growth for workers who remain in manufacturing highlighted in red. As indicated in the figure, initial M workers transitioning to relatively low-skill service industries such as staffing agencies (NAICS 56) or accommodation and food services (NAICS 72) exhibit nominal wage declines of up to 20 percent. Indeed, considering the five most common destination sectors for former manufacturing workers, shown in the left panel of Figure 5, we see that these transitions tend to disproportionately involve weak or even negative earnings growth. Among the five, transitions to two sectors (construction; wholesale) involve earnings gains just above those for continuing manufacturing workers, one (health) entails lower earnings growth, and two (admin, support, waste; education) involve outright nominal earnings declines. These outcomes are consistent with the generally lower wages paid in these sectors and the popular narrative that well-paid manufacturing workers face large drops in income when they move to service sectors (Scott et al., 2022).34 Moreover, the bottom row of Figure 4 indicates that the relative earnings losses for M workers associated with PNTR are, if anything, a bit larger for high-tenure workers due to lower earnings conditional on employment. Lastly,therightpanelofFigure6reportsquasi-medianearningsgrowthamongworkersincounties withthehighestversuslowestquartileofGapOwn . Asshowninthatpanel, earningsgrowthislower County for workers initially employed in high-exposure counties, across all destination sectors. This finding 33Quasi-mediansarebasedonmeansofgroupsofworkersaroundthemedian,asCensusBureaudisclosureavoidance procedures do not allow the reporting of true medians, which are necessarily based on one or two individuals. We caution that the estimates in Figure 1 contain a mix of voluntary and involuntary transitions, and that they may involve movement of select groups of workers. We condition on observed worker attributes in our regression analysis. 34Theseapparentwagedeclinesaredrivenbyworkertransitionsacrosssectors,ratherthandifferentialearningsgrowth over time within the sectors. According to publicly available data from the BLS, summarized in Appendix Figure A.3, theaveragehourlyearningsforproductionandnon-supervisorsinmanufacturing(NAICS3)in2000was$13.80,versus $12.00, $11.30, $10.90 and $8.10 for admin, support, waste (NAICS 56), retail (NAICS 44-5), arts, entertainment and recreation (NAICS 71), and accommodation and food services (NAICS 72). Average hourly wage growth from 2000 to 2007 in these data (which, unlike our LEHD data, do not distinguish between comers and goers), was 19 percent in manufacturing, versus 21, 17, 33 and 25 percent in the other sectors just mentioned, respectively. 21

is consistent with increased labor-market competition from manufacturing workers in high-exposure areas, as well as the scarring effects of job loss documented in Davis and von Wachter (2011) and Huckfeldt (2022). Figure 6: Median Nominal Earnings Growth Among Initial M Workers, by Transition Path (46 States) Source: LEHD, LBD, Feenstra et al. (2002) and authors’ calculations. Figure displays quasi-median 2000 to 2007 growth in nominal earnings across workers moving from manufacturing to the noted 2-digit NAICS sector between 2000 and 2007 in the 46 states for which information is available in the LEHD for these years (Alabama, Arkansas, New Hampshire Mississippi and theDistrictofColumbiaareexcluded). Leftpaneldisplaysgrowthforallworkers. Rightpaneldisplaysquasi-mediangrowthfor workersinthefirst(low)versusfourth(high)quartileofcountyexposuretoPNTR,definedinSection3. 4.4 Impact of Exposure to PNTR Across the Post-PNTR Period Further insight regarding the effects of exposure to PNTR along the supply chain can be obtained by estimating an “annual” DID specification that replaces the Post dummy variables in equation 5 with a full set of year dummies. Results from this regression are reported in Figure 7, with panels for mixed and high-tenure workers for both the NM and M sectors. In this Figure, we plot coefficient estimates for each county (upper panel) and industry (lower panel) exposure term together with their 95 percent confidence intervals. To increase readability, coefficients that are significantly different from zero at the 95 percent level have solid markers, while those that are not statistically significant at that level have hollow markers. As indicated in the figure, own-county exposure is negative and statistically significant for all four groupsof workersin theyears immediatelyafterthe changein UStradepolicy, i.e., untilat least2011 for the NM workers and 2005 for M workers. For high-tenure workers, this is the only channel of exposurethatisrelevantintheseimmediatepost-PNTRyears. However, forM workersinparticular, county downstream exposure becomes more important in subsequent years, with a notable step down in coefficient estimates toward the end of the sample. This timing for M workers – with an initial negative impact based on their county own exposure, and a later negative impact from (downstream) exposure of their customers – is consistent with the supply chain co-offshoring discussed above. That is, the delayed effect from downstream exposure can eventually lead firms to relocate to be closer to 22

their suppliers. The Figure also underscores the different effect of upstream exposure for NM versus M workers identified in Table 2. For NM workers—particularly those who are high-tenure—county upstream exposure turns positive almost immediately after 2001 and becomes statistically significant and large in magnitude in subsequent years. Industry upstream exposure also turns positive and statistically significant for NM workers in the latter portion of the sample. For M workers, upstream exposure is typically positive in post-PNTR years but is never statistically significant. These patterns are in line with NM industries benefiting from greater exposure to PNTR among their suppliers relatively soon after the change in policy, with those relative gains somewhat offset by loss of customers toward the end of the sample. M industries, by contrast, experience the latter without the former. The overlap of the negative impact of downstream exposure with the Great Recession for M workers suggests that this channel may become more prominent as the economy weakens. Figure 7: County and Industry DID Coefficients from Annual Earnings Specification - CR Transformation Source: LEHD, LBD, Feenstra et al. (2002), and authors’ calculations. Panels display the 95 percent confidence intervals for the county (top row) and industry (bottom row) DID coefficients of interest from an “annual” DID version of equation 5 that replacesthePostindicatorwithafullsetofyeardummiesandusestheCRtransformationofearningsasthedependentvariable. Foreachexposure,asolidmarkerindicatesstatisticalsignificance. Next, toillustratetheeconomicsignificanceoftheseestimates, weusetheestimatedcoefficientsto display predictions of overall relative changes in earnings based on the observed exposure of countyindustry pairs, as in Figure 2. Figure 8 displays these predictions at three-to-four year intervals from 2001 to 2014. The figure highlights important similarities and differences in the relationship between exposure to PNTR and earnings for NM and M workers. Notably, the bulk of the mass of the four red curves 23

for 2001 is to the left of zero, indicating that almost all workers – whether NMor M , mixed-tenure or high-tenure – exhibit relative earnings losses in the short-run after the policy change. This finding is intuitive given the negative relationship with own exposure in the early post-PNTR years for all workers. Over time, the implications of PNTR diverge sharply for NM and M workers. For M workers, higher exposure to the policy change is consistently associated with relative earnings losses, which become larger after the Great Recession as the negative coefficients on downstream exposure grow in magnitude and statistical significance. By contrast, NM workers experience relative earnings gains across years, particularly among high-tenure NM workers. Figure 7 shows that these gains are driven by positive effects of upstream exposure, which take several years to materialize. The absence of such an upstream boost for M workers helps explain their worsening outcomes relative to NM workers. For NM workers, the positive upstream effect peaks around 2010 before gradually fading, suggesting that the gains from cheaper inputs are front-loaded: Once firms have absorbed the initial cost reductions and competitive pressures adjust, the marginal benefits to workers diminish. Figure 8: Distribution of County-Industry Predicted Relative Earnings Growth, by Year and Sample Source: LEHD, LBD, Feenstra et al. (2002), and authors’ calculations. Figure displays the distribution of predicted relative earnings growth over time across county-industry pairs for mixed-tenure and high-tenure NM and M workers. Predictions are basedontheestimatedOwn+IO DIDcoefficientsreportedinFigure7. Foreachcounty-industrypair,theyaretheproductof thereportedcoefficientsandactualexposures. CurvesaresolidforyearsuptotheGreatRecession,anddashedafterwards. The existing literature has reached differing conclusions on the long-term implications of trade liberalization and similar shocks for earnings. On the one hand, Dix-Carneiro and Kovak (2017) find that effects of a Brazilian trade liberalization on regional earnings continue to grow 20 years after the liberalization, and Autor, Dorn, and Hanson (2021) find similarly persistent effects on US manufacturingemploymentarisingfromChineseimportcompetition. Ontheotherhand,Bloometal. (2019), find that the latter effects dissipate after 2007 in high-human-capital areas, and Kovak and Morrow (2022) show that for Canadian workers subject to larger tariff reductions in their industries, rapid transitions to industries less exposed to import competition mean that there was little effect on long-run cumulative earnings. Our results inject new findings into this debate by highlighting the centrality of differences in effects for M and NM workers, which are driven by input-output linkages. 24

5 Heterogeneous Outcomes By Worker and Firm Attributes Building on the examination of differences by high-tenure and mixed-tenure workers above, in this section, we examine whether responses to PNTR vary by workers’ initial (i.e., 1999) characteristics and those of their employing firm. We use a version of equation 5 that includes triple interactions of these attributes with own, upstream, and downstream county and industry exposure DID terms and run separate regressions for each earnings transformation and initial sector of worker, as well as each characteristic. We consider both worker and firm-level characteristics, binned into binary outcomes: females versus males; non-whites versus whites; workers aged 30 and below (“young”) versus those that are older; workers that have at least a bachelors degree (“≥ College”) versus those with less educational attainment; workers in the fourth quartile of earnings (“Q4 Earnings”) versus those in the lower quartiles; workers at firms with fewer than 50 employees (“small”) versus larger firms; workers at trading versus non-trading firms; and workers at firms that have both M and NM establishments (“diversified”) versus firms with only M or only NM plants.35 While several papers mentioned above have examined heterogeneous worker-level outcomes to Chinese import competition (Kamal et al., 2020; Kahn et al., 2022; Conlisk et al., 2022), we are able to consider a wider set of worker characteristics (existing work typically focuses on differences by gender and race), to account for detailed input-output linkages, and to examine the role of firm characteristics. Asabove,weassesseconomicsignificanceusingpredictedcounty-industryrelativeearningsgrowth, and results are displayed visually in Figures 9 and 10.36 In this case, however, we report only the differential impact of the demographic characteristic, i.e., the product of the triple-interaction DID coefficients and industry-county exposures. These distributions represent the differential predicted relative earnings growth associated with each attribute versus its left-out partner, e.g., females versus males. Distributions are organized by earnings transformation, as in earlier figures, with the upper and lower panels in each figure focused on initial NM and M workers, respectively. The figures also report statistical significance of the estimates, with distributions in bold if the underlying F-statistic of the triple interactions from which it is computed are statistically significant at conventional levels, and thin otherwise. 35Workers’initialsectorisdeterminedbytheindustrycodeoftheirestablishment. Diversificationcapturesthebroader activities of their firms. For context, Appendix Figure A.5 reports the distribution of workers in 2000 across two-digit NAICS sectors by gender, race, education level, and age using publicly available data from the LEHD extract tool. 36Estimated coefficients are relegated to Appendix Tables A.5 to A.7. Appendix Table A.8 reports the F-statistics foreachgroupoftripleinteractions. Consistentwiththepatternofresultsreportedinthelastsection,wefindthatthe countyexposuretripleinteractionsaremorelikelytobestatisticallysignificantatconventionallevelsthantheindustry exposure triple interactions. 25

Figure 9: Triple-Interaction County-Industry Predictions by Workers’ Firms’ Attributes Source: LEHD, LBD, Feenstra et al. (2002), and authors’ calculations. Figure displays differential predicted relative county-industry earnings for noted worker’s firm attribute versus those not possessing that attribute usingthetriple-interactionDIDcoefficientsdiscussedinthetextandreportedinAppendixTablesA.5toA.7. DistributionsareinboldiftheF-statisticforthecountyandindustrytriple-interactionterms,reportedinthe finaltwocolumnsofAppendixTableA.8,arejointlysignificantatthe10percentlevel. The figures convey several novel aspects of heterogeneous worker responses to trade liberalization. In particular, as shown in Figure 9, we find that initial firm characteristics are important determinants of subsequent M worker earnings outcomes. First, as shown by the bolded gray curves in the bottom row of the figure, we find that M workers initially employed at small firms have relatively better earnings outcomes than those employed at large firms. This result is consistent with Holmes and Stevens (2014)’s argument that small firms are more likely to produce customized output that is less substitutable with Chinese imports.37 By contrast, the relative benefits of working at a small firm are more limited for NM workers: The upper row shows that despite modestly higher earnings conditional on remaining employed (upper right panel), the triple interaction terms for the more comprehensive CR transformation (upper left panel) are not statistically significant as a group. Second, M workers employed at diversified firms have relatively worse earnings outcomes than those employed at firms with only manufacturing establishments (bolded green curves, lower panel). This resultissomewhatsurprising, astransitioningfromM toNM mightinprinciplebeeasierforworkers at firms that span both sectors, and allow for retention of firm-specific human capital, even if those activities are in different locations. On the other hand, a strict focus on manufacturing activities may 37Empirical evidence in line with this finding for M workers is also present in the appendix of the working paper version of Autor et al. (2014) and Kovak and Morrow (2022). 26

contribute to firms’ ability to produce the kinds of goods Holmes and Stevens (2014) have in mind.38 Finally, we find that workers at trading firms experience relatively better outcomes than those at firms that do not trade, though this result is only present for earnings conditional on employment (LN). Figure 10 examines heterogeneous responses by workers’ demographic attributes. The Figure showsthatdemographiccharacteristicstendtobemorerelevantforrelativeearningsforNM workers (toprow)thanM. First,intermsofgender(graycurves),wefindthatfemaleNM workersexperience relatively better labor market outcomes than males, in terms of all three outcomes.39 With respect to race(bluecurves),NM workerswhoarenotwhiteexhibitrelativelyworseearningsoutcomesinterms of CR, reflecting lower subsequent earnings conditional on employment and a lower but imprecisely estimatedprobabilityofbeingemployed.40 Forage, thetypicalNM workerunder30(“young,” inthe Figure) performs modestly better than older workers when considering CR, with a higher probability of employment offsetting relatively lower earnings conditional on employment. While we find some differences in terms of workers with or without bachelors degrees, there is no statistically significant difference in terms of CR, which captures both probability of employment and earnings conditional on employment.41 Lastly, perhaps the most widespread heterogeneous response we find among worker attributes relates to initial earnings. As shown by the bold red curves with substantial mass to the right of zero in every panel, we find that those with initially high earnings perform relatively better in terms of subsequent labor market outcomes than those with initially lower earnings. While this finding is consistent with results for M workers in Autor et al. (2014), here we find it holds for both M and NM workers and across all three labor market outcomes. This relatively better performance may indicate that those with initially high earnings possess skills that are more easily transferable to other industries, areas, or firms. It may also reflect a greater ability—perhaps due to savings—to be more selective in accepting a new job, resulting in a better match. 38To the extent that multinational firms are more likely to be diversified, this result is also consistent with Boehm, Flaaen, and Pandalai-Nayar (2020)’s finding that multinationals account for a disproportionate share of the decline in US manufacturing employment due to their greater ability to offshore production. 39These relatively better outcomes may, however, be the result of women entering the labor market or increasing hours to offset male partners’ lost income as in Besedeˇs, Lee, and Yang (2021). 40Kahn,Oldenski,andPark(2022)examinethepotentialfordifferentialeffectsofimportcompetitionbyworkerrace and ethnicity and find that, for a given level of exposure, trade competition has similar effects for white and minority workers. However,theover-representationofHispanicworkersinhighlyexposedindustriesimpliesthattheyexperience greater manufacturing employment losses than whites, on net. Kamal, Sundaran, and Tello-Trillo (2020) demonstrate how import competition leads to a decrease in the female share of employment, promotions, and earnings at firms covered by the Family and Medical Leave Act in comparison to those not protected by this policy. 41Greenland, Lopresti, and McHenry (2016) find that import competition is associated with increases in high school graduation rates. Ferriere, Navarro, and Reyes-Heroles (2022) find that college enrollment exhibits a relative increase in areas with greater exposure to Chinese import competition, driven by young people in the middle and top of the household wealth distribution. Building on this work, Conlisk, Navarro, Penn, and Reyes-Heroles (2022) find that enrollmentincreasesmoreforwomen,duetoalargerincreaseinthefemalecollegepremiumthatoccursinresponseto import competition. 27

Figure 10: Triple-Interaction County-Industry Predictions by Demographic Attributes Source: LEHD, LBD, Feenstra et al. (2002), and authors’ calculations. Figure displays differential predicted relative county-industry earnings for noted worker demographic attribute versus those not possessing that attribute using the triple-interaction DID coefficients discussed in the text and reported in Appendix Tables A.5 to A.7. Distributions are in bold if the F-statistic for the county and industry triple-interaction terms, reportedinthefinaltwocolumnsofAppendixTableA.8,arejointlysignificantatthe10percentlevel. 6 Conclusion While extensive research examines the implications of increased import competition for regions and industries,manyquestionsaboutoutcomesforindividualworkers–whomaymovebetweenindustries andregions,aswellasinandoutofemployment–remainunanswered. Thispaperprovidesadetailed analysisofUSworkers’responsetoalargelabormarketshockinducedbyUStradeliberalizationwith China. Our analysis uncovers novel results related to two aspects of the effects of trade liberalization that have received only limited attention: The role of input-output linkages in driving effects on earnings and a dichotomy in impact for non-manufacturing versus manufacturing workers. Using linked employer-employee data from the US Census Bureau, we introduce novel and comprehensive measures of workers’ exposure to trade liberalization with China. These measures account for the direct exposure to increased competition in the worker’s own industry, as well as competition in its input and customer markets. Importantly, due to our use of linked data, we are also able to construct measures of geographic exposure based on the workers’ counties of residence, which are useful for gauging effects of local labor market shocks. Our results indicate that geographic exposure is most relevant for driving workers’ earnings responses to the trade liberalization, highlighting the salience of spatial versus sectoral frictions and 28

implying that workers’ ultimate earnings profiles are more affected by their neighbors’ exposure to the shock than the industry in which they’re initially employed. We present circumstantial evidence that part of this effect arises from increased competition in local labor markets as displaced manufacturing workers move to – often lower-paid – service sectors. Our comprehensive measures of exposure via supply supply chains also reveal long-suspected but previously missing evidence of beneficial effects for workers of increased import competition in input markets. For non-manufacturing workers, such “upstream” exposure is associated with relative gains in earnings. For manufacturing workers, however, these gains are muted, and exposure via competition in customer markets exacerbates the negative effects of exposure in one’s own county. Due to these differences in effects of trade liberalization via input-output linkages, we show that there are substantial differences in effects for non-manufacturing versus manufacturing workers. Specifically, we show that after accounting for input-output linkages, most non-manufacturing workers experience relative earnings gains in response to PNTR due to the increased competition in input markets. For manufacturing workers, accounting for input-output linkages indicates that relative earnings losses are even larger than in specifications that do not take into account such supply chain effects. Our analysis also focuses on heterogeneous effects by worker and firm characteristics. The first finding of heterogeneous effects comes from examination of differences in outcomes for a sample of high-tenure workers – those continuously employed by the same firm – versus one that includes workers who switch employers. We show that high-tenure workers in the non-manufacturing sector perform relatively better in response to trade shocks than their lower-tenure peers, while high-tenure manufacturingworkers, ifanything, fairabitworseduetolowerearningsconditionalonemployment. Exploring outcomes by dimensions of firm heterogeneity, we show that workers at small and trading firms perform relatively better than their counterparts, while those at diversified firms fare worse. In terms of demographic characteristics, we find that women and those with initially high incomes perform relatively better in response to the trade shock. Our resultsprovide new informationon longstandingquestions related tothe responses ofworkers to increases in import competition and help elucidate some of the distributional and long-term effects of import competition examined in Autor et al. (2021), Bloom et al. (2019), and Autor et al. (2025). They are also useful for evaluating the overall impact of current proposals in the United States to raise trade barriers against China and other countries. Future research could delve deeper into the relative importance of declining aggregate demand and labor market competition in explaining the primacy of geographic exposure to the trade shock. 29

References Abowd, J.M., B.E.Stephens, L.Vilhuber, F.Andersson, K.L.McKinney, M.Roemer, andS.Woodcock (2009, January). The LEHD Infrastructure Files and the Creation of the Quarterly Workforce Indicators, pp. 149–230. University of Chicago Press. Acemoglu, D., D. Autor, D. Dorn, G. H. Hanson, and B. Price (2016). Import competition and the great us employment sag of the 2000s. Journal of Labor Economics 34(S1), S141–S198. Ad˜ao, R., C. Arkolakis, and F. Esposito (2019, February). General equilibrium effects in space: Theory and measurement. Working Paper 25544, National Bureau of Economic Research. Aghion, P., A. Bergeaud, M. Lequien, M. Melitz, and T. Zuber (2021, August). Opposing firm-level responses to the china shock: Horizontal competition versus vertical relationships? Working Paper 29196, National Bureau of Economic Research. Amiti, M., O. Itskhoki, and J. Konings (2014). Importers, exporters, and exchange rate disconnect. American Economic Review 104(7), 1942–1978. Amiti, M. and J. Konings (2007, December). Trade liberalization, intermediate inputs, and productivity: Evidence from indonesia. American Economic Review 97(5), 1611–1638. Antr`as, P. and D. Chor (2013). Organizing the global value chain. Econometrica 81(6), 2127–2204. Autor,D.,D.Dorn,andG.Hanson(2019,September). Whenworkdisappears: Manufacturingdecline and the falling marriage market value of young men. American Economic Review: Insights 1(2), 161–78. Autor, D., D. Dorn, and G. H. Hanson (2021). On the persistence of the china shock. Brookings Papers on Economic Activity 2021(2), 381–476. Autor, D., D. Dorn, G. H. Hanson, M. R. Jones, and B. Setzler (2025, January). Places versus people: The ins and outs of labor market adjustment to globalization. Working Paper 33424, National Bureau of Economic Research. Autor, D., D. Dorn, G. H. Hanson, and J. Song (2014). Trade adjustment: Worker-level evidence. The Quarterly Journal of Economics 129, 199–1860. Autor, D. H., D. Dorn, and G. H. Hanson (2013). The china syndrome: Local labor market effects of import competition in the united states. American Economic Review 103(6), 2121–68. Baldwin, R. and A. J. Venables (2013). Spiders and snakes: Offshoring and agglomeration in the global economy. Journal of International Economics 90(2), 245–254. Besedeˇs, T., S. H. Lee, and T. Yang (2021). Trade liberalization and gender gaps in local labor market outcomes: Dimensions of adjustment in the united states. Journal of Economic Behavior Organization 183, 574–588. 30

Bloom, N., K. Handley, A. Kurmann, and P. Luck (2019). The impact of chinese trade on u.s. employment: The good, the bad, and the debatable. Mimeo, Stanford University. Boehm, C. E., A. Flaaen, and N. Pandalai-Nayar (2020). Multinationals, offshoring, and the decline of u.s. manufacturing. Journal of International Economics 127, 103391. Borusyak, K., P. Hull, and X. Jaravel (2021, 06). Quasi-experimental shift-share research designs. The Review of Economic Studies 89(1), 181–213. Bown, C., P. Conconi, A. Erbahar, and L. Trimarchi (2020). Trade protection along supply chains. Technical report, Working Paper. Brandt, L., J. Van Biesebroeck, L. Wang, and Y. Zhang (2017, September). Wto accession and performance of chinese manufacturing firms. American Economic Review 107(9), 2784–2820. Caliendo, L., M. Dvorkin, and F. Parro (2019). Trade and labor market dynamics: General equilibrium analysis of the china trade shock. Econometrica 87(3), 741–835. Caliendo, L. and F. Parro (2022). Lessons from u.s.-china trade relations. Carballo, J. and R. Mansfield (2023). What drives the labor market incidence of trade shocks?: An equilibrium matching analysis of china’s wto accession. Technical report, Working Paper. Chen, J. and J. Roth (2023, 12). Logs with zeros? some problems and solutions*. The Quarterly Journal of Economics 139(2), 891–936. Conlisk, S., G. Navarro, M. Penn, and R. Reyes-Heroles (2022). International trade and gender gaps in college enrollment. Feds notes, Board of Governors of the Federal Reserve System. Costinot, A., M. Sarvim¨aki, and J. Vogel (2022). Exposure(s) to trade and earnings dynamics: Evidence from the collapse of finnish-soviet trade. Mimeo. Davis, S. J. and T. M. von Wachter (2011, December). Recessions and the cost of job loss. Working Paper 17638, National Bureau of Economic Research. Deng, L., P. Krishna, M. Z. Senses, and J. Stegmaier (2021, December). Trade, human capital, and income risk. Working Paper 29612, National Bureau of Economic Research. Dix-Carneiro, R. (2014). Trade liberalization and labor market dynamics. Econometrica 82(3), 825–885. Dix-Carneiro, R. and B. K. Kovak (2017, October). Trade liberalization and regional dynamics. American Economic Review 107(10), 2908–46. Eckert, F., T. C. Fort, P. K. Schott, and N. J. Yang (2020, January). Imputing missing values in the us census bureau’s county business patterns. Working Paper 26632, National Bureau of Economic Research. 31

Ellison, G., E. L. Glaeser, and W. R. Kerr (2010, June). What causes industry agglomeration? evidence from coagglomeration patterns. American Economic Review 100(3), 1195–1213. Eriksson, K., K. Russ, J. C. Shambaugh, and M. Xu (2019, March). Trade shocks and the shifting landscape of u.s. manufacturing. Working Paper 25646, National Bureau of Economic Research. Feenstra, R. C., J. Romalis, and P. K. Schott (2002). Us imports, exports, and tariff data, 1989-2001. Working Paper 9387, National Bureau of Economic Research. Feenstra, R. C. and A. Sasahara (2018). The ‘china shock,’ exports and u.s. employment: A global input–output analysis. Review of International Economics 26(5), 1053–1083. Ferriere, A., G. Navarro, and R. Reyes-Heroles (2022). Escaping the losses from trade. the impact of heterogeneity and skill acquisition. Mimeo, Board of Governors of the Federal Reserve System. Flaaen, A. and J. Pierce (2019). Disentangling the effects of the 2018-2019 tariffs on a globally connected u.s. manufacturing sector. Federal Reserve Finance and Economics Discussion Paper 2019- 086. Fort, T. C. (2023). The Changing Firm and Country Boundaries of US Manufacturers in Global Value Chains. Journal of Economic Perspectives. Fort, T. C. and S. Klimek (2016). The effect of industry classification changes on us employment composition. Technical report, Tuck School at Dartmouth. Goldberg, P. K., A. K. Khandelwal, N. Pavcnik, and P. Topalova (2010, November). Imported intermediate inputs and domestic product growth: Evidence from india. The Quarterly Journal of Economics 125(4), 1727–1767. Greenland, A., J. Lopresti, and P. McHenry (2016). Import competition and internal migration. Greenland, A. N., M. Ion, J. W. Lopresti, and P. K. Schott (2020, July). Using equity market reactions to infer exposure to trade liberalization. Working Paper 27510, National Bureau of Economic Research. Hakobyan, S. and J. McLaren (2016). Looking for local labor market effects of nafta. The Review of Economics and Statistics 98(4), 728–741. Handley, K., F. Kamal, and R. Monarch (2020). Rising import tariffs, falling export growth: When modern supply chains meet old-style protectionism. Working Paper 1270, International Finance Discussion Papers. Handley, K. and N. Limao (2017, September). Policy uncertainty, trade, and welfare: Theory and evidence for china and the united states. American Economic Review 107(9), 2731–83. Head, K. and J. Ries (1999). Rationalization effects of tariff reductions. Journal of International Economics 47(2), 295–320. 32

Holmes, T. J. and J. J. Stevens (2014). An alternative theory of the plant size distribution, with geography and intra- and international trade. Journal of Political Economy 122(2), 369–421. Huckfeldt, C. (2022, April). Understanding the scarring effect of recessions. American Economic Review 112(4), 1273–1310. Jacobson, L. S., R. J. LaLonde, and D. G. Sullivan (1993). Earnings losses of displaced workers. The American Economic Review 83(4), 685–709. Jaravel, X. and E. Sager (2019). What are the price effects of trade? evidence from the u.s. and implications for quantitative trade models. Working Paper 2019-068, Board of Governors of the Federal Reserve System. Javorcik, B., B. Kett, K. Stapleton, and L. O’Kane (2025). Did the 2018 trade war improve job opportunities for us workers? Journal of International Economics 158, 104125. Johnson,R.C.andA.Moxnes(2023). Gvcsandtradeelasticitieswithmultistageproduction. Journal of International Economics 145, 103796. Kahn, L. B., L. Oldenski, and G. Park (2022). Racial and ethnic inequality and the china shock. Technical report, National Bureau of Economic Research. Kamal, F., A. Sundaran, and C. Tello-Trillo (2020). Family-leave mandates and female labor at u.s. firms: Evidence from a trade shock. Technical report, Center for Economic Studies Working paper series. Keller, W. and H. Utar (2022, 03). Globalization, gender, and the family. The Review of Economic Studies 89(6), 3381–3409. Keller, W. and H. Utar (2023). International trade and job polarization: Evidence at the worker level. Journal of International Economics 145, 103810. Khandelwal, A. K., P. K. Schott, and S.-J. Wei (2013, October). Trade liberalization and embedded institutional reform: Evidence from chinese exporters. American Economic Review 103(6), 2169– 95. Kovak, B. K. (2013, August). Regional effects of trade reform: What is the correct measure of liberalization? American Economic Review 103(5), 1960–76. Kovak,B.K.andP.M.Morrow(2022,February). Thelong-runlabormarketeffectsofthecanada-u.s. free trade agreement. Working Paper 29793, National Bureau of Economic Research. Krishna, P., J. P. Poole, and M. Z. Senses (2014). Wage effects of trade reform with endogenous worker mobility. Journal of International Economics 93(2), 239–252. Long, C. and X. Zhang (2012). Patterns of china’s industrialization: Concentration, specialization, and clustering. China Economic Review 23(3), 593–612. 33

McLaren, J. (2017). Globalization and labor market dynamics. Annual Review of Economics 9(1), 177–200. McLaren, J. (2022). Trade shocks and labor-market adjustment. Pierce, J. R. and P. K. Schott (2012). A concordance between ten-digit u.s. harmonized system codes and sic/naics product classes and industries. Journal of Economic and Social Measurement 37, 61–96. Pierce, J. R. and P. K. Schott (2016, July). The surprisingly swift decline of us manufacturing employment. American Economic Review 106(7), 1632–62. Pierce, J. R. and P. K. Schott (2020, March). Trade liberalization and mortality: Evidence from us counties. American Economic Review: Insights 2(1), 47–64. Podgursky,M.andP.Swaim(1987). Jobdisplacementandearningsloss: Evidencefromthedisplaced worker survey. Industrial and Labor Relations Review 41(1), 17–29. Scott, R. E., V. Wilson, J. Kandra, and D. Perez (2022). Botched policy responses to globalization have decimated manufacturing employment with often overlooked costs for black, brown, and other workers of color: Investing in infrastructure and rebalancing trade can create good jobs for all. Stevens, A. H. (1997). Persistent effects of job displacement: The importance of multiple job losses. Journal of Labor Economics 15(1, Part 1), 165–188. Sullivan, D. and T. V. Wachter (2009). Job displacement and mortality: An analysis using administrative data. Quarterly Journal of Economics, 1265–1306. Topalova, P. (2007). Trade liberalization, poverty and inequality: Evidence from indian districts. In Globalization and poverty, pp. 291–336. University of Chicago Press. Topalova, P.andA.Khandelwal(2011). Tradeliberalizationandfirmproductivity: Thecaseofindia. The Review of Economics and Statistics 93(3), 995–1009. Traiberman, S. (2019, December). Occupations and import competition: Evidence from denmark. American Economic Review 109(12), 4260–4301. Utar, H. (2018, 10). Workers beneath the Floodgates: Low-Wage Import Competition and Workers’ Adjustment. The Review of Economics and Statistics 100(4), 631–647. Vilhuber,L.andK.McKinney(2014,June). LEHDInfrastructurefilesintheCensusRDC-Overview. Working Papers 14-26, Center for Economic Studies, U.S. Census Bureau. Wang, Z., S.-J. Wei, X. Yu, and K. Zhu (2018, August). Re-examining the effects of trading with china on local labor markets: A supply chain perspective. Working Paper 24886, National Bureau of Economic Research. Waugh, M. E. (2019, October). The consumption response to trade shocks: Evidence from the us-china trade war. Working Paper 26353, National Bureau of Economic Research. 34

Online Appendix (Not for Publication) This online appendix contains additional empirical results as well as more detailed explanations of data used in the main text. A State Coverage in the LEHD The set of states included in the LEHD varies over time as summarized in Figure A.1. We use the 46 states available as of 2000 in examining worker movement between M and NM sectors in Section 4.3, and the 19 states present from 1993 to 2014 for our regression analysis. Figure A.1: State Availability in the LEHD Source: Vilhuber and McKinney (2014). Figure displays the availabilityofstatedataintheLEHD.Dashed vertical line shows 1993, the cutoff for inclusionintheregressionsample. TableA.1reportsthe1999NM, M andtotalemployment, bystateforeachsampleusingpublicly available information from the Business Dynamics Statistics (BDS) program available at https: //www.census.gov/programs-surveys/bds.html. 35

Table A.1: Number of NM and M Workers in 1999, by State State Sample NM M Total State Sample NM M Total Alaska 19&46 0.18 0.01 0.20 Massachusetts 46 2.54 0.39 2.93 Arizona 19&46 1.63 0.19 1.82 Michigan 46 3.15 0.81 3.96 California 19&46 10.43 1.74 12.17 Minnesota 46 1.93 0.37 2.31 Colorado 19&46 1.66 0.16 1.82 Nebraska 46 0.62 0.11 0.73 Florida 19&46 5.52 0.42 5.94 Nevada 46 0.82 0.04 0.86 Idaho 19&46 0.37 0.07 0.44 NewJersey 46 3.06 0.39 3.44 Illinois 19&46 4.46 0.85 5.31 NewMexico 46 0.50 0.04 0.54 Indiana 19&46 1.95 0.63 2.58 NewYork 46 6.31 0.72 7.03 Kansas 19&46 0.91 0.19 1.10 NorthDakota 46 0.23 0.02 0.25 Louisiana 19&46 1.41 0.16 1.57 Ohio 46 3.85 0.97 4.82 Maryland 19&46 1.82 0.16 1.97 Oklahoma 46 0.99 0.17 1.15 Missouri 19&46 1.96 0.37 2.33 RhodeIsland 46 0.34 0.07 0.41 Montana 19&46 0.27 0.02 0.29 SouthCarolina 46 1.23 0.34 1.57 NorthCarolina 19&46 2.57 0.75 3.32 SouthDakota 46 0.25 0.05 0.30 Oregon 19&46 1.11 0.21 1.32 Tennessee 46 1.86 0.47 2.34 Pennsylvania 19&46 4.16 0.79 4.95 Texas 46 6.76 0.95 7.70 Washington 19&46 1.86 0.29 2.15 Utah 46 0.76 0.12 0.88 Wisconsin 19&46 1.79 0.57 2.36 Vermont 46 0.20 0.04 0.24 Wyoming 19&46 0.16 0.01 0.17 Virginia 46 2.41 0.36 2.78 Connecticut 46 1.28 0.24 1.52 WestVirginia 46 0.47 0.07 0.55 Delaware 46 0.32 0.04 0.36 Alabama Neither 1.28 0.34 1.62 Georgia 46 2.80 0.53 3.33 Arkansas Neither 0.72 0.23 0.95 Hawaii 46 0.40 0.01 0.42 DistrictofColumbia Neither 0.39 0.00 0.39 Iowa 46 1.00 0.24 1.24 Mississippi Neither 0.72 0.22 0.94 Kentucky 46 1.17 0.29 1.46 NewHampshire Neither 0.43 0.10 0.53 Maine 46 0.39 0.08 0.47 Source: BDSandauthors’calculations. Tablereportsthenon-manufacturing(NM),manufacturing(M)andtotal employment,inmillions,bystatein1999. Secondcolumnineachpanelindicateswhetherthestatesisinour19and 46-statesamples,whereallstatesinthe19-statesamplearealsointhelatter. OverallM,NM andtotalemployment forthe19-statesamplein1999is44.2, 7.6and51.8. Theanalogoustotalsfortheadditionalstatesinthe46-state sampleare45.7,7.9and53.6. Forstatesnotineithersampletheyare3.5,0.9and4.4. B Industry Variable Construction In this section we describe how the industry controls referenced in Section 4 are constructed. MFA Exposure: We control for expiration of the Multi-Fiber Arrangement, which occurs in stages during our sample period. Khandelwal, Schott, and Wei (2013) provide details on this policy: The [MFA] and its successor, the Agreement on Textile and Clothing (ATC), grew out of quotas imposed by the United States on textile and clothing imports from Japan during the 1950s. Over time, it evolved into a broader institution that regulated the exports of clothing and textile products from developing countries to the United States, European Union, Canada and Turkey...Bargaining over these restrictions was kept separate from multilateral trade negotiations until the conclusion of the Uruguay Round in 1995, when an agreement was struck to eliminate the quotas over four phases. On January 1, 1995, 1998, 2002, and 2005, the United States was required to remove textile and clothing quotas representing 16, 17, 18 and the remaining 49 percent of their 1990 import volumes, respectively. 36

Relaxation of quotas on Chinese imports did not occur until it became a member of the World Trade Organization in 2001; as a result, its quotas on the goods in the first three phases were relaxed in early 2002, and its quotas on the goods in the fourth phase were relaxed as scheduled in 2005. The order in which goods were placed into a particular phase was chosen by the United States. We calculate counties’ exposure to elimination of the MFA in three steps, as in Pierce and Schott (2020). These steps include: 1) measuring the extent to which MFA quotas were binding using the average fill rate of the industry’s constituent import products, following Khandelwal, Schott, and Wei (2013); 2) computing counties’ labor-share-weighted-average fill rate across industries for each phase; 3) cumulating the calculated fill rates as each phase of expiration takes place, so that the measure of exposure to the MFA rises over time, as additional quotas are removed. See Appendix D of Pierce and Schott (2020) for additional details. Figure A.2: Average Up- and Downstream Gaps Source: CBP,BEA,Feenstraetal.(2002),andauthors’calculations. Leftpaneldisplaysmeanindustryup-anddownstreamNTR gap, Gapup and Gapdown, across 3-digit NAICS sectors, except for 541, which is broken out by 4-digit sectors. Manufacturing industries i arehighligh i ted. Rightpanelreportsup-anddownstreamgapsforeachcountyinour19stateregressionsample,Gapu c p andGapdown,withNapa(06055)andSanMateo(06081),Californiahighlighted. Countiesareidentifiedby5-digitFIPScodes. c Changes in Chinese Policy: Chinainstitutedanumberofpolicychangesaspartofitsaccessionto the WTO, and for which we control, including reducing import tariff rates and production subsidies. As in Pierce and Schott (2016), we use product-level data on Chinese import tariffs from 1996 to 2005 fromBrandt,VanBiesebroeck,Wang,andZhang(2017)tocomputeindustry-levelchangesinChinese import tariffs. We use data from the Annual Report of Industrial Enterprise Statistics published by China’s National Bureau of Statistics (NBS) as a measure of changes in production subsidies. We construct county-level measures of exposure to each of these policy changes using labor share weights and then interact these measures with an indicator for post-PNTR years. See Appendix D of Pierce and Schott (2020) for additional details. Up- and Downstream NTR Gaps: The left panel of Figure A.2 reports the average up- and 37

downstream NTR gaps by 3-digit NAICS industry, while the right panel of the figure reports the upand downstream gap for all counties in our 19 state sample. Manufacturing sectors are highlighted in the left panel, while Napa and Santa Clara, California are highlighted in the right panel. These two counties illustrate how counties – even when geographically similar – can have very different up- and downstream NTR gaps due to their differing industrial structures. As indicated in the figure, Napa is more heavily concentrated in non-tradable services such as Retail (NAICS 44-5), Accommodation and Food (NAICS 72), and Health (NAICS 62), while Santa Clara is more heavily dependent on manufacturing, particularly Computers and Electronics (NAICS 334). Within manufacturing, Napa is concentrated in production at Wineries (NAICS 312130). C Publicly Available Data on Wages and Wage Growth by Sector Asacomplementtotheworker-leveltransitionsreportedinSection4.3,FigureA.3reportstheaverage hourly wages as of 2000, of production and non-supervisory workers, as well as wage growth from 2000-2007 by major sector using publicly available data from the US Bureau of Labor Statistics. As indicated in the figure, the average hourly wage for production and non-supervisors in Manufacturing (NAICS3)in2000was13.8dollars. TheanalogousaveragesforASW(NAICS56),Retail(NAICS44- 5),Arts, EntertainmentandRecreation(NAICS71), andAccommodationandFoodServices(NAICS 72) were 12.0, 11.3, 10.9 and 8.1, or 13, 18, 21 and 41 percent less than those in manufacturing in that year. Figure A.3: Wages and Wage Growth, by 2-digit NAICS (Public BLS Data) Source: BLSandauthors’calculations. Leftpaneldisplaystheaveragehourlywageofproductionandnon-supervisoryworkers by2-digitNAICSsectorin2000. Rightpaneldisplaysnominalgrowthintheseaveragehourlywagesfrom2000to2007. 38

D Flows from M, Alternate Time Periods (46-State Sample) Table A.2 reports beginning and ending employment at the 1-digit NAICS sector level from 2000 to 2005, a shorter time period than the 2000 to 2007 period examind in the main text. As indicated in thetable, thelargestnetoutflowsfrommanufacturingemploymentaretoNotEmployed(-.7million), Business Services (-.6 million), Wholesale, Retail, Transportation and Warehousing (.5 million), Education and Health (-.42 million), and Mining, Utilities, and Construction (-.22 million). Only two 1-digitsectors,Agriculture,Forestry,FishingandHunting,andArts,Entertainment,Accommodation and Food exhibit net inflows into manufacturing, of .04 and .05 million, respectively. E DID Estimates for E > 0 and LN Earnings Transformations This section contains baseline DID estimates for the E > 0 (Table A.3) and LN (Table A.4) earnings transformations discussed in the main text. They are presented in the same format as Table 2, which reports analogous estimates for the CR earnings transformation. 39

)elpmaS etatS-64( rotceS SCIAN tigid-1 yb snoitisnarT 5-0002 :2.A elbaT )snoilliM(tnemyolpmE nilatoT 5002nirotceS 0002 EN 9 8 7 6 5 4 3 2 1 0002nirotceS 97.1 26.0 10.0 20.0 60.0 50.0 11.0 31.0 11.0 90.0 75.0 )1(gnitnuHdnagnihsiF,yrtseroF,erutlucirgA 94.01 94.2 81.0 11.0 12.0 03.0 38.0 95.0 33.0 34.5 40.0 )2(noitcurtsnoC,seitilitU,gniniM 86.71 34.3 41.0 91.0 83.0 46.0 55.1 87.1 49.8 55.0 80.0 )3(gnirutcafunaM 81.92 80.6 43.0 05.0 91.1 87.1 22.3 38.31 92.1 88.0 90.0 )4(gnisuoheraW,noitatropsnarT,liateR,elaselohW 37.92 07.6 93.0 14.0 79.0 99.1 97.41 34.2 40.1 39.0 70.0 )5(secivreSssenisuB 41.42 53.4 83.0 52.0 44.0 00.61 14.1 58.0 22.0 22.0 20.0 )6(erachtlaeH,noitacudE 48.31 63.3 61.0 62.0 37.4 61.1 76.1 16.1 34.0 24.0 40.0 )7(noitaerceR,noitadommoccA,tnemniatretnE,strA 25.4 72.1 60.0 25.1 81.0 63.0 34.0 34.0 41.0 31.0 10.0 )8()noitartsinimdAcilbuPtpecxe(secivreSrehtO 15.5 59.0 14.3 40.0 90.0 83.0 82.0 81.0 50.0 21.0 10.0 )9(noitartsinimdAcilbuP 11.73 98.0 05.1 32.7 45.5 52.7 05.8 56.2 08.2 57.0 )EN(deyolpmEtoN 00.471 52.92 49.5 18.4 94.51 02.82 35.13 33.03 02.51 65.11 96.1 5002nilatoT leveLlaitinIfotnecrePasatnemyolpmE nilatoT 5002nirotceS 0002 EN 9 8 7 6 5 4 3 2 1 0002nirotceS 001 53 1 1 4 3 6 7 6 5 23 )1(gnitnuHdnagnihsiF,yrtseroF,erutlucirgA 001 42 2 1 2 3 8 6 3 25 0 )2(noitcurtsnoC,seitilitU,gniniM 001 91 1 1 2 4 9 01 15 3 0 )3(gnirutcafunaM 001 12 1 2 4 6 11 74 4 3 0 )4(gnisuoheraW,noitatropsnarT,liateR,elaselohW 001 32 1 1 3 7 05 8 3 3 0 )5(secivreSssenisuB 001 81 2 1 2 66 6 4 1 1 0 )6(erachtlaeH,noitacudE 001 42 1 2 43 8 21 21 3 3 0 )7(noitaerceR,noitadomoccA,tnemniatretnE,strA 001 82 1 43 4 8 9 9 3 3 0 )8()noitartsinimdAcilbuPtpecxe(secivreSrehtO 001 71 26 1 2 7 5 3 1 2 0 )9(noitartsinimdAcilbuP 001 2 4 91 51 02 32 7 8 2 )EN(deyolpmEtoN 001 71 3 3 9 61 81 71 9 7 1 5002nilatoT 5002ot)wor(0002morfsrekrowdeyolpme-tondnadeyolpmefoshtapnoitisnartehtstroperelbaT .snoitaluclac’srohtuadna,DBL,DHEL :ecruoS ,sasnakrA,amabalA .0002nignitratsDHELehtnielbaliavasinoitamrofniesohwsetats64ehtrof)sesehtnerapni(rotcesSCIANtigid-1yb)nmuloc( lenap rewoL .srekrow fo snoillim ni slevel stroper lenap reppU .dedulcxe era aibmuloC fo tcirtsiD eht sa llew sa ippississiM dna erihspmaH weN .slevellaitinifoserahsstroper 40

Table A.3: Response of Employment to PNTR LowTenureNM HighTenureNM LowTenureM HighTenureM EarningsTransformation: E>0 E>0 E>0 E>0 E>0 E>0 E>0 E>0 PostxGapOwn . . . . -.01 -.01 .01 .01 Industry . . . . .01 .01 .02 .02 PostxGapDown . -.15 . -.07 . -.01 . -.02 Industry . .1 . .09 . .02 . .03 PostxGapUp . .14 . .14 . .02 . .04 Industry . .14 . .12 . .06 . .09 PostxGapOwn -.39*** -.42*** -.25*** -.31*** -.23*** -.16 -.25*** -.15 County .08 .11 .07 .1 .06 .11 .07 .12 PostxGapDown . -.19 . -.25** . -.27* . -.47*** County . .14 . .13 . .14 . .17 PostxGapUp . .44 . .75** . .12 . .26 County . .31 . .3 . .37 . .39 R-sq .41 .41 .41 .41 .41 .41 .41 .41 Obs(000) 17,360 17,360 4,605 4,605 4,274 4,274 1,520 1,520 WorkerFE Yes Yes Yes Yes Yes Yes Yes Yes YearFE Yes Yes Yes Yes Yes Yes Yes Yes IndustryGapsF 1.002 0.446 0.468 0.212 IndustryGapsp 0.393 0.721 0.706 0.888 CountyGapsF 10.522 6.376 7.521 5.223 CountyGapsp 0 0 0 0.002 Source: LEHD,LBD,Feenstraetal.(2002),andauthors’calculations. TabledisplaysDIDtermsofinterestfromworkeryearlevelOLSregressionsofequation5forworkersinitiallyemployedinnon-manufacturing(NM)andmanufacturing (M) . The dependent variable is a dummy for whether the worker is employed, E > 0. The sample period is 1993 to 2014. Mixed-tenureworkersareemployedthroughoutthe1993to1999pre-period,butnotnecessarilybythesamefirm. High-tenureworkersareemployedbythesamefirmduringthepre-period. Postisadummyvariableforyearsafter2000. Standarderrorstwo-wayclusteredby4-digitNAICSandcountyarenotedbelowcoefficients. FinaltwopanelsreportFstatisticsforthejointsignificanceoftheindustryandcountyexposureDIDterms,respectively. ***,**,and*represent statisticalsignificanceatthe1,5and10percentlevels. 41

Table A.4: Response of ln(Earn) to PNTR LowTenureNM HighTenureNM LowTenureM HighTenureM DepVar: LN LN LN LN LN LN LN LN PostxGapOwn . . . . .05 .09* .06 .09 Industry . . . . .05 .05 .06 .06 PostxGapDown . -.14 . -.19 . -.28** . -.21* Industry . .17 . .2 . .11 . .11 PostxGapUp . .62** . .74** . -.27 . -.34 Industry . .25 . .29 . .24 . .31 PostxGapOwn -.96*** -.69*** -.69*** -.6*** -.54*** .18 -.34 .5 County .15 .18 .17 .2 .18 .24 .2 .33 PostxGapDown . -1.02*** . -.95*** . -1.11*** . -1.35** County . .27 . .36 . .33 . .53 PostxGapUp . .24 . 1.09 . -.82 . -1.77 County . .64 . .71 . 1 . 1.13 R-sq .61 .61 .63 .63 .57 .57 .56 .56 Obs(000) 15,370 15,370 4,173 4,173 3,830 3,830 1,378 1,378 WorkerFE Yes Yes Yes Yes Yes Yes Yes Yes YearFE Yes Yes Yes Yes Yes Yes Yes Yes IndustryGapsF 2.253 2.144 2.964 1.951 IndustryGapsp 0.083 0.096 0.037 0.128 CountyGapsF 22.63 9.158 6.289 3.150 CountyGapsp 0 0 0.001 0.029 Source: LEHD, LBD, Feenstra et al. (2002), and authors’ calculations. Table displays DID terms of interest from worker-yearlevelOLSregressionsofequation5forworkersinitiallyemployedinnon-manufacturing(NM)andmanufacturing(M). Thedependentvariableislogearnings,ln(Earn). Thesampleperiodis1993to2014. Mixed-tenure workers are employed throughout the 1993 to 1999 pre-period, but not necessarily by the same firm. High-tenure workersareemployedbythesamefirmduringthepre-period. Postisadummyvariableforyearsafter2000. Standard errors two-way clustered by 4-digit NAICS and county are noted below coefficients. Final two panels report F-statistics for the joint significance of the industry and county exposure DID terms, respectively. ***, **, and * representstatisticalsignificanceatthe1,5and10percentlevels. E.1 Weighting Given US Census Bureau disclosure constraints, we report the economic significance of our results in Figure 4 using county-industry pairs as a unit of analysis, and treat each county-industry equally. An alternate approach is to weight county-industry pairs in the predicted relative earnings distributions bytheirnumberofworkersinthepre-period. InFigureA.4, wecomparetheunweighteddistributions from Figure 4 (solid lines) to their weighted counterparts (dashed lines), by specification and sample. As shown in the Figure, the estimates of economic significance are qualitatively highly similar under the two weighting approaches. The primary difference is the larger mass around the center of the weighted distributions, indicating that highly-populated county-industry pairs tend to fall near the middle of the distribution of exposure to PNTR. 42

Figure A.4: Predicted Relative Earnings, Baseline versus Employment-Weighted Source: LEHD,LBD,Feenstraetal.(2002), andauthors’calculations. Figurecomparesourbaselinecountyindustrypredictedrelativeearningsdistributionstoversionsofthesedistributionswhereeachcounty-industry in the distribution is weighted by its employment in the pre-period. Distributions are plotted for for mixedtenure(MT)andhigh-tenure(HT)NM andM workers. Coefficientestimatesforthebaselinepredictionsare reportedinTables2,A.3andA.4. F Demographic Characteristics of Workers, by Sector Figure A.5 plots the distribution of workers across sectors by gender, race, age, and education in 2000 using data publicly available from the LEHD extract tool. As indicated in the first panel, femalesare relativelymore concentrated in Education(NAICS61) andHealthcare (NAICS62), while males are employed disproportionately in Construction (NAICS 23), Transportation (NAICS 48), Wholesale(NAICS48),andManufacturing(NAICS3). Non-whiteworkers(panel2)areconcentrated in Administrative Services (NAICS 56), Accommodation and Food (NAICS 72), and Healthcare (NAICS 62), while white workers work disproportionately in Construction (NAICS 23), Wholesale (NAICS 42), Education (NAICS 61), and Retail (NAICS 44). Less highly educated workers are concentrated in Administrative Services (NAICS 56), Construction (NAICS 23), Accommodation and Food (NAICS 72), Retail (NAICS 44) and Manufacturing (NAICS 3). Finally, younger workers are especially concentrated in Accommodation and Food (NAICS 72) and Retail (NAICS 44), while Education (NAICS 61) and Manufacturing (NAICS 3) skew older. G Results for Triple-Interaction Demographic Specifications This section reports estimated coefficients (in Tables A.5 to A.7) for the triple-interaction specifications discussed in Section 5. F-statistics for the significance of the county and industry triple 43

Figure A.5: Worker Demographics in 2000 (Public LEHD Data) Source: LEHD and authors’ calculations. Figure displays distribution of workers across two-digit NAICS sectors by gender, race, educational attainment, and age in 2000 from publicly available LEHD data downloadable at https://ledextract.ces.census.gov/j2j/emp. interactions are reported in Table A.8. 44

TableA.5: Triple-InteractionDemographicRegressions(CR,High-Tenure) LHS Sample Female Non-White Age<30 Bachelors HighEarner SmallFirm Diversified Trader PostxAttributexIndGap CR NM 0 0 0 0 0 0 0 0 CR NM 0 0 0 0 0 0 0 0 PostxAttributexIndUpGap CR NM 2.636** 2.183** –3.577*** 1.436* 1.006 –3.041 –1.405 -.396 CR NM 1.274 0.914 1.296 0.804 1.109 2.025 1.378 2.793 PostxAttributexIndDownGap CR NM –1.825** .09 1.409 -.688 -.445 1.326 -.806 1.613 CR NM 0.877 0.399 0.872 0.590 0.958 1.598 0.997 2.126 PostxAttributexCtyGap CR NM .296 .936 –1.382 -.093 –1.059 –2.188 –2.349 3.285 CR NM 1.346 1.939 1.277 1.310 0.811 1.371 1.545 2.173 PostxAttributexCtyUpGap CR NM 11.06** –13.63** 2.529 1.834 2.546 5.357 3.787 –13.23** CR NM 4.950 6.221 4.481 5.012 2.519 4.293 5.000 6.051 PostxAttributexCtyDownGap CR NM –4.276** 2.819 2.92 -.259 .726 2.499 3.673 –1.153 CR NM 1.923 3.755 1.911 1.777 2.292 2.191 2.509 3.226 PostxAttributexIndGap CR M .056 -.084 -.149 -.386 .089 -.236 -.192 .026 CR M 0.227 0.203 0.246 0.282 0.317 0.288 0.348 0.273 PostxAttributexIndUpGap CR M .239 1.065 .669 2.091 1.595 -.749 –1.325 .545 CR M 0.963 0.663 0.868 1.881 2.103 1.040 1.194 1.334 PostxAttributexIndDownGap CR M -.498 -.28 -.412 -.116 .722 .735 -.214 -.376 CR M 0.407 0.438 0.459 0.503 0.641 0.619 0.678 0.584 PostxAttributexCtyGap CR M -.262 –2.211 -.984 –1.913 –4.052** –6.008** –2.343 2.293 CR M 1.522 2.345 1.999 2.119 1.822 2.501 1.993 1.664 PostxAttributexCtyUpGap CR M 8.164 1.548 3.377 .205 2.343 27.06*** 20.98** –16.42** CR M 5.674 7.521 7.351 6.748 4.894 8.464 8.697 6.658 PostxAttributexCtyDownGap CR M –1.776 3.224 .604 –1.649 1.803 –1.652 –7.021* 1.222 CR M 2.536 3.484 3.498 3.076 3.492 4.729 3.812 3.219 Source:LEHD,LBD,Feenstraetal.(2002),andauthors’calculations.TabledisplaystheDIDcoefficientsofinterestfortheOLSpanelestimationofequation5thatincludes tripleinteractionswithinitial(1999)workerattributesforthenotedearningstransformation.Onlythecoefficientsofthetripleinteractionsarereported.Thefirstcolumnin eachpanelnotestheDIDexposuretermofthetripleinteraction,whilethecolumnheaderindicatesthedemographicattributewithwhichtheexposuretermisinteracted.***, **,and*representstatisticalsignificanceatthe1,5and10percentlevels.TableA.8reportscorrespondingF-statisticsforthejointsignificanceoftheseexposureterms. TableA.6: Triple-InteractionDemographicRegressions(E>0,High-Tenure) LHS Sample Female Non-White Age<30 Bachelors HighEarner SmallFirm Diversified Trader PostxAttributexIndGap E>0 NM 0 0 0 0 0 0 0 0 E>0 NM 0 0 0 0 0 0 0 0 PostxAttributexIndUpGap E>0 NM 0.210* 0.152* –0.264*** 0.165*** 0.173* –0.241 –0.105 –0.041 E>0 NM 0.116 0.087 0.099 0.058 0.092 0.195 0.111 0.202 PostxAttributexIndDownGap E>0 NM –0.151* 0.019 0.085 –0.093** –0.099 0.090 –0.099 0.160 E>0 NM 0.078 0.036 0.066 0.042 0.073 0.158 0.084 0.161 PostxAttributexCtyGap E>0 NM –0.021 0.088 –0.121 –0.016 –0.121* –0.200* –0.199 0.265 E>0 NM 0.122 0.182 0.110 0.110 0.070 0.116 0.131 0.178 PostxAttributexCtyUpGap E>0 NM 1.075** –1.343** 0.340 0.208 0.408** 0.380 0.294 –0.994** E>0 NM 0.429 0.601 0.404 0.411 0.207 0.372 0.422 0.494 PostxAttributexCtyDownGap E>0 NM –0.362** 0.292 0.275 –0 0.092 0.316* 0.374* –0.237 E>0 NM 0.179 0.349 0.167 0.156 0.193 0.189 0.213 0.274 PostxAttributexIndGap E>0 M 0.018 –0.006 –0.009 –0.036 –0.011 –0.025 –0.017 0.004 E>0 M 0.019 0.020 0.022 0.024 0.023 0.026 0.030 0.022 PostxAttributexIndUpGap E>0 M –0.023 0.100 0.084 0.174 0.165 –0.063 –0.125 0.039 E>0 M 0.082 0.066 0.069 0.135 0.146 0.091 0.102 0.106 PostxAttributexIndDownGap E>0 M –0.043 –0.020 –0.038 –0.019 0.051 0.064 –0.025 –0.043 E>0 M 0.037 0.044 0.041 0.041 0.047 0.055 0.060 0.051 PostxAttributexCtyGap E>0 M 0.015 –0.166 0.027 –0.190 –0.373** –0.421* –0.098 0.133 E>0 M 0.138 0.215 0.161 0.178 0.146 0.237 0.180 0.141 PostxAttributexCtyUpGap E>0 M 0.830 0.118 0.139 0.033 0.194 1.827** 1.401* –1.020* E>0 M 0.506 0.684 0.605 0.566 0.375 0.796 0.791 0.576 PostxAttributexCtyDownGap E>0 M –0.270 0.259 –0.022 –0.012 0.316 –0.198 –0.758** 0.136 E>0 M 0.231 0.326 0.307 0.254 0.298 0.444 0.358 0.288 Source:LEHD,LBD,Feenstraetal.(2002),andauthors’calculations.TabledisplaystheDIDcoefficientsofinterestfortheOLSpanelestimationofequation5thatincludes tripleinteractionswithinitial(1999)workerattributesforthenotedearningstransformation. Onlythecoefficientsofthetripleinteractionsarereported. Thefirstcolumnin eachpanelnotestheDIDexposuretermofthetripleinteraction,whilethecolumnheaderindicatesthedemographicattributewithwhichtheexposuretermisinteracted.***, **,and*representstatisticalsignificanceatthe1,5and10percentlevels.TableA.8reportscorrespondingF-statisticsforthejointsignificanceoftheseexposureterms. 45

TableA.7: Triple-InteractionDemographicRegressions(LN,High-Tenure) LHS Sample Female Non-White Age<30 Bachelors HighEarner SmallFirm Diversified Trader PostxAttributexIndGap LN NM 0 0 0 0 0 0 0 0 LN NM 0 0 0 0 0 0 0 0 PostxAttributexIndUpGap LN NM 0.706** 0.745*** –0.951** –0.201 –0.598 –0.774** –0.505 –0.246 LN NM 0.322 0.189 0.442 0.282 0.372 0.332 0.375 0.854 PostxAttributexIndDownGap LN NM –0.400* –0.135 0.714** 0.423*** 0.617** 0.420* 0.030 0.396 LN NM 0.227 0.125 0.316 0.157 0.255 0.223 0.240 0.420 PostxAttributexCtyGap LN NM 0.504* 0.293 –0.100 0.157 –0.500** –0.208 –0.315 0.654 LN NM 0.257 0.300 0.382 0.257 0.203 0.297 0.298 0.426 PostxAttributexCtyUpGap LN NM 0.542 –0.780 –0.381 –1.492 –0.019 1.558* 1.180 –3.899*** LN NM 0.916 1.193 1.160 0.990 0.517 0.871 1.143 1.445 PostxAttributexCtyDownGap LN NM –0.649* –0.301 –0.294 0.065 0.722 –0.785 –0.426 1.549*** LN NM 0.355 0.599 0.604 0.381 0.481 0.582 0.559 0.594 PostxAttributexIndGap LN M –0.106* –0.068 –0.099 0.050 0.321*** –0.014 –0.077 0.025 LN M 0.058 0.044 0.063 0.070 0.105 0.068 0.075 0.073 PostxAttributexIndUpGap LN M 0.324 0.054 –0.073 0.290 –0.191 0.034 0.029 0.056 LN M 0.265 0.185 0.265 0.423 0.465 0.315 0.351 0.367 PostxAttributexIndDownGap LN M –0.055 –0.019 0.025 0.094 0.147 0.109 0.040 0.046 LN M 0.094 0.088 0.103 0.117 0.152 0.167 0.155 0.126 PostxAttributexCtyGap LN M –0.295 –0.369 –1.476*** 0.040 –0.648* –1.446*** –1.151*** 1.015** LN M 0.314 0.517 0.541 0.436 0.365 0.477 0.397 0.431 PostxAttributexCtyUpGap LN M –0.249 0.246 3.106 –1.196 1.087 8.798*** 7.884*** –7.360*** LN M 1.067 1.630 2.007 1.393 0.909 1.450 1.358 1.523 PostxAttributexCtyDownGap LN M 0.444 0.520 0.768 –0.868 –0.228 –0.883 –0.954 0.350 LN M 0.626 0.796 0.757 0.917 0.679 0.907 0.730 0.702 Source:LEHD,LBD,Feenstraetal.(2002),andauthors’calculations.TabledisplaystheDIDcoefficientsofinterestfortheOLSpanelestimationofequation5thatincludes tripleinteractionswithnotedinitial(1999)workerattributedummies. ***,**,and*representstatisticalsignificanceatthe1,5and10percentlevels. TableA.8reports F-statisticsforthejointsignificanceoftheseexposureterms,bygroup. 46

Table A.8: F-Statistics for Joint Significance of Triple Interaction DID Terms High-Tenure LHS NM M FemalevsMale CR 7.25*** 1.02 Non-WhitevsWhite CR 2.25** 1.65 AgeBelow30vsOlder CR 2.34** .25 BachelorsvsLess CR 1.14 1.28 HighestEarnervsLess CR 6.61*** 10.7*** SmallFirmvsLarger CR 1.01 2* TradingvsNon-TradingFirm CR .99 2.38** DiversifiedFirmvsM CR 1.19 1.32 FemalevsMale LN 4.27*** 1.14 Non-WhitevsWhite LN 2.82*** .54 AgeBelow30vsOlder LN 2.47** 3.27*** BachelorsvsLess LN 2.23** 2.57** HighestEarnervsLess LN 4.09*** 31.79*** SmallFirmvsLarger LN 2.82*** 6.56*** TradingvsNon-TradingFirm LN .81 6.99*** DiversifiedFirmvsM LN 2.02* 6.11*** FemalevsMale E>0 6.05*** 1.71 Non-WhitevsWhite E>0 1.43 1.18 AgeBelow30vsOlder E>0 1.89* .5 BachelorsvsLess E>0 2.69** 1.44 HighestEarnervsLess E>0 18.69*** 14.62*** SmallFirmvsLarger E>0 1.06 1.27 TradingvsNon-TradingFirm E>0 1.14 2.29** DiversifiedFirmvsM E>0 1.03 .73 Source: LEHD,LBD,Feenstraetal.(2002),andauthors’calculations. Table displays the F-statistics of the triple-interaction industry and county exposure terms for noted worker or firm characteristic. There aresixexposuretermsforMand5forNM. EachpanelreportsF-stats fortheearningstransformationnotedinthesecondcolumn: CR=Chen- Roth,LN=naturallog;andE>0=linearprobabilitymodelforearnings greaterthanzero. ***,**,and*representstatisticalsignificanceatthe 1,5and10percentlevels. 47