Trade Liberalization and Mortality: Evidence from U.S. Counties

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Trade Liberalization and Mortality: Evidence from U.S. Counties Justin R. Pierce and Peter K. Schott 2016-094 Please cite this paper as: Pierce, Justin R., and Peter K. Schott (2016). “Trade Liberalization and Mortality: Evidence from U.S. Counties,” Finance and Economics Discussion Series 2016-094. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2016.094. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) 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 Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Trade Liberalization and Mortality: Evidence from U.S. Counties∗ Justin R. Pierce† Board of Governors of the Federal Reserve System Peter K. Schott‡ Yale School of Management & NBER First Draft: November 2015 This Draft: November 2016 Abstract Weinvestigatetheimpactofalargeeconomicshockonmortality. We(cid:28)ndthat countiesmoreexposedtoaplausiblyexogenoustradeliberalizationexhibithigher ratesofsuicideandrelatedcausesofdeath,concentratedamongwhites,especially white males. These trends are consistent with our (cid:28)nding that more-exposed counties experience relative declines in manufacturing employment, a sector in which whites and males are disproportionately employed. We also examine other causes of death that might be related to labor market disruption and (cid:28)nd both positive and negative relationships. More-exposed counties, for example, exhibit lower rates of fatal heart attacks. ∗SchottthankstheNationalScienceFoundation(SES-1427027)forresearchsupport. Anyopinions andconclusionsexpressedhereinarethoseoftheauthorsanddonotnecessarilyrepresenttheviewsof the U.S. Centers for Disease Control, the Board of Governors or its research sta(cid:27). We thank Lorenzo Caliendo, Belinda Chan, Steve Redding and seminar participants at various institutions for helpful comments. †20th&CStreetsNW,Washington,DC20551,tel: (202)452-2980,email:justin.r.pierce@frb.gov. ‡165 Whitney Avenue, New Haven, CT 06511, tel: (203) 436-4260, email: peter.schott@yale.edu.

1 Introduction Large literatures in economics and public health investigate the e(cid:27)ect of economic 1 shocks on physical and mental health, and on mortality. Finding exogenous sources of variation in economic conditions is an important challenge in this research agenda. Here, we explore the relationship between mortality and a plausibly exogenous change inU.S.tradepolicy(cid:21)theOctober, 2000grantingofPermanentNormalTradeRelations (PNTR) to China (cid:21) that di(cid:27)erentially exposed U.S. counties to increased international competition via their industry structure. We (cid:28)nd that counties more exposed to the change in policy exhibit higher mortality due to causes of death, particularly suicide, that have been linked to relative loss of employment and income, which we also show to be associated with the change in policy. We (cid:28)nd that relative increases in these causes of death are concentrated among whites, which is consistent with that group’s disproportionatelyhighemploymentinmanufacturing, thesectormostdirectlya(cid:27)ected by the change in trade policy, as well as recent research into rising white mortality by Case and Deaton (2015). In principle, an increase in import competition can a(cid:27)ect U.S. workers’ health positively or negatively depending upon the sector in which they are employed and the region in which they and their dependents live. On one hand, health might improve with real income in areas where production lines up with U.S. comparative advantage, and health everywhere might improve with declines in the prices of goods and services thatareimportantinputstomedicalcare. Ontheotherhand, healthmightsu(cid:27)erinareas competing most directly with imports if workers experience sharper or longer-term declines in employment and real income. We note that while our analysis contributes to a broader understanding of the distributional implications of trade liberalization, it does not constitute an assessment of PNTR’s overall e(cid:27)ect on welfare. PNTR was a non-traditional trade liberalization in that it eliminated the threat of tari(cid:27) increases on U.S. imports from China without changing the tari(cid:27) rates actually applied to Chinese goods. Speci(cid:28)cally, PNTR eliminated the need for politically contentious annual renewals of China’s Normal Trade Relations (NTR) status (cid:21) and the accompanying uncertainty (cid:21) that was needed to preserve China’s access to the low NTR rates available to most other U.S. trading partners. Removing the possibility of these potential tari(cid:27) increases made producing in China for export to the United States more attractive, e(cid:27)ectively liberlizing trade between the two countries. We de- 1 See, for example, the survey by Cutler, Deaton and Lleras-Muney (2006). 1

(cid:28)ne industries’ exposure to PNTR as the di(cid:27)erence between the higher, non-NTR rates to which tari(cid:27)s could have risen prior to PNTR and the lower NTR rates that were locked in by the change in policy. We refer to these di(cid:27)erences as (cid:16)NTR gaps,(cid:17) and, importantly for our identi(cid:28)cation strategy, show both that they exhibit substantial variation across industries and that they are unrelated to employment outcomes prior to the change in policy. Indeed, nearly all of the variation in the NTR gap is accounted for by variation in non-NTR rates, which were set by the Smoot-Hawley Tari(cid:27) Act of 1930. We compute counties’ exposure to PNTR as the labor-share weighted average 2 NTR gaps of the industries they produce. We use proprietary microdata from the U.S. Centers for Disease Control (CDC) to compute mortality rates by county, year, cause of death, gender and race. Our initial focus is on three causes of death (cid:21) suicide, accidental poisoning (which includes drug overdoses) and alcohol-related liver disease (ARLD) (cid:21) that are highlighted in Case and Deaton (2015) and that a literature described below has found to be related to labor market disruptions. We then use a di(cid:27)erence-in-di(cid:27)erences (DID) identi(cid:28)cation strategytoexaminewhethercountiesthataremoreexpostedtoPNTR((cid:28)rstdi(cid:27)erence) experience di(cid:27)erential changes in mortality and labor market outcomes after the policy is implemented (second di(cid:27)erence). We (cid:28)nd that PNTR is associated with a statistically signi(cid:28)cant relative increase in suicide, and that this result is robust to inclusion of county-level demographic and economic control variables. Coe(cid:30)cient estimates imply that an interquartile shift in counties’ NTR gaps is associated with an increase in the annual suicide rate of 4.0 percent relative to its respective average in the year 2000, the year of the change in U.S. trade policy. Across age and racial groups, we (cid:28)nd that the relationship between PNTR and suicide is concentrated among white males. We also (cid:28)nd that PNTR is associated with statistically signi(cid:28)cant relative increases in mortality from accidental poisoning, though those estimates are more sensitive to the set of county attributes included in the regression and to its speci(cid:28)cation. The evidence for a link between the policy change and mortality from ARLD is mixed, perhaps due to the longer onset 2 Pierce and Schott (2016) show that PNTR is associated with the sharp decline in U.S. manufacturing employment that began around time of its passage, as well as with increases in U.S. imports fromChina, thenumber ofdomesticand foreign-ownedChinese(cid:28)rms exportingtothe United States, the number of U.S. (cid:28)rms importing from China and the number of trading relationships between the two countries. Handley and Limao (2016) develop a theoretical model that indicates that PNTR was equivalent to a 13 percentage point permanent decline in tari(cid:27) rates. Feng, Li and Swenson (2016) document the e(cid:27)ect of PNTR on the prices and quality of goods exported to the U.S. by Chinese (cid:28)rms. 2

associated with fatal liver diseases. We examine the robustness of the DID results in several ways. First, we consider an alternate empirical speci(cid:28)cation that places no restrictions on the timing of the e(cid:27)ects of the policy change and verify that the relationship between the NTR gap and mortality from suicide for white males is only present after PNTR was passed. Second, we demonstrate that the statistical and economic signi(cid:28)cance of our results are similar after accounting for the NTR gaps of other counties within the surrounding commuting zone and controlling for potential changes in state healthcare policies. Finally, we show that PNTR has no association with other causes of death plausibly unrelated to the change in policy (e.g., deaths of unknown intent), and that it is negatively associated with fatal heart attacks, perhaps due to loss of employment in industries requiring 3 physically strenuous activity. ToverifylabormarketdisruptionasapotentialchannelthroughwhichPNTRmight a(cid:27)ect mortality, we estimate the relationship between the NTR gap and several labor market outcomes. Results from a DID speci(cid:28)cation analogous to the one described above imply that an interquartile shift in counties’ NTR gaps is associated with persistent relative increases in counties’ unemployment rates and persistent relative declines in counties’ manufacturing employment, overall employment, labor force participation 4 ratesandpercapitapersonalincome. These(cid:28)ndingssuggestthatPNTR’srelationship with mortality rates may occur at least in part via the impact of import competition on local labor markets. Finally, as a further check on the mechanism linking PNTR to mortality and to facilitate comparison of our results with those already in the literature, we perform a series of two-stage least squares estimations of county mortality rates on county unemploymentrates, usingcounties’exposuretothechangeintradepolicyasaninstrument. The resulting coe(cid:30)cient estimates suggest that a 1 standard deviation increase in the unemployment rate (2.6 percentage points) is associated with a 29.1 percent increase in the suicide rate. This elasticity is approximately an order of magnitude greater than that estimated in Ruhm (2000), which (cid:28)nds that a 1 standard deviation increase in 3 Ruhm (2000) reports a negative relationship between the unemployment rate and death due to heart attacks. Hummels, Munch and Xiang (2016) show that increased e(cid:27)ort in manufacturing jobs resulting from positive export shocks is associated with a higher rates of hospitalization due to heart attacks. 4 Complementary evidence indicating the seriousness of the labor market disruption is reported in our online appendix, where we document a relationship between PNTR and relative increases in property crime. Unlike Dix-Carneiro et al. (2015), however, we (cid:28)nd no relationship between the change in trade policy and the murder rate, or other forms of violent crime. 3

the state unemployment rate (2.1 percentage points) is associated with a 2.7 percent 5 increase in the suicide rate. Ouranalysis contributes toresearchin severalliteratures. First, the linkwe(cid:28)nd between PNTR and suicide relates to a series of papers studying the health consequences of unemployment. Two seminal contributions in this literature are Ruhm (2000), which reports a positive relationship between the unemployment rate and suicide in a panel of U.S. states, and Sullivan and von Wachter (2009), which (cid:28)nds that high-tenure workers displaced as part of a mass layo(cid:27) experience a sharp increase in their probability of 6 death. More recently, Classen and Dunn (2011) (cid:28)nd that unemployment duration is a major force in the relationship between job loss and suicide. Second, our analysis contributes to the substantial body of research examining the relationship between import competition and employment, particularly with respect 7 to China. Autor, Dorn and Hanson (2013), for example, (cid:28)nd that up to half of the decline in U.S. manufacturing employment between 2000 and 2007 is associated with rising imports from China, while Pierce and Schott (2016) show that both a decline in manufacturing employment and an increase in U.S. imports from China during this period are related to PNTR. Complementary research demonstrates that U.S. labor markets subject to larger increases in Chinese import competition experience greater declines in self-reported health outcomes (McManus and Schaur 2015a,b), reduced provision of local public goods (Feler and Senses 2015), and changes in marriage and fertility patterns (Autor, Dorn and Hanson 2015). Hummels, Munch and Xiang (2016), by contrast, (cid:28)nd that the increased job e(cid:27)ort associated with positive export demand shocks increases rates of illness and injury for Danish workers, and Bombardini and Li (2016) (cid:28)nd that higher pollution associated with expanded export production is associated with a substantial increase in infant mortality. 5 The di(cid:27)erence in magnitudes may arise from the persistence of the employment e(cid:27)ect associated with the trade shock (cid:21) as opposed to the transitory nature of a business cycle contraction (cid:21) or our focusoncounty-levelvariationinmortalityandunemploymentrates,ratherthanstate-levelvariation as in Ruhm (2000). 6 Potential reasons for the increase in mortality discussed by Sullivan and von Wachter (2009) include reduced investments in health, increased stress, and loss of health insurance. Browning and Heinesen(2012)(cid:28)ndthatworkersdisplacedbyplantclosuresinDenmarkexhibitelevateddeathrates due to mental illness, suicide and alcohol-related diseases, particularly in the short run. A number of papers in the public health literature, including Falba et al. (2005) and Deb et al. (2011), (cid:28)nd that workers facing job loss are more likely to engage in unhealthy activities. 7 See,forexample,FreemanandKatz(1991),Revenga(1992),SachsandShatz(1994)andBernard, JensenandSchott(2006). Recentpapersfocusedspeci(cid:28)callyonChinaincludeBloom,DracaandVan Reenen (2015), Ebenstein et al. (2014b), Groizard, Ranjan and Rodriguez-Lopez (2012), Mion and Zhu (2013), and Utar and Torres Ruiz (2013). 4

Finally, our research contributes to studies that take advantage of (cid:16)natural(cid:17) or actual experiments to examine the impact of shocks to healthcare coverage. Two such papers focus on the random allocation of Medicaid coverage in Oregon. Baicker et al. (2013) (cid:28)nd that coverage signi(cid:28)cantly increases use of preventative services, the probability of a positive screening for depression and diabetes and the use of diabetes medication. Finkelstein et al. (2012) (cid:28)nd that coverage leads to better self-reported physical and mental health. To the extent that the labor market disruptions associated with PNTR a(cid:27)ect access to healthcare, these (cid:28)ndings are consistent with the positive links we (cid:28)nd between the change in trade policy and mortality. The paper proceeds as follows: Section 2 describes the data, Section 3 describes our empirical strategy and mortality results, Section 4 explores mechanisms that might explain the results, Section 5 presents the two stage least squares estimates, and Section 6 concludes. An online appendix provides additional empirical results as well as information about dataset construction and sources. 2 Data 2.1 County Level Mortality Wecalculatethenumberofdeathsbycounty,demographiccategoryandcauseusingthe proprietary (cid:16)compressed all-county mortality (cid:28)les(cid:17) available by petition from the U.S. Centers for Disease Control (CDC). These data summarize all death certi(cid:28)cates (cid:28)led in 8 the United States from 1990 to 2013. Observable demographics include the deceased’s age, gender, race, county of residence and county of death. Underlying causes of death 9 are classi(cid:28)ed according to one of several hundred (cid:16)external(cid:17) or (cid:16)internal(cid:17) categories. Internal causes of death are de(cid:28)ned as those that originate within the body (e.g., liver disease) and external causes of death are de(cid:28)ned as those whose origins lie outside the body (e.g., suicide or accidental poisoning). We match year by county of residence by age by gender by race death counts to cor- 8 A public-use version of these data can be accessed at www.wonder.cdc.gov, though the extent to which mortality rates can be examined within causes of death and demographic groups over time is limited to prevent disclosure of con(cid:28)dential information. 9 Causes of death are classi(cid:28)ed according to International Classi(cid:28)cation of Diseases (ICD). The CDC data use version 10 of these codes (ICD-10) from 1999 to 2013 and version 9 (ICD-9) of these codes from 1990 to 1998. We make use of a concordance between these underlying codes and major disease categories available in Anderson et al. (2001). 5

responding population estimates compiled by the National Cancer Institute’s Surveil- 10 lance, Epidemiology and End Results (SEER) Program. We use these population estimates to compute both (cid:16)crude(cid:17) and (cid:16)age-adjusted(cid:17) mortality rates, conventionally expressed per 100,000 population. The crude death rate for a county-year is simply the total number of deaths in that county in that year divided by its total population in that year. The age-adjusted death rate for a county, by contrast, is a weighted average of the crude death rates across age categories within a county, where the shares of the 11 overall U.S. population in each age category are used as weights. We use the U.S. 12 population shares in the year 2000 for constructing age-adjusted mortality rates. Figure 1 plots the (censored) distributions of age-adjusted mortality rates across counties at four-year intervals starting in 1990. This (cid:28)gure conveys two messages. First, the leftward movement in the distributions over time indicates that overall U.S. mortality rates decline during our sample period. Second, the relatively wide support ofeachdistributionrevealsthatmortalityratesvarysubstantiallyacrosscounties. This across-countyvariationinmortalityratesisalsoapparentinTable1, whichsummarizes counties’ population-weighted average mortality rates by gender and by race for the year 2000. As indicated in the (cid:28)rst row of the table, the overall mortality rate across counties is 858, with an interquartile range stretching from 778 to 1002. The remaining rows show that mortality is higher for males than females (1047 versus 719), and higher 13 for blacks than for other racial groups. Table 2 reports the year 2000 population-weighted average age-adjusted death rates per 100,000 population for major external and internal causes of death across counties. As noted in the table, internal causes account for more than 90 percent of deaths across racial groups. The three leading causes of death are cancer, circulatory disease and respiratory ailments. 10 Eighty-onepercentofdeathsoccurinthedeceased’scountyofresidence,thefocusofouranalysis. SEER population estimates are available at http://seer.cancer.gov/popdata/download.html. 11 We use the following age categories in our baseline results: less than 1 year old, 1 to 4 years, 5 to 14 years, 15 to 19 years, 20 to 24 years..., 80 to 84 years, and greater than 85 years. 12 The SEER population weights associated with these categories are provided in Table A.1 of the online appendix. 13 In 2000, the U.S. population shares representing males, females, whites, blacks, American Indian and Asians or Paci(cid:28)c Islanders are 49, 51, 82, 13, 1 and 4 percent. Counties’ weighted average death ratesvarydependingonwhetheroverallcountypopulationordemographic-speci(cid:28)ccountypopulation isusedtoweighteachcounty. Thissensitivitycanbeseenbycomparingthe(cid:28)rsttwocolumnsofTable A.2intheonlineappendix. The(cid:28)rstcolumn(likeTable1inthemaintext)usestotalpopulationwhile the second column uses the population speci(cid:28)c to the demographic group whose weighted average is beingcomputed. Asindicatedinthetable,thelatterareclosertotheo(cid:30)cialCDCage-adjusteddeath rates for the United States as a whole. 6

Examining overall U.S. mortality rates by cause of death and demographic categories, Case and Deaton (2015) highlight a substantial rise in deaths due to suicide, chronic liver disease and poisoning (accidental and intent undetermined) among middle-aged whites starting in 1999. Figure 2 uses the CDC microdata examined here to demonstrate these trends and extend them backwards in time to the beginning of our sample period (1990), where to faciliate concordance with data from the 1990s, we focus more speci(cid:28)cally on suicide, alcohol-related liver disease (ARLD) and accidental poisoning. As indicated in the (cid:28)gure, the weighted average rates of suicide and ARLD across counties are more or less (cid:29)at during the 1990s but begin increasing around the time of the change in U.S. trade policy in the year 2000. Deaths due to accidental poisoning, by contrast, rise throughout the sample period but increase at a faster rate 14 during the 2000s. 2.2 The NTR Gap Our analysis makes use of a plausibly exogenous change in U.S. trade policy (cid:21) the U.S. granting of PNTR to China in October 2000 (cid:21) that e(cid:27)ectively liberalized U.S. imports from China. This impact can be understood by considering the two sets of tari(cid:27) rates that comprise the U.S. tari(cid:27) schedule. The (cid:28)rst set of tari(cid:27)s, known as NTR tari(cid:27)s, are generally low and applied to goods imported from other members of the World Trade Organization (WTO). The second, known as non-NTR tari(cid:27)s, were set by the Smoot- Hawley Tari(cid:27) Act of 1930 and are often substantially higher than the corresponding NTR rates. Imports from non-market economies such as China generally are subject to the higher non-NTR rates, but U.S. law allows the President to grant such countries access to NTR rates on a year-by-year basis subject to annual approval by Congress. U.S. Presidents granted China such a waiver every year starting in 1980, but Congressional votes over annual renewal became politically contentious and less certain of passage following the Chinese government’s crackdown on Tiananmen Square protests in 1989 and other (cid:29)ashpoints in U.S.-China relations during the 1990s such as China’s transfer of missile technology to Pakistan in 1993 and the Taiwan Straits Missile Crisis in 1996. Uncertainty over China’s access to NTR tari(cid:27) rates ended with Congress passing a bill granting PNTR status to China in October 2000, which formally took 14 One commonly cited explanation for the increase in death due to poisoning around the year 2000 is an increase in the misuse of prescription opioid painkillers. See, e.g., Rudd, Aleshire, Zibbel and Gladden (2016). We hope to explore a potential link between such prescriptions and labor market shocks in future drafts of this paper. 7

e(cid:27)ect upon China’s entry into the WTO in December 2001. We follow Pierce and Schott (2016) in measuring the impact of PNTR as the rise in U.S. tari(cid:27)s on Chinese goods that would have occurred in the event of a failed annual renewal of China’s NTR status prior to PNTR, NTRGap = NonNTRRate −NTRRate . (1) j j j We refer to this di(cid:27)erence as the NTR gap, and compute it for each SIC industry j using ad valorem equivalent tari(cid:27) rates provided by Feenstra et al. (2002) for 1999, the year before passage of PNTR. NTR gaps vary widely across industries, with a mean and standard deviation of 33 and 15 percentage points. As noted in Pierce and Schott (2016), 79 percent of the variation in the NTR gap across industries is due to variation in non-NTR rates, set 70 years prior to passage of PNTR, while less than 1 percent of variation is due to variation in NTR rates. This feature of non-NTR rates e(cid:27)ectively rules out reverse causality that would arise if non-NTR rates were set to 15 protect industries with declining employment or surging imports. We compute U.S. counties’ exposure to PNTR as the employment-share weighted average NTR gap across the sectors in which they are active, (cid:88) L1990 jc NTR Gap = NTRGap . (2) c L1990 j j c 16 We use employment shares from 1990, a period well before the change in policy. NTR gaps are de(cid:28)ned only for industries whose output is subject to U.S. import tari(cid:27)s, primarily in the manufacturing and agricultural sectors. For industries whose output is not subject to tari(cid:27)s, such as service industries, we set NTR gaps to zero. For each county, we also calculate the population weighted average NTR gap of the remaining counties in its commuting zone, NTR Gap . 17 cz 15 Furthermore, to the extent that NTR rates were set to protect industries with declining employment prior to PNTR, these higher NTR rates would result in lower NTR gaps, biasing our results away from (cid:28)nding an e(cid:27)ect of PNTR. 16 EmploymentbycountyandindustryareavailablefromtheU.S.CensusBureau’sCountyBusiness Patterns(CBP)database,availableathttp://www.census.gov/econ/cbp/download/. WefollowAutor etal. (2013)inimputingemploymentforcountieswhereonlyarangeofemploymentisreported. For more information, see David Dorn’s data page, at http://www.ddorn.net/data.htm. 17 We use the U.S. Department of Agriculture de(cid:28)nition of commuting zones as of 1990 (Tolbert andSizer1996)andtheconcordanceofcountiestocommutingzonesprovidedbyAutoretal. (2013). The counties in our sample are distributed across 741 commuting zones, with the number of counties per commuting zone ranging from 1 to 19 (the Washington DC area). 8

Figure 3 reports the distribution of NTR gaps across four-digit SIC industries, U.S. counties and U.S. counties’ surrounding commuting zones. Relative to the distribution across industries, the distributions for counties and surrounding labor market areas are shifted towards the left, re(cid:29)ecting the fact that most workers in most counties 18 are employed outside the manufacturing sector. Own-county NTR gaps average 7.3 percent and have a standard deviation of 6.5 percent, with an interquartile range from 2.4 to 10.6 percent, or 1.3 standard deviations. Surrounding-county NTR gaps have a similar distribution, with a mean and standard deviation of 6.5 and 4.8 percent, and an interquartile range from 3.3 to 8.8 percent, or 1.1 standard deviations. 2.3 Other Policy Variables Our empirical analysis controls for four additional variables that capture changes in U.S. or Chinese policy: the average U.S. import NTR tari(cid:27) rate associated with the goods produced by each county; the average exposure of the county to the end of quantitative restrictions on textiles and clothing imports associated with the phasing out of the global Multi-Fiber Arrangement (MFA); and changes in Chinese import tari(cid:27)s and domestic production subsidies. NTR Rates: Counties’ labor-share weighted U.S. import tari(cid:27) rates, NTR , are ct computed as in Equation 2, except that the U.S. NTR tari(cid:27) rate for industry j (in percent) is used in place of the NTR gap for industry j. The left panel of Figure A.2 in the online appendix summarizes the distribution of NTR across our sample period; ct as shown in the (cid:28)gure, it declines during the late 1990s due to implementation of tari(cid:27) 19 reductions agreed upon during the Uruguay Round. MFA Exposure: We measure counties’ exposure to the end of the MFA analogously. As discussed in greater detail in Khandelwal et al. (2013), 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 18 The distribution for industries in Figure 3 omits SIC industries that that are not imported and which therefore have NTR gaps of zero by de(cid:28)nition. 19 NTRtari(cid:27)ratesfromFeenstraetal. (2002)areunavailableafter2001andsoareassumedconstant after that year. Analysis of analogously computed (cid:16)revealed(cid:17) tari(cid:27) rates from public U.S. trade data during this interval in Pierce and Schott (2016) suggests this is an reasonable assumption that avoids having to make do with the smaller set of industries for which (cid:16)revealed(cid:17) rates are available. 9

Turkey. Bargaining over these restrictions was kept separate from multilateral trade negotiations until the conclusion of the Uruguay Round in 1995, when an agreement wasstrucktoeliminatethequotasoverfourphases. OnJanuary1,1995,1998,2002and 2005, theUnitedStateswasrequiredtoremovetextileandclothingquotasrepresenting 16, 17, 18 and the remaining 49 percent of their 1990 import volumes, respectively. 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 (cid:28)rst 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. Computation of counties’ exposure to elimination of the MFA proceeds in three steps. First, we follow Brambilla et al. (2010) in measuring the extent to which MFA quotas in industry j and phase p were binding as the import-weighted average (cid:28)ll rate of the industry’s constituent import products in the year before they were phased out, FillRate . 20 Industries with higher average (cid:28)ll rates faced more binding quotas jp and are therefore more exposed to the end of the MFA. Second, for each phase, we compute counties’ labor-share weighted average (cid:28)ll rate across industries, FillRate , cp using a version of Equation 2. Finally, we create our county-year variable of interest, MFAExposure , which, for each year t, is county c’s the weighted avearge FillRate ct cp for industries whose quotas were relaxed in the most recent phase. The right panel of Figure A.2 in the online appendix summarizes the distribution of FillRate across cp our sample period. As shown in the (cid:28)gure, (cid:28)ll rates are zero until the second phaseout, in 1998. They then step up in 2002 and again in 2005, consistent with the hypothesis in Brambilla et al (2010) that the United States placed its more (cid:16)sensitive(cid:17) textile and clothing products into the latter two phases as a means of deferring politically painful import competition as long as possible. Changes in Chinese Policy: As part of its accession to the WTO, China agreed to institute a number of policy changes that could have in(cid:29)uenced U.S. manufacturing employment, notably liberalization of its import tari(cid:27) rates and reductions of production subsidies. Following Pierce and Schott (2016) we use product-level data on Chinese import tari(cid:27)s from 1996 to 2005 from Brandt et al. (2012) to compute the average change in Chinese import tari(cid:27)s across products within each U.S. industry. For production subsidies, we use data from the Annual Report of Industrial Enterprise 20 Fill rates are de(cid:28)ned as actual imports divided by allowable imports under the the quota, and products outside the MFA have a (cid:28)ll rate of zero. 10

Statistics compiled by China’s National Bureau of Statistics (NBS), which reports the 21 subsidies provided to responding (cid:28)rms. Following Girma et al. (2009) and Aghion et al. (2015) we use the variable (cid:16)subsidy(cid:17) in this dataset to compute the change in the subsidies to sales ratio for each SIC industry between 1999 and 2005 using concordances provided by Dean and Lovely (2010). For both changes in Chinese import tari(cid:27) rates and production subsidies, we then compute the labor-share weighted average of this change across the industries each U.S. county produces. Figure A.3 in the online appendix summarizes the distribution of counties’ exposure to reductions in Chinese import tari(cid:27)s (left panel) and domestic production subsidies (right panel) 2.4 County Demographic Information Our baseline speci(cid:28)cations control for interactions of a post-PNTR indicator variable withthreeinitial-year(i.e., 1990)countyattributes: thepercentofthepopulationwithoutanycollegeeducation, medianhouseholdincomeandpercentofpopulationthatare veterans. These variables allow for the possibilities, respectively, that spurious changes in technology might have replaced low-skill workers with technology disproportionately during the 2000s, that high-income households gained better access to medical care after the 2000s, perhaps due to health insurance provided by their employers, and that an increase in destructive behaviors such as suicide might be the result of combat experience associated with post-9/11 wars in Afghanistan and Iraq (Kemp and Bossarte 2012). These attributes, summarized in Table 1, are obtained from the U.S. Census 22 Bureau’s 1990 Decennial Census. As noted in the table, the unweighted means and standard deviations across counties are 54.4 and 11.4 percent (share of population with no college education), 40.4 and 10.6 thousand dollars (median household income), and 23 14.4 and 2.4 percent (percent of population that are veterans, respectively. Table 3 reports the results of OLS regressions of counties’ NTR gaps on the inital (1990) county demographic attributes discussed in this section. As indicated in the table, counties with higher NTR gaps have greater exposure to the MFA, higher import tari(cid:27)s across the goods they produce, are exposed to larger reductions in Chinese 21 TheNBSdataencompassacensusofstate-ownedenterprises(SOEs)andasurveyofallnon-SOEs with annual sales above 5 million Renminbi (~$600,000). The version of the NBS dataset available to us from Khandelwal, Schott and Wei (2013) spans the period 1998 to 2005. 22 These data can be downloaded from the Dexter Data Extractor at the University of Missouri, available at http://mcdc.missouri.edu/. 23 These values di(cid:27)er from national averages as they are more a(cid:27)ected by counties with small populations. 11

imports tari(cid:27)s and subsidies, have lower household incomes in 1990, lower share of population with a college education in 1990, and a higher share of the population that are veterans in 1990. Counties with higher NTR gaps have lower median household income in 1990. 3 PNTR and County Mortality Rates 3.1 DID Identi(cid:28)cation Strategy Our baseline di(cid:27)erence-in-di(cid:27)erences (DID) speci(cid:28)cation examines whether counties with higher NTR gaps ((cid:28)rst di(cid:27)erence) experience di(cid:27)erential changes in mortality after the change in U.S. trade policy (second di(cid:27)erence) versus before, DeathRate = θPostPNTR × NTRGap + (3) ct t c βX +γPostPNTR ×X + ct t c δ +δ +ε , c t ct 24 The sample period is 1990 to 2013. The left-hand side variable represents an outcome in county c, for example the age-adjusted death rate for a particular cause of death and demographic group in year t. The (cid:28)rst term on the right-hand side is the DID term of interest, an interaction of a post-PNTR (i.e., t > 2000) indicator with the (time-invariant) county-level NTR Gap. X represents the two additional, timect varying controls for policy discussed in Section 2.3: the overall U.S. import tari(cid:27) rate associated with the sectors produced by the county (NTR ) and the sensitivity of the ct county to the phasing out of the global Multi-Fiber Arrangement (MFAExposure ). ct X represents the two Chinese policy variables, exposure to changes in Chinese tarc i(cid:27)s between 1996 and 2005 and exposure to changes in Chinese domestic production subsidies between 2000 and 2005, and the three initial-period county attributes, 1990 median household income, 1990 share of population without a college degree and 1990 share of population that are veterans. Including interactions of these attributes with the PostPNTR indicator allows their relationship with mortality rates to di(cid:27)er before t 24 The baseline results discussed below are robust to ending the sample period in 2007, the year before the onset of the Great Recession. 12

and after passage of PNTR. δ and δ represent county and year (cid:28)xed e(cid:27)ects. Inclusion c t of these (cid:28)xed e(cid:27)ects nets out characteristics of counties that are time-invariant, such as whether they are near the coast or inland, while also controlling for aggregate shocks that a(cid:27)ect all counties identically in a particular year. An attractive feature of these DID identi(cid:28)cation strategies is their ability to isolate the role of the change in U.S. trade policy. While counties with high and low NTR gaps are not identical, comparing outcomes within counties over time isolates the di(cid:27)erential impact of China’s change in NTR status. 3.2 Baseline DID Estimates for Suicide, Accidental Poisoning and ARLD This section examines the link between PNTR and three speci(cid:28)c causes of death (cid:21) suicide, accidental poisoning and alcohol-related liver disease (ARLD). We focus on these causes of death for several reasons: they are highlighted in Case and Deaton (2015); theyarefoundtobeimportantintheunemploymentandmass-layo(cid:27)literatures (e.g., Classen and Dunn 2011 and Browning and Heinesen 2012); their concordance acrossthecause-of-deathcodingschemesusedbytheCDCovertimeisstraightforward; and they may be more easily observable than other forms of death, particularly in the 25 case of suicide and accidential poisoning. Results from estimation of Equation 3 for suicide are reported in the (cid:28)rst four columns of Table 4, with standard errors clustered at the county level. The (cid:28)rst column reports coe(cid:30)cient estimates for a speci(cid:28)cation containing just the DID term of interest and the (cid:28)xed e(cid:27)ects. The second and third columns, respectively, add controls for policy changes and demographic variables. The fourth column includes the full set of controls. As indicated in the table, the DID point estimates of interest for suicide are positive andstatisticallysigni(cid:28)cantatconventionallevelsacrossallfourspeci(cid:28)cations, declining in magnitude as additional covariates are included in the regression. We assess the economic signi(cid:28)cance of the DID estimates of interest by computing the change in the mortality rates associated with moving a county from the 25th percentile to the 75th 25 There is reason to believe that information on death certi(cid:28)cates’ cause of death may be noisy. Kircher et al. (1985), for example, (cid:28)nds that 29 percent of 272 randomly selected autopsy reports and corresponding death certi(cid:28)cates in Connecticut in 1980 exhibit a major disagreement. The (cid:16)blue form(cid:17) instructions for completing the cause of death section of a death certi(cid:28)cate are available at http://www.cdc.gov/nchs/data/dvs/blue_form.pdf. 13

percentile of the NTR gap distribution (i.e., from 2.3 to 10.6 percent, or 1.3 standard deviations). As indicated in the bottom panel of the table, the implied increases in mortality under this counterfactual range from 0.63 = [0.089*(10.6-2.3)] per 100,000 in column 1 to 0.42 per 100,000 in column 4. These changes represent 6.0 and 4.0 percent of the of the average age-adjusted suicide mortality rates across counties in the 26 year 2000 which, as reported in the penultimate row of the table, is 10.51. In terms of the other control variables, coe(cid:30)cient estimates in column 4 indicate that counties with higher shares of the population that did not attend college and higher shares of veterans in the population experience larger increases in mortality from suicide in the post-PNTR period, relative to before. Larger declines in Chinese production subsidies are associated with lower mortality from suicide, post-PNTR. In this sense, a liberalization on the part of a U.S. trading partner is associated with a decline in mortality from suicide in the U.S. The next eight columns of Table 4 focus on accidental poisoning and ARLD. For accidental poisoning, the DID terms of interest are positive and signi(cid:28)cant in three of the four speci(cid:28)cations, including the speci(cid:28)cation that includes the full set of control variables. We do not (cid:28)nd a statistically signi(cid:28)cant relationship between PNTR and accidental poisoning when only demographic variables are included. In terms of economic signi(cid:28)cance, the impact of an interquartile shift in counties’ exposure to PNTR implied by the DID coe(cid:30)cient estimate from the speci(cid:28)cation with all controls is an increase of 27.7 percent for accidental poisoning, relative to the overall mortality rate 27 from that cause in the year 2000. By contrast, we do not (cid:28)nd a relationship between PNTR and ARLD in the full speci(cid:28)cation, possibly due to the longer onset period associated with the disease. 3.3 Baseline DID Estimates by Gender, Race and Age We examine the link between PNTR and suicide across genders and races in Table 5. We(cid:28)ndthatthepositiverelationshipbetweenPNTRandsuicideoverallisconcentrated 26 The relationship between PNTR and suicide might spuriously relate to changes in access to (cid:28)rearms across counties that occurs at the same time as the change in trade policy. Re-estimation of therelationshipaccordingtowhetherornotthesuicidesinvolvea(cid:28)rearm, however, revealsapositive and statistically insigni(cid:28)cant association at conventional levels for the former (implied impact and standard error of 0.019 and 0.013) and a positive and statistically signi(cid:28)cant association with respect to the latter (implied impact and standard error of 0.037 and 0.010). 27 The mortality rate from accidental poisoning in 2000 was 4.59 per 100,000 and 4.39 per 100,000 for ARLD. 14

inoneracialgroup(cid:21)whites(cid:21)andthatthislinkisstatisticallysigni(cid:28)cantatconventional 28 levels only for white males (p-value for white females 0.14). By contrast, we (cid:28)nd no relationship between PNTR and suicide for blacks, Asians, or American Indians. Results in Tables 6 and 7 indicate that PNTR also is associated with higher white mortality due to both accidental poisoning and ARLD. For accidental poisoning, this relationship is present for both white men and white women, while for ARLD it is 29 statistically signi(cid:28)cant for white men, but not for white women. There is generally no relationship between PNTR and mortality from ARLD or accidental poisoning for blacks, Asians, or American Indians, though for ARLD, we (cid:28)nd negative and statistically signi(cid:28)cant relationships for American Indian as well as Asian or Paci(cid:28)c Islander 30 females. Overall, the results in Tables 5, 6 and 7 provide context for the (cid:28)ndings of Case and Deaton (2015), who report a worsening of trends in mortality rates from suicide, poisoning and chronic liver disease rates among whites relative to other races. One potential explanation for the link between PNTR and white mortality (cid:21) particularly white male mortality (cid:21) is this group’s disproportionate representation among manufacturing workers, the group most directly a(cid:27)ected by exposure to PNTR. As indicated in Table A.6 of the online appendix, males accounted for 68 percent of U.S. manufacturing employment versus 49 percent of the population in 1999, and whites represented84.3percentofmanufacturingemploymentversus81.7percentofthepopulation. Moreover, withinmanufacturing, over-representationofwhitesishighestamong occupationslikelytobeearningthehighestwages(cid:21)suchasmanagerialandprofessional 31 occupations (cid:21) that might lead to largest declines in income following job separation. To examine how the above relationships between PNTR and mortality vary by age, 28 As indicated in the bottom panel of Table 5, the implied impact of an interquartile shift in the county-level NTR gap is an increase in deaths by suicide 4.8 percent of the year 2000 level for white males. 29 As indicated in the bottom panels of Tables 6 and 7, the implied impacts of an interquartile shift in the county-level NTR gap is an increase in deaths rates of 59 and 14 percent for white males and females for accidental poisoning, and of 6.7 percent for white males for ARLD. 30 Estimates for the American Indian and Asian populations may be noisy due to their small size andrelativelyunevendistributionacrosscounties. TheAmericanIndianandAsianorPaci(cid:28)cIslander populationsrepresent1.1and4.2percentoftheoverallpopulationintheyear2000. Inthatyear,these two groups have populations exceeding 50,000 in 48 and 158 counties, respectively, versus 2290 and 514 counties for whites and blacks. As reported in Figure A.4 of the online appendix, the American Indian and Asian populations also tend to inhabit counties with relatively low NTR gaps. 31 Ebenstein et al. (2014a,b), for example, (cid:28)nd that workers displaced from manufacturing on average experience wage declines in moving to another sector. As reported in Table A.6 of the online appendix, whites accounted for 90.4 percent of managers and professionals, 86.3 percent of technical, sales, administrative and service positions, and 83.0 percent of precision production positions, versus 78.9 percent among operators, fabricators, laborers and other occupations 15

we examine crude death rates by gender, race and age category. Results are displayed visually in Figure 4, which reports the 95 percent con(cid:28)dence intervals of the implied impact of an interquartile shift in counties’ exposure to PNTR on white males (left panels) and white females (right panels) for each cause of death. For comparison, the (cid:28)rstbarineach(cid:28)gurereproducesthe95percentcon(cid:28)denceintervalacrossallagesfrom thebottompanelofTables5,6and7. Asindicatedinthe(cid:28)gure,anassociationbetween PNTR and suicide is evident across several (cid:28)ve-year age bins between ages 20 and 54 32 for males, but is not statistically signi(cid:28)cant for white females in any age category. For accidental poisoning, the association between PNTR and mortality is positive and signi(cid:28)cant for both white males and white females in most age groups through 45 to 49. Finally, PNTR-related ARLD mortality is spread across most working year age bins for middle-age white males, and not evident in any age category for females. 3.4 Robustness Exercises This section describes three exercises that examine the robustness of the baseline DID results reported in the previous section. First, we use a more (cid:29)exible DID speci(cid:28)cation to examine the timing of the post-2000 changes in mortality and test for the possibility of prior trends in mortality among counties with varying exposure to PNTR. Second, we explore the e(cid:27)ect of the inclusion of additional covariates and (cid:28)xed e(cid:27)ects. And third, we examine the relationship between PNTR and other causes of death. Prior Trends and Timing: For the increase in mortality to be attributable to the change in U.S. trade policy, the NTR gap should be correlated with mortality rates after PNTR but not before. To examine whether this is the case, we estimate a version of Equation 3 that interacts the time-invariant county-level NTR gap and other county attributes with an indicator variable for each year, 32 Gemmill et al. (2016) (cid:28)nd that macroeconomic shocks appear to induce suicide among working age males, as opposed to simply moving suicides forward in time. 16

(cid:88) DeathRate = θ 1{year = t} × NTRGap + (4) ct t c t βX + ct (cid:88) γ 1{year = t} × X + t c t δ +δ +ε , c t ct Results for suicide, accidental poisoning and ARLD among white males and females are displayed visually in Figure 5. Each panel of the (cid:28)gure uses the estimated DID parameters of interest (θ ) for a particular regression to display the 95 percent t con(cid:28)dence interval associated with an interquartile shift in counties’ NTR gaps. As indicated in the (cid:28)gure, the implied impact of PNTR for suicide is generally statistically indistinguishable from zero prior to the change in U.S. trade policy but shifts upward after it is implemented. This upward shift is most clearly evident for suicide by white males in the top left panel, where the con(cid:28)dence interval for the implied impact of PNTR remains above zero after 2003. The con(cid:28)dence interval for the implied impact of PNTR on mortality from accidental poisoning for white women (cid:21) shown in the middle right panel (cid:21) is statistically indistinguishable from zero prior to 2000 but becomes positive and signi(cid:28)cant after PNTR. The equivalent con(cid:28)dence interval for white males shows a similar shift up around the time of passage of PNTR, though it is negative and statistically signi(cid:28)cant through most of the 1990s. The lower two panels of Figure 5 report results for alcohol-related liver disease for white males and females. Here, too, an upward shift is discernible for white males, though it is not statistically di(cid:27)erent from zero in any year of the sample period. For white females, there is no discernible shift. Surrounding Commuting Zones: Residents of a particular county may be a(cid:27)ected by PNTR via its impact on surrounding counties that are part of the same labor market. To account for this possibility, we calculate for each county the employmentweighted average NTR gap of the other counties in its commuting zone and augment Equation3withtheinteractionofthiscommutingzoneNTRgap(NTR Gap )andthe cz PostPNTR indicator. The (cid:28)rst two columns of the upper panel of Table 8 compare t results from this speci(cid:28)cation to the baseline results in columns 4, 8 and 12 of Table 17

4. Here, too, to conserve space we focus on the implied impact of PNTR in terms of an interquartile shift of both NTR gaps. The middle and bottom panels repeat this comparison for white males and white females, respectively, where the baseline results in the (cid:28)rst column are from Tables 5, 6 and 7. As indicated in the table, accounting for exposure via other counties in the commuting zone has little e(cid:27)ect on the results for any of the three causes of death examined. Medicaid Expansion: Sommers et al. (2012) (cid:28)nd that expansion of Medicaid in New York, Maine and Arizona in 2001, 2002 and 2006 is associated with a signi(cid:28)cant reduction in age-adjusted mortality among older adults, non-whites, and residents of poorer counties. To capture the potential in(cid:29)uence of these expansions on our results, we construct three variables that interact indicators for these states with indicators picking out the years after the expansion. To this group we add two additional variables to capture the introduction of (cid:16)Romneycare(cid:17) in Massachusetts in 2006 and the expansion of Medicaid in Oregon in 2008 that is discussed in the introduction (Baicker et al. 2013). Results with controls for these changes in health policy are reported in the third column of each panel of Table 8. As indicated in the table, including these covariates along with counties’ exposure to PNTR via their commuting zones yields results similar to those reported in column 2. A particularly stringent method of controlling for changes in state healthcare policies is the inclusion of the full set of state-by-year (cid:28)xed e(cid:27)ects. This approach captures any state-year level change in policy that might a(cid:27)ect mortality rates, and also absorbs the substantial across-state variation in the NTR gap. Results including these (cid:28)xed e(cid:27)ects are reported in the fourth column of Table 8. As indicated in the table, we continue to (cid:28)nd a positive relationship between PNTR and suicide, with the results still concentrated among white men. For accidental poisoning, the relationship with PNTR also remains positive but drops substantially in magnitude and loses statistical signi(cid:28)cance for white males and white females. Results for ARLD, by contrast, become positive and statistically signi(cid:28)cant overall once state-by-year (cid:28)xed e(cid:27)ects are added. Other Causes of Death: Finally, we investigate the relationship between PNTR and several other causes of death. The (cid:28)rst column of Table 9 investigates whether PNTR is associated with deaths from (cid:16)events of unknown intent,(cid:17) which includes, for example, poisonings, discharges of (cid:28)rearms and falls from high places that were not ruled accidental or due to suicide. We view this regression as a check on the results for suicide reported above: to the extent that events of unknown intent were not classi(cid:28)ed 18

as clear cases of suicide, we do not expect them to be related to economic conditions. As indicated in the table, we (cid:28)nd no statistically signi(cid:28)cant relationship between this cause of death and PNTR. In column two, we investigate the link between PNTR and deaths due to motor vehicleaccidents, aformofmortalityfoundtobepositivelyrelatedtoeconomicactivity in the literature. Ruhm (2000), for example, (cid:28)nds that a 1 percent increase in the unemployment rate is associated with 3 percent decline in mortality due to motor 33 vehicle accidents, and a similar relationship is found in Stevens et al. (2011). Here, however, we (cid:28)nd no association between motor vehicle fatalities and PNTR. Alargebodyofresearchintheeconomicsandpublichealthliteraturesexaminesthe potential impact of health insurance and health outcomes, hypothesizing that lack of coverage might inhibit both preventative screening and treatment of known conditions. Toward that end, columns three through (cid:28)ve of Table 9 examine links between PNTR and diabetes, which, ideally, involves consistent monitoring and treatment, and two categories of cancer found to be sensitive to preventative screening: cancer of the digestive tract, which includes colorectal cancer, and cancer of the breast, bone and 34 skin. As indicated in the table, we (cid:28)nd no relationship with respect to diabetes or the (cid:28)rst category of cancers, but (cid:28)nd positive and statistically signi(cid:28)cant relationships with respect to cancer of the digestive tract. The implied impact of an interquartile shift in a county’s exposure to PNTR for the latter is an increase in the mortality rate of 1.1 percent compared to its year-2000 levels (of 21.4). A number of papers study the link between economic shocks and circulatory dis- 35 ease in general and acute myocardial infarction (AMI, or heart attack) in particular. In columns six and seven of Table 9, we examine death due to AMI versus all other 33 The relationship between PNTR and motor vehicle accidents might be more complex if a decline in economic activity occurs as health insurance coverage decreases. Doyle (2005), for example, (cid:28)nds that the medically uninsured receive 20 percent less care and have a substantially higher mortality rate from auto accidents. Relationships might also be more complex depending on the elasticity of drinking while driving. 34 Studying the Oregon health care experiment, Baiker et al. (2013) (cid:28)nd that access to Medicaid increased the probability of being diagnosed with diabetes and increased the use of diabetes medication. Roetzheim et al. (1999), Bradley et al. (2002) and Tawk et al. (2016) (cid:28)nd that the uninsured are diagnosed with breast, skin (melanoma), colorectal and prostate cancers at later stages than the insured, reducing the chance of survival. 35 Ruhm(2000), forexample, (cid:28)ndsthata1percentincreaseintheunemploymentrateisassociated witha0.5percentdeclineindeathduetocirculatorydisease, speculatingthatthisrelationshipmight be driven by a decline in stressful activity. Browning and Heinesen (2012), on the other hand, (cid:28)nd that Danish workers displaced by plant closure are more likely to die of both heart attack and other forms of circulatory disease than workers with similar characteristics who are not laid o(cid:27). 19

forms of circulatory disease. 36 As indicated in the table, we (cid:28)nd a negative and statistically signi(cid:28)cant relationship between PNTR and AMI and no statistically signi(cid:28)cant relationship between PNTR and other forms circulatory diseases. For AMI, the implied impact of an interquartile increase in counties’ exposure to PNTR is a decrease in mortality of 3.8 percent relative to the year-2000 level (of 67.7 per 100,000). One potential explanation for this link between PNTR and AMI may be the loss of physically demanding manufacturing employment due to the trade liberalization. McManus and Schaur (2015a), for example, argue that (cid:28)rms in import-competing industries emphasize productivity at the expense of worker safety; loss of such jobs may reduce mortality due to AMI even as adverse health e(cid:27)ects may increase for those who remain employed in these industries. Relatedly, Hummels, Munch and Xiang (2016) (cid:28)nd that a rise in (cid:28)rm exports is associated with increases in injuries, severe depression and hospitalizations due to AMI and strokes. Finally, the last two columns of Table 9 summarize the relationship between PNTR and deaths due to all internal and all external causes. We (cid:28)nd a positive and statistically signi(cid:28)cant relationship in both cases. Coe(cid:30)cient estimates for all internal causes of death suggest that the implied impact of an interquartile shift in counties’ exposure to PNTR is an increase in the mortality rate of 1.7 percent versus the average mortality rate for that cause in the year 2000 (of 803 per 100,000). For external causes of death, the analogous (cid:28)gure is larger, at 6.1 percent. 4 PNTR and County Labor Markets As discussed in the introduction, one of the primary ways that trade liberalization might lead to changes in mortality rates is through its e(cid:27)ect on labor market outcomes. As illustrated in Figure 6, passage of PNTR in October 2000 is followed by a sharp decline in U.S. manufacturing employment and a pronounced increase in the U.S. 37 unemployment rate. In this section we examine the relationship between PNTR and labor market outcomes at the county level using the baseline DID speci(cid:28)cation introduced in the last section. 36 Circulatory disease is the leading cause of death during in the year 2000, with AMI accounting for one-(cid:28)fth of deaths within this category. 37 AsdiscussedinPierceandSchott(2016),U.S.valueaddedinmanufacturingcontinuedtogrowat slightly lower than the average post-WWII growth rate after PNTR. Houseman et al. (2011) provide evidence that this growth may in part be in(cid:29)ated by mismeasurement of input price indexes driven by purchases of low-cost foreign materials. 20

4.1 Employment, Unemployment and Labor Force Participation We investigate the relationship between PNTR and employment using data from the U.S. Bureau of Labor Statistics’ (BLS) Local Area Unemployment (LAU) Statistics Program and the BLS’ Quarterly Census of Employment and Wages (QCEW) 38 database. Via these data, we observe counties’ overall and manufacturing employment as well as their unemployment and labor force participation rates. The distribution of these labor market variables in the year 2000 are summarized in Table 1. For consistency, we make use of the same speci(cid:28)cation and covariates employed in our analysis of mortality rates (Equation 3). Results are reported in the (cid:28)rst three columns of Table 10, with standard errors clustered at the county level. As indicated in the table, we (cid:28)nd that both overall and manufacturing employment exhibit a negative and statistically signi(cid:28)cant relationship with county exposure to PNTR. These estimates suggest that an interquartile shift along the NTR gap distribution is associated with a relative decline in overall employment of -0.03 log points and a relative decline in manufacturing employment of -0.05 log points. We (cid:28)nd a positive but statistically insigni(cid:28)cant relationship between counties’ exposure to PNTR and non-manufacturing employment. The negative relationship with respect to manufacturing employment combined with the lack of a relationship for non-manufacturing highlights heterogeneity in the labor market implications of trade liberalization. The negative relationship with respect to overall and manufacturing employment carries through to broader measures of labor market activity and slack. The (cid:28)nal two columns of Table 10 reveal that greater exposure to PNTR is associated with a statistically signi(cid:28)cant increase in counties’ unemployment rates and a statistically signi(cid:28)cant decline in counties’ labor force participation rates. Here, the DID point estimates suggest that an interquartile shift in a county’s NTR gap is associated with a relative increase in the unemployment rate of 1.14 percentage points, or 27.9 percent of the average unemployment rate across counties in the year 2000. For the labor force participation rate, the comparable implied impact is a decline in the labor force participation rate of -1.46 percentage points, or -2.9 percent of the average labor force 39 participation rate across counties in 2000. 38 These data are available athttp://www.bls.gov/lau/ and http://www.bls.gov/cew/cewover.htm. 39 Autor et al. (2013) show that commuting zones experiencing greater increases in imports from China between 2000 and 2007 exhibit greater declines in manufacturing employment, larger increases in unemployment and greater declines in labor force participation. Their estimates imply that the 21

Figure 7 visually reports the results of regressing these labor market outcomes on interactions of the NTR gap with year dummies via Equation 4. As indicated in the (cid:28)gure, 95 percent con(cid:28)dence intervals of the estimates of θ for overall and t manufacturing employment are indistinguishable from zero prior to the change in U.S. trade policy, and decline thereafter. For the unemployment rate and the labor force participation rate, estimates of θ are indistinguishable from zero until around the t 40 change in policy, and then rise and fall, respectively, thereafter. 4.2 Personal Income, Average Annual Pay and Prices In this section we examine whether the relationship between PNTR and labor market outcomes documented in the previous sub-section is also manifest in residents’ income. This analysis helps to determine whether residents of a(cid:27)ected counties experienced changes in income if PNTR led them to change (cid:28)rms or industries following a job loss, even if these switches were not accompanied by periods of unemployment. Income lossescouldoccur, forexample, duetothelossofaccumulated(cid:28)rm-orindustry-speci(cid:28)c human capital. We use data from two sources: the U.S. Bureau of Economic Analysis’s (BEA) Local Area Personal Income (LAPI) database, which tracks counties’ overall and per capita personal income; and the BLS’ Quarterly Census of Employment and Wages (QCEW) database, which contains information on counties’ average annual pay inside 41 and outside of manufacturing. In both cases, data are expressed in current dollars. Absent the availability of county-level consumer price indexes, we de(cid:29)ate the nominal series for each county by their corresponding BLS regional Consumer Price Index for $1,840 actual increase in imports per worker from China from 2000 to 2007 decreases the labor force participation rate by 1.42 percentage points. 40 Additional evidenceregarding the severity of theshock to labormarkets comes from examination of the link between PNTR and crime. In Section B of the online appendix, we demonstrate a positive linkbetweenexposuretoPNTRandpropertycrimeaswellasanegativebutstatisticallyinsigni(cid:28)cant associationbetweenexposuretoPNTRandbirthrates. Autoretal. (2015),bycontrast,(cid:28)ndadecline in natality among commuting zones most exposed to rising imports from China. 41 Personalincomeisde(cid:28)nedasincomereceivedfromalldomesticandinternationalsources,including wage income, income from assets and government transfers, but excluding realized or unrealized capital gains or losses. Annual pay include bonuses, stock options, severance pay, pro(cid:28)t distributions, cashvalueofmealsandlodging,tipsandothergratuitiesand,forsomestates,employercontributions to deferred compensation plans. The LAPI data are available at http://www.bea.gov/regional/. Detailed discussions of the de(cid:28)nitions of personal income and wages are available on the BEA and BLS websites. 22

42 all urban consumers (CPI-U). The base year for each real series is 2000. As indicated in the (cid:28)rst two columns of Table 11, we (cid:28)nd negative associations between PNTR and counties’ real personal income and real per capita personal income, though only the latter is statistically signi(cid:28)cant at conventional levels. Its DID point estimateimpliesthataninterquartileshiftinacounty’sexposuretoPNTRisassociated with a drop in per capita personal income of -0.021 log points. Results in columns three and four indicate a negative relationship between PNTR and average annual pay and a positive association with respect to average annual pay in manufacturing. Neither, 43 however, is statistically signi(cid:28)cant at conventional levels. Figure 8 visually reports the results of regressing real per capita income and real average annual pay on interactions of the NTR gap with year dummies via Equation 4. In contrast to the employment results in Figure 7, the estimates of θ for real personal t income, real per capita personal income and real average annual pay exhibit an upward trend during the 1990s before beginning pronounced declines starting in the year 2000. One potential explanation for the rising pre-trends is that the CPI de(cid:29)ators used here do not adequately account for di(cid:27)erential changes in prices across counties during this period. 5 2SLS Estimates of Mortality on the Unemployment Rate To further verify labor market outcomes as a mechanism behind the relationship between PNTR and mortality rates, and to facilitate comparison of our estimates to 42 Resultsforthenominalseriesaresimilarbutexhibitslightlyhighermagnitudes,whichisintuitive given the behavior of the CPIs noted below. Results are also similar if state GDP de(cid:29)ators are used in lieu of the CPIs (see below). See Section A of the online appendix for a more detailed discussion of the regional CPI de(cid:29)ators. Table 1 reports the distribution of these series across counties for the year 2000. As indicated in the table, the population weighted average per capita personal income in the year 2000 has a mean and standard deviation of 30.5 and 9.2 thousand dollars, and ranges from 10.2 to83.2thousanddollars. Themeansandstandarddeviationsforcounties’averageannualoveralland manufacturingpayare32.9and9.9thousanddollars,and41.8and13.4thousanddollars,respectively. 43 Greater import competition might push the average wage in manufacturing up by driving out the lowest skill workers and push it down by subjecting remaining workers to greater competition. The sign pattern observed here is consistent with results in Autor, Dorn, Hanson and Song (2014), which (cid:28)nds that increased exposure to imports from China leads to a reduction in U.S. workers’ cumulativeearnings,andEbensteinetal. (2014a,b),which(cid:28)ndsapositiverelationshipbetweenimport competition and U.S. wages in manufacturing. In complementary research, McLaren and Hakobyan (2010)(cid:28)ndthatbluecollarworkersintheU.S.industriesmostvulnerabletoimportcompetitionfrom NAFTA experience wage declines. 23

those already in the literature (e.g., Ruhm 2000), we estimate a series of two-stageleastsquaresregressionsofdeathduetosuicide, ARLDandaccidentalpoisoningonthe unemployment rate, using counties’ NTR gaps as an instrument for the unemployment rate. The plausible exogeneity of PNTR satis(cid:28)es the standard exclusion restriction for instruments, and the relationship between PNTR and the unemployment rate, documented above, demonstrates its explanatory power. Results are reported in Table 12. As indicated in columns one, three and (cid:28)ve, we (cid:28)ndapositiverelationshipbetweentheunemploymentrateandallthreecausesofdeath when using OLS, though the result is not statistically signi(cid:28)cant for ARLD in column (cid:28)ve. Two-stage least squares results, reported in columns two, four and six, indicate a positive relationship between the unemployment rate and all three causes of death, though here, too, results for ARLD are statistically insigni(cid:28)cant at conventional levels. Point estimates for suicide and accidental poisoning imply that a 1 standard deviation increase in the county unemployment rate (2.6 percentage points) is associated with a 29.1 (=2.6*1.176/10.51) percent increase in suicides and an 84.2 (=2.6*1.487/4.59) percent increase in accidental poisonings vis a vis their year 2000 levels. This magnitude of the e(cid:27)ect on suicide is substantially higher than that reported by Ruhm (2000), where a 1 standard deviation increase in the state unemployment rate (2.1 percentage 44 points) is associated with a 2.7 percent increase in the suicide rate. The di(cid:27)erence in estimates may be driven by the di(cid:27)erent levels of aggregation in the two analyses (cid:21) with variation here at the county-level, compared to state-level variation in Ruhm (2000) (cid:21) as well as the nature of unemployment associated with exposure to the change in trade policy, which may be more persistent than unemployment due to more typical cyclical (cid:29)uctuations. 6 Conclusion This paper examines the relationship between county-level mortality rates and exposure to an important economic shock, the trade liberalization associated with the U.S. granting of Permanent Normal Trade Relations to China. We calculate exposure to PNTR as the employment-weighted average exposure of the industries active in each county. We then estimate the relationship between PNTR and mortality using a 44 In Ruhm (2000), the standard deviation of unemployment rates across states during the 1972 to 1991 sample period is 2.1. 24

di(cid:27)erences-in-di(cid:27)erences framework that nets out any time-invariant county characteristics, as well as annual shocks that a(cid:27)ect counties identically. We (cid:28)nd that exposure to PNTR is associated with an increase in mortality due to suicideandrelatedcauses, particularlyamongwhites. Theseresultsareconsistentwith that group’s relatively high employment in manufacturing, the sector most a(cid:27)ected by the change in trade policy. We (cid:28)nd that these results are robust to various extensions, including an alternate empirical speci(cid:28)cation that places no restrictions on the timing of the e(cid:27)ects of the policy change as well including controls for changes in state health care policy and exposure of other counties in the surrounding labor market. While the results in this paper do not provide an assessment of the overall welfare impact of PNTR, they do o(cid:27)er a broader understanding of the distributional implications of trade liberalization. References [1] Aghion, Philippe, Mathias Dewatripont, Luosha Du, Ann Harrison and Patrick Legros. 2015. (cid:16)Industrial Policy and Competition.(cid:17) American Economic Journal: Macroeconomics 7(4): 1-32. [2] Anderson, Robert N., Arialdi M. Minino, Donna L. Hoyert, and Harry M. Rosenberg. 2001. Comparability of Cause of Death Between ICD(cid:21)9 and ICD(cid:21)10: Preliminary Estimates. National Vital Statistics Reports 49(2):1-32. [3] Anukriti, S., and Todd J. Kumler. 2012. "The E(cid:27)ects of Trade Liberalization on Fertility and Child Health Outcomes in India." http://www.columbia.edu/~tjk2110/Trade_Kumler_Anukriti.pdf. [4] Autor, David H., David Dorn and Gordon H. Hanson. 2013. (cid:16)The China Syndrome: Local Labor Market E(cid:27)ects of Import Competition in the United States.(cid:17) American Economic Review 103(6): 2121-68. [5] Autor, David H., David Dorn, Gordon H. Hanson and Jae Song. 2014. (cid:16)Trade Adjustment: Worker Level Evidence.(cid:17) Quarterly Journal of Economics 129(4): 1799-1860. [6] Autor, David H., David Dorn and Gordon H. Hanson. 2015. The Labor Market and the Marriage Market: How Adverse Employment Shocks A(cid:27)ect Marriage, 25

Fertility, and Children’s Living Circumstances. Unpublished working paper at http://economics.mit.edu/(cid:28)les/11562. [7] Baicker, Katherine, Sarah L. Taubman, Heidi L. Allen, Mira Bernstein, Jonathan H. Gruber, Joseph P. Newhouse, Eric C. Schneider, Bill J. Wright, Alan M. Zaslavsky and Amy N. Finkelstein. 2013. The Oregon Experiment - E(cid:27)ects of Medicaid on Clinical Outcomes. New England Journal of Medicine 368 (May 2, 2013):1713-1722. [8] Becker, Gary S. 1960. An Economic Analysis of Fertility. http://www.nber.org/chapters/c2387. [9] Becker, Gary S. 1965. A Theory of the Allocation of Time. Economic Journal 75(299): 493-517 (September). [10] Ben-Porath, Yoram. 1973. Economic Analysis of Fertility in Israel: Point and Counterpoint. Journal of Political Economy 81(2): S202-S233 (March-April) [11] Bernard, Andrew B., J. Bradford Jensen, and Peter K. Schott. 2006. (cid:16)Survival of the Best Fit: Exposure to Low-Wage Countries and the (Uneven) Growth of US Manufacturing Plants.(cid:17) Journal of International Economics 68 (1): 219-237. [12] Bloom, Nicholas, Stephen Bond and John Van Reenen. 2007. (cid:16)Uncertainty and Investment Dynamics.(cid:17) Review of Economic Studies 74: 391-415. [13] Bloom, Nicholas, Mirko Draca and John Van Reenen. 2015. (cid:16)Trade Induced Technical Change: The Impact of Chinese Imports on Innovation, Di(cid:27)usion and Productivity.(cid:17) Review of Economic Studies, forthcoming. [14] Bombardini, Matilde and Bingjing Li. 2016. Trade, Pollution and Mortality in China. NBER Working Paper 22804. [15] Bradley, Cathy J., Charles W. Given and Caralee Roberts. 2002. Race, Socioeconomic Status, and Breast Cancer Treatment and Survival. Journal of the National Cancer Institute 94(7): 490-6. [16] Brambilla, Irene, Amit K. Khandelwal and Peter K. Schott. 2010. (cid:16)China’s Experience Under the Multi(cid:28)ber Arrangement (MFA) and the Agreement on Textiles and Clothing (ATC).(cid:17) In China’s Growing Roll in World Trade, edited by Robert Feenstra and Shang-Jin Wei. Chicago: University of Chicago Press. 26

[17] Brandt, Loren, Johannes Van Biesebroeck, Luhang Wang and Yifan Wang. 2012. (cid:16)WTO Accession and Performance of Chinese Manufacturing Firms.(cid:17) Center for Economic Policy Research Working Paper 9166. [18] Browning, Martin and Eskil Heinesen. 2012. E(cid:27)ect of job loss due to plant closure on mortality and hospitalization. Journal of Health Economics 31(4): 599-616. [19] Case, Anne and Angus Deaton. 2015. Rising Morbidity and Mortality in Midlife Among White Non-Hispanic Americans in the 21st Century. Proceedings of the National Academy of Sciences doi: 10.1073/pnas.1518393112. [20] Che, Yi and Xu, Xun. 2015. The China Syndrome in US: Import Competition, Crime, and Government Transfer. Mimeo, University of Munich. [21] Classen, Timothy J. and Richard A Dunn. 2012. The e(cid:27)ect of Job Loss and Unemployment Duration on Suicide Risk in the United States: A New Look Using Mass-Layo(cid:27)s and Unemployment Duration. Health Economics 21(3):338- 350.David, Angus Deaton and Adriana Lleras-Muny. 2006. The Determinants of Mortality. Journal of Economic Perspectives 20(3): 97-120. [22] Deb Partha, William T. Gallo , Fletcher Ayyagari and Jody L. Sindelar. 2011. The e(cid:27)ect of job loss on overweight and drinking. Journal of Health Economics. 2011 Mar; 30(2): 317-27. Epub 2011 Jan 14. PMCID: PMC3086369. [23] Dettling, Lisa J., and Melissa S. Kearney. 2014. House Prices and Birth Rates: The Impact of the Real Estate Market on the Decision to Have a Baby. Journal of Public Economics 110 (2014): 82-100. [24] Dix-Carneiro, Rafael, Rodrigo R. Soares and Gabriel Ulyssea. 2015. (cid:16)Local Labor Market Conditions and Crime: Evidence from the Brazlian Trade Liberalization.(cid:17) Mimeo. [25] Doyle, Joseph J. Jr. 2005. Health Insurance, Treatment and Outcomes: Using Auto Accidents as Health Shocks. Review of Economics and Statistics 87(2): 256- 270. [26] Ebenstein, Avraham, Ann Harrison, Margaret McMillan and Shannon Phillips. 2014a. Estimating the Impact of Trade and O(cid:27)shoring on American Workers using 27

the Current Population Surveys, The Review of Economics and Statistics, MIT Press, vol. 96(3), pages 581-595, October. [27] Ebenstein, Avraham, Ann Harrison and Margaret McMillan. 2014b. "Why are American Workers getting Poorer? China, Trade and O(cid:27)shoring." Mimeo. [28] Falba Tracy, Hsun-Mei Teng, William T. Gallo and Jody L. Sindelar. The E(cid:27)ect of Involuntary Job Loss on Smoking Intensity and Relapse. Addiction 2005 Sep; 100(9):1330-9. [29] Fazel, Seena, Achim Wolf, Zheng Chang, Henrik Larsson, Guy M Goodwin, and Paul Lichtenstein. 2015. Depression and violence: a Swedish population study. The Lancet 2(3): 224-232. [30] Feenstra, Robert C., John Romalis and Peter K. Schott. 2002. (cid:16)U.S. Imports, Exports and Tari(cid:27) Data, 1989-2001.(cid:17) NBER Working Paper 9387. [31] Feng, Ling, ZhiyuanLiandDeborahSwenson.2014.TradePolicyUncertaintyand Exports: Evidence from China’s WTO Accession. CESifo Working Paper Series 4780, CESifo Group Munich. [32] Feler, Leo and Mine Z. Senses. 2015. (cid:16)Trade Shocks and the Provision of Local Public Goods.(cid:17) Mimeo. [33] Finkelstein, Amy, SarahTaubman, BillWright, MiraBernstein, JonathanGruber, Joseph P. Newhouse, Heidi Allen, Katherine Baicker. (2012). (cid:16)The Oregon Health Insurance Experiment: Evidence from the First Year.(cid:17) The Quarterly Journal of Economics 127(3): 1057-1106. [34] Freeman, R., Katz, L., 1991. (cid:16)Industrial Wage and Employment Determination in an Open Economy, in Immigration, Trade and Labor Market,(cid:17) edited by John M. Abowd and Richard B. Freeman. Chicago: University of Chicago Press. [35] Galbraith, Virginia L and Dorothy S. Thomas. 1941. Birth Rates and the Interwar Business Cycles. Journal of the American Statistical Association, 36(216): 465- 476. 28

[36] Gemmill, Alison, April Falconi, Deborah Karasek, Terry Hartig, Elizabeth Anderson and Ralph Catalano. 2016. Do macroeconomic contractions induce or ‘harvest’ suicides? A test of competing hypotheses. J Epidemiol Community Health 2015;69:1071-1076 doi:10.1136/jech-2015-205489. [37] Girma, Sourafel, Gong, Yundan, G(cid:246)rg, Holger and Yu, Zhihong (2009). (cid:16)Can Production Subsidies Explain China’s Export Performance? Evidence from Firmlevel Data.(cid:17) The Scandinavian Journal of Economics 111: 863(cid:21)891. [38] Hummels, David, Jakob Munch and Chong Xiang. 2016. No Pain, No Gain: The E(cid:27)ects of Exports on E(cid:27)ort, Injury, and Illness. Unpublished manuscript available at http://www.krannert.purdue.edu/faculty/cxiang/cvwp/ExpHealth.pdf. [39] Khandelwal, Amit K., Peter K. Schott and Shang-Jin Wei. 2013. (cid:16)Trade Liberalization and Embedded Institutional Reform: Evidence from Chinese Exporters.(cid:17) American Economic Review 103 (6): 2169-95. [40] Handley, Kyle and Nuno Limao. 2016. (cid:16)Policy Uncertainty, Trade and Welfare: Evidence from the U.S. and China.(cid:17) Mimeo. [41] Houseman, Susan, Christopher Kurz, Paul Lengermann and Benjamin Mandel. 2011. (cid:16)O(cid:27)shoring Bias in U.S. Manufacturing.(cid:17) Journal of Economic Perspectives 25 (2): 111-132. [42] Hoyert and Anderson. 2001. (cid:16)Age-adjusted Death Rates: Trend Data Based on the Year 2000 Standard Population. National Vital Statistics Reports 49 (9): 1-7. [43] Iyer, Lakshmi and Petia B. Topalova. 2014. (cid:16)Poverty and Crime: Evidence from Rainfall and Trade Shocks in India.(cid:17) Mimeo. [44] Kemp, Janet and Robert Bossarte. 2012. Suicide Data Report. Department of Veteran A(cid:27)airs, Mental Health Services, Suicide Prevention Program. [45] Kircher, T., J. Nelson and H. Burdo. 1985. The autopsy as a measure of accuracy of the death certi(cid:28)cate. New England Journal of Medecine 313(20): 1263-9. [46] McLaren, John and Shushanik Hakobyan. 2010. (cid:16)Looking for Local Labor Market E(cid:27)ects of NAFTA.(cid:17) NBER Working Paper 16535. 29

[47] McManus, T. Clay and Georg Schaur. 2015a. (cid:16)The E(cid:27)ects of Import Competition on Health in the Local Economy.(cid:17) Mimeo. [48] McManus, T. Clay and Georg Schaur. 2015b. (cid:16)The E(cid:27)ects of Import Competition on Worker Health.(cid:17) Mimeo. [49] Mincer, Jacob. 1963. "Market prices, opportunity costs, and income e(cid:27)ects." Measurement in Economics (1963): 67-82. [50] Mion, GiordanoandLinkeZhu.2013.(cid:16)ImportCompetitionFromandOutsourcing to China: A Curse or a Blessing for Firms.(cid:17) Journal of International Economics 89 (1): 202-215. [51] Pierce, Justin R. and Peter K. Schott. 2016. (cid:16)The Surprisingly Swift Decline of U.S. Manufacturing Employment.(cid:17) NBER Working Paper 18655. [52] Pierce, Justin R. and Peter K. Schott. 2012. (cid:16)Concording U.S. Harmonized System Codes Over Time.(cid:17) Journal of O(cid:30)cial Statistics 28(1): 53-68. [53] Revenga, A.L. 1992. (cid:16)Exporting Jobs? The Impact of Import Competition on Employment and Wages in U.S. Manufacturing.(cid:17) Quarterly Journal of Economics 107 (1): 255-284. [54] Roetzheim, Richard G., Naazneen Pal, Colleen Tennant, Lydia Voti, John Z. Ayanian, Annette Schwabe and Je(cid:27)rey P. Krischer. 1999. E(cid:27)ects of Health Insurance and Race on Early Detection of Cancer. Journal of the National Cancer Institute 91(16): 1409-1415. [55] Rudd, Rose A., Noah Alshire, Jon E. Zibbell and Matthew Gladden. 2016. (cid:16)Increases in Drug and Opioid Overdose Deaths - United States, 2000-2014.(cid:17) Morbidity and Mortality Weekly Report. Centers for Disease Control and Prevention. Available online at http://www.cdc.gov/mmwr/preview/mmwrhtml/mm6450a3.htm. Accessed May 11, 2016. [56] Ruhm, Christopher. 2000. (cid:16)Are Recessions Good for your Health?(cid:17) Quarterly Journal of Economics 115(2 May): 614-650. [57] Ruhm, Christopher. 2015. (cid:16)Recessions, Healthy No More?(cid:17) Journal of Health Economics 42: 17-28. 30

[58] Shatz Howard J. and Je(cid:27)rey D. Sachs. 1994. (cid:16)Trade and Jobs in U.S. Manufacturing.(cid:17) Brookings Papers on Economic Activity 1994 (1): 1-69. [59] Schott, Peter K. 2008. The Relative Sophistication of Chinese Exports. Economic Policy 53: 5-49 (January). [60] Silver, Morris. 1965. "Births, marriages, and business cycles in the United States." The Journal of Political Economy (1965): 237-255. [61] Sommers, Benjamin D., Katherine Baicker and Arnold Epstein. 2012. Mortality andAccesstoCareAmongAdultsAfterStateMedicaidExpansions.NewEngland Journal of Medicine 367: 1025-1034. [62] Stevens, Ann, Douglas L. Miller, Marianne E. Page and Mateusz Filipski. 2011 (cid:16)The Best of Times, the Worst of Times: Understanding Pro-cyclical Mortality(cid:17) NBER Working paper 17657. [63] Sullivan, Daniel and Till von Wachter. 2009. (cid:16)Job Displacement and Mortality: An Analysis Using Administrative Data.(cid:17) Quarterly Journal of Economics August 2009: 1265-1306. [64] Tawk, Rima and Adrian Abner, Alicestine Ashford, and Clyde Perry Brown. 2016. Di(cid:27)erences in Colorectal Cancer Outcomes by Race and Insurance. International Journal of Environmental Research and Public Health. 2016 Jan; 13(1): 48. [65] Tolbert, Charles and Molly Sizer. 1996. U.S. Commuting Zones and Labor Market Areas: A 1990 Update. Rural Economy Division, Economic Research Service, U.S. Department of Agriculture Sta(cid:27) Paper No. AGES-9614. [66] Utar, Hale and Luis B. Torres Ruiz. 2013. (cid:16)International Competition and Industrial Evolution: Evidence form the Impact of Chinese Competition on Mexican Maquiladoras.(cid:17) Journal of Development Economics 105: 267-287. 31

Table 1: Summary Statistics 32

Table 2: Average Death Rates by Major Causes of Death 33

Table 3: 1999 NTR Gap versus Other County Attributes 34

Table 4: PNTR and Suicide, ARLD and Accidental Poisoning 35

Table 5: PNTR and Suicide 36

Table 6: PNTR and Accidental Poisoning 37

Table 7: PNTR and Alcohol-Related Liver Disease 38

Table 8: Robustness Exercises 39

Table 9: PNTR and Other Causes of Death 40

Table 10: PNTR and Employment Outcomes (LAU and QCEW) 41

Table 11: PNTR and County Per Capita Personal Income (LAPI and QCEW) 42

Table 12: Mortality and Unemployment (2SLS) Figure 1: Distribution of Overall Mortality Rates 43

Figure 2: Death Rates for Non-Hispanic Whites Figure 3: Distribution of 1999 NTR Gaps Across Counties 44

Figure 4: Implied Impact of PNTR on Death by Suicide and Alcohol-Related Liver Disease, by Age Category 45

Figure 5: Implied Impact of PNTR on Death Rates Using Annual DID Speci(cid:28)cation (Equation 4) 46

Figure 6: Post-War U.S. Manufacturing Employment Figure 7: Implied Impact of PNTR on Employment Outcomes Using Annual DID Speci(cid:28)cation (Equation 4) 47

Figure 8: Implied Impact of PNTR on Income Using Annual DID Speci(cid:28)cation (Equation 4) 48

Online Appendix This online appendix contains additional empirical results and information on data creation referenced in the main text. A Regional Price Indexes The BLS produces CPIs for four regions: the northeast (Maine, Massachusetts, New Hampshire,NewJersey,NewYork,Pennsylvania,RhodeIslandandVermont),themidwest (Illinois, Indiana, Iowa, Kansas, Michigan, Minnesota, Missouri, Nebraska, North Dakota, Ohio, South Dakota, Wisconsin), the south (Alabama, Arkansas, Delaware, District of Columbia, Florida, Geogia, Kentucky, Louisiana, Maryland, Mississippi, North Carolina, Oklahoma, South Carolina, Tennessee, Texas, Virginia and West Virginia ) and the west (Alaska, Arizona, California, Colorado, Hawaii, Idaho, Montana, Nevada, New Mexico, Oregon, Utah, Washington, Wyoming). For each region, the BLS produces indexes for three sets of cities: the overall CPI for urban consumers (CPI-U), the CPI for metropolitan urban areas with population above 1.5 million (class A) and the CPI for metropolitan urban areas fewer than 1.5 million (class B/C). These indexes are compared in Figure A.5 of this online appendix. B Complementary Evidence In this section we make use of the same di(cid:27)erence-in-di(cid:27)erence speci(cid:28)cation used to studymortalitytodemonstratethatcountiesmoreexposedtoPNTRexperiencehigher rates of property crime and lower birth rates. B.1 Crime WeexaminetherelationshipbetweenPNTRandcrimeratesforthreereasons. First,an increase in crime could a(cid:27)ect mortality directly, e.g. via homicides, though it turns out that we do not (cid:28)nd evidence for that channel. Second, an increase in crime contributes to a lower quality of life and thereby might contribute to depression or other conditions consistent with the increases in mortality noted above. Finally, a link between PNTR and crime rates provides additional evidence of the seriousness of the labor market 49

disruptions documented in the main text (Iyer and Topalova 2014, Dix-Carneiro et al. 45 2015). Our analysis makes use of county-level crime rate statistics per 100,000 residents available from the Federal Bureau of Investigation (FBI) via the Uniform Crime Re- 46 porting(UCR)database. Thesedata,availablefrom1990to2006,breakoverallcrime rates into two main categories, violent and property crime, and eight sub-categories: 47 murder, rape, robbery, assault, burglary, larceny, auto theft and arson. Table A.7 reports the results. As indicated in the (cid:28)rst and second columns of the table, counties’ exposure to the change in U.S. trade policy has a positive relationship with both overall violent crime and overall property crime, but this relationship is only statistically signi(cid:28)cant at conventional levels for overall property crime. The DID point estimate for the property crime regression implies that an interquartile shift in counties NTR gap is associated with an increase in the rate of property crime per 100,000 residents of 144.2, or 5.6 percent of the average property crime rate across counties in the year 2000 (2592 per 100,000). These results are consistent with Feler and Senses (2015) who note that counties more exposed to imports from China experienced small increases in property crime, while the least exposed counties experienced a substantial reduction in crime. The remaining columns of Table A.7 illustrate positive and statistically signi(cid:28)cant relationships between counties’ exposure to PNTR and several sub-categories of property crime, including robbery, larceny, motor vehicle theft and arson. Coe(cid:30)cient estimates suggest interquartile shifts in counties’ exposure to PNTR are associated with increases in the rates of these crimes of 15.8, 5.1, 11.6 and 21.9 percent compared to their year-2000 levels. B.2 Birth Rates Inprinciple,theassociationbetweenPNTRandlocallabormarketconditionsdiscussed above might a(cid:27)ect birth rates in at least two ways. On one hand, to the extent that workers the declines in income and employment as temporary, they might perceive a drop in the opportunity cost of having children and the birth rate might rise. On the 45 Che et al. (2015) examine the link between Chinese imports and U.S. crime across commuting zones. 46 These data are available at https://www.fbi.gov/about-us/cjis/ucr/ucr. 47 Burglary is de(cid:28)ned as theft (i.e., larceny) combined with unlawful entry. Robbery is de(cid:28)ned as forcible theft from a person. 50

other hand, to the extent that PNTR results in long-term reductions in income and 48 employment, birth rates might decline. 49 We make use of county-level data on births available from CDC. Using these data and population estimates from SEER, we compute both the birth rate for each county, de(cid:28)ned as births per population. We caution that the county-level birth data are often suppressed, and that they are available to us for years 1992 to 2006 only. The number of counties included in the regression results (i.e., those with observations both before and after 2000) rises from an average of 457 between 1995 and 2000 to 501 between 2001 and 2006. These counties are among the largest, representing an average of 76 percent of the total U.S. population across the sample period. Table A.8 reports the results of baseline DID speci(cid:28)cations similar to those used above. The (cid:28)rst column report results for the birth rate per 100,000 population, while the second column reports results for the log number of births. As indicated in the table, the DID coe(cid:30)cients of interest are negative but statistically insigni(cid:28)cant in both 50 columns. C PNTR and Import Prices To the extent that imports are an important input into the production of healthcare, trade liberalization with China might improve workers’ health via lower prices for health-related goods. Moreover, a general reduction in prices associated with trade liberalization may lead to welfare improvements for U.S. consumers. We investigate 51 this link using customs data from the U.S. Census provided by Schott (2008). We employ a generalized triple di(cid:27)erences speci(cid:28)cation that compares products with varying NTR gaps ((cid:28)rst di(cid:27)erence) before and after PNTR (second di(cid:27)erence) 48 These trade-o(cid:27)s and the potential cyclicality of birth rates are discussed in Ben-Porath (1973), Becker (1960, 1965), Galbraith and Thomas (1941), Mincer (1963) and Silver (1965). Dettling and Kearney (2014) provide a concise discussion of this literature. Anukriti and Kumler (2012) (cid:28)nd that an increase in import competition in India associated with the end of the License Raj in 1991 raised birth rates among women with low socioeconomic status but had the opposite a(cid:27)ect among women of high socioeconomic status. 49 These data can be downloaded from http://wonder.cdc.gov/wonder/help/natality.html. 50 Autor et al. (2015), by contrast, (cid:28)nd a decline in natality among commuting zones most exposed to rising imports from China. 51 Thesedataareavailablefordownloadathttp://faculty.som.yale.edu/peterschott/sub_international.htm. 51

and across source countries (third di(cid:27)erence) for the years 1992 to 2007: ln(UnitValue) = θ1{c = China} ×PostPNTR ×NTRGap + (A.1) hst s t h +λTariff +δ +δ +δ +α+ε . hst st sh ht hst The left-hand side variable represents the log of the average unit value observed for ten-digit HS product h from source country s in year t. 52 The (cid:28)rst term on the righthand side is the term of interest: a triple interaction of an indicator for China, an indicator for the post-PNTR period, and the NTR gap for product h that captures the impact of the change in U.S. policy. Tariff represents the U.S. revealed import hct tari(cid:27) for product h from country c in year t, computed as the ratio of duties collected to dutiable value using publicly available U.S. trade data. δ , δ and δ represent ct ch ht country×year, country×product and product×year (cid:28)xed e(cid:27)ects. α is the regression constant. Results are reported in Table A.9. As indicated in the table, the NTR gap has a negative and statistically signi(cid:28)cant relationship with import unit values. The point estimate in the (cid:28)rst row of the table implies that Chinese imports after PNTR are 0.18 log points lower vis a vis imports of products from other source countries. We use the following back-of-the-envelope procedure to gauge the potential impact of the decline in Chinese import unit values on health-related versus other goods within the United States. First, we use the results in Table A.9 to compute the predicted relative impact of PNTR for each HS import product. Second, we take the average of these impacts across HS products within NAICS industries. Third, we merge these NAICS-level mean log changes into the 2007 U.S. total requirements input-output matrix, whose coe(cid:30)cients indicate the amount of the (cid:16)input(cid:17) NAICS industry needed to produce one dollar of the (cid:16)using(cid:17) industry. Fourth, we compute the weighted average implied relative log unit value changes across the input industries for each using industry, using the IO coe(cid:30)cients as weights. Finally, we examine the changes associated with healthcare-related NAICS industries. These industries are identi(cid:28)ed by having one of the following key words in their description: health, care, pharmaceutical, drug, hospital, medical, surgical, medicine, and imaging. 52 The trade data report both value and quantity for each transaction and we use the ratio of these two variables as a proxy for the price. We omit products whose units change over time, and make use of concordances provided by Pierce and Schott (2016) to ensure product codes are consistent over the sample period. Further details on data construction are provided in the online appendix. 52

ThedistributionoflogunitvaluedeclinesacrossusingindustriesisdisplayedinFigure A.6; themean and standarddeviation across industriesis -0.069 and0.095. The declines associated with the health industries identi(cid:28)ed in the last paragraph are reported inTableA.10, alongwiththeaverageforthoseindustriesversusallothers. Asindicated in the table, four healthcare-related industries have sizable weighted-average changes: surgical instruments (-0.080), surgical appliances (-0.066), electromedical manufacturing (-0.060) and pharmaceutical preparation manufacturing (-.054). Weighted-average changes for the remaining industries in the table are far lower. Intuitively, this is due to their relatively high share of labor versus other goods. Appendix Tables and Figures Table A.1: Distribution of U.S. Population Across Age Categories in 2000 53

Table A.2: Sensitivity of Weighted Average Death Rates Across Counties to Population Weights 54

Table A.3: PNTR and Suicide by White Males, By Age Group 55

Table A.4: PNTR and Suicide by White Females, By Age Group 56

Table A.5: PNTR and Alcohol-Related Liver Disease for White Males, By Age Group 57

Table A.6: Share of Whites and Males Among Occupations in Manufacturing, 1999 58

Table A.7: PNTR and Crime Rates per 100,000 Population (UCR) 59

Table A.8: PNTR and Birth Rates per 100,000 Population 60

Table A.9: PNTR and U.S. Import Prices 61

Table A.10: Unit Value Declines Weighted by Health-Industry IO Coe(cid:30)cients 62

Figure A.1: Counties’ Own versus Surrounding Commuting Zone NTR Gaps Figure A.2: Distribution of Counties’ Exposure to MFA Phase-Outs (MFAExposure ) and Counties’ NTR Tari(cid:27)s (NTR ) ct ct 63

Figure A.3: Distribution of Counties’ Exposure to Reductions in Chinese Tari(cid:27)s and Domestic Production Subsidies Figure A.4: County NTR Gaps by Racial Group 64

Figure A.5: Consumer Price Indexes by Region and Size Class 65

Figure A.6: Weighted Average ln(Unit Value) Declines by (cid:16)Using(cid:17) NAICS Industries 66