Trade and the geographic spread of the great recession

Journal of International Economics 119 (2019) 169–180

Contents lists available at ScienceDirect

Journal of International Economics journal homepage: www.elsevier.com/locate/jie

Full Length Articles

Trade and the geographic spread of the great recession☆ Sebastian Stumpner Université de Montréal & Banque de France, Département de sciences économiques, C-6020 Pavillon Lionel-Groulx, 3150 Rue Jean-Brillant, Montréal, Québec H3T 1N8, Canada

a r t i c l e

i n f o

Article history: Received 30 October 2017 Received in revised form 2 April 2019 Accepted 4 April 2019 Available online 17 April 2019 Research data related to this submission: https://data.mendeley.com/datasets/ w8h87vsfjp/draft?a=07a16c94-a41c-4af48f53-adb7efde9c2a.

a b s t r a c t I study the role of trade between U.S. states in the regional propagation of local consumer demand shocks during the Great Recession. To identify the trade channel empirically, I make use of heterogeneity in the direction of trade flows across industries in the same state: Industries that depended relatively more on final demand from states with housing boom-bust cycles grew by more before the crisis and declined faster from 2007 to 09. A one standard deviation difference in the exposure to demand shocks during the recession explains a 2.9 percentage point difference in employment growth. © 2019 Elsevier B.V. All rights reserved.

JEL codes: F14 F16 Keywords: Interregional trade Regional propagation Great recession

1. Introduction While the initial increase in unemployment during the Great Recession was concentrated in areas with housing busts, subsequently unemployment spread across space. By 2009, it was above pre-crisis levels in almost all U.S. counties. Fig. 1 maps this “geographic spread” of the crisis.1 How did local shocks diffuse through the economy, causing business cycle co-movement across U.S. states? This paper argues that trade across U.S. states contributed substantially to the spread of the crisis across space. To the extent that producers of tradable goods across the U.S. depend on markets experiencing a housing bust and consumption collapse, they face a shock to their market size. I empirically trace the effect of these demand ☆ I am grateful to Atif Mian, Yuriy Gorodnichenko, Andres Rodriguez-Clare, and Ted Miguel for their advice and encouragement. For useful comments and suggestions I thank Arnaud Costinot, Ben Faber, Pierre-Olivier Gourinchas, Simon Hilpert, Amir Kermani, Ross Levine, Brent Neiman, Matt Notowidigdo, Steve Redding, David Romer, Frank Schilbach, Reed Walker, and particularly Robin Burgess. I am also grateful to Ralph Ossa (editor) and two anonymous referees for their valuable feedback, and to seminar participants at LSE, UC Berkeley, University of Montréal, HEC Montréal, Bank of Canada, NY Fed, IIES, Bonn, Yale SOM, PUC-Rio, FGV-Rio, HEC Paris, the West Coast Trade Workshop, SED 2014, NBER SI (ITM), and the Econometric Society World Congress for comments.

1

E-mail address: [email protected]. Similar maps can be found in Fogli et al. (2012).

https://doi.org/10.1016/j.jinteco.2019.04.001 0022-1996/© 2019 Elsevier B.V. All rights reserved.

shocks through the trade and input-output network that connects U.S. states at the industry level. Industries that depended more on markets with larger declines in consumer demand experienced a substantially larger fall in employment. This effect is driven by both direct sales to consumers and indirect effects through input-output relationships. I exploit differences in trading patterns across industries that are located in the same state to separate the trade channel from other potential contagion mechanisms. Within a state, industries differ in their shocks to market size to the extent that they depend on markets experiencing a consumption collapse. I measure this dependence using data on trade flows from the Commodity Flow Survey (CFS) and intermediate goods relationships from BEA IO tables. The resulting measure captures both direct links to consumers in affected states and indirect links through intermediate goods. State-specific consumer demand shocks are captured using data on pre-crisis household leverage, as in Mian and Sufi (2012). The empirical approach relies on the identification assumption that industries that depend relatively more on states experiencing a housing bust are not relatively more affected by other shocks. Thus, other potentially confounding shocks (such as credit supply shocks or expectation shocks) are controlled for to the extent that they do not affect industries differentially in a way that is correlated with the direction of trade flows. My empirical analysis finds a sizable role for trade in the transmission of the crisis: First, I find that a one standard deviation in the variable measuring exposure to demand shocks causes a 2.9 percentage point

170

S. Stumpner / Journal of International Economics 119 (2019) 169–180

Fig. 1. Yearly change in unemployment rate across US counties, 2006–07, 2007–08, and 2008–09. Notes: The maps show the year-on-year change in the unemployment rate from 2006 to 07, 2007–08, and 2008–09 across U.S. counties. Data are from the BLS.

S. Stumpner / Journal of International Economics 119 (2019) 169–180

difference in 2007–09 employment growth, which corresponds to 18% of the total dispersion in employment growth among tradable producers. This result is robust to focusing only on variation in trade flows that arises from different transportation costs across industries. Moreover, it is specific to trade flows to, but not trade flows from highly leveraged states. Second, I study the dynamic evolution of the industries over a longer (10 year) horizon. Industries selling particularly to high-leverage states were booming before the crisis, thus benefitting from the housing and consumption boom in these states. This pattern reverses with the beginning of the recession in 2007 and reaches a low in 2009. With the recovery starting in 2009, the differential effect across industries slowly converges back to zero. I thus find evidence that strongly supports the view that trade in goods is important for linking business cycle fluctuations across states. Two tests of heterogeneous adjustment across industries further support the main results. First, I find that the effect is stronger for industries producing durable goods. This is in line with existing empirical evidence that demand for durable goods fell by more during the recession. Second, the effect is more pronounced in industries producing more differentiated goods. This is consistent with the idea that industries producing more homogeneous output can more easily offset a shock to a particular market by increasing their market share at other destinations. Next, I split up the trade shock variable into a component capturing direct sales to consumers in affected states, and a component capturing indirect dependence on these states through intermediate goods relationships. I find that both effects contributed to the decline in employment. The estimated coefficients are of similar magnitude, implying that only an industry's total exposure to demand shocks matters, but not its division into direct and indirect linkages. The main result holds in several further robustness tests. Among others, I control explicitly for credit supply shocks, using data on the geography of bank lending from bank balance sheets and the location of bank branches. While I also find some role for the credit channel in reducing employment, consistent with previous literature (ChodorowReich (2014), Greenstone et al. (2014)), the estimates of the trade channel do not change when the credit channel is controlled for. This project relates to several strands of literature. First, the paper is closely linked to the work by Caliendo et al. (2018), who study the regional propagation of local shocks in a calibrated model of the U.S. economy. Similarly, Eaton et al. (2016) and Monte (2014) have studied the general equilibrium effects of regional shocks in trade models. I view this paper as complementary to this line of work, especially since its focus is an empirical investigation of the regional propagation of shocks. Second, this paper is closely connected to the literature on trade, volatility, and business cycle comovement. Frankel and Rose (1998) were the first to highlight that countries that trade more with each other tend to have more correlated economic outcomes. Following this work, a long literature has investigated the relationship between trade openness, volatility, and output comovement. On the theory side, Kose and Yi (2006) have shown that in a standard international business cycle model, trade does not lead to the high degree of comovement that is observed in the data (the trade-comovement puzzle).2 On the empirical side, many papers have focused on the association between trade openness and the variance of output growth or the correlation of output growth across countries.3

2 A recent contribution to that literature is Johnson (2014) who studies the role of intermediate goods for the trade-comovement puzzle. 3 Recent examples are di Giovanni and Levchenko (2009) and di Giovanni and Levchenko (2010). Clark and van Wincoop (2001) show that synchronization is higher among U.S. regions than among EU countries.

171

The paper also relates to the literature on consumer demand shocks during the Great Recession. On the theory side, Midrigan and Philippon (2011) and Eggertsson and Krugman (2012) provide models of the recession driven by a collapse in aggregate demand. On the empirical side, Mian et al. (2012), using micro consumption data, document the fall in consumer demand as a result of the housing bust. Mian and Sufi (2014) show that this collapse in consumption was the main driver of employment losses in cities that experienced housing boom and bust cycles. I empirically trace the effect of these local demand shocks on other regions through the within-U.S. trade network. Finally, this paper is related to the literature on the contagion of crises, such as van Rijckeghem and Weder (2001), Glick and Rose (1999), Kaminsky and Reinhart (2000), and Forbes (2004).4 This line of literature has focused on estimating the channels of contagion across countries using data on cross-country financial and trade linkages. More recently, some papers have analyzed international contagion in the context of the Great Recession, such as Aiyar (2012) and Centorelli and Goldberg (2011). In contrast to this literature, I focus entirely on within-country contagion.5 The remainder of the paper proceeds as follows. Sections 2 and 3 present the empirical strategy, and discuss data and summary statistics. Section 4 covers the main empirical results and section 5 presents several robustness checks. Section 6 concludes. 2. Empirical strategy The identification approach exploits within-state across-industry heterogeneity in exposure to demand shocks. I first define a measure of exposure to demand shocks through trade, which captures both direct links to consumers and indirect links through intermediate goods. I then use differences in trading patterns across industries in the same state to identify which industries should be relatively more affected by demand shocks. To derive an expression summarizing the effect of demand shocks in different markets on production in state i ∈ {1, … , N} and industry s ∈ {1, … , S}, I build on the multi-country, multi-industry input-output framework used in Bems et al. (2010) and Johnson and Noguera (2012). In particular, consider the following market-clearing condition for the product from industry s located in state i that is used either as intermediate good or for final consumption.

Y is ¼

N X S X

XM ijst þ

j¼1 t¼1

N X

X Cijs

j¼1

Yis denotes sales of state i, industry s, XM ijst denotes the value of output from state i, industry s, used as intermediate good in the production of industry t in state j, and XCijs denotes consumption expenditure in state j on the good produced by state i and industry s. Let aijst ≡

XM ijst Y jt

denote expenditure on inputs from state i, industry s, as

a fraction of gross output in the destination industry. Collecting coefficients aijst in the matrix 2

aij11 6 aij21 6 Aij ¼ 4 ⋮ SS aijS1

aij12 ⋱ …

… aij1S …

3 7 7 5

aijSS

4 Another closely related paper is Gorodnichenko et al. (2012) that studies the role of the Soviet collapse for the Finnish depression in a small open economy models. Enders and Peter (2015) use a small open economy model to study the international contagion of the Great Recession. 5 The focus on U.S. states is also interesting because they form a monetary union, which implies that individual regions cannot buffer against external shocks through exchange rate changes. The results in this paper may thus be particularly interesting for understanding the transmission of shocks among Euro-zone countries.

172

S. Stumpner / Journal of International Economics 119 (2019) 169–180

the regional input-output matrix of direct requirements coefficients collects all bilateral input-output relationships between US states. It is given by: 2

A

NSNS

A11 6 A21 6 ¼4 ⋮ AN1

A12 ⋱

…

…

… ANN

A1N

3 7 7 5

The system of market clearing conditions can then be written as Y ¼ AY þ

X

X Cj ¼

X

j

ΩX Cj ;

ð1Þ

j

where Ω ≡ (I − A)−1 is the Leontief inverse, XCj denotes the vector of final expenditure by state j: XCj = [XC1j1, XC1j2, … , XCNjS]′, and Y = [Y11, Y12, … , YNS]′. The following lines derive a measure for the pass-through of demand shocks through the input-output network. The empirical strategy uses a demand shock that varies only at the state-level. If Δ log (XCijt) = Δ log (XCj ), and holding the coefficients of the input-output matrix fixed, then the percentage change in sales of industry s in state i is6: Δ logðY is Þ ≈

N φ X ijs j¼1

Y is

  Δ log X j

S C where φijs ≡ ∑N k=1∑t=1ωikstXkjt and the weights

ð2Þ φijs Y is

sum to one. The

term ωikst is the (s, t) element in the (i, k) block of Ω. It gives the amount of gross output of sector s in state i that is used, directly and indirectly, to produce one unit of output of sector t in state k. Multiplying by final consumption of state j, and summing over all sectors t and intermediate states k, the term φijs gives the total amount of output by sector s in state i that is consumed, directly and indirectly, in state j. The weight φijs Y is

is the corresponding share of output that depends on final demand

in state j. Eq. 2 says that the decline in output is a weighted average of the demand decline in the destination states, where the weights are the shares of output that, directly or indirectly, are used to satisfy final demand in the particular destination states. φ

The weights Yijsis depend on the entire structure of the regional inputoutput network. In the data, however, I do not observe trade flows separated by their use as intermediates or for consumption, but only total C trade flows Xijs = ∑St=1XM ijst + Xijs. This means that the regional IO matrix cannot be taken directly from the data, but needs to be constructed using data on trade flows and the aggregate input-output matrix as well as a proportionality assumption.7 All the details of this construction are deferred to the appendix. As mentioned earlier, eq. 2 assumes that the coefficients of the regional IO matrix do not change as a result of demand changes. If the sectoral production function features an elasticity of substitution equal to one both across intermediate goods sectors and within intermediate goods sectors (across origin states), as e.g. in Jones (2013), then this assumption holds exactly. In general, however, changes in regional factor costs and prices may cause a reallocation of expenditure shares, and therefore changes to the regional input-output matrix. To the extent that these general equilibrium effects are induced by the changes in final demand, the estimated coefficient on the trade demand shock variable measures the total effect of demand changes on employment. In general equilibrium, growth of industry sales and consumer expenditure are jointly determined. In a regression framework of Eq. 2, this leads to a simultaneity problem. In order to estimate the role of 6 The derivation of this equation and more details on the interpretation can be found in the appendix. 7 In particular, I assume that for a given destination state and a given industry, the use of incoming shipments (across intermediate use sectors and between intermediate use and consumption) is the same for different origin states.

trade in transmitting the crisis, I need exogenous variation in expenditure growth. Work by Mian et al. (2012) has shown that pre-crisis household leverage had a strong effect on expenditure growth during the recession. In 2006, there were large differences in household leverage across U.S. states. A large part of these differences were due to differences in leverage growth from 2002 to 2006. Strong house price growth in some U.S. states during that time led to a buildup of leverage, driven to a large extent by home equity withdrawals (Mian and Sufi (2011)). At the peak of the housing bubble in 2006, states with rapid house price growth during 2002–06 also tended to have the most highly indebted households. With the reversal of house prices starting in 2006, highly leveraged households experienced significantly lower expenditure growth, compared to households with low leverage (Mian et al. (2012). I therefore use pre-crisis household leverage as an initial shock to expenditure growth, which is not subject to simultaneity concerns.8 I thus define the trade demand shock at the state i, industry s level as follows: TDSis ¼

N φ X ijs j¼1

Y is

Lev j

ð3Þ

That is, TDS is the weighted sum of destination-state pre-crisis household leverage, where the weight for state j is state i – industry s specific output share that depends (directly or indirectly) on final demand in state j. I then estimate the reduced form effect of the trade demand shock variable on industry employment9: Δ logðLis Þ ¼ β0 þ β1 TDSis þ γ i þ α s þ εis

ð4Þ

I also consider employee earnings and the average wage (defined as earnings over employment) as dependent variables. By adding a state fixed effect γi, the estimation makes use of differences in trading patterns across industries within a state. The industry fixed effect αs controls for shocks that hit all producers in a specific industry. The main endogeneity concern for the estimation of Eq. 4 is potential omitted variable bias: In addition to a demand shock that is transmitted through trade, a particular state-industry pair may be subject to other shocks that also affect economic outcomes. For instance, employment may decline due to a contraction in local credit supply, or due to expectation shocks. To the extent that these other shocks vary only across states or across industries, they are absorbed by the state and industry fixed effects. However, credit supply shocks, for instance, likely differed across U.S. states and industries. To the extent that these other shocks are specific to a state-by-industry pair, they are part of the error term εis. The validity of the identification assumption depends on what drives the variation in the measure of the trade demand shock. That is, why do trading patterns differ across industries that are located in the same state? I argue that these differences are to a large part driven by transportation costs. While greater distance between two states reduces the trading volume in every industry, the effect of distance is much smaller in industries with naturally lower transportation costs. In section 5 I show this pattern in the data by estimating a gravity model on bilateral industry-level trade flows. The effect of distance systematically varies with an industry's value-to-weight ratio, a common (inverse) measure of transportation cost. The effect of transportation costs on trade implies that in highleverage states, industries with lower transportation costs should 8 In Mian et al. (2013) the authors use the household fall in net worth induced by falling house prices instead of household leverage to explain the fall in consumer demand. All my main results hold if I use their measure of the housing net worth shock instead of pre-crisis household leverage. 9 I only present results for the reduced form, because of problems in reliability for consumer expenditure data at the state level. For a summary of concerns raised about the consumer expenditure survey, see for instance Cantor et al. (2013).

S. Stumpner / Journal of International Economics 119 (2019) 169–180

experience a lower demand shock (compared to industries with high transport costs), because they ship more out-of-state. Similarly, in low leverage states, they should be relatively more affected than hightransportation cost industries. For other shocks (like credit supply or expectation shocks) to play a confounding role, they would have to have similar cross-sectional effects. Since transportation costs are a technological characteristic of an industry, there is no immediate reason to believe that other shocks would exhibit this particular cross-sectional pattern. In a robustness exercise, I use the predicted trade flows from the gravity model to compute the trade demand shock. This allows me to focus only on the variation in trade flows that can be traced back to differences in transportation costs across industries. Additionally, I also test robustness to the inclusion of a measure of credit supply. Both robustness tests do not affect the main results of the paper. Finally, one may worry that cross-border state flows are correlated with cross-border company holdings, and that local shocks are transmitted through a network of company ownership rather than goods trade. I argue that this concern is not very important in the present setting. First, work by Atalay et al. (2012) using the CFS micro-data has shown that shipments between plants of the same company only account for a small share (16%) of all shipments. The large majority of shipments is therefore not related to common ownership. Second, if only common ownership of plants mattered, the direction of trade flows would not be important for spillovers to other states. However, I show in a robustness test that it is only shipments to states with high leverage that played a role in the contagion, not shipments from these states. This lends support to the view that the shocks are transmitted via trade flows, and not through cross-border company holdings. The variation in the trade demand shock variable is generated both by variation in household leverage across states and by variation in trade shares across industries within a state. Across states, household leverage ranges from 0.87 in North Dakota to 2.45 in California, thus providing substantial variation across states. Within a state, industries differ in the trade demand shock because of a different composition of trade shares. Fig. 2 visualizes the within-state variation of the trade demand shock. Each vertical bar in the graph represents one state, and states are sorted on the horizontal axis by their 2006 household leverage. For each state, the graph then plots the 25th–75th percentile, and the min and max values of the trade demand shock. On average, industries in states with higher leverage face higher trade demand shocks. This is a natural consequence of the home bias in shipments. However, this variation across states is controlled for by a state fixed effect. To identify the trade channel, I use instead the variation across industries within a state, which is represented in the graph by the vertical spread of the boxes.10

173

Fig. 2. Intra-state heterogeneity in the trade demand shock. This figure shows a box plot of the trade demand shock at the state-industry level vs. state-level 2006 household leverage. Each vertical bar represents one state. The thick boxes show the range of values spanned by the 25th and 75th percentile. The thin lines show the range spanned by the extreme values.

Some caveats of the data require more detailed discussion. First, for some combinations of origin, destination, and industry, the Census Bureau only observes very few shipment records. In case the precision of the estimates is too low, the CFS withholds information and these entries appear as missing values in the data. I use the state-to-state industry-level table in which these missing entries only account for 15% of the total value of trade flows.12 For the empirical implementation, I set these missing entries to zero. For D.C., Alaska, and Hawaii, the table features many missing trade flows, which is why I exclude these states from the sample. Finally, to arrive at a more homogeneous sample of industries, I focus on 39 of the 45 NAICS industries covered in the CFS, namely all manufacturing (21) and all wholesale trade (18) industries.13 Second, the CFS data also capture shipments that are related to international exports. For exports, the CFS records the U.S. port of exit as domestic destination and in my data I cannot distinguish between exports and domestic shipments. While the CFS does collect data on exports, these information are not made available in the state-to-state industry files. However, data on exports are available at the aggregate level. According to this information, exports account for only 8.1% of the total value of shipments in the 2007 CFS. A robustness test in the appendix shows that the main results do not change when I exclude the most export-oriented industries from the sample.

3. Data 3.1. Trade flow data

3.2. Other datasets

To measure trade flows across states I use data from the Commodity Flow Survey (CFS). The CFS is conducted every five years by the Bureau of Transportation Statistics and the U.S. Census Bureau as part of the Economic Census. A sample of firms in manufacturing, wholesale trade, and a few other industries is asked to report individual shipments of goods throughout the year.11 From these individual shipments, the Census Bureau estimates trade flows at a more aggregate level. In particular, the sampling frame of the CFS allows to estimate representative state-by-state trade flows at the industry level. I use the 2007 publicly available table on state-to-state trade flows at the industry level to measure pre-crisis trade flows between states.

Annual data for employment and wage payments of industries at the state and county level come from the U.S. Census County Business Patterns (CBP). For monthly data on employment of industries I turn to the publicly available files from the BLS Quarterly Census of Employment and Wages (QCEW). The level of aggregation of industries is higher in the CFS data compared to the employment and earnings data, and therefore the CFS data determine the level of the analysis. A full list of industries used in the analysis is provided in table A.1 in the appendix.

10 A regression of the trade demand shock on state and industry fixed effects delivers an R2 of 0.72. 11 The CFS only reports trade of physical goods, but not trade in services.

12 Total outgoing flows for a state-industry are reported separately. This allows calculating the fraction of total flows that missing values account for. 13 I therefore exclude Mining (NAICS 212), Electronic Shopping and Mail-Order Houses (4541), Warehousing and Storage (4931), Newspaper, Periodical, Book, and Directory Publishers (5111), Fuel Dealers (45431), and Corporate, Subsidiary, and Regional Managing Offices (551114).

174

S. Stumpner / Journal of International Economics 119 (2019) 169–180

Table 1 Summary statistics.

Employment 2007 Earnings 2007 (Mil. USD) Av. Wage 2007 Employment Growth 2007–09 Earnings Growth 2007–09 Wage Growth 2007–09 Trade Demand Shock

N

Mean

Median

StdDev

10p

90p

1620 1620 1620 1620 1620 1620 1620

11,442 564 46,728 −0.08 −0.062 0.017 1.428

5023 226 44,724 −0.083 −0.08 0.008 1.374

18,039 1038 13,375 0.164 0.2 0.101 0.254

823 32 32,045 −0.273 −0.292 −0.08 1.153

28,889 1379 63,420 0.118 0.176 0.123 1.762

This table shows summary statistics for variables at the state-by-industry level. The trade demand shock is defined as the weighted average of destination-state household leverage ratios. For state i and industry s, it is given by TDSis ¼ ∑ j

φij ðsÞ Y i ðsÞ Lev j ,

where

φij ðsÞ Y i ðsÞ

is the share of output by industry s in state i that depends, directly or indirectly, on final demand in state

j, and Levj is household leverage of state j.

Finally, I use data on county-level household leverage in 2006 from the Federal Reserve Bank of New York Consumer Credit Panel. These data are based on a 5% random sample of individuals who have credit reports with Equifax. Leverage is computed as county-level debt divided by income, where debt is measured as the sum of mortgage debt, auto debt, and credit card debt, and income is adjusted gross income from 2006 county IRS data. Mortgages are by far the most important debt category, accounting for roughly 75% of total debt (on average across counties). The NY Fed data are only available for roughly two thirds of U.S. counties, as counties with an estimated population of less than ten thousand consumers with credit reports are excluded. However, the counties in the sample account for roughly 80% of the 2006 U.S. population. I then use population weights in order to arrive at a measure of pre-crisis household leverage at the state level. Leverage is highest in California, Nevada, Arizona, and Florida and is lowest in North Dakota, Mississippi, and West Virginia. 3.3. Summary statistics Data from the County Business Patterns provide information on 1743 state-by-industry cells of manufacturing and wholesale trade industries. However, some of these cells are very small and the CFS data do not show any shipments. This leaves me with a sample of 1620 observations (837 from manufacturing and 783 from wholesale trade) that account for over 99% of employment. Total 2007 employment in these industries amounted to 18.6 million jobs, of which 70% are in manufacturing and the remaining 30% in wholesale trade. Summary statistics of employment, earnings, and the trade shock are presented in Table 1. On average, employment fell by 8.0% between 2007 and 2009, with substantial heterogeneity across industries (standard deviation of 16.4%). Moreover, variation in employment growth accounts for most of the variation in earnings growth. Fig. A.1 in the appendix shows that total employment growth declined much more strongly in states with high household leverage, and therefore reaffirms the role of leverage as a measure of initial demand shocks. Finally, if trade was in fact important for the regional propagation of the recession, then we would expect that tradable industries account for a larger share of the employment decline in states without initial demand shocks (i.e. low-leverage states). Fig. A.2 shows that this is indeed the case. It plots the jobs lost in tradable industries (defined as manufacturing and wholesale trade) as a share of the total jobs lost against pre-crisis household leverage. It shows that tradables account for half of the jobs lost in low-leverage states, but only for a low share (b20%) of jobs lost in high-leverage states. 4. Results In this section, I first document that the trade demand channel had a significant negative effect on employment and earnings during the recession. Next, I show that this co-movement was not restricted to the crisis period. The same industries selling to states with housing boombust cycles grew relatively faster during the housing boom period

preceding the recession. Together, these results show that trade links business cycles across space. 4.1. Main results I estimate Eq. 4 using growth in employment, earnings, and the average wage as left hand side variables. By definition of the average wage, the coefficients in the employment and wage regressions have to sum up to the coefficient obtained in the earnings regression. For each dependent variable, I run an unweighted and a weighted least squares specification, using 2007 employment of the state-industry cell as weight. Standard errors are twoway clustered at the state and industry level. If the trade channel is relevant for transmitting shocks, we would expect β1 b 0. Table 2 shows a negative and large effect of the trade demand shock on state-industry level employment and earnings. The point estimate implies that a one standard deviation increase in the trade demand shock causes a reduction in employment growth by approximately 2.9 percentage points (= − 0.114 * 0.254). This corresponds to 18% of the standard deviation of employment growth. The fall in employment accounts for most of the earnings adjustment (70%–80%), while the remainder is accounted for by the average wage.14 The employment coefficient of −0.11 is considerably larger than the estimates in Mian and Sufi (2012) who look at the relationship between household leverage and local nontradable employment growth and report a coefficient of −0.02. This difference likely reflects two effects. First, there are compositional changes in demand, i.e. a shift in spending away from tradable goods. In the context of international trade, these compositional changes have already been noted by Levchenko et al. (2010). Second, my estimates document a relative employment decline across industries (as opposed to across counties) and therefore capture not only net job losses, but also worker reallocation across industries. The main result is visualized in Fig. 3 which shows averages for employment growth and the trade demand shock for 25 quantiles of the trade demand shock. Both variables are demeaned by state and industry fixed effects before computing employment-weighted averages for each bin. The graph demonstrates the systematic negative relationship between these two variables. 4.2. The trade channel over the recent business cycle So far, the analysis has focused on the 2007–09 recession. In this subsection, I expand the view to a ten-year window, including the prerecession and post-recession periods. To track the recent business cycle, I estimate Eq. 4 for a sequence of rolling windows of two-year employment growth. I use monthly employment data available from the BLS for a better alignment with the key events during the Great Recession. I start with the window of Jan 2002–Jan 2004, and end with Dec 14 The average wage adjustment could reflect either changes in the hourly wage or the number of hours worked.

S. Stumpner / Journal of International Economics 119 (2019) 169–180

175

Table 2 The Effect of the Trade Demand Shock on Industry Growth.

TDS Observations R-squared Industry FE State FE Weights

(1) Employment

(2) Employment

(3) Earnings

(4) Earnings

(5) Wage

(6) Wage

−0.103** (0.040) 1620 0.354 ✓ ✓

−0.114*** (0.025) 1620 0.559 ✓ ✓ ✓

−0.123** (0.048) 1620 0.384 ✓ ✓

−0.163*** (0.031) 1620 0.541 ✓ ✓ ✓

−0.020 (0.023) 1620 0.226 ✓ ✓

−0.049*** (0.014) 1620 0.278 ✓ ✓ ✓

This table shows results for regressions of industry growth on the trade demand shock. An observation is a state-industry cell. Weights are 2007 employment levels. Standard errors are clustered at the state and industry level.

Fig. 3. Trade Demand Shock and Employment Growth. The graph shows employmentweighted averages of the trade demand shock and employment growth for 25 quantiles of the trade demand shock variable. Before dividing them into bins and computing weighted averages, both variables are demeaned by state and industry fixed effects.

2009–Dec 2011 to include the evolution of the coefficient during the years of the credit boom, the crisis, and the recovery. If trade links business cycles across states, then the coefficient should be positive before the crisis. As Mian and Sufi (2011) document, high pre-crisis leverage was largely a result of high growth in leverage during the years 2002–2006. In places with high growth of house prices, homeowners extracted new debt from the rising value of their homes to finance consumption growth. We would then expect that industries that depended heavily on final demand of states with high pre-crisis leverage were booming in the pre-crisis period. Fig. 4 shows that this pattern holds in the data. It is only in 2007Q4 that the coefficient turns negative (roughly 1.5 years after the CaseShiller house price index peaks). The lowest point is reached in 2009Q3, shortly before the peak of the national unemployment rate. Afterwards, the coefficient slowly reverses and approaches zero during the period of the recovery. By 2011, industries selling primarily to high-leverage states show no significant difference in growth from the control industries. Fig. 4 is also evidence that the results obtained in the previous subsection are not caused by different pre-existing trends. If that was the case, industries selling to high-leverage states would have been growing more slowly even before the crisis. The fact that the coefficient is positive confirms instead the role of trade in transmitting shocks across space. Finally, the timing of the employment decline in Fig. 4 also suggests that the coefficient does not pick up the decline in international trade. Eaton et al. (2016) document that international trade collapsed particularly in 2008Q4 and 2009Q1, which is clearly after the decline in the

Fig. 4. The transmission of boom and bust through trade. This graph shows a one-standard deviation effect of the trade demand shock on employment growth, estimated from regressions of 2-year intervals of employment growth on the trade demand shock and state and industry fixed effects. The 95% confidence intervals are computed using twoway-clustered standard errors. The vertical lines show important events during the Great Recession: (i) The time at which the Case-Shiller house price index peaks (May 2006), (ii) the fall of Lehman (Sept 2008), and (iii) the month at which the national unemployment rate peaks at 10% (Oct 2009).

coefficient on the trade demand shock variable starts and after it turns negative for the first time. 4.3. Heterogeneous effects I now present evidence from two sets of tests that focus on heterogeneous effects across industries. Results from both tests strengthen the interpretation of the previous results as demand shocks being transmitted through trade.15 4.3.1. Differentiated vs. homogeneous goods The main idea of contagion through trade is inherently related to the idea that producers cannot substitute their customers easily. If an industry's output was sold on a single spot market, then heterogeneity in demand shocks across regions should not introduce heterogeneity in producer growth: Producers could substitute costlessly between buyers, and demand shocks are averaged out between producers. I use this intuition to test for the idea that the employment effect should be stronger in industries that sell differentiated products. Rauch (1999) classifies products into three categories: Those sold on organized exchange markets, products that are reference-priced, and all other 15 The analysis in this section is restricted to manufacturing, since both indicators used here are not available for wholesale trade industries.

176

S. Stumpner / Journal of International Economics 119 (2019) 169–180

products that are labeled “differentiated”. I code these products as 1 for differentiated products and 0 for all others. I then use product-industry concordance tables to compute an average degree of differentiation at the level of NAICS industries.16 I then consider the following specification: Δ logðLis Þ ¼ β0 þ β1 TDSis þ β2 TDSis  Diff s þ γ i þ α s þ εis

ð5Þ

Table 3 Differential Effects: Product Differentiation and Product Durability.

TDS TDS ⋅ Diff

(1)

(2)

(3)

Employment

Earnings

Employment

Earnings

−0.074* (0.041) −0.080*** (0.019)

−0.113** (0.055) −0.093*** (0.036)

−0.103*** (0.037)

−0.148*** (0.050)

−0.057*** (0.016) 837 0.537 ✓ ✓ ✓

−0.060** (0.024) 837 0.501 ✓ ✓ ✓

TDS ⋅ Durable

where Diffs denotes the product differentiation measure. Results are in columns 1 and 2 of Table 2. As expected, the estimations indicate that the effect is stronger for industries producing more differentiated products. 4.3.2. Durable vs. nondurable goods The size of the effect might also differ between industries producing durable vs. nondurable goods, since the demand for durable goods tends to be more cyclical.17 I use the definitions by the BLS to divide manufacturing industries into durable and nondurable goods producers and then re-run eq. 5, replacing the measure of differentiation by an indicator for durable goods. Results are in columns 3 and 4 of Table 3, and show that the effect is roughly 40–60% stronger for durable goods industries.

Observations R-squared Industry FE State FE Weights

Table 4 Direct and Indirect Effects.

Direct TDS

where Θ ≡ A + A2 + A3 + …. The element ωikst of Ω thus equals ωikst ¼ 1k¼i;t¼s þ θikst , where θikst is an element of Θ and 1k¼i;t¼s denotes an indicator variable that equals one if k = i and t = s. We can then re-express the share of gross output of sector s in state i that depends on final demand in state j as the sum of two terms:

φijs ¼ þ Y is Y is

Indirect TDS Observations R-squared Industry FE State FE Weights

(1)

(2)

(3)

Employment

Employment

Earnings

(4) Earnings

−0.094** (0.039) −0.110** (0.047) 1620 0.354 ✓ ✓

−0.098*** (0.035) −0.130*** (0.024) 1620 0.560 ✓ ✓ ✓

−0.122** (0.049) −0.123** (0.055) 1620 0.384 ✓ ✓

−0.158*** (0.053) −0.169*** (0.027) 1620 0.541 ✓ ✓ ✓

This table shows results for regressions of industry growth on the components of the trade demand shock capturing direct and indirect linkages to states of final demand. An observation is a state-industry cell. Weights are 2007 employment levels. Standard errors are clustered at the state and industry level.

Ω ¼ ðI−AÞ−1 ¼ I þ Θ;

X Cijs

837 0.502 ✓ ✓ ✓

This table shows results for regressions of growth of employment and earnings at the state-industry level on the trade demand shock and its interaction with the Rauch index of product differentiation (columns 1 and 2) and the index of product durability (columns 3 and 4). Weights are 2007 employment levels. Standard errors are clustered at the state and industry level.

4.4. Direct and indirect effects The main result reflects both pass-through of demand shocks through direct linkages to high-leverage states, and indirect linkages through intermediate goods. In this subsection, I separately estimate the role of direct and indirect effects by splitting up the RHS variable into two components. This decomposition starts from the idea that the Leontief inverse can be rewritten as

837 0.537 ✓ ✓ ✓

(4)

The trade demand shock is then decomposed as follows: TDSis ¼

N XC X ijs

Lev j þ

N S N X ∑k¼1 ∑t¼1 θikst X Ckjt

!

Y is |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}

Lev j Y is |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}

Direct Exposure

Indirect Exposure

j¼1

ð6Þ

j¼1

N S ∑k¼1 ∑t¼1 θikst X Ckjt

Y is

The first term on the RHS captures direct exposure to demand shocks in state j through the sale of consumption goods. It simply equals the fraction of total sales of state i, industry s, that goes towards consumption of state j. The second term captures indirect exposure to final demand in j through first- and higher-order intermediate goods linkages. The output of state i, industry s is used (through first and higher-order linkages) as input in the production of state k, industry t, whose output is then consumed in state j. The fraction

θikst X Ckjt Y is

measures

the share of state i industry s output that depends on final demand in state j through this trajectory. This share is then summed over all using industries t and all third states k to arrive at the total indirect exposure to demand in state j. 16 The industry measure is lowest for Petroleum and Coal Products (0.12), Primary Metal Manufacturing (0.14) and Paper Manufacturing (0.19), and highest for Transportation Equipment, Furniture, Apparel, and Printing (all with a score of 1). 17 The role of compositional changes of GDP for international trade are also discussed in Eaton et al. (2016) and Bems et al. (2011).

I use both these variables on the RHS of the estimation. Results are in Table 4. The coefficients are negative and significant for both variables, showing that both direct and indirect exposure were important in the transmission of demand shocks. The magnitudes of the two effects are very similar and very close to the estimates obtained in the main set of regressions. This suggests that only an industry's total exposure to demand shocks matters for its employment outcomes, but not the division into direct and indirect linkages. 5. Robustness This subsection presents a set of robustness checks, each addressing a specific concern one may have about the previous estimations. Further robustness checks can be found in the appendix. 5.1. Controlling for the credit channel The presumably most important alternative channel that may have contributed to the geographic spread of the recession is the credit channel. In particular, several papers have documented the decline in bank lending and resulting employment losses, following in particular the

S. Stumpner / Journal of International Economics 119 (2019) 169–180

collapse of Lehman Brothers in the fall of 2008.18 Ivashina and Scharfstein (2010), for instance, show that new loans to large borrowers fell by 79% between the second quarter of 2007 and the fourth quarter of 2008 in their sample of syndicated loans. Cornett et al. (2011) report that banks that held more illiquid assets on their balance sheets were more likely to reduce lending, indicating that at least part of this reduction in lending was caused by a fall in credit supply. In this section, I control for credit supply shocks, using data on banklevel lending and the location of bank branches. I use these information to compute the growth in state-industry-level bank loans. Using both OLS and IV estimation, the results provide some evidence for the role of the credit channel in lowering employment, consistent with previous literature (Gozzi and Goetz (2010), Greenstone et al. (2014), Chodorow-Reich (2014)). Most importantly, the estimates of the trade channel remain completely unchanged once the credit channel is controled for. Loan growth at the state-by-industry level is constructed using data from three sources. Data on the universe of bank branches are provided by the FDIC. In particular, I use information on branch location, the identity of the bank owning the branch, and the amount of deposits held in the branch to construct a measure of bank pre-crisis (2007) market shares.19 I then match these data with bank financial information from Call Reports. To account for potential risk sharing across banks within the same holding company, banks are aggregated to the bank-holdingcompany (BHC) level. Finally, to avoid changes in balance sheet variables that are due to M&A as opposed to the normal bank operations, I delete all banks that were engaged in M&A across different bank holding companies. The final bank branch dataset consists of 43,000 branches belonging to 6300 banks (or bank holding companies). I focus on the growth of commercial and industrial loans on bank balance sheets and first compute loan growth at the county (c) level as follows: Loan Growthc ¼

X

scb Loan Growthb ;

b

where scb is the market share of bank b in county c. I then merge these information on county-level loan growth with information on the within-state location of an industry. The resulting measure of state-industry level loan growth is computed as follows: Loan Growthis ¼

X Lics c

Lis

Loan Growthc

That is, it is a weighted average of county-level loan growth, where the weights come from an industry's distribution of employment across counties within a state.20 The idea is that if an industry in a state is predominantly located in counties populated by banks with negative credit supply shocks, then this industry is more exposed to these negative credit supply shocks than if it was located elsewhere in the state. A potential concern with this measure is that distance between a bank branch and a borrowing firm might play no role in the matching between banks and firms. Even though the role of distance for business lending has fallen over the last decades, however, it still seems to play a significant role. Petersen and Rajan (2002) report that although the average borrower-lender distance has declined over the period 1973–1993, the median distance in 1993 still remained very low 18 A detailed description of events in financial markets can be found in Brunnermeier (2009). 19 I delete all branches of banks that do not file Call reports (prior to 2012, Office of Thrift Supervision (now OCC) institutions filed a quarterly Thrift Financial Report), all branches that do not offer full services (like drive-through facilities, administrative offices, etc.), and branches of banks that specialize in activities other than commercial lending (i.e. agriculture, credit cards, mortgages, consumer lending, and other specializations). 20 For small industries and counties, employment information is withheld by the Census Bureau. In that case, I impute employment information using data on the size class of establishments provided at the county-industry level.

177

(5 miles). In more recent work, Brevoort et al. (2010) study data until 2003. They find that there was only a modest increase in borrowerlender distance over this period and that the trend of growing borrower-lender distance has come to a halt in the second half of their study period (1998–2003). Overall, they conclude that “distance still matters”. I add state-industry loan growth as an additional control to the main regression, leading to the equation Δ logðLis Þ ¼ β0 þ β1 TDSis þ β2 Loan Growthis þ γ i þ α s þ ε is

ð7Þ

Given that loan growth is determined by both credit demand and credit supply, I also consider IV specifications. In particular, I use the pre-crisis values for the capital ratio, the share of non-performing loans in total loans, the share of illiquid assets in total assets, and offbalance-sheet commitments in the form of open credit lines as instruments for loan growth.21 This follows the evidence in Cornett et al. (2011) who show that these variables were important predictors of bank lending during the recession. Results can be found in Table 5. Columns 1–4 show results from a least squares estimation, while columns 5–8 use IV. As expected, loan growth is positively related to employment and earnings growth. Most importantly, including the credit channel control variable leaves the estimate of the trade channel unchanged, both qualitatively and quantitatively. 5.2. Restricting the variation in trade flows In this exercise, I limit the variation in trade flows to the part that can be explained by different transportation costs across industries. I compute predicted trade flows from a gravity framework, using state-tostate distance and the industry-specific value-to-weight ratio to measure transportation costs. Using these predicted trade flows, I recompute the trade demand shock.22 In a final step, I use the newly constructed variable as an instrument for the trade demand shock and find that the main results are unchanged.23 The value-to-weight measure is calculated from the CFS data as the aggregate value of all shipments of an industry divided by the aggregate tonnage of shipments. Differences in the value to weight ratio are enormous as it ranges from 110 per ton shipped (Nonmetallic Mineral Products) to 71,000 per ton (Computer and Electronic Products). Distance is defined as the great circle distance between the population-weighted centers of two states.24 I follow the literature on the estimation of gravity equations (Head and Mayer (2013)). A wide variety of trade models yields an expression for trade flows that can be written in log form as follows:       log X ijs ¼ logðGÞ þ logðSis Þ þ log M js þ log ϕijs As before, Xijs stands for the value of sales from state i to state j in industry s. Sis denotes all factors that promote exports of industry s in state i to all destinations, and Mjs all factors that promote imports. Finally, the variable ϕijs captures trade costs, and G is a constant. To estimate the 21 That is, I first compute these variables at the state-by-industry level using the same procedure used to compute loan growth at the state-industry level, and then employ these variables as instruments. The first-stage F-stat is 22.2 22 This step also includes recomputing the regional input-output matrix, as detailed in section 2 and in the appendix. 23 A similar strategy has been used by Do and Levchenko (2007). 24 The log of distance is then undefined for within-state shipments. In order to include within-state shipments in the estimation, I make use of additional data in the CFS. For many (though not all) origin-destination-industry cells, the CFS reports the average distance traveled of shipments. While these data are often missing when the origin state is different from the destination state, they are almost always available for within-state shipments (because of their size). For within-state shipments, I therefore use this number as a measure of distance. Results are very similar, however, when I exclude within-state shipments.

178

S. Stumpner / Journal of International Economics 119 (2019) 169–180

Table 5 Controlling for the Credit Channel. (1)

(2)

(3)

(4)

(5)

(6)

(7)

Employment

Earnings

Employment

Earnings

Employment

Earnings

Employment

Earnings

0.231*** (0.073) 1620 0.535 ✓ ✓ ✓ LS

−0.107*** (0.025) 0.185*** (0.068) 1620 0.563 ✓ ✓ ✓ LS

−0.156*** (0.033) 0.205*** (0.078) 1620 0.544 ✓ ✓ ✓ LS

0.239** (0.093) 1620 0.535 ✓ ✓ ✓ IV

−0.108*** (0.021) 0.177* (0.091) 1620 0.563 ✓ ✓ ✓ IV

−0.156*** (0.027) 0.208** (0.096) 1620 0.544 ✓ ✓ ✓ IV

TDS Loan Growth Observations R-squared Industry FE State FE Weights Specification

0.202*** (0.067) 1620 0.556 ✓ ✓ ✓ LS

0.198** (0.088) 1620 0.556 ✓ ✓ ✓ IV

(8)

This table shows results for estimations of employment and earnings growth at the state-by-industry level on the trade demand shock and state-industry specific loan growth. Columns 1– 4 use least squares estimation, while coloumns 5–8 use 2SLS, instrumenting for loan growth. The instruments used are pre-crisis values for the capital ratio, the share of non-performing loans in total loans, the share of illiquid assets in total assets, and the value of off-balance-sheet commitments as a fraction of total assets. Weights are 2007 employment levels. Standard errors are clustered at the state and industry level.

effect of trade costs on trade flows, I employ a fixed effects model which is standard in the gravity literature. To that end, I use exporter-industry and importer-industry fixed effects to control for Sis and Mjs. Modeling trade costs using an interaction between distance and the value-to-weight ratio parsimoniously captures the heterogeneous effect of distance on trade flows across industries. More specifically, I model trade costs as       log ϕijs ¼ β log Distanceij þ δ log Distanceij  logðValue−to−weights Þ;

5.3. The spread of the recession across state Borders Next, I employ two tests to investigate specifically the diffusion of the crisis across state borders. For a first test, I define a new variable which only captures variation in demand shocks through out-of-state trade flows: External TDSis ¼

X

φijs

j≠i

∑ j≠i φijs

Lev j

ð8Þ

where Value − to − weights is the value-to-weight ratio of an industry in $1000 per ton. To see that this model does a good job at fitting the heterogeneity in trade costs across industries, consider a nonparametric approach: S X     log X ijs ¼ μ þ α is þ γ js þ βv  log Distanceij  Iv þ εijs v¼1

That is, I first estimate the effect of distance on trade flows industryby-industry, where Iv is a dummy for industry v. Fig. 5 plots the βv coefficients against the log of the value-to-weight ratio (in 1000 per ton) of the respective industry. As expected, the graph shows a positive relationship between the two: Distance is less of a barrier for trade flows in industries with a higher value-to-weight ratio. Moreover, the graph shows that a linear fit does a good job at describing the heterogeneity of the effect of distance across industries. This suggests the following estimation equation:

Just as the variable TDS, this variable is measured in leverage points, since the weights sum up to one. I then re-do the main estimations using this variable as regressor. Results can be seen in columns (3) and (4) of Table 5. The effect is negative and significant for both employment growth and earnings growth. An alternative test for cross-border effects decomposes the trade shock into a component reflecting demand shocks in the “Home” state vs demand shocks in “Foreign” states. TDSis ¼

φiis Levi Y is |fflfflfflffl{zfflfflfflffl}

}Home} Demand

þ

N φ X ijs

Lev j Y is j≠i |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}

ð9Þ

}Foreign} Demand

    log X ijs ¼ μ þ α is þ γ js þ log ϕijs þ εijs : with trade costs log(ϕijs) specified as in eq. 8. Estimating this equation, I ^ ¼ −1:43 (std 0.019) and ^δ ¼ 0:145 (std 0.012), with N find values of β = 27,836. As expected, there is a strong negative effect of distance on trade flows, which is muted in industries with high value-to-weight ratios. I then compute predicted trade flows using only the part of the variation in trade flows that can be explained by heterogeneous transport costs across industries.     ^ ijs ¼ μ^ þ log ϕ ^ log X ijs With these predicted trade flows in hand, I recompute the regional input-output matrix, and use the resulting variable as an instrument for the original trade demand shock. The first stage delivers a coefficient value of 0.659 (std = 0.080), and an F-stat of 68. Table 6, Columns 1–2 shows the results for the second stage. For both employment and earnings, the coefficients stays negative and significant.

Fig. 5. Distance Effect on Trade vs. Value-to-Weight Ratio. This figure plots the distance effect on trade flows against the industry-level value-to-weight ratio. The vertical axis plots the estimated βv coefficients for the regression log(Xijs) = μ + αis + γns + ∑Sv=1βv × log (Distanceij) × Iv + εijs. Xijs are trade flows from state i to state j in industry s. The variable Iv is a dummy variable for industry v. The horizontal axis measures the log of the value-to-weight ratio (in 1000 per ton) for an industry.

S. Stumpner / Journal of International Economics 119 (2019) 169–180

179

Table 6 Robustness.

TDS

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Employment

Earnings

Employment

Earnings

Employment

Earnings

Employment

Earnings

−0.212*** (0.044)

−0.297*** (0.037)

−0.109*** (0.030)

−0.167*** (0.041)

−0.008 (0.036) 1620 0.559 ✓ ✓ ✓

0.007 (0.054) 1620 0.541 ✓ ✓ ✓

−0.101*** (0.037)

External TDS

−0.166*** (0.057) −0.123*** (0.028) −0.103*** (0.026)

Home Demand Shock Foreign Demand Shock

−0.182*** (0.036) −0.142*** (0.034)

Reverse TDS Observations R-squared Industry FE State FE Weights IV

1620 0.553 ✓ ✓ ✓ Pred. Trade

1620 0.533 ✓ ✓ ✓ Pred. Trade

1620 0.556 ✓ ✓ ✓

(8)

1620 0.539 ✓ ✓ ✓

1620 0.560 ✓ ✓ ✓

1620 0.543 ✓ ✓ ✓

This table shows results for several robustness exercises. Columns 1–2 show results using predicted trade flows, instead of actual trade flows to construct the RHS variable. The RHS variable in columns 3–4 uses only exposure to demand shocks in other states. Columns 5–6 show results from splitting up the TDS into a “Home State” and a “Foreign States” component. Finally, columns 7–8 show results when adding a variable (Reverse TDS) that measures demand shocks in shipment origin, instead of shipment destination states. Weights are 2007 employment levels. Standard errors are clustered at the state and industry level.

Columns (5) and (6) of Table 6 show the results from this test. The coefficients are negative and significant for both components of the trade demand shock, and of similar size. These tests therefore provide further evidence that the crisis was transmitted across state borders due to trade. 5.4. Placebo test based on reverse trade flows If demand shocks in housing boom and bust states triggered the crisis, then it should be exports to, but not imports from these states that transmit the crisis across space. I use this idea to run a placebo test, constructing the right-hand-side variable with incoming, instead of outgoing trade flows. Based on incoming trade flows, I recompute the regional inputoutput matrix and the RHS variable. This new variable is labeled Reverse TDS and captures the weighted leverage of shipment origin states. If most of bilateral trade is intra-industry (instead of inter-industry), then exports and imports would be highly correlated. As a consequence, the variables using either incoming or outgoing trade flows as weights would yield a high correlation. In contrast, the two variables may differ if a substantial part of trade is inter-industry. After controlling for industry and state fixed effects, I find that this new variable is only moderately correlated with the original trade demand shock measure (ρ = 0.29). I use the original variable and the newly constructed one in a joint regression to separate the effects. Columns (7) and (8) of Table 5 presents the results of this exercise. It shows that the coefficient on the original variable is unchanged, while the coefficient on the variable using reverse trade flows is indistinguishable from zero. This test therefore strengthens the interpretation that trade-transmitted demand shocks spread the crisis across space. 5.5. Further robustness A table with additional robustness tests can be found in the appendix to this paper. Among others, I present results from a test in which the local housing supply elasticity (Saiz (2010)) is used as an IV for household leverage, and a test that excludes industries that rely more on international exports. All of these results support the main findings presented earlier. 6. Conclusion This paper studies the role of trade in the regional propagation of local shocks in the context of the US Great Recession. I argue that this

is a particularly good setting for this question. First, local housing cycles and high household debt have been shown to be at the origin of the recession. This is important because it implies that the initial shocks arose from outside the tradable goods sector and exhibited substantial geographic variation. Second, focusing on within-country rather than across-country contagion presents several advantages. U.S. states are highly integrated, trade frictions between states do not change as a result of the recession, and data availability allows to better control for other potential contagion mechanisms, most notably the credit channel. The main finding is that trade contributed substantially to the regional propagation of the recession. Empirically, I identify the trade channel by comparing economic outcomes of industries with different shipment patterns that are located in the same state. Industries that depended relatively more on final demand in states with housing boom-bust cycles grew by more before the crisis and declined faster during 2007–09. This relative effect of the trade channel is sizable: One standard deviation in the trade demand shock explains a 2.9 percentage point difference in 2007–09 employment growth between industries. Further results suggest that this effect is transmitted through both direct sales to high-leverage states and indirect effects through input-output linkages. The main result holds in a large range of additional tests. Most importantly, it is robust to controlling for shocks to credit supply, and it holds when I restrict attention to the variation in trade flows that arises from different transportation costs across industries. Consistent with the idea of trade-transmitted demand shocks, the results are also specific to trade flows to, but not trade flows from high-leverage states. This paper therefore provides evidence that trade is still important for transmitting crises across space, even within a country like the U.S., where the manufacturing share in GDP has declined for years, and is low in international comparison. Moreover, the estimates may understate the importance of trade because the data only cover trade in manufactured goods but not in services. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.jinteco.2019.04.001.

References Aiyar, Shekhar, 2012. From financial crisis to great recession: the role of globalized banks. American Economic Review P & P 102 (3), 225–230.

180

S. Stumpner / Journal of International Economics 119 (2019) 169–180

Atalay, Enghin, Hortaçsu, Ali, Syverson, Chad, 2012. Why Do Firms Own Production Chains? University of Chicago mimeo Bems, Rudolfs, Johnson, Robert C., Yi, Kei-Mu, 2010. Demand spillovers and the collapse of trade in the global recession. IMF Economic review 58 (2), 295–326. Bems, Rudolfs, Johnson, Robert C., Yi, Kei-Mu, 2011. Vertical linkages and the collapse of global trade. The American Economic Review P&P 101 (3), 308–312. Brevoort, Kenneth P., Holmes, John A., Wolken, John D., 2010. Distance still matters: The information revolution in small business lending and the persistent role of location, 1993-2003. Board of Governors of the Federal Reserve System Finance and Economics Discussion Series (08). Brunnermeier, Markus K., 2009. Deciphering the liquidity and credit crunch 2007-08. J. Econ. Perspect. 23 (1), 77–100. Caliendo, Lorenzo, Parro, Fernando, Rossi-Hansberg, Esteban, Sarte, Pierre-Daniel, 2018. The impact of regional and sectoral productivity changes on the U.S. economy. Review of Economic Studies 85 (4), 2042–2096. Cantor, David, Mathiowetz, Nancy, Schneider, Sid, Edwards, Brad, 2013. Redesign Options for the Consumer Expenditure Survey. BLS and WESTAT mimeo. Centorelli, Nicola, Goldberg, Linda S., 2011. Global banks and international shock transmission: evidence from the crisis. IMF Economic Review 59 (1), 41–76. Chodorow-Reich, Gabriel, 2014. The employment effects of credit market disruptions: firm-level evidence from the 2008-09 financial crisis. Q. J. Econ. 129 (1), 1–59. Clark, Todd E., van Wincoop, Eric, 2001. Borders and business cycles. J. Int. Econ. 55, 59–85. Cornett, Marcia M., McNutt, Jamie John, Strahan, Philip E., Tehranian, Hassan, 2011. Liquidity risk management and credit supply in the financial crisis. J. Financ. Econ. 101 (2), 297–312. di Giovanni, Julian, Levchenko, Andrei A., 2009. Trade openness and volatility. Rev. Econ. Stat. 91 (3), 558–585. di Giovanni, Julian, Levchenko, Andrei A., 2010. Putting the parts together: trade, vertical linkages, and business cycle Comovement. American Economic Journal: Macro 2 (2), 95–124. Do, Quy-Toan, Levchenko, Andrei A., 2007. Comparative advantage, demand for external finance, and financial development. J. Financ. Econ. 86 (3), 796–834. Eaton, Jonathan, Kortum, Samuel, Neiman, Brent, Romalis, John, 2016. Trade and the global recession. Am. Econ. Rev. 106 (11), 3401–3438. Eggertsson, Gauti B., Krugman, Paul, 2012. Debt, deleveraging, and the liquidity trap: a fisher-Minsky-Koo approach. Q. J. Econ. 127 (3), 1469–1513. Enders, Zeno, Peter, Alessandra, 2015. Global Banking, Trade, and the International Transmission of the Great Recession. mimeo. University of Heidelberg. Fogli, Alessandra, Hill, Enoch, Perri, Fabrizio, 2012. The geography of the great recession. NBER Int. Semin. Macroecon. 9 (1), 305–331. Forbes, Kristin J., 2004. The Asian flu and Russian virus: the international transmission of crises in firm-level data. J. Int. Econ. 63 (1), 59–92. Frankel, Jeffrey A., Rose, Andrew K., 1998. The endogeneity of the optimum currency area criteria. Economic Journal 108 (449), 1009–1025. Glick, Reuven, Rose, Andrew K., 1999. Contagion and trade: why are currency crises regional? J. Int. Money Financ. 18, 603–617.

Gorodnichenko, Yuriy, Mendoza, Enrique G., Tesar, Linda L., 2012. The Finnish great depression: from Russia with love. Am. Econ. Rev. 102 (4), 1619–1643. Gozzi, Juan Carlos, Goetz, Martin, 2010. Liquidity Shocks, Local Banks, and Economic Activity: Evidence from the 2007–09 Crisis. Working paper. Greenstone, Michael, Mas, Alexandre, Nguyen, Hoai-Luu, 2014. Do Credit Market Shocks affect the real economy? Quasi-Experimental Evidence from the Great Recession and Normal Economic Times. mimeo. Princeton University and University of Chicago. Head, Keith, Mayer, Thierry, 2013. Gravity Equations: Workhorse, Toolkit, and Cookbook. CEPR Discussion Paper No. 9322. Ivashina, Victoria, Scharfstein, David, 2010. Bank lending during the financial crisis of 2008. J. Financ. Econ. 97 (3), 319–338. Johnson, Robert C., 2014. Trade in intermediate inputs and business cycle Comovement. Am. Econ. J. Macroecon. 6 (4), 39–83. Johnson, Robert C., Noguera, Guillermo, 2012. Accounting for intermediates: production sharing and trade in value added. J. Int. Econ. 86 (2), 224–236. Jones, Charles I., 2013. Misallocation, economic growth, and input-output economics. In: Acemoglu, D., Arellano, M., Dekel, E. (Eds.), Advances in Economics and Econometrics. Vol. 2. Cambridge University Press, pp. 419–455. Kaminsky, Graciela L., Reinhart, Carmen M., 2000. On crisis, contagion, and confusion. J. Int. Econ. 51, 145–168. Kose, M. Ayhan, Yi, Kei-Mu, 2006. Can the standard international business cycle model explain the relation between trade and comovement? J. Int. Econ. 68, 267–295. Levchenko, Andrei, Lewis, Logan, Tesar, Linda, December 2010. The collapse of international trade during the 2008-2009 crisis: in search of the smoking gun. IMF Economic Review 58 (2), 214–253. Mian, Atif, Sufi, Amir, August 2011. House prices, home equity-based borrowing, and the U.S. household leverage crisis. American Economic Review 101 (5). Mian, Atif, Sufi, Amir, 2012. What explains high unemployment? The aggregate demand channel. NBER Working Paper No. 17830. Mian, Atif, Sufi, Amir, 2014. What explains the 2007-09 drop in employment. Econometrica 82 (6), 2197–2223. Mian, Atif, Rao, Kamalesh, Sufi, Amir, 2012. Household Balance Sheets, Consumption, and the Economic Slump. UC Berkeley mimeo. Mian, Atif, Rao, Kamalesh, Sufi, Amir, 2013. Household balance sheets, consumption, and the economic slump. Q. J. Econ. 128 (4), 1687–1726. Midrigan, Virgiliu, Philippon, Thomas, 2011. Household Leverage and the Recession. mimeo. New York University. Monte, Ferdinando, 2014. Local Transmission of Trade Shocks. mimeo. John Hopkins University. Petersen, Mitchell, Rajan, R.G., 2002. Does distance still matter? The information revolution in small business lending. J. Financ. 57 (6). Rauch, James E., 1999. Networks versus Markets in International Trade. J. Int. Econ. 48, 7–35. Saiz, Albert, 2010. The geographic determinants of housing supply. Q. J. Econ. 125 (3), 1253–1296. van Rijckeghem, Caroline, Weder, Beatrice, 2001. Sources of contagion: is it finance or trade? J. Int. Econ. 54, 293–308.