The Agglomeration of Exporters by Destination

Andrew J. Casseya

aSchool of Economic Sciences, Washington State University, 101 Hulbert Hall, P.O. Box 646210, Pullman, WA99164, USA. Phone: 509-335-8334. Fax: 509-335-1173. Email: cassey@wsu.edu.

Katherine N. Schmeiserb,∗

bDepartment of Economics, Mt. Holyoke College, 117 Skinner Hall; 50 College St, South Hadley, MA 01075, USA.Phone: 612-910-2538. Email: kschmeis@mtholyoke.edu.

Abstract

Precise characterization of informational trade barriers is neither well documented nor understood,

and thus remain a considerable hindrance to development. Using spatial econometrics on Russian

customs data, we document that destination-specific export spillovers exist for developing countries,

extending a result that was only known for developed countries. This result suggests behavior

responding to a destination barrier. To account for this fact, we build on a monopolistic competition

model of trade by postulating an externality in the international transaction of goods. We test the

model’s prediction on region- and state-level exports using Russian and U.S. data. Our model

accounts for up to 40% more of the variation than in gravity-type models without agglomeration.

This finding has important development implications in that export policy that considers current

trade partners may be more effective than policy that focuses only on the exporting country’s

industries.

Keywords: agglomeration, distribution, exports, development, location, Russia

JEL: D23, F12, L29

∗Corresponding author

Preprint submitted to . September 21, 2011

1. Introduction1

One of the unsolved problems in economic development is the lack of knowledge about the2

size and nature of barriers to trade, and how these barriers affect growth and income. Great3

theoretical and empirical strides have been made by country-level and firm-level studies, yet the4

precise characterization of barriers to trade continues to elude researchers. A better understanding5

of the size and nature of barriers to trade may lead to better trade-focused development policy.6

Using spatial econometrics on Russian customs data, we document that firms exporting to7

the same destination are located near each other. This agglomeration is beyond the clustering of8

exporters around domestic ports (Glejser, Jacquemin, and Petit 1980; Nuade and Matthee 2007)9

or gross domestic product (GDP). We document that destination-specific export spillovers exist10

in a developing country such as Russia. Previous empirical studies first established that export11

spillovers exist and are destination specific, but the data are from developed countries such as12

the United States (Lovely, Rosenthal, and Sharma 2005), France (Koenig 2009), and Denmark13

(Choquette and Meinen 2011). We also document destination-specific exporter spillovers using14

spatial econometrics instead of a logit estimator. Therefore our results confirm the previous work15

using a different method, and extend it to developing countries.16

Additionally, we develop a theory to account for this destination-specific agglomeration in a17

trade setting. We assume there are economies of scale in the transaction cost of international trade18

at the firm level. In principle, destination-specific export spillovers could work through the fixed19

or variable costs of trade. Econometric results by Choquette and Meinen (2011) provide evidence20

for export spillovers working through variable costs as well as fixed costs. Furthermore, though21

Koenig, Mayneris, and Poncet (2010) do not find robust support for spillovers working through22

the variable cost, they do find evidence of this when the total number of exporters is small. Our23

Russian data shows that the number of exporters selling to the same country from the same region24

in Russia is often small, and thus are more likely to have the destination-specific export spillovers25

working through variable costs. Notwithstanding these results, both Choquette and Meinen and26

Koenig et al. measure agglomeration exclusively as the ratio of similar exporting firms to total27

firms within some geographic area. With this measure, they find that destination-specific export28

2

spillovers work primarily through the fixed cost. We try alternative measures of agglomeration and29

find that variables of group activity such as the aggregate region-country weight of shipments work30

more through the variable costs of trade.31

The Russian customs data show that, by far, most shipments are small in weight. As long as32

shipping prices to exporters include a fixed cost per container (buying the container) or per desti-33

nation (permit to unload) and are based on weight (Micco and Perez 2002), then there are shipping34

cost savings in grouping small exports together on a boat or plane heading to the same destination.35

Alternatively, aggregate export weight may be thought of as the result of an informational spillover36

about foreign market access. The exact nature of the externality is not important for our model or37

results. However, we do offer some initial evidence in distinguishing between the two. We test our38

theory using region-level data from Russia and state-level data from the United States.39

We find that using the region-level export data from Russia, our model with agglomeration40

functioning through the variables costs of trade accounts for 10% more of the variation in export41

share than the benchmark model. Using U.S. state export data, our model accounts for 40% more42

of the variation. Therefore, we have uncovered empirical evidence of a barrier to trade large enough43

for firms to respond to it by agglomerating and we provide theoretic justification for it.44

In our model, an exporter may skip over a destination that has more appealing demand in order45

to achieve decreased transactions costs in selling to a less preferred destination. All else the same,46

this creates variation in the destination choices of otherwise identical firms in different regions.47

Lawless (2009) documents such skipping of “preferred” destinations at the national-level, but her48

evidence is inconsistent with the theory developed in Eaton, Kortum, and Kramarz (Forthcoming).49

Our theory better accounts for the geographic pattern of exporting by creating a regional hierarchy50

of export destinations that differs from the national hierarchy.51

That exporter clustering by destination exists for both developed and developing suggests firms52

are responding to destination-specific transaction costs of trade. This finding offers insight into53

improvement of trade-focused development policy in that the policy should perhaps respect the54

destination of exports in addition to the exporting country’s industrial base and comparative ad-55

vantage.56

3

The transaction-level data that motivates our theory are from Russia. We choose to study57

Russian exporters because Russia is a developing country that is geographically large and diverse.58

It borders fourteen countries (with China and Ukraine being major trading partners), has four59

coasts (Arctic Ocean, Pacific Ocean, Baltic Sea, and Black Sea) and diverse economic subregions.60

Finally, the laws in Russia limit the movement of firms regionally. (See Appendix A for a more61

detailed description). This allows us to model a firm’s export destination decision without also62

modeling its location decision within Russia. Furthermore, though Russia has a unique history63

that may affect development (relationship with former Soviet countries, corruption, and remnants64

of central planning to name a few), Cassey and Schmeiser (2011) show that the Russian data65

qualitatively and quantitatively match firm-level data sets from Colombia and France. Therefore66

the development lessons from this work apply more generally than Russia alone.67

The findings of Cassey and Schmeiser (2011) notwithstanding, because we are aware that Rus-68

sian data may be different from other countries’ data, we test the macro-level predictions of our69

theory on the World Institute for Strategic Economic Research’s (WISER) Origin of Movement70

U.S. state export data (OM). These data contain export information for all 50 U.S. states and the71

same 175 countries that we use with the Russian data. Furthermore, our U.S. data are a panel72

whereas the Russian customs data are a cross-section. Having the time dimension allows us to73

conduct robustness tests on the U.S. data that we cannot do with the Russian data.74

2. Facts on the location & agglomeration of Russian exporters75

In this section, we show that 1) exporters are clustered regionally given output and the location76

of ports and 2) exporters shipping to the same destination are further clustered. The purpose of77

this section is to establish these two agglomeration facts using a spatial econometrics method. Our78

empirical method differs from that used by Koenig (2009) and Choquette and Meinen (2011), who79

use a logit estimator to see if the odds that a firm exports to a particular country depends on the80

ratio of other firms in the same area also exporting to that country to total firms, and also from81

that of Lovely et al. (2005) who put an industry concentration variable on the left hand side of82

a linear model with the difficulty of reaching an export destination on the right hand side. Our83

4

interest differs from the previous work in that we use data from a developing country instead of84

a developed country. Therefore, one way of interpreting the purpose of this section is establishing85

the generality of the above-mentioned agglomeration facts across countries in various stages of86

development. Because of the uniqueness of our data, we begin by describing them.87

2.1. The Russian data88

To document the existence of destination-specific agglomeration of exporters, we use 2003 data89

from the Russian External Economic Activities (REEA) set reported by Russian customs. The90

REEA data are shipment-level customs data of uniquely identified firms, providing value of goods91

shipped (in 2003 USD), weight of shipment, exporter location within Russia, and destination coun-92

try. See Cassey and Schmeiser (2011) for complete description of these data, as well as descriptive93

statistics and stylized facts.94

We only consider manufacturing exports because firms that exported natural resources were95

necessarily located in the region endowed with those resources (Bradshaw 2008). In addition, we96

only use manufacturing data so that we can later have a direct comparison with the OM data for97

the United States.198

From information on the customs form, we use data from firms who are making the export99

decision—who may or may not be the firm that manufactured the final good. We do this because100

we want to use the location of the exporter and not the producer (if they are different). Therefore,101

we include manufacturers, distributors, and wholesalers who are coded as the principal agent in102

interest. We do not include freight-forwarding firms because these firms could bias results if they103

are located somewhere other than the true exporter’s location. Also, we eliminate observations in104

which the “originating firm” is located outside of Russia since we believe they indicate re-exports105

rather than true Russian exports, as well as to be consistent with the U.S. data.106

To check for consistencies in the data, we refer to Cassey and Schmeiser (2011) who compare107

the REEA data with product data from the United Nations and find close agreement. They108

1Cassey (2009) shows that for the United States, manufacturing data are the only data that can be reliably usedfor the state of production of the export instead of the state where the export began its journey abroad, whereasagriculture and mining data are not reliable.

5

also compare the REEA data to statistics obtained with firm-level data from Colombia (Eaton,109

Eslava, Kugler, and Tybout 2008) and France (Eaton et al. Forthcoming). The export patterns110

are qualitatively and quantitatively similar. We are therefore confident that our data are not111

greatly influencing or biasing our interpretation or generality of results, in particular towards any112

geographic region.113

We combine the REEA data with information about the the gross domestic products (GDP) of114

each region using Russia: All Regions Trade & Investment Guide (CTEC Publishing 2004, 2006).115

(We do not have data on the number or size of domestic firms.) We also use the World Economic116

Outlook Database (IMF 2006) for GDP information for the 175 countries in our sample. Finally,117

we calculate the great circle distance from the capital of each Russian region to the capital of each118

country in the world.119

Our data have two large drawbacks that limit the forthcoming empirical analysis. First, though120

we have transaction-level data on shipments that we can match to unique firms, we are not able to121

pinpoint exporter location by exact physical address. The best we can do is assign each exporter122

to one of Russia’s 89 federal regions.2 This limitation forces us to conduct clustering tests on123

the region-level (and therefore to later perform econometric tests of the model at the region-level124

rather than the firm-level). See Appendix A for a detailed description of the political organization125

of Russia and also for a table listing the regions. Second, we only have the customs data and126

thus we are not able to match Russian firms with their characteristics. We do neither know the127

employment, age, or total sales of the firms in our data nor do we know anything about firms that128

are not exporting. Therefore we cannot compare the clustering of Russian manufacturing exporters129

to that of the Russian manufacturing industry or subsectors therein, nor systematically exclude130

other sources of firm clustering such as natural, labor, or infrastructural advantages. Given this131

limitation, we instead show that Russian firms exporting to the same country are non randomly132

located within Russian regions, controlling for variables such as distance, importer GDP, and133

preferences of the importer for particular goods. This kind of clustering cannot be fully explained134

2Russia’s federal regions are somewhat similar in geographic scope to U.S. states. The number of regions has decreasedsince 2003 because of mergers. It is common to study Russia at this level of geographic disaggregation. See Broadmanand Recanatini (2001) for example.

6

without destination-specific export spillovers.135

2.2. Russian exporters and exports136

First, consider the number of exporting firms by region. Figure 1 shows that the regional137

concentration of Russian manufacturing exporters is diverse. The majority of exporters are located138

either around Moscow and St. Petersburg or the Kazakhstan/Mongolia/China border. Nine regions139

did not have any reported exporting firms in 2003. Figure 1 shows that in Russia, as elsewhere,140

many exporting firms are located near major ports. However, the figure also shows that there are141

regions with many exporting firms that are not located near a major port. A map of Russia with142

major rivers and railroads is in Appendix A.143

The top destinations for Russian exports in 2003 were, in order: China, United States, United144

Kingdom, Japan, Ukraine, Kazakhstan, Turkey, Netherlands, Germany, and Iran. Of these, only145

Ukraine and Kazakstan were part of the Soviet Union. Additionally, Russia does not currently146

have a Warsaw Pact nation as a top destination.3 Because we do not see evidence of favored trade147

relations with former Soviet or Eastern bloc countries, we do not treat them any differently than148

Western countries.149

Federal laws in Russia severely limit the ability of Russian firms to change region (Botolf 2003).150

Because of this, firms choose how much to export to each country in the world, taking their location151

as given. Furthermore, because the Soviet Union did not trade with western countries, we believe152

it is likely that the Soviet central planners chose firm locations for reasons other than the expansion153

of trade (see Huber, Nagaev, and Worgotter (1997) and Bradshaw (2008)). Other evidence that154

international trade was not important in Soviet economic planning is that the large Pacific port of155

Vladivostok, home of the Soviet Pacific Fleet, was closed to foreign vessels until 1991.156

2.3. The geography of Russian exporters157

To show that exporters are clustered, we first use the customs data to calculate the number of158

firms in each region that export anywhere in the world. Figure 1 shows this as well as the location159

3Soviet trade was with countries of similar ideology such as Yugoslavia and Poland.

7

Fig

ure

1:

Num

ber

ofex

port

ing

firm

s,by

regio

n.

Cut-

off

sco

rres

pond

toquin

tile

sso

that

ther

eare

an

equalnum

ber

ofre

gio

ns

inea

chca

tegory

.See

App

endix

Afo

ra

map

of

Russ

iaw

ith

the

regio

ns

lab

eled

and

railw

ays

indic

ate

d.

8

of ports, bodies of water, and neighboring countries, without modification for population or GDP.160

Cut-offs for the groups are determined so that each group has the same number of regions.161

Next, we weight the exporter count observations by regional GDP. We compare each region’s162

exporter count to the overall mean. Figure 2 highlights regions with a concentration of exporters163

far from the mean. Darker colors indicate regions whose exporter count, weighted by GDP, is much164

larger than the mean. As seen in the figure, there exist regions in Russia that have statistically165

significantly more exporters than expected based on regional GDP. Furthermore, many of these166

export-concentrated regions do not have a major port. Figure 2 also shows there are regions167

(weighted by GDP) with less exporting firms than the mean. These regions, indicated with the168

lightest shades, are not nearly as far below the mean as the regions much larger than the mean.169

We use this information to calculate the Moran’s I statistic for spatial autocorrelation. The170

expectation of the statistic is −1/(89− 1) under a GDP-weighted random distribution of exporters171

to Russia’s 89 regions. With our data, the Moran’s I statistic for spatial autocorrelation is 0.16∗ (z−172

score = 6.37) indicating that we reject spatial (GDP-weighted) randomness with 99% confidence.173

Additionally, the clustering is not based on regions that have major ports. In Appendix C, we174

conduct two other tests to confirm this result.175

To see what the clustering is based on, consider a specific example: the location of Russian176

firms that export to the United States and the location of Russian firms that export to Canada.177

We choose these two countries because they have similar economies (though the United States is178

much larger), distance from Russia, and consume similar goods (at the 2 digit level, six of the179

top ten imported industries from Russia are the same). The Moran’s I for the location of Russian180

exporters shipping to the United States is 0.04∗∗∗ (z = 1.93) and it is 0.03∗∗∗ (z = 1.68) for Canada.181

Thus there is clustering for exporters to the United States and Canada with 90% confidence. To182

see that it is not the same regions (and therefore the same firms within a region) that specialize in183

exporting to both countries, see figure 3.184

Figure 3 shows the deviation from the mean number of firms that export to the United States185

and the mean number of firms that export to Canada. Notice that there are more regions in Russia186

that have exporters shipping goods to the United States than Canada, but that is not surprising187

9

ALL EXPORTERS

Figure 2: Deviations from the average number of exporting firms, by region, weighted by GDP. Moran’s I (89) = 0.16∗

(z − score = 6.37).

given the difference in GDP and that the United States is the second largest receiving country of188

Russian exports. The second fact is that the regions where exporters shipping to the United States189

are most concentrated (shaded black in figure 3 and indicating more than 2.5 standard deviations190

from the mean) are not the same regions as those in which exporters to Canada are concentrated.191

Therefore, Russian exporters to the United States are clustered in a different pattern than the192

Russian exporters to Canada.193

To support this claim, we perform a bilateral Moran’s I statistic for these two countries. The194

statistic differences the number of exporters in each region i shipping to country c from the average195

number of exporters also shipping to c and compares it to the difference between the number of196

exporters in another region j shipping to country d from the average also shipping to d. Because197

of its construction with respect to the other country d, the bivariate Moran’s I is not symmetric.198

A statistic different from E(I) suggests a spatial pattern (clustering if positive) between exporters199

10

UNITED STATES

CANADA

Figure 3: Deviations from the average number of exporting firms, by region, weighted by GDP. For the Unites States,Moran’s I (89) = 0.04∗∗∗ (z = 1.93) and for Canda, Moran’s I (89) = 0.03∗∗∗ (z = 1.68).

11

Table 1: Moran’s I for Location of Firms Exporting to Select CountriesCanada China Germany Great Japan Poland United Ukraine

Britain States

Canada .0345∗ −.0012 .0243 .0313 .0300 .0196 .0194 .0340China .0089 .0023 −.0448∗ −.0458∗ −.0147 −.0450∗ −.0375∗ −.0408∗

Germany .0249 −.0472∗ .0324∗ .0610∗∗ .0248 .0547∗ .0406 .0637∗∗

Great Britain .0215 −.5160∗ .0677∗∗ .0886∗∗∗ .0676∗∗ .0277 .0783∗∗ .0916∗∗∗

Japan .0221 −.0164 .0319 .0533∗ .0227 .0337 .0593∗∗ .0677∗∗

Poland .0146 −.0514∗ .0581∗ .0370 .0446 .0115 .0450 .0658∗∗

United States .0189 −.0367∗ .0413 .0571∗ .0525∗∗ .0307 .0369∗ .0751∗∗

Ukraine .0258 −.0426∗ .0593∗ .0801∗∗ .0626∗∗ .0520 .0773∗∗ .0220

Source: Cassey (Forthcoming)Notes: Observations on the diagonal are the Moran’s I for the location of exporters shippingto that country. Off-diagonals are the bivariate Moran’s I representing the degree of spatialcorrelation between the exports to the “row” countries and exports to the “column” coun-tries. Observations have been normalized with respect to regional GDP. Bold observationsare the bivariate statistic where both countries have a statistically significant univariate Moran’s I.

∗, ∗∗, and ∗∗∗ indicate p-values less than .1, .05, and .01.

shipping to country c and those shipping to country d. The Moran’s I for Canada–U.S. is 0.0194 and200

for U.S.–Canada it is 0.0189. Neither of these statistics is different from E(I) indicating that the201

geographic location of exporters shipping to country c is statistically different from those shipping202

to country d.203

This pattern of exporter agglomeration holds when we expand the sample. Consider table 1,204

taken from Cassey (Forthcoming). Observations on the diagonal are starred if there is statistically205

significant clustering, indicating that Russian exporters are clustered around the destinations of206

Canada, Germany, Great Britain, and the United States. The lack of stars on the off-diagonal207

indicates the clustering pattern differs to these two countries if there are stars on the diagonal. (Stars208

on the off-diagonal indicate the spatial autocorrelation of Russian exporters is the same for the two209

countries.) Therefore, the pattern of clustering is not the same for seven of twelve possibilities (in210

bold in table 1). One advantage of using the Moran’s I to find exporter agglomeration is that is is211

easier to identify for which specific countries there is clustering. Our evidence contrasts somewhat212

with that in Lovely et al. (2005) in that we find clustering around relatively open countries rather213

than relatively closed countries that are difficult to export to. Perhaps this is a difference between214

a developing country such as Russia and a developed country such as the United States or perhaps215

this is due to Russia’s relationship with countries otherwise considered relatively closed.216

12

To verify that this clustering by destination of exported shipments holds more generally, we217

consider the location of the destination relative to Russian firms using the gravity equation. Though218

the gravity equation takes several forms, we use the version derived by Anderson and van Wincoop219

(2003). In this formulation, the exponent on exporter GDP, Yi, and importer GDP, Yj , must be220

one, so we divide both sides by exporter and importer GDP. Aggregate region-country exports are221

Xij and great circle distance are Dij . As recommended by Feenstra (2004, p. 161), we insert an222

export-specific set of binary variables, Si, and an import-specific set of binary variables, Tj , to223

account for the unobservable multilateral resistance terms. Using Ordinary Least Squares (OLS),224

we estimate,225

logXij

YiYj= − 7.72

(1.033)∗∗∗− 1.32

(.110)∗∗∗logDij +

89∑i=2

κiSi +

175∑j=2

δjTj + εij (1)

N = 2985, R2 = 0.58, RMSE = 2.03

where ∗∗∗ on the robust standard errors indicates the corresponding coefficient is significantly differ-226

ent from zero at with 99% confidence. The regression results suggest that the overall explanatory227

power of the gravity equation to account for Russian regional exports is low compared to the228

R2 ≥ 0.7 common in country-level data sets that often do not use binary variables to increase the229

goodness of fit.4230

We consider the relatively low adjusted R2 from using subnational data (compared to the R2231

on national data) in a gravity equation as evidence that something is missing from this equation232

that applies to the regional level, but not at the geographic scope of the national level. Given that233

many theories of trade predict a gravity equation-like relationship—Anderson (1979) and Chaney234

(2008) for example—we have documented a weakness in existing theory applied to subnational235

data. See Appendix D for the estimation of gravity nonlinearly to address the fact that only 19%236

4The estimates for a “naive” gravity equation are

logXij = − 0.84(0.580)

+ 0.56(0.043)∗∗∗

log Yi + 0.29(0.020)∗∗∗

log Yj − 1.17(0.065)∗∗∗

logDij + εij

N = 2985, R2 = 0.15, RMSE = 2.45

13

of our possible observations are nonzero.237

To check the robustness of our results, we use a similar method to Koenig (2009) and Choquette238

and Meinen (2011) on our Russian data. We estimate the following logistic regression:239

Eij = 4.576(.069)∗∗∗

+ 0.078(.004)∗∗∗

logMij − 1.456(.007)∗∗∗

logDij − 0.108(.004)∗∗∗

log Yi + 0.252(.003)∗∗∗

log Yj + εij (2)

N = 3, 979, 007

where Eij = 1 if a firm in region i exports to country j and Mij is the number of firms in region i240

exporting to country j. These findings support our results using the Moran’s I, that agglomeration241

exists and is destination-specific.242

Using customs data from Russia, we have shown that Russian exporters are clustered but that243

these clusters differ depending on the country receiving the exports. Next, we develop a theory to244

account for this fact.245

3. A theory of exporter agglomeration based on shipment destination246

We develop a theory of trade where exporters agglomerate with other exporters shipping to the247

same location due to international transaction costs. Our model is a monopolistic competition trade248

model in the spirit of Melitz (2003) and Chaney (2008) but modified to incorporate agglomeration249

as in Ottaviano, Tabuchi, and Thisse (2002) and Fujita and Thisse (1996).250

The export agglomeration by destination we documented could be achieved either as a reduction251

in the fixed costs of exporting or the variable costs of exporting. We base our firm-level theory on252

the agglomeration being achieved through the variable cost because of evidence by Choquette and253

Meinen (2011) of this mechanism. Koenig et al. (2010) finds non robust evidence of agglomeration254

occurring through the variable cost as well, in particular when the number of firms in the area255

exporting to the same country is small, as it often is in Russia. We therefore choose to explore this256

avenue, while acknowledging that it is not the only path to take.257

14

3.1. Motivation and evidence258

Our theory may take a few forms in the real world. The first version is a traditional treatment259

of an externality. As in Head, Ries, and Swenson (1995), agglomeration occurs within a region260

because of knowledge spillovers that are limited in geography. In our case, the knowledge that is261

spilled is about exporting to a specific country (because of good contacts, translation, or effective262

bribes paid as a percent of shipment value, for example) regardless of the exporter’s product.263

A second story is that firms benefit from economies of scale in transportation and shipments.264

Because of containerization in rail, ship, and air transportation, there is a fixed size to the shipment.5265

As documented in Besedes and Prusa (2006), most shipments are small.266

This is true in our Russian data as well. Table 2 shows the quartiles for shipments by weight.267

As the table shows, 50% of shipments weighed less than 185 kg (407 lbs) and 25% or less than 14268

kg (31 lbs). Thus most shipments are small, with a few huge shipments.269

Table 2: Weight of Transaction-level Shipments, by Quartile

Quartile Weight(%) (kg)

25 1450 18575 4716

100 1,597,020,500

N 115,133

Because most shipments are small, exporters cannot fill a standardized container or a ship on270

their own. Hence if firm A is exporting to China, firm B can add its goods to the shipment. Either271

the transportation company would be willing to sell its remaining room to firm B below average272

cost or firm A would be willing to sell its extra container room. Thus both firm A and B pay a273

lower transportation cost than if they shipped alone. Micco and Perez (2002) find evidence of this274

by estimating that a doubling of trade volume weight from a particular port reduces transportation275

costs on that route by 3–4%. They write: “In general, even though most of these economies of276

5Korinek and Sourdin (2010) show that shipping companies quote transportation rates in cost per container.

15

scale are at the vessel level, in practice they are related to the total volume of trade between two277

regions.”278

Regardless of which of these motivating tales is the most reasonable, the modeling of them is the279

same. The particular stories above should not be taken too seriously. For example, we know that280

some countries require that imports in containers only come from one firm. But we think these two281

stories exemplify the spirit of why we posit an externality in transaction costs. Nonetheless, though282

not a focus of our paper, we do an initial test on the two competing mechanisms in section 4.4.283

3.2. The model284

Our model is a variation of Melitz (2003) as expanded by Chaney (2008). There are N non-285

symmetric destination countries spread spatially across the world. We do not model the character-286

istics of countries other than that each is a passive consumer of Russian goods and each differ in size287

and location. The representative consumer in each country j has standard taste-for-variety utility288

on the consumption of an aggregate good composed of the varieties of Russia goods available in289

that county. Countries differ exogenously in their income endowment Yj and their spatial location.290

They differ endogenously in the availability of Russian goods, their aggregate price index, and their291

aggregate consumption.292

There is a continuum of Russian exporters distributed spatially throughout Russia. These293

exporters differ in manufacturing productivity in addition to their location in space. Firm produc-294

tivity, ϕ, is drawn from a Pareto distribution and cannot be changed. Each firm, identified by their295

productivity and the region where they are located, produces a unique good.296

There is a per-unit cost of production wi that is common to all firms in a region, but differs297

across regions. (We believe this is plausible given the relative lack of labor mobility within Russia.)298

Because of the rigidness of Russian law, we assume that firms are not free to change location.299

Finally, there is a fixed cost, fij , associated with exporting from each region i in Russia to each300

country j in the world.301

Given their productivity, the cost of production, and the fixed cost to export, a firm in region302

i maximizes profits in each market j it sells to by choosing the price, pij(ϕ), and quantity, xij(ϕ),303

16

in that market:304

πij(ϕ) = pij(ϕ)xij(ϕ) − wi

(xij(ϕ)

ϕτij(Aij , ·) + fij

). (3)

Notice there is a region-destination variable transportation cost, τij(Aij , ·). This cost is not305

a constant, but a function of the agglomeration term Aij , among other variables such as physical306

distance which will be specified later and are now represented with a dot. Turning τij into a function307

of other exporter activities based on the activities of the entire region is the primary theoretical308

contribution of this work. For technical purposes, the form of τij(Aij) can be very general, though309

we think it economically reasonable to suppose τij(Aij) > 0, τ ′ij(Aij) < 0, and τ ′′ij(Aij) > 0. If the310

solution to (3) is not positive, then the firm does not export to j.311

3.3. Equilibrium312

Our equilibrium concept uses Bertrand competition between the monopolistic competitors.6

Bertrand competition gives the price for each firm’s output in each destination:

pij(ϕ) =σ

σ − 1

wiϕ

× τij(Aij , ·)

where σ > 1 is the elasticity of substitution in the utility function. Unlike the constant markup313

common in Melitz-style models, the per-unit price charged by each firm in a destination also depends314

on our agglomeration term Aij such that the greater the agglomeration, the lower the transaction315

cost, and thus the lower the price.316

From the zero profit condition of Bertrand competition, there exists a threshold productivity317

for each region exporting to each destination. The equilibrium aggregate price index for country318

j is given by the prices of firms from each region that are above the productivity threshold and319

export to country j. Finally, equilibrium aggregate exports to country j from region i are:320

Xij = YiYj

(wiτij(Aij , ·)

θj

)−γ(wifij)

1− γ(σ−1)λ (4)

6Results are the same under Cournot competition because there is a continuum of goods.

17

where θj may be interpreted as a weighted trade barrier as in Chaney (2008) or multilateral resis-321

tance as in Anderson and van Wincoop (2003), γ > σ−1 is the parameter on the Pareto distribution322

of firm productivities, and λ is a constant.323

Our agglomeration effect Aij shows up in equation (4) in two places. It shows up directly in324

τij(Aij , ·) but also indirectly in θj =[∑

i Yi(wifij)1−γ/(σ−1)[wiτij(Aij , ·)]−γ

]−1/γ.325

4. The benchmark gravity equation and our model326

We derived our theoretical counterpart to gravity in equation (4). We take this prediction to327

the Russian data at the region-level. To do this, we convert (4) into reduced form by taking logs,328

logXij = log YiYj−γ log τij(Aij , ·)+

(1 − γ − γ

σ − 1

)logwi+γ log θj+

(1 − γ

σ − 1

)log fij+log λ.

We rewrite to get,329

logXij

YiYj= α+ β1 log τij(Aij , ·) + β2 logwi + β3 log θj + εij (5)

where α = log λ, β1 = −γ, β2 = 1 − γ − γσ−1 , β3 = γ, and εij = (1 − γ

σ−1) log fij .330

In order to estimate (5), we assume a functional form for how the agglomeration term, Aij ,331

relates to bilateral trade costs, τij . We assume trade costs are a function of physical distance, Dij ,332

and the agglomeration term, but nothing else. Often the gravity literature uses variables such as333

common language, colonial history, location of overseas trade offices, and exchange rates as barriers334

to trade. But these variables are not nearly as important in our data because there is not much335

variation across regions within Russia. Therefore, while acknowledging some of these variables may336

be of second-order importance, we assume:337

τij(Aij , ·) = Dij ×A−ηij . (6)

This technology’s modeling nests the gravity model by setting η = 0. Later, we test if η = 0.338

We take (5) to our Russian data. We perform our experiments at the region-level rather than339

18

the firm-level to document the increased explanatory power that the agglomeration term provides340

for empirical trade flows. By doing so, we do not directly distinguish whether the agglomeration341

works through the variable or fixed cost at the firm-level, since both would be realized as an increase342

in aggregate exports. Also, we cannot directly distinguish what the correct agglomeration term is.343

We consider two possibilities. The first is the number of exporters, similar to the agglomeration344

measure used by Koenig (2009) and Choquette and Meinen (2011). The second agglomeration345

measure we consider is aggregate weight of exports from each region to each country.346

4.1. Agglomeration as measured by the number of exporters347

We have data on Xij , Yi, Yj , Dij , and the number of exporters Mij . We do not have reliable data348

on wi and fij , and therefore cannot calculate θj . We use a set of region-specific binary variables349

Si to account for wi and any other unobserved unilateral region features. We use a set of country-350

specific binary variables Tj to account for θj and any other unobserved unilateral country features.351

We let the bilateral fixed cost be the error term.352

logXij

YiYj= − 7.133

(.910)∗∗∗− 1.327

(.111)∗∗∗logDij + .032

(.048)logMij +

89∑i=2

κiSi +

175∑j=2

δjTj + εij (7)

N = 2985, R2 = 0.58, RMSE = 2.03

The results indicate that the number of firms in region i exporting to country j is not statistically353

different from zero in accounting for aggregate exports. This variable brings no additional explana-354

tory power over standard gravity, which is reflected in the estimates being nearly identical to that355

in (1) and identical adjusted R2.356

4.2. Agglomeration as measured by aggregate weight357

As an alternative to measuring agglomeration by the number of exporters, we consider another358

way of measuring agglomeration, aggregate weight. At the industry or product level, it must be359

the case that an increase in shipment weight will cause an increase in shipment value. However,360

this mechanical relationship between weight and exports is less strict across industries. That is,361

19

the value of a shipment of computers and steel is not as strongly related to the combined weight362

as either one or the other of these products. We choose to aggregate our data and focus on all363

manufacturing industries at once and thus we are less concerned about a purely mechanical effect364

driving our results than otherwise.7365

Let Wij be the aggregate weight of exports shipped from region i to country j. First consider366

a logistic regression similar to (2) except replacing the number of exporting firms with aggregate367

weight:368

Eij = 0.567(.074)∗∗

+ 0.230(.002)∗∗∗

logWij − 0.764(.008)∗∗∗

logDij − 0.329(.004)∗∗∗

log Yi + 0.076(.003)∗∗∗

log Yj + εij (8)

N = 3, 979, 007

Thus the estimate on aggregate weight is statistically significant and larger than the estimate on the369

number of exporters from (2). Because of this evidence, we continue onto our preferred specification370

of the gravity equation, which includes the agglomeration term as measured by aggregate weight.371

logXij

YiYj= − 14.787

(.986)∗∗∗− .813

(.101)∗∗∗logDij + .253

(.014)∗∗∗logWij +

89∑i=2

κiSi +175∑j=2

δjTj + εij (9)

N = 2985, R2 = 0.64, RMSE = 1.88

Compare (9) to the original gravity equation (1). In addition to our agglomeration term being372

significant at the 99% level, the adjusted R2 improves from .58 to .64. Therefore our model with373

an agglomeration term based on the aggregate weight of shipping costs accounts for 10.2% more374

variation in export share than (1).375

Because of our choice of how weight enters the transportation cost function (6), our model376

predicts that the coefficient of physical distance Dij is equal to −γ and the coefficient on aggregate377

export weight Wij is equal to γη. From these two predictions, we calculate η = 0.310, which is378

economically significant. We conduct a nonlinear Wald test of whether η = 0. We reject that η = 0379

7Koenig et al. (2010) does not find evidence of exporting clustering by industry, though they do find “spillovers on thedecision to start exporting are stronger when specific, by product and destination.” But since we focus on aggregateexports, we do not report the results when industry is included.

20

with 99% confidence (F (1, 2755) = 46.35), so η is statistically significant as well. This suggests380

our choice of technology is reasonable. Note that when we use the number of exporters as our381

agglomeration term, η = 0.024 and we cannot reject that η = 0.382

Another endorsement of our model comes from our estimate of γ. We cannot reject that γ = 1383

with 99% confidence which agrees with Axtell (2001) who reports that γ is near one for many384

countries. Furthermore, notice that the estimate for γ in (1) without the agglomeration term is385

larger than Axtell.386

Using our estimate on γ and Chaney’s estimate for γσ−1 , we calculate σ = 1.40. This is consistent387

with the requirement that γ > σ − 1 and that σ > 1. At first, σ seems too low for an estimate388

of the elasticity of substitution, since it implies very large markups in price over marginal cost.389

Additionally our σ is low compared to the 3 to 8 range estimated by Hummels (2001), but Hummel’s390

estimates are within a product category whereas ours is for all manufactured tradeables. Crozet391

and Koenig (2010) estimates an average across trade-weighted industries of 2.25. Furthermore, as392

argued by Rauch (1999), if there are informational costs to trade in the data that do not appear393

as physical distance as in equation (6), then the estimate for σ will be low. When we use the394

estimates from (7), σ = 0.34 indicating that goods are complements. We find this further evidence395

that measuring agglomeration by aggregate weight is appropriate.396

Though we derived our reduced form aggregate equation from firm-level theory, there is a397

potential problem with the estimates in (7) due to the mechanical relationship between aggregate398

shipment weight and export value. Though the relationship between export weight and value across399

industries is far less strong than within industries, there still exists the possibility of an endogeneity400

problem. To address this endogeneity concern, our approach is to follow Koenig (2009) and use401

aggregate export weight from the previous year, 2002. Unfortunately, we only have the 2003 cross-402

section of Russian customs data. Thus in order to proceed, we switch to U.S. state-level export403

data for which we have a panel.404

4.3. Application to U.S. state exports & robustness405

Though we established the facts about agglomeration of exporters around the destination of406

their shipments and motived our model using transaction-level customs data from Russia, there407

21

is nothing in the reduced form equation (5) that requires customs data. Therefore, we take our408

model to another regional export data set. The data set is the Origin of Movement export data for409

U.S. states. We do this to check that our previous results are applicable to a wider set of countries410

than Russia and to address the possible endogenity due to using aggregate shipment weight on the411

right-hand side.412

A detailed description of the OM data is in Cassey (2009). Because of his findings on data413

quality, we limit our observations to manufactured exports only. We also restrict our sample to414

2003 to match the Russian data. The same sample of countries is used except that Russia as a415

destination replaces the United States. Also there are 22 observations which have positive export416

sales but a weight of zero. We do not use these observations. Beginning with the benchmark gravity417

equation in (1) and using OLS, we find:418

logXij

YiYj= 2.21

(1.26)− 2.05

(.137)∗∗∗logDij +

50∑i=2

κiSi +175∑j=2

δjTj + εij (10)

N = 7061, R2 = 0.54, RMSE = 1.28

The estimated coefficient on distance is much larger than for the Russian data, and the constant419

is not statistically significant at the 1% level. For us, the relevant statistic is R2 = .54.420

Applying the U.S. data to (5) yields:421

logXij

YiYj= −10.09

(.950)∗∗∗− .953

(.102)∗∗∗logDij + .433

(.006)∗∗∗logWij +

50∑i=2

κiSi +175∑j=2

δjTj + εij (11)

N = 7061, R2 = 0.75, RMSE = 0.93

The R2 increases from .54 in the benchmark to .75 in our model, an increase of 40.2%. (The results422

for the data pooled over 1997 to 2008 are essentially the same.) We estimate η = 0.454∗∗∗. As423

with the Russian data, we can confidently reject that η = 0 (F (1, 6835) = 81.87). Furthermore, the424

estimated coefficients in equation (11) are similar to those using the Russian data in (9) whereas425

the same is not true for the benchmark. Another encouraging result is that our estimate for γ is426

again not statistically different from one, as before. Furthermore, our point estimate is close to427

22

the 0.99 to 1.10 range for U.S. data reported in Axtell (2001) whereas the benchmark is too high.428

Finally, with the U.S. data, we estimate σ = 1.5.429

One concern with the results so far is that they are driven by the mechanical relationship430

between export value and weight. Because our data is across manufacturing industries, we do431

not think this relationship is strong. Nevertheless, we repeat the estimation using aggregate weight432

lagged one year to 2002. This severs any mechanical relationship since past aggregate weight cannot433

affect current export value in any way other than a reduction of the transaction costs of trade. We434

lose some observations compared to 2003 because those state–country pairs did not trade in 2002.435

logXij

YiYj= − 4.11

(1.177)∗∗∗− 1.50

(.124)

∗ logDij + .243(.010)∗∗∗

log lag Wij +50∑i=2

κiSi +175∑j=2

δjTj + εij (12)

N = 6617, R2 = 0.60, RMSE = 1.15

In this case, we find that our model accounts for 12% more variation in the data than the bench-436

mark. The estimate on the aggregate weight term is statistically significant, though smaller than437

previously. Therefore, we find that the agglomeration term is important even if it is lagged one438

year.439

Applying the OM state export data to our model strengthens the results achieved with the440

Russian data since the U.S. data was not used in the spatial econometrics that guided the theory.441

4.4. Explorations on clustering mechanism442

Now that we have shown that agglomeration as measured by aggregate weight is a statistically443

and economically significant variable in the gravity equation of trade, we do exploratory analysis to444

determine which mechanism the agglomeration term works through. Our model cannot distinguish445

agglomeration occurring because of an informational spillover about how to export to a specific446

destination or economies of scale in shipping. However, we expect that savings from consolidat-447

ing shipments to the same country would be more prevalent for small firms than big firms, who448

presumably have their own distribution networks and can fill a container by themselves.449

We re-estimate (9) for separate subsamples based on firm size. We split firms into size based450

23

on total exports to the world.451

Small firms: logXij

YiYj= − 22.03

(1.737)∗∗∗− 0.55

(.150)∗∗∗logDij + .325

(.036)∗∗∗logWij +

89∑i=2

κiSi +175∑j=2

δjTj + εij

N = 851, R2 = 0.42, RMSE = 2.43

Medium firms: logXij

YiYj= − 21.40

(1.628)∗∗∗− .744

(.120)∗∗∗logDij + .400

(.030)∗∗∗logWij +

89∑i=2

κiSi +

175∑j=2

δjTj + εij

N = 1075, R2 = 0.45, RMSE = 2.40

large firms: logXij

YiYj= − 25.10

(.777)∗∗∗− .717

(.075)∗∗∗logDij + .601

(.019)∗∗∗logWij +

89∑i=2

κiSi +175∑j=2

δjTj + εij

N = 2624, R2 = 0.53, RMSE = 2.54

These estimates show that the effect of the agglomeration term as measured by aggregate weight452

is greater for large firms than medium or small firms, though the coefficient is still statistically453

and economically significant for all sizes. This evidence suggests that destination specific export454

agglomeration is working though an informational spillover rather than consolidation of shipments.455

5. Conclusion456

The theoretical and empirical international trade literature typically studies flows at the country-457

level. However more recent papers such as Eaton et al. (Forthcoming); Bernard and Jensen (2004);458

Tybout (2003); and Arkolakis and Muendler (2010) study firm-level export decisions. All of these459

articles study the production and technology characteristics of the firm and their export decision—460

what to export, how much to export, and to which countries. The findings from this work indicate461

there are barriers to exporting whose precise characterization continues to elude researchers.462

The work in Koenig (2009), Koenig et al. (2010), and Choquette and Meinen (2011) shows463

export spillovers are important for firm export decisions in developed countries if those firms are464

exporting to the same destination. We use the physical location of exporting establishments in465

Russia to see if there is exporter agglomeration by destination in a developing country. Using the466

Moran’s I statistic from spatial econometrics, we find there is agglomeration of exporters, not only467

24

in general and around ports, but also in terms of the destination to which exports are shipped.468

This evidence is supported by a logistic regression.469

We believe this clustering of exporters by destination indicates firm behavior confronting economies470

of scale in the transaction cost of trade to a specific destination. We posit that this agglomeration471

of exporters by destination is an externality from how-to-export information spillovers or economies472

of scale in shipping costs. We measure this agglomeration two ways: by the number of firms ex-473

porting and by the aggregate weight of exported shipments. We find that the agglomeration term474

as measured by aggregate weight increases the explanatory power of the gravity model of trade.475

Using aggregate weight, our model’s prediction is verified with the Russian data and outside476

sources, and our model accounts for more of the variation in export share than the benchmark477

Anderson and van Wincoop (2003) or Chaney (2008) gravity-type model. We then test our model478

on U.S. state export data and are able to account for 40% more variation in export share than the479

benchmark.480

We perform exploratory tests to find evidence through which mechanism the externality op-481

erates. We find that the agglomeration term is more important for large firms than small firms,482

indicating that destination-specific knowledge on how to export is important.483

The documenting of agglomeration by destination and the confirmation of our model takes a step484

toward understanding the size and nature of the international barriers to trade. Our combination485

of localization techniques and externality modeling from industrial organization with the tools from486

firm-level international trade models yields new empirical and theoretical insights into the nature487

and size of barriers to international trade, and the behavior of firms to overcome these barriers.488

The implications for trade-focused development is that policy perhaps should consider where extent489

exports are operating when promoting exports rather than just particular industries or countries.490

Acknowledgements491

The authors thank Thomas Holmes, Yelena Tuzova, Anton Cheremukhin, and seminar participants492

at Vasser College, Washington State University, Kansas State University, University of Scranton,493

University of Richmond, and the New York Economics Association annual meeting. Cassey thanks494

25

Qianqian Wang and Pavan Dhanireddy for research assistance, and Jeremy Sage for help with495

ArcGIS. Portions of this research supported by the Agricultural Research Center Project #0540 at496

Washington State University.497

Appendix A. The political organization of Russia498

In 2003, Russia had 89 federal regions. These regions were separated into 49 oblasts (provinces),499

21 republics, 6 krais (territories), 10 okrugs (autonomous districts), 1 autonomous oblast, and 2500

federal cities (Moscow and St. Petersberg). The divisions are based on ethnic groups and history.501

Each region has equal representation in the Federal Assembly, though there are differences in502

autonomy. However these differences are minor (and decreasing over time) because each region503

is attached to one of eight federal districts administered by an envoy of Russia’s President. The504

districts are designed to oversee regional compliance with federal laws.505

Most taxes are set federally. In 2003, regions could set corporate property tax, gambling business506

tax, and transport tax within bounds established federally (CTEC 2006, p. 955). Exports of goods507

from Russia are subject to a value-added tax of 0% throughout Russia (p. 960), except on some508

exports of oil and natural gas. Export customs duty rates are set federally as well (p. 961).509

Appendix B. The location of products and ports510

We show that the location of Russian exporters can neither be accounted for by the location of511

export commodities nor the location of borders and ports.512

Russia’s major manufacturing industries are: all forms of machine building from rolling mills513

to high-performance aircraft and space vehicles; defense industries including radar, missile produc-514

tion, and advanced electronic components, shipbuilding; road and rail transportation equipment;515

communications equipment; agricultural machinery, tractors, and construction equipment; elec-516

tric power generating and transmitting equipment; medical and scientific instruments; consumer517

durables, textiles, foodstuffs, and handicrafts.518

Russia’s largest export commodities are: petroleum products, wood and wood products, metals,519

chemicals, and a wide variety of civilian and military manufactures. Products such as petroleum,520

26

_

Gulf o

f Oma

n

Kolpo

s Kas

sand

ras

Arab

ian Se

a

Marm

ara D

eniz

Philip

pine S

ea

65

15

63

73

30

20

17

4

39

86

514

3244

13

34

103

74

7

36

70

53

28

76

59

49

2678

84

48

54

58

56

33

8361

666

29

18

38

7742

24

50

40

69

9

72

31

378

5260

65

85

67

47 62

88

87

71

12

64

16

5779

2

8035

4182

25

46

75 435521

45

81

151

1127

23

22 51

68

Ü

_Ca

pital

Rail

Regio

n Typ

eAu

tonom

ous R

egion

(1)

Auton

omou

s Prov

ince (

10)

City (

1)Te

rritory

(7)

Regio

n (48

)Re

publi

c (21

)

Fig

ure

A.1

:R

uss

ia’s

Reg

ions

in2003.

The

Cit

yof

Mosc

owR

egio

nis

incl

uded

inth

em

ap

only

as

the

star

and

not

list

edse

para

tely

.

27

wood, metals, and chemicals need to be processed to become manufactured goods. Therefore it521

is possible that the location of exports is the same as the location of the refineries and processing522

plants for these goods.523

Figure 1 shows that though many exporters are located in the same region as a major port, there524

are regions that are not close to a major port in the top quintile of the number of exporting firms.525

The largest ports are Azov (Rostov, Black Sea), Kaliningrad (Kaliningrad, Baltic Sea), Nakhodka526

and Vostochny (Primorsky, Sea of Japan, largest), Novorossiysk (Krasnordar, Black Sea, largest),527

Primorsk (Leningrad, Baltic Sea, largest, lots of oil), and Saint Petersburg (Saint Petersburg, Baltic528

Sea). Cassey (2010) documents that access to water is an important predictor of U.S. state exports.529

Appendix C. Exports in the context of output530

If external geography played no role in the location of exporters, the pattern of region-country531

aggregate exports (Xij) would be identical to the pattern of regional production (Yi) with some532

randomness. In this case, the destination of Russian sales would not matter for the location of533

Russian exporters. To show that this does not hold, we conduct two simple tests. The first test is534

to run the following regression:535

logXij = − 5.40(0.386)

+ 0.44(0.046)∗∗∗

log Yi + εij (C.1)

N = 2985, R2 = 0.03, RMSE = 2.62

The index i runs over the set of Russia’s regions and the index j runs over the set of countries536

in the world. Standard errors are robust and ∗∗∗ indicates the estimated coefficient is significantly537

different from zero with 99% confidence. The extremely low explanatory power as given by the R2538

indicates that the pattern of production in Russia does not account for the pattern of exports in539

the data.540

The second test is to compare the ratio of exports from each Russian region to the ratio of GDP

from each region to Russia. We modify the familiar concept of location quotient to put Russian

28

exports by region into the context of overall economic activity,

LQi =Xi

X

/YiY.

If the LQ is greater than 1 then the region’s export share to the world is larger than it’s share541

of GDP. This indicates the region is relatively specialized in exporting in comparison to overall542

economic activity. The strength of this approach is that it controls for the geography of economic543

activity within Russia,544

The average LQi is 1.454 with standard deviation 0.489. But though the average LQi is not545

significantly different from one, the regions with the largest LQi are several times the standard546

deviation. Figure C.1 shows the histogram of LQi. Though most region’s LQi is around 1, there547

are many that are in the right tail of figure C.1. Thus there are regions in Russia that are heavily548

concentrated in exporting compared to overall output and regions that do not export anything at549

all.550

Because we only consider manufacturing activity, we repeat the proceeding exercise by scaling551

the regional GDP data by the share that is from industry. The results (not reported) are not552

qualitatively different from before. Therefore, we have documented that Russian exporters are not553

simply located in the geographic pattern of industrial economic activity.554

Appendix D. Robustness555

One feature of the Russian export data is the high frequency of zeros. There are 89 regions in556

Russia and we have 2003 GDP data (IMF 2006) and distance for 175 countries. Therefore, there557

are 15,575 possible observations. Of these, only 2991 have positive exports, or 19%. (We also lose558

some observations because of missing regional GDP data from Russia.) Santos Silva and Tenreyro559

(2006) show that (2) with OLS can bias estimates when there are many zeros in the data. We560

29

010

2030

40Percent

0 10 20 30 40LQ_exports_GDP

Figure C.1: LQ Histogram

30

follow them and use the poisson quasi-maximum likelihood estimator.561

Xij

YiYj= exp

(−20.06(1.189)

∗ × logD−1.04(0.150)

∗

ij +89∑i=2

κiSi +175∑j=2

δjTj

)× εij .

N = 15400, AIC = 526.00

Xij

YiYj= exp

(−12.695∗

(1.231)× logD

−.804(0.119)

∗

ij × logW.116

(0.027)

∗

ij +

89∑i=2

κiSi +

175∑j=2

δjTj

)× εij .

N = 2985, AIC = 456.00

Though the estimates for the regression with the agglomeration term change quantitatively, the562

are the same qualitatively. The agglomeration term is economically and statistically significant.563

Importantly, the model with the agglomeration term has a lower Akaike Information Criterion564

than the model without the term, indicating the model with agglomeration is preferred. However,565

region-country observations of zero exports also have zero weight for those exports. Santos Silva566

and Tenreyro do not consider the model where zeros occur on both the left and right hand side of567

the regression.568

References569

, 2004, 2006. Russia: All Regions Trade & Investment Guide. CTEC Publishing LLC, Moscow.570

Anderson, J. E., 1979. A theoretical foundation for the gravity equation. American Economic571

Review 69 (1), 106–116.572

Anderson, J. E., van Wincoop, E., 2003. Gravity with gravitas: A solution to the border puzzle.573

American Economic Review 93 (1), 170–192.574

Arkolakis, C., Muendler, M.-A., 2010. The extensive margin of exporting goods: A firm level575

analysis, NBER working paper no. 16641.576

Axtell, R. L., 2001. Zipf distribution of U.S. firm sizes. Science 293 (5536), 1818–1820.577

Bernard, A. B., Jensen, J. B., 2004. Why some firms export. Review of Economics and Statistics578

86 (2), 561–569.579

31

Besedes, T., Prusa, T. J., 2006. Product differentiation and duration of US import trade. Journal580

of International Economics 70 (2), 339–358.581

Botolf, P., 2003. Divergence and dispersion in the Russian economy. Europe Asia Studies 55 (8),582

1165–1185.583

Bradshaw, M., 2008. The geography of Russia’s new political economy. New Political Economy584

13 (2), 193–201.585

Broadman, H. G., Recanatini, F., 2001. Where has all the foreign investment gone in Russia?, world586

Bank Policy Research working paper no. 2640.587

Cassey, A., Schmeiser, K. N. L., 2011. Multilateral export decompositions, Washington State Uni-588

versity and Mt. Holyoke College.589

Cassey, A. J., 2009. State export data: Origin of movement vs. origin of production. Journal of590

Economic and Social Measurement 34 (4), 241–268.591

Cassey, A. J., 2010. State export behavior and barriers, School of Economic Sciences, Washington592

State University working paper no. 2010-14.593

URL http://www.econ.umn.edu/~cassey/Appendix/State_export_behavior/state_594

behavior_july2007.pdf595

Cassey, A. J., Forthcoming. An application of the Ricardian trade model with trade costs. Applied596

Economics Letters.597

Chaney, T., 2008. Distorted gravity: The intensive and extensive margins of international trade.598

American Economic Review 98 (4), 1707–1721.599

Choquette, E., Meinen, P., 2011. Export spillovers and the extensive and intensive margins of600

trade, unpublished.601

URL http://www.etnpconferences.net/sea/sea2011/PaperSubmissions/602

Submissions2011/S-F-22.pdf603

32

Crozet, M., Koenig, P., 2010. Structural gravity equatons with instensive and extensive margins.604

Canadian Journal of Economics 43 (1), 41–62.605

Eaton, J., Eslava, M., Kugler, M., Tybout, J. E., 2008. Export growth in Colombia: Firm-level606

evidence. In: Helpman, E., Marin, D., Verdier, T. (Eds.), The Organization of Firms in a Global607

Economy. Harvard University Press, Camdridge, pp. 231–272.608

Eaton, J., Kortum, S., Kramarz, F., Forthcoming. An anatomy of international trade: Evidence609

from French firms. Econometrica.610

Feenstra, R. C., 2004. Advanced International Trade. Princeton University Press, Princeton.611

Fujita, M., Thisse, J.-F., 1996. Economics of agglomeration. Journal of the Japanese and Interna-612

tional Economies 10 (4), 339–378.613

Glejser, H., Jacquemin, A., Petit, J., 1980. Exports in an imperfect competition framework: An614

analysis of 1,446 exporters. Quarterly Journal of Economics 94 (3), 507–524.615

Head, K., Ries, J., Swenson, D., 1995. Agglomeration benefits and location choice: Evidence from616

Japanese manufacturing investments in the United States. Journal of International Economics617

38 (3–4), 223–247.618

Huber, P., Nagaev, S., Worgotter, A., 1997. The changes in location of Russia industry in early619

transition 1987–1993, Institute for Advanced Studies, Vienna, East European Series no. 42.620

Hummels, D., 2001. Toward a geography of trade costs, Center for Global Trade Analysis, Purdue621

University GTAP working paper no. 1162.622

URL http://www.mgmt.purdue.edu/faculty/hummelsd/research/toward/TGTC.pdf623

IMF, 2006. World Economic Outlook Database. International Monetary Fund, Washington, DC,624

http://www.imf.org/external/pubs/ft/weo/2006/01/data/index.htm (accessed December625

15, 2006).626

Koenig, P., 2009. Agglomeration and the export decisions of French firms. Journal of Urban Eco-627

nomics 66 (3), 186–195.628

33

Koenig, P., Mayneris, F., Poncet, S., 2010. Local export spillovers in France. European Economic629

Review 54 (4), 622–641.630

Korinek, J., Sourdin, P., 2010. Martime transport costs and their impact on trade. Applied Eco-631

nomic Perspectives and Policy 32 (3), 417–435.632

Lawless, M., 2009. Firm export dynamics and the geography of trade. Journal of International633

Economics 77, 245–254.634

Lovely, M. E., Rosenthal, S. S., Sharma, S., 2005. Information, agglomeration, and the headquarters635

of U.S. exporters. Regional Science and Urban Economics 35 (2), 167–191.636

Melitz, M. J., 2003. The impact of trade on intra-industry reallocations and aggregate industry637

productivity. Econometrica 71 (6), 1695–1725.638

Micco, A., Perez, N., 2002. Determinants of maritime transport costs, Inter-American Development639

Bank working paper no. 441.640

URL http://www.iadb.org/res/publications/pubfiles/pubWP-441.pdf641

Nuade, W., Matthee, M., 2007. The geographical location of manufacturing exporters in South642

Africa, United Nations University–World Institute for Development Economics Research working643

paper no. 2007/09.644

Ottaviano, G., Tabuchi, T., Thisse, J.-F., 2002. Agglomeration and trade revisited. International645

Economic Review 43 (2), 409–436.646

Rauch, J., 1999. Networks versus markets in international trade. Journal of International Economics647

48 (1), 7–35.648

Santos Silva, J. M. C., Tenreyro, S., 2006. The log of gravity. Review of Economics and Statistics649

88 (4), 641–658.650

Tybout, J. R., 2003. Plant- and firm-level evidence on “new” trade theories. In: Choi, E. K.,651

Harrigan, J. (Eds.), Handbook of International Trade. Basil-Blackwell.652

34

WISER, Various years. Origin of Movement State Export Data. World Institute for Strategic Eco-653

nomic Research, Holyoke, MA, http://www.wisertrade.org (initially accessed Nov. 15, 2005).654

35

Table A.1: Geographic Regions

Region Type GDP Region Type GDP(millions) (millions)

Adygeya Republic 352.8 Mordovia Republic 1281.7Agin-Buryat Autonomous Okrug 77.8 Moscow Federal City 84742.3Altai Krai 3130.4 Moscow Oblast 15517.4Altai Republic 269.6 Murmansk Oblast 2834.3Amur Oblast 1901.5 Nenets Autonomous Okrug 876.1

Arkhangelsk Oblast 3735.1 Nizhny Novgorod Oblast 7720.1Astrakhan Oblast 1884 North Ossetia-Alania Republic 726.8Bashkortostan Republic 54851.7 Novgorod Oblast 1356.1Belgorod Oblast 2766.4 Novosibirsk Oblast 5797.2Bryansk Oblast 1686.2 Omsk Oblast 4363

Buryat Republic 1685.9 Orenburg Oblast 4345.8Chechen Republic 346.4 Oryol Oblast 1594.1Chelyabinsk Oblast 7995.8 Penza Oblast 1706.1Chita Oblast 1859.4 Perm Oblast 8058.3Chukotsky Autonomous Okrug 11325.8 Permyakia Autonomous Okrug 4305.3

Chuvash Republic 1741.8 Primorsky Krai 1052.1Dagestan Republic 2326.3 Pskov Oblast 9707.2Evenkiysky Autonomous Okrug 141.8 Rostov Oblast 6366.7Ingushetia Republic 167.4 Ryazan Oblast 2301.8Irkutsk Oblast 5999 Saint Petersburg Federal City 15122.6

Ivanovo Oblast 1250.9 Sakha Republic 4526.4Jewish Autonomous Oblast 298.3 Sakhalin Oblast –Kabardino-Balkar Republic 938.3 Samara Oblast 9541Kaliningrad Oblast 1783.5 Saratov Oblast 4557.5Kalmykia Republic 4770.8 Smolensk Oblast 1794.7

Kaluga Oblast 1852.9 Stavropol Krai 3823.1Kamchatka Krai 1016.8 Sverdlovsk Oblast 11133.6Karachayevo-Cherkessia Republic 412.5 Taimyrsky (Dolgano-Nenetsky) Autonomous Okrug 36269Karelia Republic 1668.1 Tambov Oblast 1765.6Kemerovo Oblast 5948.7 Tatarstan Republic 11075

Khabarovsk Krai 4253.7 Tomsk Oblast 3600.3Khakassia Republic 997.4 Tula Oblast 2683.8Khanty-Mansi Autonomous Okrug 26409.8 Tuva Republic 166.6Kirov Oblast 2165.5 Tver Oblast 2573Komi Republic 3941.4 Tyumen Oblast 2145.2

Koryaksky Autonomous Okrug 150.6 Udmurtia Republic 3393.1Kostroma Oblast 1144.2 Ulyanovsk Oblast 2025Krasnodar Krai 9573.7 Ust-Ordynsky Buryatsky Autonomous Okrug 145.4Krasnoyarsk Krai 9548.8 Vladimir Oblast 2326.3Kurgan Oblast 1386.9 Volgograd Oblast 4770.8

Kursk Oblast 2066.6 Vologda Oblast 3962.7Leningrad Oblast 4595.1 Voronezh Oblast 3646.4Lipetsk Oblast 3406.5 Yamalo-Nenets Autonomous Okrug 11325.8Magadan Oblast 797.6 Yaroslavl Oblast 3626.9Mari El Republic 853.2

Notes: In 2003, Russia had 49 oblasts (provinces), 21 republics, 6 krais (territories), 10 okrugs (autonomous districts), 1autonomous oblast, and 2 federal cities (Moscow and St. Petersberg).

36