Transportation gateways and trade: how accessibility to the border shapes the spatial concentration of commerce

ABSTRACT Based on a micro-founded gravity equation and resorting to a new database of interregional trade flows within European Union countries, we study the asymmetric concentration of inflows in importing regions when exports come from regions with different geographical advantages in terms of their accessibility to the international border. As a case study we consider the Spanish–French border characterized by two distinct coastal gateways that funnel most road freight. Differentiating between domestic and international distances we conclude that the intensity of trade depends on the relative accessibility to the transportation network. We perform robustness checks for different sectors, and multimodal and re-exporting schemes.


INTRODUCTION
Recent contributions have investigated the relationship between international trade performance, firm location and the geographical advantages of exporters considering their accessibility within the country with respect to coastlines, gateways or certain transport infrastructures (Bleakley & Lin, 2012;Coşar & Fajgelbaum, 2016;Duranton et al., 2014). In a related literature reviewed by Brülhart (2011), it has been considered how accessibility to international borders affects the international concentration of economic activity within a country. Although this literature is inconclusive, it is more plausible that once the heterogeneity of space is considered (asymmetric regions), trade liberalization increases spatial inequality within a country in favour of regions with better access to new markets (Gallego & Zofío, 2018;Krugman & Elizondo, 1996). Most of the empirical literature tends to explain how liberalization promotes the agglomeration of economic activity in regions with geographical advantages with respect to international markets. An agglomeration which is accompanied by a simultaneous increase in the corresponding trade flows, but whose spatial distribution has received scarce attention. Particularly, the literature has failed to study the spatial concentration of the export flows from agglomerating regions into the importing regions of the destination countries. In this regard, our main contribution is the analysis of the spatial patterns of trade flows once a trade liberalization process takes place, which is represented by the incorporation of Spain into the European Union (EU) in 1986. The aim is to analyse the presence of different trade intensities arriving at regions of an importing country (i.e., France) while controlling for the transport network characteristics of the exporting regions of the country of origin (i.e., Spain).
This empirical analysis is theoretically grounded in the postulates of spatial economic models (generally known as new economic geography and new trade theory, NEG/ NTT). Here we follow Gallego and Zofío (2018) who introduce a model that allows for any intra-and international trade network topology to study the redistribution of regional economic activity and trade flows both within and between countries resulting from trade liberalization. Their model explicitly considers the geographical advantage of border regions behaving as gateways for interacting with other countries and allows splitting the distance between two trading regions into two legs: the internal (domestic) distance from the exporting region to the frontier, and the distance from the frontier to the final destination in a European region. In this study we rely on the demand function derived by these authors, representing the trade value of the imports consumed at the regional level, to define the associated gravity equation and the corresponding econometric specification. The relevance of their model is that it provides rationale to the existing patterns of trade in favour of better located border regions that enjoy a privileged position in the trading network. Conversely, border regions whose adverse geographical situation has resulted in limited accessibly do not benefit from the increased openness between countries.
We explore the real case of asymmetric regions and heterogeneous space in Europe. Our analysis centres on actual interregional trade flows by road between Spanish regions and the regions of its six main European partners, with special consideration of the Spanish-French border. This border is of particular interest, given that the Pyrenees behave as a natural barrier for the economic interaction between Spain and the European core, finding two main gateways (La Jonquera in Catalonia and Irun in the Basque Country) located considerably far apart from each other (more than 600 km). Thanks to our singular dataset, we can investigate the spatial pattern of inflows in the importing countries, while considering the actual trading regions on both sides of the national border. First, we use kernel regressions and spatial autocorrelation indices to identify patterns of spatially concentrated/disperse trade, depending on first nature geographical advantage of regions close to the main gateways (i.e., La Jonquera and Irun) from Spain to Europe through the French border. We then use the micro-founded gravity equation to determine whether the spatial concentration of trade flows differs statistically depending on the proximity of a region to the French border measured as the actual distance to the border and through other indicators related to its accessibility. We also perform several robustness checks, repeating the estimation with sector-specific flows and addressing the potential bias introduced by multimodal and re-exportation schemes.
Our results complement other articles studying border effects that nevertheless do not account for the transportation network configuration (Lafourcade & Paluzie, 2011), suggesting that not all border regions benefit from their geographical situation, but mainly those that serve as gateways. Hence, as predicted by the model, gateway regions are those exhibiting the highest trade agglomeration patterns in both the exporting and importing countries. Exports from Spanish regions agglomerate around the French regions nearest to the gateway, swamping these markets more intensely than a standard gravity model would predict. For these border regions, we find that trade decays abruptly as distance from the border increases. To the best of our knowledge, this paper is the first to analyse the concentration of region-to-region trade flows and place it in the context of accessibility to international borders in Europe. Indeed, very few papers analyse the role of borders using region-to-region flows in the EU (Gallego & Llano, 2014, 2015Zofío et al., 2020). Two recent papers have also analysed the border effect for the EU regions, but using an alternative approach to the gravity equation for modelling bilateral trade flows (Capello et al., 2018a(Capello et al., , 2018b.
The remainder of the paper is structured as follows. The next section introduces the theoretical framework used in this study to obtain the demand equations explaining the trade flows and incorporating the accessibility to the border. Section 3 describes the econometric specification, historical background and research hypotheses. Section 4 revises the dataset and provides an exploratory analysis of trade flows. Section 5 analyses the econometric results using several specifications of the gravity equation and different subsamples. Section 6 concludes by drawing the main conclusions.

THE MODEL: A GRAVITY SPECIFICATION OF INTERREGIONAL TRADE FLOWS BETWEEN COUNTRIES
The theoretical model from which we derive the import demand equations underlying the gravity regression specifications corresponds to that proposed by Gallego and Zofío (2018). They assume a world economy, in our case the EU, with several regions situated in different countries, which are denoted as r e i , with the subscript referring to a specific region i ¼ 1, … , j, … , n, and the superscript to the particular country it belongs to: e ¼ 1, … , u, … , z. Following the postulates of spatial economics (NEG/NTT), preferences for heterogeneous (differentiated) goods are characterized through a constant elasticity of substitution (CES) utility function. Production employs a single (labour) input and is subject to increasing returns to scale in tradable sectors within a monopolistic competition market structure. Trade takes place over a transport network connecting all regions and countries. Trade costs between regions are a function of distance over the transport network, and ad valorem tariffs when trade takes place between regions of different countriesto study the effect of trade liberalization. In our case, from the perspective of the results obtained by Gallego and Zofío (2018) we assume that this latter element is removed because of the market integration that took place when Spain joined the EU in 1986. It is then assumed that the spatial distribution of trade flows reflected by the data is the result of the trade openness that Spain experienced when entering the single market.
Household preferences in region j and country u are defined over a group of sectors s ¼ 1, … , S, each consisting of a continuum of horizontally differentiated varieties (V e i,s ), according to the Cobb-Douglas function: where D u j,s stands for the aggregate consumption of varieties in sector s; 0 , a s , 1 is the share of income spent on each sector; and s a s = 1. In turn, the aggregate consumption of all varieties within each sector s is given by a CES function: where d eu ij,s (f) is the individual consumption in region j of country u of sector-s variety f produced in region i situated in country e, including that to which the destination region j belongs, that is, e ¼ u, representing the case of intranational trade; and V e i,s is the set of sector-s varieties produced in i country e. The multipliers b eu ij,s (f) represent the preference parameters for each of the varieties. The parameter s s . 1 measures the constant price elasticity of demand and the elasticity of substitution between any two varieties. Let p eu ij,s (f) be the price of sector-s variety f produced in region i in country e and consumed in region j in country u; and let w u j denote the wage rate in this latter location. Maximizing (2) subject to the budget constraint yields the following individual demands: 1 where: is the CES price index in sector s of region j in country u. Consumption of the differentiated good can be sourced locally or imported from other regions within the same country or from foreign countries. In the commodity specific demands (3), destination prices in the numerator are equal to: Regarding this expression, on the one hand, as a result of the pricing rule associated with the assumption of a market structure under monopolistic competition, the mill (factory) price in the exporting region equals p e i,s = s s s s − 1 w e i,s , where the salary w e i,s constitutes the marginal cost of production and s s /(s s − 1) is the mark-up reflecting the degree of market power. On the other hand, t eu ij,s represents the trade costs between the exporting and importing regions. Trade of the differentiated goods is costly. We follow standard practice and assume that transport related costs are of the iceberg form: d eu ij,s ≥ 1, representing the number of units that must be dispatched from region i in country e to region j in country u to have one unit to arrive (Barbero et al., 2018;Barbero & Zofío, 2016). Besides transport costs, shipping goods between countries is normally subject to non-transport frictions. These normally include tariff barriers and non-tariff barriers: red tape, administrative delays, different product standards, etc. (Behrens et al., 2007). Contrary to transport frictions between all regions, regardless the country they belong to, these non-transport barriers are country pair specific, and we denote them by r eu ij,s , with r eu ij,s . 0 if region i and j belong to different countries e ≠ u, whose trade is subject to tariffs (i.e., the case of Spanish regions exporting to EU countries before 1986), and r eu ij,s = 0 when these countries form a free trade area. These barriers are considered as an ad valorem tariff in addition to transport costs, and therefore, total trade frictions between any two pair of regions are given by t eu ij,s = (1 + r eu ij,s )d eu ij,swith t ee ii,s = 1 when trade is intra-regional because r ee ii,s = 0 and d ee ii,s = 1. Substituting equilibrium prices and trade costs in equation (5) results in the following destination prices:

Trade costs and the transportation network
Concerning the geography of transport costs relevant for this study, we follow Gallego and Zofío (2018) and define a trade matrix that captures the specific configuration of the spatial topology both within and between countries. The whole trade cost structure between all n regions belonging to all z countries including transport (network related) and non-transport frictions is described by way of the following symmetric matrix T, where each element represents trade frictions between a specific pair This is a both symmetric and partitioned matrix, whose elements in the diagonal correspond to the intra-regional transport costs; that is why they are always equal to 1, which is equivalent to costless trade and therefore origin and destination prices are the same. The first and last elements of the transport cost matrix are related to the transport costs within the first country 1 and the last country z. The off-diagonal elements of the up-right and down-left matrices represent the cross-country trade costs between countries 1 and z and z and 1, respectively, which are equal if tariffs are reciprocal and the bilateral optimal distances between 1 and z and between z and 1 are the same. Within this matrix, the elements representing the trade costs between region i in country e and region j in country u characterize both the geographical properties of the spatial network related to transport costs d eu ij,s , and the existence of non-transport related costs r eu ij,s , which have no relationship whatsoever with the transport network. When shipping goods it is assumed that firms minimize the transport costs between any two regions choosing the least cost itineraries (Zofío et al., 2014;Persyn et al., 2020). Finally, adapting the general model to our empirical study that focuses on the role of distance on trade flows from Spanish to European regions, we further divide their distance in that between the exporting region (i) in Spain (e) to the Spanish frontier and that from the frontier to the importing region (j) in the EU country (u). Considering also that shipping distance does not differ across sectors, this final qualification results in:

ECONOMETRIC SPECIFICATION, HISTORICAL BACKGROUND AND RESEARCH HYPOTHESES
We express the aggregate sectoral demand of all varieties ϕ (equation 3) in value terms, T eu ij,s , by multiplying both sides by origin-destination prices. Also, under the monopolistic competition assumption, this demand can be related to the individual firm h exporting each variety multiplied by the number of symmetric firms m operating in the exporting industry; that is, T eu ij,s = p eu ij,s m e i,s d eu ij,h,s = p eu ij,s d eu ij,s . 2 Then, multiplying (3) by (6) as presented in the second equality, and taking natural logs of the resulting equation, yields the following gravity equation for imports: The standard econometric strategy, starting with Anderson and Van Wincoop (2003), exploits that all variables except the bilateral preferences, b eu ij,s , and trade costs, t eu ij,s = (1 + r eu ij,s )d eu ij,s , are either importer or exporter specific. Therefore, their effects on bilateral trade can be captured through importer and exported-specific fixed effects. Denoting by m e i and m u j the vectors of exporter and importer regional fixed effects, results in the following specification: 3 ln T eu ij, s = m e i,s + m u j,s + s s ln b eu ij, s + (1 − s s ) ln ((1 + r eu ij,s )d eu ij,s ), s = 1, . . . , S.
The commodity-specific preference parameters b eu ij,s represent idiosyncratic characteristics that may affect trade between the importer and exporter. It is customary to include adjacency (contiguity) to capture the additional intensity of trade between neighbouring regions. In our case we use different dyadic and monadic dummy variables to control for the higher intensity of trade between the border regions as well as the gateways regions in both countries, for 4 example: We now discuss the rationale behind the research hypotheses related to trade flows accumulation in key gateway regions after the liberalization of trade, considering the specific case represented by Spanish experience.

Background on the Spanish-French border
The French-Spanish border is of great interest for anyone studying the effect of geography (first nature) and history (second nature) on trade and the actual allocation of 540 Nuria Gallego et al. economic activity. Historically, the Spanish-French border has been a field of controversy and intense military conflicts. It extends over 656.3 km through south-west France and north-east Spain. The frontier runs along the Pyrenees, an abrupt mountain range (with more than 50 peaks higher than 3000 metres above sea level and more than another 30 higher than 1000 metres above sea level) that extends for about 491 km, from the Mediterranean Sea to the Bay of Biscay in the Atlantic Ocean. The whole border offers two main access points by highway: through La Jonquera (Catalonia) and Irun (Basque Country). Figure 1 maps the border between Spain and France along with neighbouring regions. Although there are other roads connecting the two countries, the amount of shipping through these routes is negligible. Alternative transport modes have been almost impossible for many years. A well-known example is the development for Spanish and Portuguese railways of an exclusive 'Iberian gauge' (1668 mm between the rails), which differs from the most common European standard (1435 mm). The Iberian standard was set in 1845 and ratified in 1955, specifically to deter French invasions. Thus, the Pyrenees behaves as a natural wall for the economic interaction between Spain and the core of the EU, with the main exceptions of the two highways indicated above. The level of isolation and underdevelopment of Spain from the 18th century until the Francoist dictatorship ended in 1975 is summarized in the following phrase attributed to both Napoleon and France's Louis XVI: 'Europe ends at the Pyrenees.' In this respect, it is easy to underestimate the changes that joining the EU in 1986 brought to Spain at all socio-economic levels.

Research hypotheses
The theoretical model described before predicts that liberating trade, as it happened, first, after the Spanish Transportation gateways and trade: how accessibility to the border shapes the spatial concentration of commerce 541 autarky ended in the late 1960s, and definitively in 1986, when Spain joined the EU and entered the single market, favours the location of economic activity in bordering regions with better accessibility in the transportation network. 5 This can be likened to a gateway effect that is set into motion through the so-called home-market and price effects. For brevity, the spatial equilibria underlying the research hypotheses are discussed in Appendix A1 in the supplemental data online. Given the existing transportation network in 1986 favouring La Jonquera (Catalonia) and Irun (Basque Country), the removal of tariffs made these regions more appealing for the location of exporting firms, which find more profitable to serve EU markets from the nearest locations to the new markets, and to foreign bordering regions (i.e., Languedoc-Roussillon and Aquitaine). In particular, Gallego and Zofío (2018) find that once trade openness takes place, border locations start agglomerating economic activity thanks to dispersion forces that, based on relatively lower trade costs, favour these regions. In fact, if a country such as Spain increases its supplies to EU markets, producers will seek to locate closer to the border to reduce transport costs and be more competitive. Consequently, they will leave inner locations, where production was initially concentrated, in favour of better located places. Clearly, as shown in the gravity equation (8), as the number of firms m e i,s increases, so does the exports from these regions to the new markets; in particular, those to neighbouring foreign regions. Symmetrically, French regions at the border also experience the same benefits of trade liberalization, with increased levels of exports and imports.
From the perspective of the trade flows (exports) from Spanish regions to EU regions (imports), our empirical strategy aims to answer the following interrelated hypotheses: Hypothesis 1. After trade liberalization, outflows (exports) agglomerate at border regions. Spanish regions with better accessibility by road to the European core enjoy geographical advantage, mainly due to lower transportation and information costs. Thus, consistent with the theoretical model (see Appendix A1 in the supplemental data online), and as observed with region-tocountry data for France by Zofío et al. (2020), we expect a larger concentration of exports to Europe in the Spanish regions closest to the French border. Such hypothesis will be tested by analysing the sign, magnitude and significance of a wide range of dummy variables capturing accessibility.
Hypothesis 2: But there are asymmetric effects at border regions depending on their transport accessibility. If a country's specific geography matters, and considering the natural barrier imposed by the Pyrenees, being a Spanish border region with France may not result in a locational advantage unless you also enjoy better accessibility to the gateways. Thus, although there are four Spanish regions along the border with France, the Pyrenees clearly reduce the accessibility of two of them (Aragon and Navarre) compared with Catalonia and the Basque Country, where the main gateways are found and which channel most freight trade by road.
Hypothesis 3: While inflows become more intense with neighbouring foreign regions, it is also expected that these two gateway Spanish regions will thrust singular spatial agglomeration of inflows in the French regions closest to the border. Therefore we expect to find higher intensities of trade between Catalonia and the Languedoc-Roussillon, as well as between the Basque Country and Aquitaine; and less clearly for the Spanish regions of Aragon and Navarre, and the French Midi-Pyrénées. This is related to the 'proximity' of Catalonia with the south of the Languedoc region, that is, the departments of Pyrénées-Orientales and Aude (language, culture, etc.), and the Basque Country with Aquitaine (including the French Basque Country). In this case, this hypothesis will be tested by analysing the sign, magnitude and significance of the linear and quadratic terms capturing the effect of the international distance on trade flows, that is, Dist u fj and (Dist u fj ) 2 , and by a range of dummy variables capturing accessibility.
Hypothesis 4: Finally, regarding the product dimension, it is expected that products with lower degree of transportability (low value/volume ratios) have larger shares of concentration just after crossing the border. We address this question by exploiting the product heterogeneity of our dataset.
The different spatial patterns can be explained by several factors. Our theoretical model stresses the role of centripetal effects driving the reallocation of economic activity and trade flows in favour of better located regions. Trade liberalization triggers the main agglomerative forces in the form of the so-called market size effect. Also, beyond the general equilibrium model, there exist social and business networks and selection bias of Spanish exporting firms, accumulated in the best accessible Spanish regions through history.
With these hypotheses, our empirical analysis combines a set of methodologies such as kernel regressions, indices of local spatial autocorrelation (hot-cold maps) and different specifications of the gravity model. For this latter purpose, we make use of the split distance between the exporting and the importing regions and the pseudo-Poisson maximum likelihood estimator (PPML). 6

Data on international trade flows and optimal distances
Currently there are no official data on region-to-region international trade flows for any country in the EU. Gallego et al. (2018) describe a methodology for estimating region-to-region international flows between 17 Spanish regions and the regions of six European countries across the Spanish-French border (see Appendix A2.1 in the supplemental data online for a summary). This combines region-to-region freight statistics on Spanish trucks with international price indices for each region-country variety drawn from official trade data. 7 The result is a unique dataset with information on region-to-region flows for international trade between Spanish regions (NUTS-2) and the regions of Spain's six main European partners whose trade flows are shipped across the Pyrenees. As the relocation effects brought about by trade liberalization started many years ago, we consider that the trade flows available in the 2004-07 period already reflect the spatial changes in trade flows.
The estimated trade flow T eu ij,t (€, thousands) corresponds to the aggregated freight (including agriculture, energy, intermediates, equipment and consumption) shipped by road. Note that if e ¼ Spain, the flow captures an export from Spanish region i to region j located in a European country u included in our sample (see Appendix 2.2 in the supplemental data online). Hence, consistent with the gravity equation (9) and the underlying model, we regress the imports of these EU regions. 8 For the distance variable, we use the STATA command GEOROUTE. 9 With this command we build a distance variable Dist eu ij (km, thousands) for each ij pair in the period 2004-07. Each bilateral distance can be split into two legs: from the exporting region i to the frontier f, Dist e if ; and from the frontier f to the importing region j, Dist u fj . To do so, for each of the Spanish regions (NUTS-2) we have computed an optimal travel distance to the gateways, obtained as a weighted average of the corresponding optimal distance from the provinces (NUTS-3) in each NUTS-2 region (weighted by their population) to them. Thus, Dist e if measures the computed optimal distance travelled from Spanish region i to the closest main gateway at the French border f. Meanwhile, we obtain Dist u fj as a residual between the total computed optimal distance and the computed optimal distance to the border (Dist u fj = Dist eu ij − Dist e if ). It is convenient to remark that GEOROUTE uses the centroid of each polygon of the corresponding NUTS-3 region to compute the bilateral optimal distance. For robustness we have compared the results obtained with GEOROUTE with alternative distances computed using similar route optimization routines in ArcGIS, finding no relevant differences. Finally, we take regional gross domestic products (GDPs), y e i and y u j , and other regional indicators from Eurostat.

Kernel distribution for outflows
Next, we examine the distribution of Spanish exports, corrected by the product of the origin and destination GDPs, 10 against the international distance (Dist u fj ) using kernel regressions. Figure 2 plots the kernel distribution of the 15 Spanish regions considered, clustered in three groups. The rationale behind this grouping is to differentiate between the simple fact of being a Spanish border region with France, despite the presence of the Pyrenees, and the consideration of the accessibility to the gateways in a more flexible way, using Dist e if . As we will see in the results obtained in our baseline specification, the effect of the accessibility to the eastern and western gateways is different. Thus, we explore the possibility of finding geographical specializations of the exporting regions in the use of the west versus east corridor. With this aim, the first panel corresponds to the kernels of group 1, including regions with a favoured access through the eastern gateway, that is, Catalonia, the Valencian Community, Murcia and Andalusia. We then plot the kernels for a second group, which includes the regions with preferential access to the western gateway: the Basque Country, Navarre, Aragon and La Rioja. A third group plots the kernels for the remaining Spanish regions. The grouping is based on the shape of the kernels and the European 'TEN-T' road network shown in Figure A4.2 of Appendix A4 in the supplemental data online.
For groups 1 and 2, the kernels show a clear concentration of exports over short distances. Conversely, the rest of the regions gathered in group 3 show more heterogeneous profiles, whose trade appears to be less concentrated right across French border regions. This analysis supports the third research hypothesis related to the concentration of imports in French bordering regions. These three groups will be then used in the econometric analysis (Table 2), showing how the different spatial patterns of the imports match the theoretical model and the variables used to control for the main accessibility indicators.

Spatial concentration of Spanish exports in importing regions
Next, further studying the spatial concentration of trade flows, we wonder: (1) if there is a group of Spanish regions with a significant spatial concentration of exports to the main EU regions; 11 and (2) how important accessibility to the frontier with the European core (border with France) is for explaining said concentration.
Thus, by means of the local G eu i -statistic (Getis & Ord, 2010), we perform a hot-cold analysis for each of the Spanish exporting regions using average bilateral flows for the entire period 2004-07 (see Appendix A2.3 in the supplemental data online). This index measures local spatial autocorrelation by comparing the value of a variable in one spot with the value of the same variable in neighbouring spots.
The results from the G eu i -statistic are reported in Figure 3. The hot-cold map captures the spatial concentration of the main European regional clients for each of the 15 Spanish exporting regions. For each of them, the regions coloured in red are the ones with statistically significant high relative levels of imports, contiguous to other intensive importing regions. The regions in intense blue are the ones with statistically significant low levels of imports, surrounded by other regions with low levels of imports. Regions in pale blue have no spatial correlation (non-significant G eu i statistic but positive inflows). The spatial concentration of the flows themselves and the shape of the areas coloured in red can be likened to a swamped area after a dam breaks. Moreover, the silhouette drawn by the G eu i -statistics clearly shows spatial clustering in regions close to the Spanish-French border (accessibility-driven exports). In some cases, the inflows in the French gateway regions do not appear in red (i.e., Comunidad Valenciana (ES52), Navarre (ES22), Región de Murcia (ES62), Castile-La Macha (ES42), Castile and León (ES41), Andalusia (ES61), Aragon (ES24)). This is caused by the fact that the G-statistics uses a contiguity matrix W, Transportation gateways and trade: how accessibility to the border shapes the spatial concentration of commerce 543 and these French border regions are contiguous to the Spanish regions, which have no imports. To avoid misunderstandings, we also plot, in the last panel, the overall concentration of the Spanish exports by regions, using the natural breaks map, which does not depend on such W, providing a clearer visualization of our research hypotheses. This shows how Spanish exports concentrate mainly in the French border regions located close to the gateways and around the TEN-T corridors (Hypothesis 3), with the intensity of exports in the closest French locations being larger in Catalonia and the Basque Country, than in Aragon and Navarre (Hypotheses 1 and 2)as well as around Paris region (Île-de-France) and other regions with capital cities. Kernel regressions: international exports relative to gross domestic product (GDP) on Dist u fj by region. Note: T eu ij,t /Y e i,t Y u j,t . NUTS-2 region-to-region data; euros; average, 2004-07. Source: Authors' own elaboration. 544 Nuria Gallego et al.

Benchmark specifications
In this section we explore the above results modelling the relation between Spanish exports and a relevant set of explanatory variables, particularly the role played by internal (domestic) and international (foreign) distances. As trade takes place after Spain joined the EU's single market in 1986, non-transport related costs (tariffs in particular) have been removed: r eu ij,s = 0, and therefore the only trade frictions are those related to distance: Considering the product dimension, the analysis is conducted initially for the aggregate flows (sector-specific regressions are discussed below), and uses several variations of the theory-based gravity equation defined in (9), adopting the following specifications: where, recalling our notation, T eu ij,t is the value of trade flows from region i in country e ¼ Spain to region j in country u (EU) in year t. The variables ln Y e i,t and ln Y u j,t represent, respectively, the logarithm of the nominal gross domestic product (GDP) of the exporting and the importing regions in year t. As explained above, the variable Dist eu ij between Spanish exporting region i and European importing region j has been split in two parts: Dist e if , which captures the distance from the Spanish region to the optimal gateway 'en route' to the destination market through the Spanish-French border (La Jonquera or Irun); and Dist u fj , obtained as a residual (Dist u fj = Dist eu ij − Dist e if ), which captures the distance from the optimal gateway to the final destination.
The two proposed gravity equations correspond to the basic specification in empirical trade studies (Anderson & Van Wincoop, 2003;Head & Mayer, 2014;Silva & Tenreyro, 2006). In the case of equation (11), it is equivalent to the gravity equation (10) once the exporter × time and importer × time fixed effects are substituted by the corresponding GDPs (m e i,t = ln Y e i,t and m u j,t = ln Y u j,t ). Such elements reflect the multilateral resistance terms and capture regionspecific characteristics. However, this equation departs from its traditional counterparts in two distinct ways: first, by differentiating between domestic and international distances; and second, by capturing non-linearities through the squared value of the latter distance (Dist u fj ) 2 . Although squaring the distance represents a standard approach to model non-linearities in the trade-distance relationship, this is always applied to the whole distance Dist eu ij and not that corresponding to the international leg, Dist u fj . The quadratic term is expected to capture the non-linear relation between trade flows and distance travelled in the destination country, with special interest in France. It will therefore capture the agglomeration of inflows over the shortest distance just after crossing the border, allowing us to test the third hypothesis, which is supported by Figures 2 and 3. The interpretation of the three distance variables is straightforward: (1) we expect a negative and direct effect of Dist e if and Dist u fj on trade; and (2) in addition, for those regions whose export flows are highly concentrated in regions close to the French border (Languedoc-Roussillon, Midi-Pyrénéss, and Aquitaine), we expect a positive sign for the square of the distance from the border to the importing region (Dist u fj ) 2 . Moreover, in line with the most standard specifications of the literature on border effect, we include a variable Contiguity eu ij defined as a dyadic dummy variable taking value 1 when the trading regions are border regions and contiguous, independently of whether they harbour any of the main gateways; and 0 otherwise. We conclude this section by remarking on the rationale of the treatment given to the distance variable. On the one hand, our approach wants to innovate by splitting the internal and external distance, and by considering the quadratic term just for this second stretch of the distance, we intend to capture the concentration of imports arriving to the French regions closer to the two main gateways. On the other hand, we consider the logarithmic transformation of the distance in other specifications, understanding that by doing so we increase the comparability of our results with those obtained with more classical approaches.
5.1.1. Accessibility variables for geographical analysis As aforementioned, the variable capturing bilateral preferences, b eu ij,s , is proxied through different dummies measuring relatedness. In equation (11) and all the extended versions described below, such variables are gathered in the matrix ACCESS_VARS. Such variables are intended to capture the relative intensity in trade with the French regions that are closer to the main border gateways of La Jonquera and Irun (i.e., our research hypotheses). Note here the strong geographical, cultural and historical ties between these Spanish-French border regions (which can be found in other border regions of many European countries). 12 The relatedness between border regions is incorporated into our model through the following set of variables that we briefly define, offering more detail in Appendix A2.4 in the supplemental data online: . Dist e if /Dist u fj captures the ratio between the internal and external distances for every trade flow, as an attempt to capture the relative value of the internal distance to the French border and the second leg to the destination. . Contig i e is a monadic dummy variable taking value 1 when the exporting Spanish regions are border regions to the Spanish-French border, independently of whether they harbour any of the two main gateways; and 0 otherwise. . Contig j u is a monadic dummy variable taking value 1 when the French importing region is a border region with Spain, independently of whether they harbour any of the two main gateways; and 0 otherwise. . Gateway ij eu is a dyadic dummy variable taking value 1 when the trade is between the two border regions at both sides of the Spanish-French border for the two main gateways; and 0 otherwise. . East_Gate ij eu is a dyadic dummy variable taking value 1 when the trade is between the two border regions at both sides of the French border for the eastern gateway (La Jonquera), that is, Catalonia and Languedoc-Roussillon; and 0 otherwise. . West_Gate ij eu is a dyadic dummy variable taking value 1 when the trade is between the two border regions at both sides of the Spanish-French border for the western gateway (Irun), that is, Basque Country and Aquitaine; and 0 otherwise. . Gate i e is a monadic dummy variable that takes the value 1 when the Spanish exporting region i is one of the two Spanish border regions with France where the two Transportation gateways and trade: how accessibility to the border shapes the spatial concentration of commerce 547 main gateways are located, that is, Catalonia and Basque Country; and 0 otherwise. . Gate j u is a monadic dummy variable that takes the value 1 when the destination j of a Spanish export is any of the two French border regions where the two main gateways are located, that is, Languedoc-Roussillon and Aquitaine; and 0 otherwise. . Capital i e is a monadic dummy variable that takes value 1 when the Spanish exporting region is the Madrid region; and 0 otherwise. . Capital j u is a monadic dummy variable that takes value 1 when the EU importing region has the capital city of the country; and 0 otherwise.
Finally, a word of caution regarding the existence of reverse causality by which higher intensities of trade of regions that are closer to the gateways will not be driven by accessibility per se, but by the fact that the main gateways were built precisely in the two regions with higher exporting capacity (Catalonia and Basque Country). Although this question is in general valid for any analysis considering trade and infrastructures, we emphasize the fact that the Spanish-French border is of special interest given the natural wall represented by the Pyrenees and the huge cost difference of building an alternative access point to La Jonquera and Irun (e.g., between Aragon and Midi-Pyrénées). The existing gateways are natural axes for cross-border communication, now and in the past. Indeed, the layout of the current TEN-T corridors, promoted by the European Commission (DG Move), basically follows the original design of the Roman roads. Consequently, the current highways, likened to 'second nature' factors, are nothing but the expected evolution of a 'first nature' geographical advantage since these corridors can be traced back to ancient times. We are then confident that the use of accessibility variables related to 'first nature' advantage allows us to control for the fact that two main gateways have not followed trade intensities. Table 1 reports the results for eight alternative specifications of the gravity equations using the regional Spanish exports to the European regions considered. All regressions are run using the whole sample (with 34% of zero flows) and the PPML estimator. We start with the traditional contiguity variable, which is then complemented with the accessibility variables. The difference between the internal (domestic) and external (foreign) distance is always considered, using two alternative approaches: (1) in specifications [1], [7] and [8], by incorporating the log of the internal distance; or (2) incorporating the ratio Dist e if /Dist u fj , which is a dyadic variable, in models [2][3][4][5][6]. With respect to the external distance, all the specifications include it in levels along with its second quadratic term.

Regression results
The results in column [1] correspond to the first model presented in equation (10), using just time fixed effects and including the GDPs of the trading regions. The coefficients for all the variables are significant and have the expected signs. The coefficients for the log of the GDPs are positive and close to unity. This specification shows that Spanish exports decrease with the domestic distance, Dist e if , as well as the international one, Dist u fj . The quadratic term (Dist u fj ) 2 obtains a positive and highly significant coefficient (0.899***), which indicates a more than proportional agglomeration of trade over short distances after crossing the French border. The contiguity variable, Contiguity eu ij , shows a positive and significant coefficient (0.599), suggesting that exports agglomerate between contiguous partners.
The next specifications follow equation (11), where we include the multilateral resistance terms (origin-time and destination-time fixed effects) and the ACCESS_VARS. In model [2], the non-linear effect of the external distance seems to be clearer than in [1], with a higher negative effect in the external distance coefficient (−4.617 versus −2.926) and a higher positive effect in the quadratic term (1.272 versus 0.899). However, when we introduce the log of the internal distance [7,8], the quadratic term becomes not significant, and the negative effect of the external distance decreases (−2.778 and −2.616). About the contiguity effect, Contiguity eu ij , it seems sensitive to the way in which the domestic distance is specified (ratio of distances versus the log of the internal distance). However, as expected, what fully drives a change in its value and significance is the introduction of the specific ACCESS_VARS.
Regarding the interpretation of Dist e if /Dist u fj , a negative sign for the coefficient (as in model [2] and the regressions by trade corridors reported in Table 2) indicates that the lower the ratio the higher the intensity of trade. A low ratio might be derived from a short internal distance and/or a long external one. In this case, for a given destination region (fixing Dist u fj ), which is not a border region or a gateway, the intensity of imports from farther Spanish regions decreases. The other way round, given a Spanish exporting region of origin (fixing Dist e fj ), which is not a border region or a gateway, the intensity of trade increases for longer external distances. Nevertheless, in this last approach where the destination changes, we should simultaneously consider the negative and significant effect of the external distance.
Focusing on the other ACCESS_VARS, Capital in origin and destination always shows positive and significant coefficients.
with the two main gateways, finding now a positive and significant coefficient, without altering the other coefficients. Next, specification [5] considers the eastern and western gateways separately, finding that in both cases the coefficients are positive and significant, with a higher value for the west (Irun). Then, model [6] includes the monadic variables Gate i e and Gate j u , similarly to model [8], which uses the log of domestic distance. The positive and significant effects of the Gateway eu ij and Gate i e are similar in both specifications, indicating again that the Spanish gateway regions are the ones trading above the average, while the French gateway regions do not. Remarkably, Contig i e effect remain positive and significant in model [8] even after controlling for gateways.
These results confirm the patterns of trade depicted in Figures 2 and 3 and predicted by the theoretical model: the regions with the best accessibility through the main gateways trade above the average and agglomerate a large part of their exports in the border regions of France. Finally, robustness results for equivalent specifications, but including as regressor the log of the domestic distance, are reported in Appendix A3 in the supplemental data online.
We now delve deeper into the asymmetric results obtained for the eastern and western gateways, a fact that might have direct implications for the development of the TEN-T corridors.  [6] show the results for the remaining regions (group 3). The two specifications included in each group just vary in the treatment of the foreign distance, while the rest of variables basically reproduce the model [6] of Table 1. Transportation gateways and trade: how accessibility to the border shapes the spatial concentration of commerce 551 The results for the distance variable in group 1 are as expected. The high negative value for the coefficient of Dist u fj and the positive and significant coefficients for (Dist u fj ) 2 remark, once again, the agglomeration of inflows in the French border regions across the frontier, mainly through La Jonquera. In this case, being Catalonia the only region that is contiguous to the eastern gateway, some of the ACCESS_VARS enter in conflict with the fixed effects, so the results are not illustrative or even reported. Capital j u reaches a positive and significant coefficient, indicating how these regions export above the average to the capital regions of the EU. Regarding group 2, the coefficients for the distance variables reflect the agglomeration of flows in the shortest distance, mainly through Irun. However, the coefficient for Contiguity eu ij becomes negative and significant, while the ones for Contig i e and Gate i e are positive and significant. Finally, group 3 obtains slightly different results. Although ln Dist u fj reaches a negative and significant coefficient, those for Dist u fj and (Dist u fj ) 2 are not. Another key difference is the positive and significant coefficient for Capital i e (corresponding to Madrid), the negative and significant coefficient for Contig j u , and the positive and significant coefficient for Gate j u , which indicates that these regions trade below average with the French border regions but above the average with the ones having the two main gateways. Such results remark the accessibility burden imposed by the Pyrenees and the polarization of flows in the two main gateways for certain regions such as the Midi-Pyrénées or Aragon, which are contiguous but do not have the same connectivity than the regions harbouring the gateways, thereby confirming Hypotheses 1 and 2.
It is also remarkable: (1) that the ratio of distances Dist e if /Dist u fj is just significant for group 1, obtaining negative coefficients; and (2) the high positive coefficients for Capital i e in group 3, where also that coefficient for ln-Dist u fj reaches lower values than group 1.
In conclusion, all these results suggest the existence of geographical specialization of Spanish regions when exporting to Europe (France specifically) by trucks. Interestingly, the division does not follow the criterion suggested in the literature for border regions versus non-border regions (north-south divide in our case), but an alternative criterion following either the eastern or western axis, which is perfectly aligned with the TEN-T corridors described in Appendix A4 in the supplemental data online. Far from contradicting our hypotheses, this enriches it by introducing the idea of corridors or river basins; again, if Spanish exports behaved like water, passing through the two existing extreme funnels (La Jonquera and Irun) to bypass the physical barrier of the Pyrenees, the spatial agglomeration of inflows into France must depend more on the proximity to a gateway than to a region's mere presence on the border. This reflects the geographical specialization of the eastern-western regions with respect to the closest French regions, which can be aptly described with our water metaphor as the effect of an obstacle on the course of a fluid, or the effect of different slopes on the surfaces carrying flows to the importing country.

Results by products and robustness checks:
re-exporting and multimodality So far, we have studied all trade flows without differentiating among sectors in our dataset. In this section we explore new insights focusing on the sectoral dimension and address how modern logistics might affect our previous results.

Product specific analysis
To exploit the sectoral dimension of the dataset, we rely on equation (12), recalling the subscript s for the dependent variable.
where lnTransp k captures the transportability of the products in terms of their value-to-volume ratio in logs. The definition of this variable is described at length in Appendix A2.4 in the supplemental data online. The value-to-volume ratio is defined as the average of the trade flows' value measured in monetary units over shipped quantities measured in physical units for each product k along the whole period. Note that index k refers to the 24 NST products presented in Appendix A2.5 online, which are then grouped into five higher level sectors s (1 ¼ Agriculture, 2 ¼ Energy, 3 ¼ Intermediates, 4 ¼ Equipment and 5 ¼ Consumption). Table 3 reports the sectoral results, which convey a relevant analytical novelty, that is, the explicit consideration of the degree of transportability of the products. It includes two panels. In the left, we describe the results for five product categories built upon our 24 NST product-specific flows; in all these independent regressions the new variable lnTransp k is included, thereby controlling for the heterogeneous transportability of the products included in each sector basket. Note that since such variable is product k specific, its values vary even within each sectoral s category. The right panel of Table 3, labelled Transportability, offers a complementary view: we first divide the whole sample of flows (origin-destination-product), creating a group of High transportability (above the median) and Low transportability (below the median). By doing so, we avoid the ad-hoc statistical classification categories used in the left panel, and consider a continuum of products, which are classified according to the reported value/ volume ratio. An alternative analysis was obtained using the mean as cut-off obtaining very similar results. In all cases, we use the Dist e if /Dist u fj variable to control for the internal distance.
. Starting with the first panel (sectors): first, we highlight that the distance variables lose significance in most of the sectoral subsamples, except for S4 Table 3. Results for product specific flows; PPML; region-to-region Spanish exports to six EU countries, 2004-07; equation (12). . Note that more than 71% of the observations included in our full sample are controlled by such dummy variables, given the high concentration of exports with origin in the bordering regions (Catalonia, Aragon, Navarre and Basque Country) and Madrid. Thus, when the whole sample is split in the five sectoral subsamples, the effect of distance is partially absorbed by these wide set of dummies, as well as the sectoral heterogeneity including the lnTransp k variable. Thus, the effect of the distance variables persists, being negative and significant, for S4 and S5, the two groups of products that are more representative in the remaining 29% of observations with origin in other Spanish regions, not being border or capital regions. . Regarding the ACCESS_VARS, Contiguity eu ij presents a positive and significant coefficient just for sector 3 (Intermediate), but the one for Gateway eu ij is positive and significant, as well as the one for Gate i e , which is positive and significant for all sectors (except Gateway eu ij for S1, Agriculture). It is also remarkable that Gate j u only exhibits a positive and significant coefficient for S1, indicating that just for these products the inflows to Languedoc-Roussillon and Aquitaine (France) are above the average. . Noticeably, in four sectors (S2-S5) the Capital i e dummy for Madrid region has a positive and significant coefficient, while the one for S1 (Agriculture) is negative and significant. By contrast, agricultural products obtain a positive and significant coefficient for the importing capital regions Capital j u , showing how valued are certain agricultural Spanish products in distant and highly competitive EU regions (Paris, Brussels, Berlin, etc.). We also obtain positive and significant coefficients for this variable in S4 (Equipment) and S5 (Consumption), representing the typical example of more sophisticated products travelling longer distances. . Remarkably, the transportability variable reaches significant results for each subsample, obtaining a negative coefficient for S1 (Agriculture), and positive for the rest. The latter indicates that the higher the value/ volume ratio for each Spanish variety exported, the higher the intensity of trade by road, after controlling for all the other factors.
As a complement, the second panel (Transportability) reports also interesting results. In the case of Low transportability varieties, Dist u fj and (Dist u fj ) 2 exhibit, respectively, the expected negative and positive coefficients that indicate a significant concentration of flows in the nearest French regions to the border, even after controlling for the contiguous and gateway regions. Conversely, the products with High transportability, obtain non-significant coefficients for the two variables related to the external distance. The sign for Contig i e is the opposite for High (positive) and Low (negative) groups. The same happens for Capital i e , indicating that Madrid's exports are mainly concentrated in products with higher value-to-volume relations. Moreover, Gate i e , obtains positive and significant coefficients for both subsamples, while Gate u j for none of them.
Summing up, the sectoral analysis confirms Hypothesis 4, which indicates that the higher concentration of imports in the most accessible French regions after crossing the border is heterogeneous by products, and it is mainly driven by the value-to-volume ratio (transportability). Despite this result, the spatial advantage of the Spanish gateway regions persists.

5.2.2.
Logistic variables controlling for multi-modal, hub-and-spoke and potential re-exportation schemes We now turn to discuss whether the spatial agglomeration of imports is driven by the presence of hub-and-spoke and re-exportation schemes in the EU regions of destination. The analysis is based upon equation (13), which is rooted in a wider investigation developed in Appendix A4 in the supplemental data online.
First, our dataset just captures flows departing from Spain by road, without the possibility of distinguishing if the region of destination is the endpoint of consumption, denoted as region j, or an intermediate stop, from where there may exist a subsequent re-exportation to a third destination (in Europe or elsewhere) using road or any other transport mode. With this limitation, we include new explanatory variables capturing the potential capacity of the destination regions of performing as logistic hubs. Although it can be contended that such regional characteristic is captured by the destination fixed effects, we wish to study this possibility explicitly. For brevity, we consider a set of additional variables related to presence of multi-modal connections and transport and storage capacity, grouped in the matrix LOGISTICS VARS (see Appendix A4.2 in the supplemental data online): Air_loaded jt u measures the total freight loaded in regional airports (tons, thousands); Mar_loaded jt u measures the total freight loaded on maritime ships in each region (tons, thousands); SBS_transport j,t u records the number of establishments in each region dedicated to the transport sector; Canals j,t u , Rivers j,t u , Motorways j,t u , Roads j,t u and Railway j,t u are five variables containing the number of kilometres of each of these transport infrastructures in each EU region; and Airport j,t u is a dummy variable that takes the value 1 if there is at least one airport in the region, and 0 otherwise. 13 To avoid conflicts between these regional variables and the multilateral resistance   terms, we introduce regional (NUTS-2) fixed effect interacted with time for the origins (m e i,t ), but country (NUTS-0) fixed effects interacted with time for the destinations (m u t ).
The results for this final check are reported in Table 4, which in all cases points out to a very low risk of inclusion of re-exporting or multimodality schemes affecting our flows. Eight out of the nine variables controlling for the logistic capacity of the importing regions have a positive and significant coefficient, but the economic relevance of these coefficients is negligible. The only exception is Airport j u , with a coefficient of 0.5, but whose definition is not crisp enough to conclude a definite relationship, since it is a dummy variable controlling for the presence of an airport in the region, something that is highly frequent in the EU regions at the NUTS-2 level. However, coupled with the results for Air_loaded j,t u , whose definition is more adequate to capture the possibility of re-exportation, by combining road-air modes, and whose coefficient is almost zero, we conclude that this mode does not play a role in shipments freighted by road. Finally, as a robustness check accounting for dimensional standards in freight road transportation, we perform a complementary analysis in Appendix A4 in the supplemental data online, where we split the dependent variable in two, distinguishing between products that are subject or not to be 'unitized', that is, cargo shipped either in containers or pallets.
To conclude, Table 4 includes a last column (pro memoria) with the results obtained in the reference specification reported in Table 1 (i.e., replicating column [6]). By doing so we want to ease the comparison of the new results obtained when the logistic variables are included. Regarding the distance variables, having in mind the reference value of −2.056**, all the new specifications obtain negative coefficients, which are significant when we introduce Air_loaded j,t u [1], Canals j,t u [3], Roads j,t u [7] and Airport j, t u. [9]. Concerning (Dist u fj ) 2 , we obtain non-significant coefficients, as in the reference specification, except in [8], which obtain a slightly significant coefficient of −0.641*. As for (Dist e if /Dist u fj ), the coefficient in the reference regression was 0.488*, while in all the specifications from [1] to [9] this coefficient is positive but not significant. Finally, the results obtained for the ACCESS_VARS are very similar and robust with the ones obtained in the reference specification (Table 1, [6]). The most salient difference is observed in specification [3], where Capital j u shows a negative and significant coefficient of −0.458**, opposite in sign to those obtained in all other regressions.

CONCLUSIONS
In a world of heterogeneous space and asymmetric regions, some locations might benefit from geographical advantage when exporting to foreign markets. To date, the literature has analysed mainly the internal effects of openness to trade, by measuring the extent to which the liberalization of goods and mobility factors triggers the internal agglomeration of firms in the regions better located in the transportation network. By contrast, little attention has been paid to the spatial concentration of the flows themselves as they reach importing countries. To fill this gap, we resort to a theoretical model grounded on the paradigm of Spatial Economics that provides the rationale for the asymmetric distribution of trade in border regions depending on their relative accessibility in the transportation network. The model allows us to state our research hypotheses and provides the foundation for the gravity equations.
Then, focusing on the case of Spanish-EU trade, we explored the presence of different spatial patterns in the concentration of inflows in importing regions depending on the locational advantage of the exporting regions, defined in terms of their proximity to the most convenient gateway in the French border. Our results confirm the research hypotheses predicted by our theoretical model. We provide new evidence showing that better connected border regions in a country (Spain) tend to trade more intensively with the nearest regions in another country (France). We conclude that, in accordance Hypotheses 1 and 3, there is a higher intensity in trade among pairwise border regions along the eastern gateway of La Jonquera, that is, Catalonia (ES) with Languedoc-Roussillon (FR), and the western gateway of Irun, that is, the Basque Country (ES) with Aquitaine (FR). Conversely, as stated in Hypothesis 2, worse connected regions on both sides of the border trade substantially less: Navarre (ES) and Aragon (ES) with Midi-Pyrénées (FR). Moreover, we have shown that this geographical advantage is non-dichotomic since non-border regions close to the two main strategic gateways also show significant spatial concentration. These results suggest that higher spatial concentration in exports is associated with the geographical advantages and better accessibility of some regions through the two main gateways. However, for regions with no geographical advantages, we found more dispersed spatial patterns. For these regions, one may expect exports (at least those delivered by road) to be less driven by geographical advantages than by other factors, such as pure market size (GDP), social and business connections or sectoral complementarity. Our empirical analysis partially corroborates this extent, by controlling, for example, for capital regions and other accessibility variables, but further research is needed. Finally, our analysis shows how the concentration of trade in the shortest distance is mainly driven by the transportability of products (Hypothesis 4). These findings have been made possible thanks to a new dataset that estimates interregional trade flows shipped by road between Spanish regions to the regions of Spain's six main European partners. This dataset also considers the distance associated with the shipment of the flows, which in our case is split into two legs: the distance from the exporting region to the frontier and from there to the importing region in the foreign country. This specific differentiation between domestic (internal) and foreign (external) distances, new to the literature, is what allows us to model and substantiate the above conclusions regarding the spatial distribution of trade flows Transportation gateways and trade: how accessibility to the border shapes the spatial concentration of commerce 557 and test the main hypotheses. As ancillary results we have shown that, first, our main conclusions above are robust to sector-specific analyses, and second, that the spatial agglomeration of the trade flows is not particularly driven by the presence of hub-and-spoke and re-exportation schemes at the border regions, given the absence of complex warehouse logistics and multimodal transportation in our regions of interest. We also provide relevant qualifications when differentiating trade flows by dimensional standards, that is, whether cargo is unitized using pallets or containers.

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors. 2. We assume that all firms within a given region operate with the same technology and face the same input costs. Consequently, for simplicity, we drop the firm-specific subscript h in the following expressions. 3. Therefore, in this final econometric specification of the gravity equation, the number of firms or varieties, along with the preference parameters, and any originspecific determinants, are eventually swept out by the fixed effects capturing export-only characteristics. Correspondingly, the importer region's price index, expenditure and any other destination-specific determinants are also swept out by the importers' fixed effects. 4. For the specific case study related to the Spanish-French border, in the empirical section we include a series of dummy variables capturing several levels of contiguity and border relations between Spanish and French border regions. This allows controlling for the gateways of La Jonquera (Catalonia) and Irun (Basque Country) to Europe. 5. In the specific case of Spain, there was already a strong concentration of industrial activity in the two main border regions, Catalonia and the Basque Country, which dates back to the 19th century (see Martinez-Galarraga, 2012;