Modeling inter-provincial flows in the presence of hub-spoke. structures and multimodal flows: a spatial econometrics approach 1

Transcription

1 Modeling inter-provincial flows in the presence of hub-spoke structures and multimodal flows: a spatial econometrics approach 1 Díaz-Lanchas, J., (jorge.diaz@uam.es) ( ) Gallego, N. (nuria.gallego@uam.es) ( ) Llano, C. (carlos.llano@uam.es) ( ) De la Mata, T., (tdelamata@iese.edu; tamara.delamata@uam.es) (*) ( ) Departamento de Análisis Económico: Teoría Económica e istoria Económica. Facultad de Ciencias Económicas y Empresariales. Universidad Autónoma de Madrid. Campus Cantoblanco Madrid. ; and L. R. Klein Institute, Universidad Autónoma de Madrid, Cantoblanco, Madrid. * IESE Business School. Barcelona. Strategic Management Department. IESE Business School, Barcelona, C/ Arnús i Garí, , Barcelona; and L. R. Klein Institute, Universidad Autónoma de Madrid, Cantoblanco, Madrid. First Draft. October 2013 JEL Codes: F14, R40, C21 Keywords: border effect, distance, inter-provincial trade, multimodal transport flows, logistics. Abstract: Trade and transport flows are usually considered synonyms. Consequently, it is standard to identify the origin of a flow with the point of production and the destination with the point of consumption. owever, economic and logistical complexity may lead to a number of misrepresentations in the observation of specific dyadic records. Ambushed flows is used in this paper to denote distortions due to transit flows, themselves due to the presence of hubspoke structures or multimodal flows. The presence of such hidden flows may introduce bias in several types of analysis. From the econometric viewpoint, they can be considered an additional source for cross-sectional correlation between dyadic flows. The aim of this paper is twofold: first, to analyze transport-mode competition in inter-provincial deliveries in Spain and, second, to analyze some of these potential biases, such as multimodal and transit flows. To this end, we use a novel dataset with fifty-two inter-provincial, sector-specific flows occurring via four alternative transport modes. We then feed this dataset into various specifications of a gravity model that incorporates cross-sectional dependence attributable to intermodal and sectoral linkages. 1 Acknowledgements: This paper was developed in the context of three research projects: the C-intereg Project ( founded by seven Spanish regional governments; the TransporTrade Program S2007/UM/497, funded by the Directorate for Education and Universities from the Comunidad Autónoma de Madrid ( and (ECO ) from the Spanish Ministry of Economics and Innovation. Nuria Gallego and Jorge Díaz express their gratitude to the UAM and the Ministry of Education and Science respectively for their generous FPI and FPU scholarships. Muchappreciated comments from various colleagues have served to improve the quality and scope of previous versions of this manuscript. owever, any errors herein are entirely the responsibility of the authors. We all want to express our sincere gratitude to J.P. Lesage for gratefully sharing his Matlab routines in which served as a base for our personal developments. Similarly, we want to thank to R. Patuelli and G. Arbia their help and generosity, and their effort of leading the book where this paper will be published. 1

2 1. Introduction In most countries, information concerning interregional commodity flows is scarce and incomplete (Jackson et al., 2006). One of the most common ways to obtain it is to look at transport flows within the country. Many such analyses are found in the literature on border effects, where the trade intensity within one country (or region) is measured against that within one or more other countries (or regions of the same country). Using a combination of inputoutput tables and bilateral shipments, the seminal paper by McCallum (1995) compared interprovincial flows in Canada with interstate flows in the US. Many authors have since repeated the exercise with other countries 2 and other spatial units, be they regions, provinces or even zip codes. For the European Union (EU), similar analyses have been produced at the subnational level. These have taken a country s regions (or provinces) as their point of reference and counted how many more times they traded with the rest of the country (as a whole) than with other countries (Gil et al., 2005; Ghemawat et al., 2010). Other papers have focused on the national level and computed a country s home bias or border effect, defined as how much more trade some region (province) conducts with itself than with any other region (province) of the same country. Wolf (1997, 2000) and Yilmazkuday (2012) did it for the United States; elliwell for Canada (1996, 1997, 1998); elble (2007) and Nitsch (2002) for Germany; Combes et al. (2005) for France; and Garmendia et al. (2012) for Spain. In most of these studies, interregional transport flows by a specific mode of transport (mainly road) or deliveries as reported by the exporting firm (illberry and ummels, 2008) are used as the best proxy for internal bilateral trade (as in fact it is). In principle, a commercial transaction implies the shipment of a good, but shipment can mean single-mode or multimodal transport. It might be simple door-to-door delivery, or it might encompass a range of intermediation activities, such as multimodal nodes, logistics platforms, 2 Japan (Okubo, 2004), United States (Wolf, 2000; illberry, 2002; illberry and ummels, 2003, 2008; Millimet and Osang, 2007), the European Union (Chen, 2004; Nitsch, 2000, 2002; Evans, 2003), Germany (Shultze and Wolf, 2008), Russia (Djankov and Freund, 2000) and Brazil (Daumal and Zignago, 2008), among others. 2

3 warehouses and wholesale operations. The same single trade transaction could be registered in any number of ways by different trade and transport surveys; there could thus be coherence or divergence between the surveyed records of the various intervening agents (exporter, importer, transporter, customs, etc.). Assuming a good match between transport and trade information, it is standard to identify the origin of a transport flow with the point of production and the destination with the point of consumption. In this regard, recent literature on transport economics, geography, global management and logistics points to a rise in economic and logistical complexity (Rodrigue et al., 2009; Rodrigue, 2003; esse and Rodrigue, 2004; Rodrigue and Notteboom, 2010; Woudsma, 1999; van Klink and van den Berg, 1998), which may introduce a number of misrepresentations in the observation of specific dyadic records, for both international and intra-national transport flows. Thus behind any given interprovincial flow across a country a number of factors could contravene the previous hypothesis. Indeed, it is a commonplace to call into question anyone s work in this literature for the bias induced by bilateral transport flows within a country (or between countries) precisely because of this logistical complexity. In this paper we use the term ambushed flows to denote such distortions. In general, such complexity in logistics has been barely considered in international and interregional trade analysis. And therefore, the points of origin and destination have been usually identified with the points of production and consumption. For example, McCallum (1995) commented in his data appendix that: In principle, to quote from the basic data source for this study, "the point of origin is where the good is produced or removed from inventories of producers, wholesalers and retailers.... The point of final destination is where goods are purchased for final consumption or for use in the production of other commodities or added to inventories'" (Statistics Canada, 1992 p. 3). owever, as the document acknowledges, for want of precise information, the actual estimates do not always accord with the theoretical definitions 3

4 Similarly, enderson and Millimet (2008), when analyzing nonlinearities affecting the gravity equation using inter- and intrastate trade flow data from the US Commodity Flow Survey (CFS), collected by the Bureau of Transportation Statistics within the US Department of Transportation, they remark that: There are two important limitations of the CFS data worth noting. First, the CFS tracks all shipments, not just shipments to the final user. For example, shipments of a single good from a factory to a warehouse and then to a retail store would each be included in the data. The fact that wholesale trade is included in the CFS data will likely affect inferences pertaining to the existence and size of the state border effect (illberry, 2001). Second, ( ) that utilization of the aggregate CFS data may affect the interpretation of the border effect. The aim of this paper is to analyze the complexity of modal cooperation and competition behind apparently standard deliveries within a country. We will evaluate some of the aforementioned biases attributable to logistics infrastructures, warehouses and wholesale activities as well as to hub-spoke structures and multimodal internal and international transit flows. We will use a novel dataset that contains sixteen sector-specific flows, by four modes of transport (truck, train, ship, and aircraft), between fifty provinces. Our dataset which includes 160,000 origin-destination deliveries (50*50*4*16) for the single year of 2007, along with a comprehensive set of regressors and rich distance measures based on Spain s actual transportation network is notable both for its size and for its quality. We then introduce a novel index to address the presence of hub-spoke structures as well as the transport-mode specialization/diversification of each province in its trade with the other provinces of the same country. We analyze this index with GIS techniques using with the actual transport network for each mode. Finally, we feed our data into various specifications of the gravity model that incorporate spatial dependence attributable to intermodal and sectoral linkages. Our hypothesis is that ambushed flows constitute an additional source of cross-sectional correlation between dyadic commodity flows to the sources already described in the literature (LeSage and Pace, 4

5 2008, 2009; Fischer and Griffith, 2008). To the best of our knowledge, no one has analyzed this type of relationship in the same detail. Spain, moreover, is a good experimental field in which to explore these topics, because of its complex geography (two island provinces, two autonomous cities in Africa; a privileged enclave for interactions via the Atlantic Ocean), its diversified sectoral structure (with agriculture, refineries / chemical clusters, automotive-equipment-machinery industrial clusters), its relatively isolated location with respect to Europe (on a peninsula, with only two current highway links to the rest of Europe and almost no international train connections before the 1970s) and, at the same time, its Centuries-old geographic advantage towards the Mediterranean Sea, America and Africa. The rest of the paper is structured as follows: Section 2 describes some important links between trade, transportation and logistics and briefly reviews the state of the art in the analysis of these complexities in Spain and elsewhere. Section 3 lays out our empirical strategy for dealing with ambushed flows. Section 4 describes the dataset. Section 5 comprises our descriptive and econometric analysis. A final section concludes. 2. Trade, transport flows and logistics esse and Rodrigue (2004) have argued that freight distribution is widely underrepresented in regional science and geographical research. This is surprising insofar as traditional spatial theory has been developed in relation to transport costs and trade. The emergence of global production networks and structural changes in retail, however, is promoting new research in several fields. As certain authors (Rodrigue, 2003; esse and Rodrigue, 2004; Rodrigue and Notteboom, 2010) have posited, freight flows at the local level result from global and regional economic processes. Internationally, with the division of the production chain and the development of door-to-door distribution schemes, distribution networks have expanded. The development of hub structures, gateways and corridors is also connected as global supply chains have sprouted up, firms have sought out optimal locations in more stable and free-entry 5

6 situations for foreign investment and multi-location firms have surged. These developments have modified freight-distribution systems and spurred a proliferation of new transport terminals and distribution centers in new locations. As Rodrigue (2003) succinctly put it: From their traditional location around central areas with prevalent port and rail linkages, transport terminals and distribution centres have shifted to peripheral locations where road and airport linkages are more prevalent. The geography of freight at this scale is mainly derived from operational considerations aimed at servicing the requirements of local distribution with well-known strategies such as just-in-time and door-to-door. Locally/internally, the urban structure has evolved from a nodal single centre structure to a multi-nodal one. From their traditional location around central areas with prevalent port and rail linkages, transport terminals and distribution centres have shifted to peripheral locations where road and airport linkages are more predominant. The same author (Rodrigue, 2003) defines an articulation point as a location that promotes the continuity of circulation in a transportation system servicing a supply chain. Such concept expands the traditional hub concept as it includes the consideration of terminal facilities, distribution centres, warehouses and finance. e also suggests that an articulation point s relevance depends on the volume and nature of the traffic it handles, which in turn suggest that the characterization of such points depends on flow type (international versus intranational) as well as on the transport modes in use at them. In line with such concepts, esse and Rodrigue (2004) provide an insightful description of the evolution of logistics as it pertains to the core dimensions of transport geography (flows, nodes/locations and networks). In a related paper, Rodrigue and Notteboom (2010) compared the North American and European freight transport systems, showing that the two regions were following alternative patterns of transport and logistics networks. To explain this polarity, the authors cited three main factors: globalization, economic integration and intermodal transport. Intermodalism results from high integration between transport modes and tends to produce complex transport networks. The authors highlighted the impact of global maritime shipping networks on the 6

7 emergence of hubs, gateways and corridors. They contended, for example, that with the accumulation of such terminal infrastructures as ports, rail terminals and freight-distribution centers, gateways had become significant logistical clusters for the maritime/land interface. owever, the role of such logistical infrastructures and articulation points varied by geography, history and sectoral specialization at the regional and local level. In their analysis of freight structures in Western Europe, the authors found that the hinterland was intense not only along the coastline but also in the interior. Cases in point were the Rhine River system, Bavaria in southern Germany, the economic centers around Milan in northern Italy and Madrid in central Spain. European gateways therefore did not consist only of major markets (mega-urban regions) important as such large industrial centers were but could often be intermediary locations. The hinterland was accessed from coastal gateways (Rotterdam, Antwerp, amburg, Barcelona, Marseille) by various transport-mode combinations. The authors also suggested that the role of gateways in shaping functional freight regions within Europe was more recent and profound than in America, and they agreed on the lower specialization of European gateways in specific foreland regions. So much for international freight flows within large regional areas (Europe vs. North America). What about equivalent systems within single countries? These, it turns out, are similarly complex, although, because of a lack of statistical sources, perhaps more opaque. Our paper focuses on this more opaque part of the problem. Another interesting result in the literature on logistics and transport geography is the idea of transport-mode competition and cooperation. Some stimulating papers have analyzed these relationships within Spain with micro-founded general equilibrium and model-choice approaches, both for freight and person movements (Álvarez-SanJaime et al., 2013a, 2013b; Cantos-Sánchez et al., 2009; Monzón and Rodríguez-Dapena, 2006). Similarly, Feo-Valero et al. (2011) focuses on the empirical evidence regarding modal-choice determinants for the inland leg of maritime shipments. The paper analyzes modal choice between road and rail transport on the inland leg of Spanish maritime freight shipments through an error components mixed logit discrete choice model, with the aim of evaluating the potential of rail transport to connect ports 7

8 with their respective hinterlands. In most European countries, rail freight departing from or bound for ports accounts for the bulk of the rail-freight market. But its share is much smaller in Spain. The authors main contribution is to have produced an empirical study of door-toport/port-to-door freight traffic. Their results confirm the role that frequency plays in the relative competitiveness of rail transport. Similar analyses in the field of trade and spatial economy are scarce, although increasing in number and scope. For example, related articles are found in the field of trade, urban economics and the new economic geography, where there is a flourishing discussion about the link between a country s openness to trade and the internal allocation of its economic activity (Brülhart, 2011; Brülhart, et al., 2004; anson, 1996, 2001). Some of these papers consider the presence of non-neutral space, in which certain regions enjoy locational advantages thanks to purely geographic factors (coast location, temperature, ruggedness, etc.) or to better accessibility, which can result from, say, transport infrastructures (ports) or proximity to a border that gives on to a large market (gateway status). Such advantages, stemming from firstor second-nature geography, are the forces behind the development of hub-spoke structures or, more generally, the aforementioned articulation points. Moreover, some recent papers have investigated the relationship between openness to international markets, firms location choice and geographical advantages associated with coastlines or certain transport infrastructures (Duranton et al., 2012; Atkin and Donaldson, 2012; Cosar and Fajgelbaum, 2012; Rappaport and Sachs, 2003; Fujita and Mori, 1996). Certain other papers, taking a more empirical approach, have combined gravity equations and spatial econometric techniques in order to deal with some of the issues mentioned above. We will discuss two examples here. First, Lesage and Polasek (2009) modeled commodity flows with an extended gravity model that incorporates information on the highway network into the spatial connectivity structure of the spatial autoregressive econometric model. Their results pointed to better model fits and higher likelihood-function values. They also found different types of origin and destination connectivity between regions when considering the links between trading regions and their neighbors, as defined by the actual highway network. 8

9 Second, and more recently, Alamá-Sabater et al. (2013) analyzed whether transport connectivity affects trade flows. Using first-order contiguity and incorporating logisticsnetwork-structure dependence in a spatial autoregressive model, they considered flows between fifteen inner regions (Nuts 2) of Spain and sector-specific flows by road. Their results provided evidence for the role of logistics-platform location in satisfying existing demand for transport structures in Spain. Their approach relied mainly on the definition of two different connectivity concepts. First, they considered a purely geographical effect, using contiguous regions as an adhoc structure for spatial autocorrelation effects. Then, they built an index that captured the relative presence of logistics platforms in each region. By considering the endowment of such platforms in regions adjacent regions to trading partners, they expected to capture part of the previously defined logistical complexity. In line with previous papers (Lesage and Llano, 2013), the results confirmed the presence of spatial-dependence effects for the Spanish case. In addition, the weighted matrix defined by the index of logistics platforms generated poor results for aggregate flows but interesting insights for certain industry-specific deliveries. The authors have also developed a working paper that uses similar techniques, as well as transport flows by road, to analyze the same effects at NUTS-3 level in Spain. Our paper explores a wider set of flows (fifty NUTS-3 provinces and four transport modes instead of one) and considers alternative effects (hub-spoke structures, international transit flows and wholesale activity). We also check for the presence of interactions between flows and logistics infrastructures, gratefully borrowing the authors rich variable for the current allocation of logistics platforms in Spain. 3. A conceptual framework for competition and cooperation between transport modes Let us start by considering a country with I provinces (I = 50 for Spain). Without loss of generality, all types of bilateral trade flows from province i to province j could be defined ast. If i = j we have intra-provincial trade flows, while if i j we have inter-provincial, intranational trade flows. International exports and imports from each province can be denoted T irow 9

10 and T RoWj respectively. Note that such flows can correspond to products delivered to the point of final or intermediate demand. Let us now consider that each bilateral aggregated flow can be decomposed into a set of sector-specific flows for every set of tradable goods k 3, which can be delivered by four alternative transport modes (m): road (R), train (T), ship (S) and aircraft (A). Therefore, for the most general case of aggregate trade flow, T can be decomposed as: k kr kt ks ka T = T + T + T + T (1) where kr T is the bilateral trade flow in current of commodity k delivered from province i to province j within Spain. The capital letters R, T, S, A denote the transport mode used for the delivery. In principle, an inter-provincial trade flow by road of product k from province i to kr province j (that is, T ) corresponds to a transport movement with similar characteristics, kr which could itself be denoted F (F for freight). Under normal circumstances it is standard to k k assume that T = F. owever, as described in the previous section, economic complexity and modern logistics can introduce certain alterations into this match between trade flows (usually non-observable within a country) and transport flows for each mode (sometimes observable within a country). In some cases k k T F, because behind two or more apparent transport flows k k k k ( F, F,... F ) there is just one trade flow T. in nm cj 3.1. Transport mode competition 3 Although in some parts of this paper the terms goods, products and sectors are used interchangeably, we prefer the term sector-specific flows. This reason is the limited number of products (fifteen) available in our dataset. For the sake of rigor, we have discarded alternative expressions such as: industry-specific flows (because of the inclusion of agricultural products) and commodity flows (because commodity can be identified with raw materials and non-transformed products, which are not included here). 10

11 Before we analyze the main reasons why k k T F, and more specifically the case of multimodal flows (transport-mode cooperation), it is convenient to briefly discuss the concept of transport competition. Firms located in province i have access to a range of transport-mode mixes with which to deliver a given product k to some other province j, whether within the country or abroad. All these transport modes differ in price, security and speed. Since each product type has its own nature and characteristics (volume, value, perishability, demand, etc.), there is a preferred transport mix by sector and province. Because of the geographical location of firms, final markets and actual transportation networks, however, similar products can be shipped in parallel by alternative transport modes. The likelihood of competition will differ by product k and dyad i-j. When it comes to island provinces, for example, ships and aircraft may compete for certain products but not for others. For a given pair of inner provinces i-j, competition between road and train will depend on the actual transport network and the nature of the products. Moreover, the transport network will be endogenous to preexisting demand for a certain transport mode between the countries of a dyad. <<<Figure 1 about here>>> 3.2. Transit flows and transport-mode cooperation Transit flows We define a transit flow as the general case where a trade flow (transaction) is linked to more than one concatenated transport flow (physical movement of a good), and therefore k k where T F. In this paper, a transit flow can imply multimodality or not. Generally, then, transit flow refers to flows that satisfy either of the following two conditions: The transport flow s point of departure (F k j) is not the point of production but the ub (). In this case, logistics platforms, warehouses and wholesale activities located in the province of delivery accumulate certain amounts of a product k, which has been 11

12 produced elsewhere within the country (intra-national transit flow) or abroad (international transit flow). The transport flow s point of destination (F k i) is not the point of consumption but the ub (). That is, shipments arrive at a province where a logistics platform, warehouse or wholesale activity is located. We can thus depict a transit flow with the following scheme: <<<Figure 2 about here>>> In principle, any province can serve as a ub, for both intra-national and inter-national flows. With intra-national flows, the presence of two (or more) transport flows corresponding to the same trade operation leads to a magnification of intra-national trade estimates, and therefore of home bias within the country. With inter-national flows, if the international trade flow s internal transport is registered as an interprovincial movement separately from the international operation, intra-national trade flows will rise with respect to international trade deliveries. Again, home bias at the country level will be an upward bias. The effect on the internal border effect will depend on whether the transit flow takes places with the same ub province or with another in the country. In summary, in this paper the term transit flow implies that two or more concatenated flows correspond to the same commodity (k 1 = k 2 ). In this general definition of the hub-spoke relationship, no additional conditions are set regarding the transport mode used for the initial inflow to the ub and the subsequent outflow from the ub. Multiple combinations are possible, each explicable in terms of location (i,, j) and the specific product in question (k). For example, we can account for (transit) flows entering or leaving a province () in at least one of two ways: (i) by the presence of logistics platforms, warehouses and wholesales 12

13 infrastructures 4 or (ii) by the presence of multimodal platforms. The latter case will be described to some extent in the next section Multimodal flows k k As previously suggested, one of the main reasons that T F is the presence of transport cooperation or multimodality, which occurs during a single transaction (one trade flow between i and j) when some number of units of k, being transported by two or more modes from point of production i to final destination d, is unloaded and stored temporarily (for an unknown period of time) at an intermediate location N. The situation can be expressed as k kr ks T = F F, where a single transaction implies two related freight flows with one in Nj modal switch (in this case, from road to ship). In principle, multimodal connection could occur through all possible transport-mode combinations. owever, the literature shows that the main multimodal connections within a mid-size country like Spain occur between road and the other three, mainly because of complementarities between these modes and the peculiarities of certain regions and products. Figure 3 shows a sample illustration of multimodal flows within a country: <<<Figure 3 about here>>> ere we assume that two or more transport flows correspond to just one trade flow kr ks which uses two different transport modes for each delivery ( 1 2 ) F F. k T, kr ks F F in Nj Flow1_ by _ road: Inflow _ in_ N Flow2_ by _ ship: Outflow _ from_ N. k T 4 Moreover, inflows to and outflows from the warehouse could be different in nature: (a) the warehouse could be owned by a logistics company offering services to the producer or to retailers. Therefore, although there is no transformation of product k when it is delivered to the warehouse, the outflows and inflows truly correspond to trade flows ; (b) however, if the warehouse is owned by the producer or the retailer, product k s entry into or departure from that infrastructure does not necessarily imply a trade flow between i and j but, perhaps, an intra-firm displacement of products with the aim of getting said products closer to the final market. An additional problem is that the products could be produced abroad and only moved within the country. 13

14 We also imply that commodity k is the same for both deliveries, since a pure transit flow implies no transformation at the node (N). We expect such behavior when N has a main transport infrastructure, such as a maritime port or an airport, but it could also occur at certain multimodal platforms connecting any modes (e.g., train and road). For simplicity, we can consider that the arrival of commodity k at j does not imply an additional flow from or within j. In reality, however, we are likely to find an additional flow (mainly by road) in the hinterland : i.e., a capillarity delivery from the port/station/airport to the final market of consumption. In conclusion, although it is generally assumed that in most countries a large share of k mk trade flows corresponds to just one transport flow ( T = F ), we may nevertheless be very k k likely to find that T F given the aforementioned definitions. Such anomalies should occur mainly in small countries (regions) with geographical advantages (coastlines, proximity to borders) and a tendency to operate as corridors to bigger markets (e.g., the Netherlands) or intermodal transport hubs (e.g., the port of Algeciras, the Spanish province of Zaragoza). The omission of this reality may lead to double counting and thus to a bias in the estimation of intranational flows and the subsequent home bias or border effect. We emphasize that, in principle, any province of a country can serve as a ub or a Node. In fact, any i and any j in a given system of bilateral flows can play either role. For simplicity, we will assume that there is no interconnection between transit flows in other words, that a flow i 1 --j 1 is not connected to another flow i 2 --j 2 such that j 1 = i 2 becomes a new ub (), and so on. As we will see in the next sections, spatial econometrics allow us to identify and control for potential connections between flows departing from or arriving at any i and j from any other location in the system. To this end, our empirical approach assumes a typical situation: the statistical system of a developed country, where trade and transport bilateral flows are reported separately, with little connection between them. Moreover, transport bilateral flows within the country are stated on an isolated basis by transport mode, with no information about their 14

15 relationship with the actual point of production/consumption or about any previous or subsequent delivery using the same or an alternative mode (Boonstra, 2011). Perfect identification of such links between concatenated flows would, of course, require a very detailed dataset on the products and specific locations of each flow. Since such data is, to the best of our knowledge, unavailable anywhere, we have developed our empirical strategy in such a way as to describe an average EU country while taking full advantage of a very rich dataset for Spain. This is the next best thing, and far better than nothing at all. 4. The empirical strategy In this section we define the empirical models we will use to embody the main concepts described before regarding competition and cooperation between the transport flows within a country (multimodal connections and hub-spoke relationships) The gravity equation in the presence of cross-sectional dependence The simplest gravity equation for modelling commodity flows between a set of n provinces within a country is described in Eq. (2): Fi j = αin + Xiβ1 + X jβ2 + dβ3 + Intra β4 + εi j (2) Where F is the bilateral flow in current euros with origin in province i and destination in province j. Flows F have been organized into an n n origin-destination (i-j) flow matrix, which contains intra-regional flows in the on-diagonal elements, and inter-regional in the offdiagonal ones. A dummy variable, Intra, is included to control for the potentially different ii nature of flows occurring within a province and between provinces. This dummy variable takes the value one if the flow s origin and destination are the same province and zero otherwise. The anti-log of this dummy is the internal border effect or home bias discussed in the introduction, where α i is constant; N X is the regressor explaining the production capacity of i province i (GDP i ); X is the regressor explaining the absorption capacity in province j; and j 15

16 d is the bilateral distance between the exporting and importing province (the best available proxy for transport cost by transport mode and commodity type). Note that all variables are expressed as logs. It is common to depart from this basic equation and introduce variables that help account for other factors affecting the direction and intensity of flows. Inadequate specification of the forces driving flow intensity can induce problems of cross-sectional autocorrelation, rendering the estimation method unsuitable. To use least-squares is to assume that the bilateral flows are independent. The validity of this assumption has long been questioned by various authors (Griffith and Jones, 1980; Black, 1992; Bolduc et al., 1992; LeSage and Pace, 2008), who have shown how potential spatial and network dependence can affect different types of bilateral flows. LeSage and Pace (2008) suggested, among other potential causes, the variable-omission problem. According to Black (1992), network and spatial autocorrelation could bias the classical estimation procedures for spatial interaction models. e suggested that autocorrelation may [ ] exist among random variables associated with the links of a network. Bolduc et al. (1992) suggested that classical gravity models did not consider the socioeconomic and network variables adjacent to the bilateral origin-destination regions i and j, arguing that these too should be incorporated into the relationship that attempts to explain flows ( ) between these regions. They emphasized that the omission of neighboring variable values would give rise to spatial autocorrelation in regression errors. Sources of spatial autocorrelation among errors included model misspecification and omitted explanatory variables to capture effects related to the physical and economic characteristics (distances between zones, zone sizes, lengths of frontiers between adjacent zones, etc.) of a region. More recently, LeSage and Pace (2008) challenged the assumption that the classical gravity model s origin and destination (OD) flows contained in the dependent variable vector exhibited no spatial dependence. owever, for most socioeconomic spatial interactions (migration, trade, commuting, etc.), there are several explanations for these effects. For example, neighboring origins (exporting provinces) and 16

17 destinations (importing provinces) could exhibit estimation errors of similar magnitude if underlying latent or unobserved forces were at work, so that missing covariates exerted a similar impact on neighboring observations. Agents located at contiguous provinces could experience similar transport costs and profit opportunities when evaluating alternative nearby destinations. This similar positive/negative influence among neighbors could also be explained in terms of common factor endowments, complementary/competitive sectoral structures, etc. As explained in the previous section, the link between trade and transport flows is not always simple and straightforward. It should be clear, too, that in many cases two or more bilateral transport flows might be anything but independent, the end of the one being the beginning of the next. Transport infrastructures (the network itself, but also ports, airports and logistics facilities) are usually not included as explanatory variables in trade-flow models. Moreover, the actual transport mode used for delivery and transport-mode competition/cooperation schemes is often overlooked in models of aggregate flows and sectorspecific deliveries. All these omitted variables can easily become additional sources of spatial and network autocorrelation effects affecting spatial interaction models. In the next section, we formally test an extended gravity-model specification that accounts for such cross-sectional autocorrelation effects in models of inter-provincial flows with Spain. The extended model subsumes models that exclude such dependence as special cases of the more elaborate model. Taking a cue from the literature, our empirical model is based on several alternative gravityequation specifications embedding spatial econometric terms. For the sake of clarity, each specification is defined along with each of the previously described possible relationships between competing or cooperative transport modes Transport mode competition To consider the presence of transport-mode competition within the country, we first estimate a non-spatial model, as in Eq. (3): 17

18 F = αi + X β + X β + d β + Intra β + Adj β + F β + F β + F β + ε (3) kr kt ks ka N i 1 j Where kr F is the bilateral flow of product k (or the aggregate) by road with origin i and destination j. Elements α i, N X, i however, add a dyadic dummy, X, j d and Intra are the same as in Eq. (2). We do, Adjac, to control for trading-partner adjacency. Otherwise, the main novelty here is that the regressors in some specifications include sector-specific variables such as the gross output of the k sector in exporting province i (GO k i) so as to explain the production capacity of product k in province i; certain variables to capture the logistical capacity of trading partners (logistics platforms, warehouses, wholesale activities); and some other important controls for international transit flows. For simplicity, we now define these variables only for exporting province i (there is an equivalent set of variables for importing province j): Ln(wholesales pc i ): the ratio between the number of wholesale activities in province i (La Caixa, Anuario Económico, 2007) and the population of i. It is expressed as a log and used as an alternative to the platform infrastructure in Alamá-Sabater et al. (2013). Island i : a dummy variable identifying the three island provinces of Spain (Islas Baleares, Las Palmas and Santa Cruz de Tenerife) as exporting regions. Border int. core EU i : a dummy variable identifying Spanish provinces that share an international border with France and Andorra in the north. This variable taking the value one for border provinces and zero otherwise is meant to control for expected higher flows departing from / arriving at these provinces as gateways for trade with the EU core. A positive and significant coefficient for the variable should be interpreted as a symptom that these border provinces are behaving as ubs for international flows; in other words, their exports exceed expected inter-provincial flows (relative to their size, remoteness, etc.) because they are receiving from the EU core international imports of product k, which will be subsequently re-exported domestically (generating 18

19 an apparent inter-provincial flow). Note that by including an equivalent dummy for importing provinces j, we also control for the potential of border provinces to behave as ubs and receive domestic imports for subsequent re-exportation to the EU core. Border int. other i : a dummy variable meant to control for the same effect as the previous variable but for Spanish provinces that share an international border with countries to the southeast: that is, Portugal and countries in Africa. Although the aim is equivalent, the importance of the EU core and these other markets, along with the size of Spain s border provinces, make it worthwhile to consider the effects separately. Ln(imp. int. all i ): a further variable added to take into account the bias introduced by ambushed international flows within domestic flows. In addition to controlling for Spanish provinces bordering on foreign markets, this variable is the logarithm of the total products, measured in euros, imported by i (origin province of the domestic flow) regardless of transport mode. The idea is to acknowledge that even a non-border province can behave as a ub (in this case, not as a gateway ), thanks to, say, a large maritime port (e.g., Barcelona, Bilbao, Valencia). Thus a positive and significant coefficient for this variable indicates that interprovincial product exports from i are positively associated with the province s absorption capacity for international imports. Note that by including an equivalent variable, Ln(exp. int. all j ), for importing provinces j, we also control for the contrary case, where the large capacity for importing interprovincial Spanish flows of a province j is associated with its high intensity of international exports regardless of transport mode. Moreover, for those of our specifications that use sector-specific flows as endogenous variables, the previous variables Ln(imp. int. all i ) and Ln(exp. int. all j ) are redefined much more precisely as Ln(imp. int.all k i) and Ln(exp. int. all k j). In these cases, although all transport modes are considered together, international exports and imports are k- specific. We thus expect to disentangle the very cases of hub-spoke structures (which implies international re-exportation of exactly the same product k that was previously unloaded in the exporting province and produced elsewhere in the country), and an 19

20 alternative situation where a province is an important international exporter (of all products) and a great importer of a specific product k produced in another province within the country. kt ks ka In addition, Eq. (3) includes three new elements, F ; F ; F, each corresponding to equivalent (same k) flows between i and j by alternative transport modes (T = train, S = ship, A = aircraft). 5 By means of these new elements we expect to measure whether the trade flows of one mode are, on average, compatible with those of the others for the full range of products (aggregate flow) and for each separately (k-specific). Next, the previous model is extended to a spatial version to determine whether competing structures exist in neighboring provinces (neighbors to the origins and destinations of the observed transport flows). The economic rationale behind this extension is the following: if a province i delivers products to a province j by a specific transport mode (e.g., train), it would raise the price of that mode for this specific trip (dyad i-j). Thus neighboring provinces of i (n i ) [or of j (n j )] could gain by using an alternative transport mode (e.g., road) for their deliveries to j (or from i) 6. We therefore test the data for such effects with a spatial autocorrelation model (SAR), described in Eq. (4): F = αi + X β + X β + d β + Intra β + Adj β + F β + F β + F β + ρwf + ε (4) kr kt ks ka kr N i 1 j This model includes all the explanatory variables from the previous models, allowing them to subsume non-spatial regression models as special cases. Following LeSage and Pace (2009), we 5 Similarly, we estimate alternative combinations such as: kt kr ks ka = αin + Xiβ1 + X jβ2 + d β3 + Intra β4 + Adj β5 + β6 + β7 + β8 + ε F F F F ks kt kr ka = αin + Xiβ1 + X jβ2 + d β3 + Intra β4 + Adj β5 + β6 + F β7 + β8 + ε F F F ka kr kt ks = αin + Xiβ1 + X jβ2 + d β3 + Intra β4 + Adj β5 + β6 + F β7 + β8 + ε F F F 6 Or, on the other hand, it could be that, thanks to economics of scale in freight (by train in this example), the very fact that a province i delivers product k to j by train could increase the probability that i or j will become a ub, since neighboring regions to i or j might prefer to ship their products to i in order to enjoy less expensive deliveries by train to j, than to ship them directly by an alternative mode (say, road). If this were the case, the preceding model would fail (non-significant results would be expected), and we would have a ub-spoke structure (see the next family of models described in this paper). 20

21 depart from an OD flow matrix with destinations in rows and origins in columns. Note that this is opposite to the standard direction. By vectorizing the matrix, we obtain a column vector F with dimensions n 2 1; is an n 2 1 vector of ones, is the n n matrix of interregional distances transformed into an n 2 1 vector, are n 2 1 vectors containing the explanatory variables appropriate for each bilateral flow, and is an n 2 1 vector of normally kr distributed constant variance disturbances. The main novelty lies in the spatial lag term ρ 1, WF where represents a spatial weight matrix of the form suggested by LeSage and Pace (2008). In a typical cross-sectional model with n provinces, where each pair of provinces represents an observation, spatial regression models rely on an n n non-negative weight matrix that describes the connectivity structure between the n provinces. For example, W > 0 if province i is contiguous to province j. By convention, W ii = 0 to prevent an observation s being defined as a neighbor to itself, and the matrix W is typically row-standardized. In the case of bilateral flows, where we are working with N = n 2 observations, LeSage and Pace (2008), Chung (2008), Chun and Griffith (2011) and Fischer and Griffith (2008) suggest using, where represents an N N spatial weight matrix that captures connectivity between the importing province and its neighbor, and is another N N spatial weight matrix that captures connectivity between the exporting province and its neighbor 7. We row-standardize the matrix to form a spatial lag of the N 1 dependent variable. LeSage and Pace (2008) note that the spatial lag variable captures both destination - and origin -based spatial dependence relations using an average of flows from neighbors to each origin (exporting) and destination (importing) province. Specifically, this means that flows from any origin to a particular destination may exhibit dependence on flows from the origin s neighbors to the same destination. LeSage and Pace (2008) call this origin-based dependence. The spatial lag matrix,, also captures destination-based dependence, the term used in 7 We use the symbol to denote a Kronecker product. 21

22 LeSage and Pace (2008) to reflect dependence between flows from a particular origin province to neighbors of the destination province. As suggested above, origin- and/or destination-based dependence could be present for purely spatial reasons (log-lat, temperature, factor endowments, etc.), or because of unobserved factors that are also conditioned by space. Moreover, as some papers (de la Mata and Llano, 2013) have described, other non-pure spatial correlations could also appear. In the next sections we address certain such correlations connected with the idea of hub-spoke relations and multimodal links between subsequent trips. To help elucidate the next sections, we remind the reader that, as in LeSage and Pace (2009), the spatial model in Eq. (4) can be described as filtering for spatial dependence related to the destination- and origin-based effect, as described in Eq. (5): ( I ρ W )( I ρw) F = αi + X β + X β + d β + Intra β + Adj β + F β + F β + F β + ε kr kt ks ka N j j N i i N i 1 j (5) Again, the N N weight matrix W is obtained by adding the destination- and origin-based matrices ( ). Specifically, W j is a row-stochastic n n matrix built from W s that describes spatial connectivity between n provinces. This W s matrix is symmetric, so it has a real eigenvalue and n orthogonal eigenvectors. The matrix W j can be expressed as a Kronecker product, as in Eq. (6): Ws 0n 0n 0n Ws W j In W = s = 0n Ws Similarly, origin-based connectivity can be defined as (6). An important characteristic of these expressions is that, for example, in a system of five provinces, where province 2 is contiguous to province 1, W s is structured as follows 8 : 8 Remember that the OD matrix used for the endogenous variable is expressed with destinations in rows and origins in columns 22

23 W s = (7) 4.3. ub-spoke structures and transit flows As previously described, a hub-spoke structure exists where an apparent transport flow i-j is connected with a preceding flow (with i serving as a ub: i = ) or a subsequent one (with j serving as a ub: j = ). For simplicity, we assume that there is no concatenation of two transit flows. We also assume that transit flows can be associated with any combination of transport modes (R, T, S, A). With this approach, any i or j in a system of N (n n) bilateral flows i-j can behave as a ub. With this idea in mind, we formulate an approach linked to the gravity equation and the spatial autocorrelation effects described before. This strategy for modeling transit flows in hubspoke structures assumes that every province j is a potential ub (). Thus, much as in the previous case, if we wish to explain a flow by road from i (Barcelona) to j (Zaragoza), we should consider the possibility that j (Zaragoza) will become that is, that j will re-export the same load, or some part of it, to a third province d (Madrid). This approach is identical with the previous, but the focus is now on the apparent destination of a given flow (probability of j s becoming ). Thus, j s absorption capacity (i.e., Zaragoza s GDP) should be reinforced with variables explaining the possibility of j s being (Zaragoza). This possibility is based on, among other factors, the number of potential re-exporting flows of the same product k that is delivered from j to other destinations in the country (in this case, Madrid as a final destination). As we will see in Eq. (8), we can easily put this intuition to use in the standard structure of an SAR model, but using a special version of the weight matrix, labeled W ( for ub ): F = αi + X β + X β + d β + Intra β + Adj β + ρ W F + ε (8) kr kr N i 1 j

24 where ρ2w weight matrix kr F i j is a spatial lag term similar to the one in Eq. (4). The main difference lies in W, which, unlike W = W h i + W j (defined on a purely spatial basis), is defined on the basis of actual trade connectivity for a given sector k and a given transport mode m. The goal of W is to capture all flows departing from a given destination j to any other province d in the system. To be more restrictive, we consider all potential (final) destinations d with the exception of i. Thus, we exclude the possibility that an initial flow of product k from region i to region j will be re-exported afterwards (with no further transformation) from j to i. The procedure to obtain this W matrix is as follows: First, it is convenient to remember once again that the structure of all of our observations (dyadic flows against a set of dyadic and monadic variables) has a destinationorigin layout. Thus, for a given k, the first fifty observations correspond to interprovincial flows imported by province j = 1, the next fifty to imports arriving at j = 2, and so on. Thus, as expressed in Eq. (6), the weighted matrix W j associating flows arriving at j with flows arriving at its neighbors (n j provinces contiguous to province j) is defined on a block diagonal matrix. Similarly, we produce a on-diagonal matrix referring to each importing province j. This W matrix that, like W j, is a block diagonal matrix with each 24 W matrix is defined in such a way as to capture every flow departing from the destination province j of a given dyad i j. We thus explore the hypothesis that j is a potential ub, which re-exports the same load elsewhere in the system. Three alternative criteria are used to fill in each on-diagonal matrix W j : 1. As described in Eq. (13), W contains a weight matrix where all potential flows departing 1 from j to any third destination d (except i) have a one, or otherwise a zero. The matrix thus links every flow i j with every flow j d (d being any province in the system except i) that uses the same transport mode. For example, in a system of five provinces, flows from province 3 to province 2 (3 2) are connected to all flows from 2 to {1, 4, 5} with the exclusion, that is, of the intra-provincial flow (2-2) and the perfect loop (3 2 3). Such assumption can be accomplished through a block matrix like W, with W in the on- 2 diagonal block.

25 W W 1 0 0n 0 0 n W n 2 = n W with W 2 = (9) Note that with this procedure, and for every i-j flow in the system, we consider j a potential ub. Therefore W can be described as a variation on a pure destination-based dependence like W j in Eq. (6). It is also important to note that the previous matrix, W, gives the same weight to all potential re-exporting flows from j (the potential ub ) to any subsequent destination d (except i). Such an approach could be considered naive, since the farther a location is from j the less likely it is to receive a re-exportation flow from it. 2. To refine this approach, Eq. (10) shows how to compute W with W on the on-diagonal 2 elements, which contain ones for the flows from j to contiguous regions n j, and zeros otherwise. We thus link every flow i j with every flow j n j that uses the same transport mode. As in the previous example, a flow from 3 2 is connected only to the flow from 2 to its contiguous province {1}. This situation is captured by a block matrix like W contains only one element with a one (columns = destinations = 2), and zeros 2 elsewhere 9 : W where W W 1 0 0n 0 0 n W n 2 = n W with W2 = (10) 3. Our final refinement is based on Eq. (11), where each block matrix in W contains the inverse of the distance from j (the ub) to d (the final destination). ere as in W, we link 9 Note that, by means of W, flow i j is associated with (subsequent) flow j n j. This spatial autocorrelation scheme, which is explained by re-exporting activity within a hub-spoke structure, is different from the typical destination-base spatial autocorrelation scheme, where, by means of matrix W j, flow i j is associated with the flow from the same origin i to the neighbor of destination j (here labeled n j ). 25

26 every flow i j with every flow j d (except i) but assign a different weight to each potential re-exportation flow. The weight structure is based on the inverse of the distance to the ub (j). Thus the closer is j is to d, the larger the weight for the potential re-exporting flow j d. Our previous example s flow from 3 2 is now connected with the flows from 2 to {1, 4, 5}, again with the exclusion of the intra-provincial flow (2-2) and the perfect loop (3 2 3). Now, however, the weight of each of these potential re-exporting flows depends on the inverse of the distance to the potential ub (d 21 ; d 24 ; d 25 ). W W 1 0n 0n 0 W 0n W 50 n 2 h = with 0 1/ d W2 = / d / d (11) Note that in none of these analyses are we considering the possibility of using a combination of transport modes between two concatenated flows. Full coverage, although straightforward, is beyond the scope of this paper, and will be made in future research. Instead, we devote special attention to multimodal connections in the following section. 5. Data The dataset used in this paper to feed in the endogenous variable (flows) is based on the most accurate data on Spanish transport flows of goods by transport mode (road, train, ship, aircraft), in addition to fifty specific export price vectors, one for each province of origin, transport mode and product type. This rich dataset was collected and filtered in accord with the methodology described in Llano et al. (2009) and published as part of the C-intereg project ( Included in the methodology is a process for debugging the original transport-flow database. This process makes it possible to identify and reallocate multimodal transport flows and international transit flows that may initially be hidden within interregional flows. The result is initial estimates of interregional trade flows in current euros, based on a combination of the transport and price databases. At each stage up to the final aggregation into fifteen product types, the methodology relies on the lowest level of disaggregation available. 26

27 Transport flows in particular are based on classifications that range from forty to 160 product types (depending on which of the four transport modes are available) and price data on 11,000 types. For each transport mode, the goal is to estimate a set of origin and destination matrices (OD matrices) that capture all deliveries while maintaining the greatest possible product disaggregation. The flows are described by k F elements that capture flows F of product k with origin province i and destination province j. As previously stated, we obtain aggregate flows by transport mode by the aggregation of kr F, kt F, ks F ka F from the equivalent OD matrices for each mode (R = road, T = train, S = ship, A = aircraft). Each of these elements can be kmod e k mod e decomposed into quantities ( Q ) and unit prices ( P ), as in Eq. (12): ( * ) ( * ) ( * ) ( * ) k kr kr kt kt ks ks ka ka i i i i F = Q P + Q P + Q P + Q P (12) The largest common product disaggregation for all transport modes is for fifteen product types (R-15), in keeping with the official NACE classification. Again, the procedure for filtering and harmonizing this vast amount of data is described in Llano et al. (2009). Table 1 provides a summary of the original information used for each transport mode, and the Annex briefly explains the estimation procedure for this dataset. We also use a rich set of regressors. For brevity, these variables are described in Table 2. i << Table 1 about here>> << Table 2 about here>> 6. Descriptive Analysis In recent decades, and particularly since 1990, international trade has significantly expanded, thanks in part to fast industrialization in developing economies. From 1990 to 2012 international exports have increased about 10%. The top transport mode for international trade has been ships (89.8% in 2008), followed by land (road and, to a lesser extent, trains, with a 27

28 total of about 9.96%). Ships are simply the most efficient mode for long distances, while road is the most convenient for door-to-ship/ship-to-door hauls and for distribution between and within countries. Also, if we take into consideration the geographical location of developing countries, their main partners (suppliers of raw materials) or their main consumers, maritime transport is sometimes the only alternative. In the context of intra-european trade, the modal split is characterized by the preeminence of road haulage, with a concomitant overuse of motorway networks. Indeed, European authorities are faced with the dilemma of finding a way to foster a shift towards the optimal alternative of road-train combinations, which suppose only a 20% and it usually fails in being too costly. The inland freight transport in European on average- road supposed a 74% for the period , while trains were second-most-in-demand, with 19.7%. For Spanish inner trade, road is the transport mode with the highest weight, representing on average approximately 83% of trade value in (Spanish Ministry of Public Works, 2012), with ships representing around 19% and trains 3% for the same period. The road-train combination achieves its highest share (16.6%) for distances greater than 800 kilometers. Several factors can explain this polarized modal split. On the one hand, short-distance distribution of goods by truck represents a significant share of total trade within a country. On the other, the country s geographical characteristics natural impediments in some areas, limited territorial extension limit the maximum road distance to about 1,230 kilometers, at times making trains or the train-road combinations very costly. The Spanish Ministry of Public Works has estimated, for instance, that a profitable road-train combination trip must exceed 600 kilometers. aving taken this snapshot of the state of the art in current European and Spanish logistics, we now move on to describe the dataset. Our initial exploration aims to describe as thoroughly as possible the logistical complexity previously introduced. We will define some general indexes and explain our use, through GIS software, of some very rich information about the actual transportation network. The starting point of this analysis is found in Figure 4, which depicts total inter-provincial flows for each province in

29 <<<Figure 4 about here>>> As shown, the main inter-provincial flows stem from Spain s richest provinces: Madrid, Barcelona, Valencia, Sevilla and Zaragoza. The northeast shows high trade volume; the west and south, low. As we will see later, this localization of trade flows is driven by network infrastructure (road, port and rail networks), which determine the transport mode used in each province. This previous map breaks down trade flows by transport mode composition in order to bring out the transport specialization of each province. Our exploratory analysis aims to determine which provinces maintain high-level exporting and importing flows with either single or multiple transport modes. A high intensity of inflows and outflows for a specific sector could be due to intra-industry trade, where interchanges between the production and consumption of goods occur within the same place and sector, or to product re-exporting flows, where third regions act as intermediate trade points between the actual origin and destination provinces (i.e., regions which receive and dispatch the same sector s inflows and outflows through the same or different transport modes). To this end, we propose the Intra-Mode Re-exporting Index (IMRI) described in Eq. (13). It identifies provinces that can be considered re-exporting points (hub-regions). The IMRI represents provinces that maintain a significant level of intra-mode trade flows: i.e., that move significant exports and imports by the same transport mode (m). IMRI F F (13) km km ( Fj. + F j ) K km km km km m 1 Fi. + F. i i.. i i = 1 km km K k= 1 max. Fi. + F. i This index is a version of the well-known Grubel and Lloyd (1975) index, which is the standard tool for measuring intra-industry trade between countries. ere, instead of international trade flows (regardless of transport mode), we use inter-provincial freight flows split by transport mode, where the variable F i. represents total freight flows with an origin in province i 29

30 and a destination elsewhere in the country, freight flows within i itself excluded. Conversely, the variable F.i represents the total inter-provincial freight flows that arriving at that province from elsewhere in the country. As a first step, the index focuses strictly on each transport mode m in sector k. The authors proposed their index to evaluate the extent to which intra-industry trade was more important than inter-industry trade between countries. As our definition of the index shows, the Grubel-Lloyd index is implemented on the last term on the right-hand side of the equation. We add a second expression to weight the importance of exports and imports, in freight flows, of region i in sector k by transport mode m, relative to the maximum level of exports plus imports observed in the data for sector k and transport mode m. km If total importing freight flows for province i in sector k ( F. i ) are equal to province i s km total exporting freight flows ( F i. ), the ratio of the second term of the IMRI index is zero, and the whole term equals one. On the other side, if a province receives or sends freight flows only within the same sector and transport mode i.e., buys or sells goods only from sector k the second term of the IMRI index equals zero. Finally, to summarize the information, we take the average mean for the fifteen sectors to obtain one unique index by transport mode. If the index for a specific province and transport mode approaches the value one, the province is, in relation to the rest of Spain, a re-exporting province (potential ub ) with a high intensity of trade flows by that transport mode. By contrast, if the index is closer to zero, the province could be a site of consumption or of production without re-exporting flows. The following maps show this index, highlighting each transport mode. Figure 5, showing the main road network in Spain (highways and national routes), reflects the IMRI for road by province. At a glance, the index identifies three main ubs (Madrid, Barcelona and Valencia) when the road transport mode is considered alone. These regions present the widest road-crossing intensity, with a positive correlation between re-exporting flows by road and the actual road network, especially for regions crossed by main highways. Other provinces (Zaragoza, Sevilla, Valladolid and the three provinces of the País Vasco), which show a high 30

31 volume of trade flows in Figure 4, are not as important in terms of re-exporting intensity. As we will see later, this could be because other transport modes account for their trade-flow intensity. <<<Figure 5 about here>>> Figure 6 reflects the IMRI for rail and the main Spanish rail network. Again, there exists a high correlation between the rail network and re-exporting trade flows by train. ere the potential ubs are Madrid, Barcelona, Valencia and Vizcaya in the north of Spain, with important provinces to the northeast and the south. Indeed, there exists a positive correlation between provinces that re-export by rail and provinces that re-export by road. <<<Figure 6 about here>>> As for trade flows by ship, as expected, only provinces with ports (the islands, Cádiz in the south and, at lower intensity, Valencia and Barcelona on the Mediterranean coast) appear to be potential ubs, since they receive high-intensity inflows from the rest of Spain for export outflows to the islands (Islas Baleares and Islas Canarias ) or to other provinces with ports. Anecdotally, it is interesting to highlight that some inner provinces show slight (but not null) intensity of inflows/outflows by ship. This is because trade-flow statistics with the Islas Canarias are, as explained in the Annex, official statistics from the AEAT, which collects the data that establishes the main transport mode used in the freight flow. That is, a freight flow originating in an inner province and proceeding by road to a port before taking a ship is nevertheless recorded as a delivery by ship from the inner province. <<<Figure 6 about here>>> Finally, Figure 7 presents the IMRI for aircraft flows, along with the location of the main Spanish commercial airports. Again, Madrid, Valencia and Barcelona appear to be the potential 31

32 ubs, with Madrid s Barajas airport, one of the largest in Europe, playing a particularly important role and maintaining high-frequency freight flows with the Islas Canarias (Las Palmas and Santa Cruz de Tenerife). ere the Islas Baleares and the Islas Canarias are also important hub-provinces for airplanes, because most of their re-exporting flows occur between the small islands forming each province (e.g., the single province of the Islas Baleares, which includes such islands as Mallorca, Menorca, Ibiza and Cabrera) and use planes as the most efficient transport mode. For interior provinces with airplane freight flows but no airport, we advance the same explanation as in the case of ship flows, although freight-flow volumes tend to be insignificant in these provinces. <<<Figure 7 about here>>> Taking a further step in our analysis, we define an additional index to take into account the number of transport modes that each region uses to export or import. Following Rodrigue (2003), we determine the relevance of an articulation point (ub) by the variety of transport modes connected to it that is, we assume that a province with intense inflows and outflows by different transport modes is more relevant as a ub than a province that trades through only a single transport mode; the first is more diversified than the second. We define our Inter-Mode Diversification Index (IMDI) as: IMDI F F. (14).... mk mk k k 1 K k k min( Numi., Num i ) Fi. + F. i i.. i i = 1 k k k k K k = 1 4 max( Fj + F j ) Fi + Fi This index is an extension of the previous one, the IMRI. Its second term is similar to the IMRI s, but we have added a second expression to weight the minimum number (Num) of mk mk different modes used to export ( Num i. ) or import ( Num. i ). The minimum number of transport modes is divided by the total number of modes available: four. We can thus distinguish provinces with diverse transport modes from provinces specialized in a single mode. 32

33 Provinces with the maximum number of transport modes (hub-regions) will be multiplied by one, while provinces with only one mode will be weighted by one-quarter. Additionally, since we are weighting by the amount of trade developed by the province in relation to the maximum trade flow observed in the data, the index has two properties: the number of transport modes used to export/import and the volume of trade developed by the province. The higher a province s transport diversification, the greater the weight that the potential hub-spoke structure receives. Finally, we take the average mean between the fifteen sectors to obtain one unique index that includes the information for all four transport modes. Like the IMRI, the IMDI can take values from zero to one. The closer it is to zero, the less diversified the province in terms of transport mode. An IMDI approaching one indicates a province with high trade volume and multiple transport modes 10. The Figure 8 presents the IMDI at the province level. <<<Figure 8 about here>>> The IMDI allows us to pinpoint the main and most diversified hub-provinces in Spain, since it takes into account the re-exporting situation of each province by transport mode. Indeed, we thus observe that only maritime regions, regions with high GDP or regions with great interprovincial trade flows (Madrid, Barcelona and Valencia) are likely to perform as ubs for inter-provincial trade within Spain. Interior provinces are insignificant, and provinces with high volumes of inter-provincial flows, such as Zaragoza and Sevilla, turn out to be less important than expected once we take into account the number of transport modes in use. Certain provinces could be re-exporting regions (hub-regions) for a specific transport mode, but once 10 We have performed a set of correlation tests between the IMDI and the IMRI for each transport mode. The IMDI is positively correlated with the IMRI for ships and the IMRI for aircraft (planes) but shows no significant correlation with the IMRI for road or for trains. There is, however, a positive correlation between the IMRI for road and the IMRI for trains, but they are not significantly correlated with the IMRI for other transport modes. Finally, the IMRI for aircraft and the IMRI for ships are correlated between themselves but do not show correlations with road or trains. 33

34 we control for the other modes, and as long as we take into account the number of different transport modes used (transport diversification), this hub-characteristic is reduced. As mentioned above, those regions with high GDP present high IMDI value. Indeed, the two variables show a strong correlation. Figure 9 presents the spatial autocorrelation (Local Moran s I) between the IMDI and per-capita provincial GDP 11. <<<Figure 9 about here>>> As we can see, the north of Spain is characterized by provinces with a high per-capita GDP surrounded by provinces with a high or low IMDI, and the center of Spain shows no significant spatial autocorrelation between both variables. As we head south, we find provinces with low IMDI that are also surrounded by provinces with low per-capita GDP the exceptions being Sevilla and Murcia, which present high IMDI but are surrounded by provinces with low percapita GDP. 7. Econometric Analysis In this section we present estimates for the expressions in Section 4. The analysis is divided into two parts. First, we present a set of models to consider flows that are aggregated from the sectoral perspective but disaggregated from the transport-mode perspective. Then, before dealing with hub-spoke structures, we complement this analysis with an additional section where sector specific flows are analyzed in light of our main models. For each model it is important to consider the entry of endogenous variables, as well as aggregate or sector-specific factors considered regressors Aggregate flows by modes Our analysis begins with basic specifications of the gravity equation, with aggregate flows and OLS estimation procedures. The results are reported in Table 3 and correspond to Eq. (1) 11 We have determined Moran s I between the IMDI and per capita GDP, and it shows positive spatial autocorrelation. 34

35 and Eq. (2). The first five columns (M1 M5) present the results for aggregate trade flows (M1) and each transport mode separately, with the basic gravity model taken into account. Models M6 and M8 reflect estimations for an augmented gravity model, including, among other variables, the logistics-infrastructure variable (M6) and the per-capita wholesale variable (M8) for the origin and the destination. M7 and M9, meanwhile, present the same augmented gravity model plus the results for Eq. (2), which tests cross-relationships between transports modes (i.e., road trade flows against those of other transport modes). <<<Table 3 about here>>> The results for aggregate flows are in line with the standard results of the gravity-model literature for interregional trade flows; we find significant and negative coefficients for distance and positive and significant coefficients for the GDP of trading provinces. The positive and significant result for the intra dummy also confirms that intra-provincial flows are much higher than inter-provincial flows; this result is consistent with the border-effect literature for almost all countries and circumstances. For transport-mode specific (but still sectorally aggregated) flows, and for flows by road and train, we find consistent results with M1, although for road flows (M2) island geography and distance have a much higher negative impact on trade. Indeed, it is remarkable that the intra-provincial dummy variable is non-significant in both regressions. For ship (M4) and aircraft (M5) flows, we obtain new and promising results. In these regressions, the distance and island variables present signs opposite to those obtained previously in the literature. We find a positive effect for both variables, thereby confirming the heterogeneity of trade, distance, transport costs and transport-mode competition. This result, in our view, confirms that use of these two modes increases with the distance between origin and destination, something that usually occurs for an island. Concretely, for some locations (islands) aircraft and ships are the only transport modes available for trade. Consequently, the distance and island dummy variables have a positive sign. The intra-region dummy variable also presents a positive sign, because ships and aircraft are the main transport modes used to supply other locations (smalls islands) within the same province. 35

36 The M7 and M8 regressions present the augmented gravity model specifically including international trade flows (exports and imports) and dummy variables that indicate a border with a European core-country (basically France) or with other countries to the south, such as Portugal and Morocco. Both columns show consistent results, where GDP, distance and islands have the same qualitative impact on inter-provincial trade. More interestingly, international trade flows do not show significant impact on inter-provincial flows, which might indicate that we are correctly isolating inter-provincial flows from the rest of the world. Also, country dummy variables have no significant impact on trade except if the origin trade-flow province shares a border with a country other than a European core country (in which case the effect is in fact negative). This result is crucial for anyone using the C-intereg dataset, or the official freight flows considered in this paper, since it indicates that the risk incurred when including international transit flows in the inter-provincial flows is non-significant for the Spanish case. Moreover, the negative sign found in many cases (significant only for Border int. other i ) also suggests a certain degree of international isolation in the Spanish border provinces, which indeed show a negative correlation between their international importing/exporting capacity and their inter-provincial exporting/importing deliveries. This result makes sense in light of the political tensions some of these border provinces have suffered with their international neighbors (for centuries, a province sharing a border with France was more likely to suffer invasion than enjoy a geographical advantage for trade) and their tendency to be overshadowed by other, larger nearby Spanish provinces (e.g., Girona vs. Barcelona, Guipuzcoa vs. Vizcaya, Cáceres vs. Sevilla). We move now to the other variables of interest in these models: namely, those that capture the role of logistics, warehouses and wholesale activity in each trading province. In M6 (aggregate flows), the logistics infrastructure variable has a positive impact whether it pertains to the origin or the destination. Similarly, M8 (wholesale per capita for the exporting province) shows a positive impact on trade even greater than that of the logistics infrastructure variable. owever, the same variable for the importing province is non-significant. The results obtained 36

37 when just road flows are considered (M6 and M9) are equivalent in this regard, and thus indicate the prevalence of this transport mode throughout the aggregate. Finally, M7 and M9 perform regressions based on the relationship between road flows and flows for the other transport modes. Remarkably, train and aircraft flows show a positive and significant effect on road flows, while ship flows present no significant impact. Such results suggest that pairs of provinces with intense flows by road can also have intense bilateral flows by train and aircraft and could be a sign of competition between transport modes, at least when all products are analyzed together. All these regressions lead us to perform a set of tests to determine the existence of spatial autocorrelation in our model. Table 4 reflects various spatial tests, such as Moran s I, the LM error test, the LM lag test, the robust version of these tests and a combination of them. We have performed them for Eq. (2) the M6 M9 regressions, using the three pure spatial weight matrices separately (W i, W j, W i + W j ). As we can observe, all tests reject the null hypotheses for the nonexistence of spatial autocorrelation effects based on the origin, the destination and both. We have tested alternative spatial models on the basis of these results and found the spatial lag model to be the most suitable. <<<Table 4 about here>>> <<<Table 5 about here>>> Table 5 reports the results for modeling aggregate flows from a sectoral perspective with spatial autoregressive effects considered. The first four models (M10 M13) use pure spatial dependence structures, defined by W = Wi + Wj matrices. In all these models, then, we are testing for the presence of origin- and destination-based spatial autocorrelation effects affecting flows by road (M10), train (M11), ship (M12) and aircraft (M13). In all cases, ρ is positive and significant, indicating that the exporting and importing province s neighbors exert, on average, an enhancing effect on the trading dyads bilateral flows. This result is consistent with previous papers and similar datasets in Spain (LeSage and Llano, 2013; de la Mata and Llano, 2013; Alama et al., 2013). Results for the standard variables also meet expectations based on previous 37

38 specifications. Among the exceptions is the variable Border int. core EU for the exporting and importing province, which, being negative and significant in more cases than previously, reinforces our previous conclusions. The special results for ships and aircraft also find confirmation. More interestingly, the four models include the flows of the other three. For road (M10) we find enhancing effects, as before, with trains and aircraft. For trains (M11), we find a positive correlation with road, ships and aircraft. As expected, for ships (M12) we find positive coefficients with trains and aircraft but not with road. Finally, for aircraft (M13) we obtain positive results with all other modes. The last three columns of Table 5 reports the results when road flows are modeled with Eq. (8), designed specifically to deal with hub-spoke structure. The only difference between the three models is the weight matrix used for capturing the re-exporting scheme of importing province j when it can be assumed to be serving as a ub within the country. M14 uses M15 uses W, W, and M16 uses W. In all three cases a positive and significant ρ is obtained. In all humility and prudence, one cannot affirm when using aggregate flows that this result unequivocally proves the presence of a hub-spoke structure in the country. What it indicates is that, on average, with three alternative measures for the re-exportation scheme, there is a positive and significant relation between inflows received by road at province j and exports dispatched by road from j to other provinces. Again, with aggregate flows, this could be a sign of a hub-spoke structure, but it could also be an effect of intra-industry trade or an indication of strong urban provinces with intense inflows and outflows. Clearly, a more detailed analysis using sector specific flows is needed Product-specific flows In Tables 6 and 7 we report the results of using Eq. (8) and, respectively, the W and W to model sector-specific flows by road. For brevity, we focus on this mode and this specification, the most interesting one. Similar results with W are reported in the Annex 38

39 (Table A2). Moreover, we have produced a number of alternative estimates, using the same models for the other three transport modes and testing alternative specifications. The results are available upon request. <<<Table 6 about here>>> We now focus on Table 6. Remember that W is the narrowest definition of a hub- spoke structure considered here, since by means of this weight matrix, for every given i j dyad, only re-exporting flows of product k from j to its contiguous provinces n j will be considered. The results show a certain heterogeneity, which is a sign of the richness and complexity of the phenomenon being modeled at the sectoral level. To facilitate this analysis we schematically focus on the main variables: As in the case of aggregate flows (M15, Table 5), ρ is significant and positive for every sector. This indicates that, on average, provinces that are strong importers of product k from other provinces in the country are strong exporters of the same product k to neighboring provinces (such is our narrow concept of re-exporting when we use W ). Size of trading regions: for most sectors, Gross Output has a positive and significant effect on the intensity of inter-provincial flows. S4, S5 and S6 generate strange results, where the coefficient for Ln(GO k i) is non-significant or even negative. In our opinion, this can be explained by a certain conflict between this variable and others included in the regression, such as Ln(wholesale pci). The performance of GDPj as a proxy for the absorption capacity of the importing province is unquestionably valid in all cases. Variables capturing the relevance of trading provinces in terms of logistics, warehouses and wholesale facilities Ln(wholesale pc i ) and Ln(wholesale pc j ) generate interesting results. In the case of the exporting province (Ln(wholesale pc i )), with all other variables controlled for, all sectors but one show significant and positive coefficients, indicating that provinces with large ratios of such facilities per capita are positively associated with a larger capacity for export to other provinces than one might expect given their gross output in this sector. The only exception is S4, one of the sectors for which Ln(GO k i) shows a non-significant 39

40 coefficient. By contrast, the equivalent variable for the importing province (Ln(wholesale pc i )) generates completely different results. None of the inter-provincial sector-specific flows are enhanced (or hampered) by a high/low per-capita ratio of wholesale facilities in the importing province. Results obtained with variables controlling for international trade flows entering or living the country are also worth mentioning: o Results obtained with the four border-province variables (Border int. core EU i, Border int. core EU j, Border int. other j, Border int. other i ) confirm that border provinces are not operating as gateways to foreign markets or, better yet, their inflows and outflows are not inflated by their border geography, because of international transit flows of the same product k. o The results of contrasting inter-provincial flows with international flows for the same k are not so clear. For example, with Ln(imp int.all k i) most sectors show positive and significant associations between their important inflows of product k by all transport modes and, at the same time, their strong interprovincial exports within Spain. This could be a sign that international transit flows affect road transport or can be induced by strong demand for intermediate products imported from abroad (usually included under the same rubric at this rough, fifteen-sector disaggregation). At least in the first case (S1: agricultural products), both effects can be dismissed, since the coefficient obtained is negative and significant. Further research is needed on this point. o By contrast, the results obtained for Ln(exp.int.all k j) are less uniform and straightforward. In any case, high intensities of inter-provincial imports in province j are associated with strong international exports of k from j. In fact, in most cases the coefficients are negative, and in just four cases both significant and negative. o Finally, it is interesting to comment briefly the heterogeneous results obtained for the Distance, Intra and Adjacency variables. Distance is negative and significant in all cases. owever, the elasticity shows clear heterogeneity, even when we consider 40

41 only one transport mode and a medium-size country like Spain. Further research will analyze the variability of mixing the four available transport modes and the nature of each product-specific flow (value/weight; transportability, etc.). Intra, the home bias, is also positive and significant in all cases, with large values for some sectors, as in previous papers using similar datasets (Garmendia et al., 2012). Adjacency remains an enhancing force for inter-provincial flows in many sectors, with exceptions in certain sectors (S4, S5, S7 and S14). <<<Table 7 about here>>> As a robustness check, we offer a quick review of the results in Table 7, which correspond to those obtained when we model sector-specific flows by road with Eq. (8) and W. Remember, we take a prudent approach with this matrix, considering that potential hub provinces can re-export to any other provinces in the system except i and pondering such flows with the inverse of the distance to the ub. Interestingly, the ρ term is significant and positive as before, with little variation. The rest of the results are in general similar to the previous ones, although with some interesting differences. Although Ln(GO k i) for S4 is now positive and significant, four other results prove strange. Moreover, the variable Ln(wholesale pci) shows negative and significant results for some sectors, contradicting the positive effect in Table 6. The rest of the variables behave similarly, confirming the intuitions of the previous table. 8. Conclusions Local freight flows result from global and regional economic processes. Internationally, distribution networks have expanded, in line with the division of the production chain and the development of door-to-door distribution schemes. The proliferation of hub structures, gateways and corridors is growing in most parts of the world. Such developments are behind the rise of efficiency in logistics. owever, their complexity is compounding the already difficult task of producing accurate trade analyses within and between countries. Many papers comparing international and interregional flows use transport flows as the best proxy for internal bilateral trade, assuming no transit flows or re-exportation activities. 41

42 owever, very few have seriously addressed the difficulties that the growing complexity of logistics poses for the identification of production and consumption sites with the points of origin and destination of flows. The aim of this paper is to analyze the complexity of modal cooperation and competition, modeling interprovincial flows within a country and in the presence of logistics infrastructures, warehouses and wholesale activities, as well as hub-spoke structures and multimodal internal and international transit flows. It develops various gravity models incorporating spatial and network autocorrelation effects, and conducts tests upon a rich dataset with aggregate and sector-specific flows between fifty provinces by four alternative transport modes (road, trains, ships, and aircraft) in Spain. The results are rich and promising, but not conclusive. The bottom line is that our strategy, although apparently a clever way to control for the growing complexity in logistics, clearly requires a very detailed dataset, and such datasets are hard to come by. Our paper uses a rich dataset for Spain, but the reader may still wonder whether the significant and positive spatial and network autocorrelation effects we obtain truly confirm the presence of hub-spoke structures. The answer, strictus sensus, is that we still do not know. owever, we can at this point draw at least four important conclusions: (i) certain transport modes compete with other modes within the country (road, trains and aircraft); (ii) the effect of distance on trade becomes positive and significant when the focus is ships and aircraft (i.e., the preferred modes for long distances and access to islands); (iii) as our results suggest, our dataset is free of international transit flows, for we find no inflation of inter-provincial flows at border provinces attributable to their status as gateways or corridors to important foreign markets; (iv) on average, with aggregate and sector-specific flows, provinces receiving strong inter-provincial imports in Spain are also strong exporters to other provinces in the country. This important result is robust with the definition of hub-spoke structures and the use of aggregated and disaggregated data. owever, rather than clearly confirming the presence of hub-spoke structures, it could also be symptomatic of other, equally probable phenomena, such as intra-industry trade or intermediatefinal linkages formed by the presence of a multiple-stage production process within the country. 42

43 Further research is needed to refine the modeling strategy, as well as the empirical assessment for Spain and other countries. In the meantime, our results can be considered a first approach to a complex topic. 9. References Alamá-Sabater L., Márquez-Ramos L. & Suárez-Burguet C. (2013a), Trade and transport connectivity: a spatial approach, Applied Economics, 2013, 45, Alamá-Sabater L., Márquez-Ramos L. & Suárez-Burguet C. & Navarro-Azorín J.M. (2013b): Interregional Trade and Transport Connectivity: An Analysis of Spatial and Networks Dependence by Using a Spatial Econometric Flow Model at NUTS3 level. Working paper. Álvarez-SanJaime O., Cantos-Sánchez P., Moner-Colonques R., Sempere-Monerris J.J. (2013) Vertical Integration and Exclusivities in Maritime Freight Transport. Transportation Research Part E 51 (2013) Álvarez-SanJaime O., Cantos-Sánchez P., Moner-Colonques R., J. J. Sempere-Monerris J.J. (2013) Competition and orizontal Integration in Maritime Freight Transport. Transportation Research Part E 51 (2013) Anderson, J.E. and Van Wincoop, E. (2003). Gravity with Gravitas: A Solution to the Border Puzzle, American Economic Review, 93, 1, Anselin, L. (1988), Spatial Econometrics: Methods and Models, Kluwer Academic Publishers. Atkin D. and Donaldson D. (2012): Who s Getting Globalized? The Size and Nature of Intranational Trade Costs. Working paper. Autant-Bernard, C. & LeSage, J. (2011), Quantifying Knowledge Spillovers Using Spatial Econometric Models. Journal of Regional Science, 51 (3), Behrens, K., Ertur, C & Koch, W. (2010), Dual Gravity: Using Spatial Econometrics to Control for Multilateral Resistance, Journal of Applied Econometrics, first published online: 12 Dec Bensassi, S., Márquez-Ramos, L., Martínez-Zarzoso, I. and Suárez-Burguet, C. (2011) Spanish Logistics and Trade, mimeo, Universitat Jaume I. Black, W. R. (1992), Network Autocorrelation in Transport Network and Flow Systems, Geographical Analysis 24(3), Bolduc, D., Laferriere, R. and Santarossa, G. (1992), Spatial Autoregressive Error-Components in Travel Flow Models, Regional Science and Urban Economics 22(3), Boonstra. J.. (2011): Macro-Integration with Inequality Constraints: An Application to the Integration of Transport and Trade Statistics. Statistica Neerlandica (2011) Vol. 65, nr. 4, pp Brooks, M.R., North American Freight Transportation: The Road to Security and Prosperity. Edward Elgar, Cheltenham, UK. 238 pp. Brülhart M. (2011): The Spatial Effects of Trade Openness: A Survey. Review of the World Economy, 147: Brülhart, M., Crozet, M., & Koenig, P. (2004). Enlargement and the EU periphery: The Impact of Changing Market Potential. World Economy, 27(6), Cantos-Sánchez P., Moner-Colonques R., Sempere-Monerris J.J, Álvarez-SanJaime O. (2009) Alternative Pricing Regimes in Interurban Passenger Transport with Externalities and Modal Competition. Regional Science and Urban Economics 39 (2009) Chen, N. (2004), Intra-national versus International Trade in the European Union: Why Do National Borders Matter?, Journal of International Economics 63(1), Chun, Y. (2008). Modeling Network Autocorrelation within Migration Flows by Eigenvector Spatial Filtering. Journal of Geographical Systems, 10(4), Chun, Y., & Griffith, D. A. (2011). Modeling Network Autocorrelation in Space Time Migration Flow Data: An Eigenvector Spatial Filtering Approach. Annals of the Association of American Geographers, 101(3), Combes, P. P., Lafourcade, M. and Mayer, T. (2005), The Trade-Creating Effects of Business and Social Networks: Evidence from France, Journal of International Economics 66(1), Cosar A.K. and Fajgelbaum P. D. (2012): Market Access and Regional Specialization in a Ricardian World. Mimeo. 43

44 Daumal M, Zignago S (2008) Border Effects of Brazilian States. Working Papers , CEPII Research Center De la Mata, T. and Llano C. (2013), Social Networks and Trade of Services: Modelling Interregional Flows with Spatial and Network Autocorrelation Effects. Journal of Geographical Systems, 15(3), Djankov S, Freund C (2000). Disintegration and Trade Flows: Evidence from the Former Soviet Union, Policy Research Working Paper Series The World Bank Duranton G.; Morrow P. M.; Turner M. A. (2012): Roads and Trade: Evidence from the US. CEPR Discussion Papers 9393, C.E.P.R. Discussion Papers. Evans, C.L. (2003). "The Economic Significance of National Border Effects", American Economic Review, 93, 4, Feo-Valero M., García-Menéndez L., Sáez-Carramolino L., Furió-Pruñonosa S. (2011): The Importance of the Inland Leg of Containerised Maritime Shipments: An Analysis of Modal Choice Determinants in Spain. Transportation Research Part E 47 (2011) Fischer M.M.; Griffith, D. A. (2008), Modeling Spatial Autocorrelation in Spatial Interaction Data: An Application to Patent Citation Data in the European Union. Journal of Regional Science, 48, (5), Fujita and Mori (1996), "The Role of Ports in the Making of Major Cities: Self-Agglomeration and ub- Effect," Journal of Development Economics 49, pp Garmendia, A., Llano-Verduras, C. and Requena-Silventre, F. (2012) Networks and the Disappearance of the Intranational ome Bias, Economic Letters 116, Ghemawat, P., Llano, C. and Requena, F. (2010), Competitiveness and Interregional as Well as International Trade: The Case of Catalonia, International Journal of Industrial Organization 28(4), Gil-Pareja, S., Llorca-Vivero, R., Martínez Serrano J.A. (2006): The Border Effect in Spain: The Basque Country Case. Regional Studies, 40.4, anson, G.. (2001). US-Mexico Integration and Regional Economies: Evidence from Border-City Pairs. Journal of Urban Economics, 50(2), anson, G.. (1996), "Economic Integration, Intraindustry Trade, and Frontier Regions," European Economic Review 40, pp elble (2007). Border Effect Estimates for France and Germany, Review of World Economics, Vol. 143 (3) elliwell JF (1998). ow Much Do National Borders Matter? Brookings Institution Press, Washington, DC. elliwell, J. F. (1997), National Borders, Trade and Migration, Pacific Economic Review 2(3), elliwell, JF (1996). Do National Borders Matter for Quebec s Trade?, Canadian Journal of Economics, 29(3), enderson D. J. And Millimet D.L. (2008), Is Gravity Linear? Journal of Applied Econometrics, 23: esse, M., Rodrigue, J.-P., The Transport Geography of Logistics and Freight Distribution. Journal of Transport Geography 12 (3), illberry, R. and ummels, D. (2008). Trade Responses to Geographic Frictions: A Decomposition Using Micro-Data", European Economic Review, 52, 3, Jackson, R., Walter R. Schwarm, Yasuhide Okuyama, Samia Islam (2006): A Method for Constructing Commodity by Industry Flow Matrices. Annals of Regional Science, 40, Ledyaeva, S. (2009), Spatial Econometric Analysis of Foreign Direct Investment Determinants in Russian Regions, The World Economy 32(4), LeSage J.P. and Llano C. (2013), A Spatial Interaction Model with Spatially Structured Origin and Destination Effects. Journal of Geographical Systems, 15(3), LeSage J.P & Fischer M. M. (2008): Spatial Growth Regressions: Model Specification, Estimation and Interpretation, Spatial Economic Analysis, 3:3, LeSage, J. P. and Polasek, W. (2008) Incorporating Transportation Network Structure in Spatial Econometric Models of Commodity Flows, Spatial Economic Analysis, 3, LeSage, J. P. & Pace, R. K. (2008), Spatial Econometric Modeling of Origin-Destination Flows, Journal of Regional Science 48(5), LeSage, J. P. & Pace, R. K. (2009), Introduction to Spatial Econometrics, Chapman & all/crc Press. Llano, C., Esteban, A., Pulido, A., Pérez, J (2010): Opening the Interregional Trade Black Box: The C- intereg Database for the Spanish Economy ( ). International Regional Science Review, 33,

45 Llano-Verduras C.; Minondo A., Requena-Silvente F. (2011). Is the Border Effect an Artefact of Geographical Aggregation? The World Economy, 34, 10, Markus esse, Jean-Paul Rodrigue (2004): The Transport Geography of Logistics and Freight Distribution. Journal of Transport Geography 12 (2004) Márquez-Ramos, L., Martínez-Zarzoso, I., Pérez-García, E. and Wilmsmeier, G. (2011) Special issue on Latin-American research maritime networks, services structure and maritime trade, Networks and Spatial Economics, 11, McCallum, J. (1995), National Borders Matter: Canada-U.S. Regional Trade Patterns, The American Economic Review 85, Millimet D, and Osang Th (2007). Do State Borders Matter for U.S. Intranational Trade? The Role of istory and Internal Migration, Canadian Journal of Economics, 40(1), Monzón A., and Rodríguez-Dapena A. (2006): Choice of Mode of Transport for Long-Distance Trips: Solving the Problem of Sparse Data. Transportation Research Part A 40 (2006) Nitsch, V., (2002). Border Effects and Border Regions: Lessons from the German Unification. Banakgesellschaft, Berlin. Okubo, T., (2004) The Border Effect in the Japanese Market: A Gravity Model Analysis". Japan International Economics, 18:1 11 Porojan, A. (2001), Trade Flows and Spatial Effects: The Gravity Model Revisited, Open Economies Review 12(3), Rappaport, J. and Jeffrey D. Sachs (2003),"The United States as a Coastal Nation," Journal of Economic Growth 8, pp Requena, F. and Llano, C. (2010), The Border Effects in Spain: An Industry-Level Analysis, Empirica 37(4), Rodrigue J.P. (2003): Freight, Gateways And Mega-Urban Regions: The Logistical Integration of the Bostwash Corridor. Tdschrift voor Economische en Sociale Geografie 2004, Vol. 95, No. 2, pp Rodrigue J-P., Notteboom T. (2010): Comparative North American and European Gateway Logistics: The Regionalism of Freight Distribution. Journal of Transport Geography 18 (2010) Rodrigue, J.-P., Comtois, C., Slack, B., The Geography of Transport Systems. Routledge, New York. Shultze, M.S. and N. Wolf (2008), On the Origins of Border Effects: Insights from the absburg Empire, Journal of Economic Geography, van Klink,. & G. van den Berg (1998), Gateways and Intermodalism. Journal of Transport Geography 6, pp Wolf,. C. (2000), Intranational ome Bias in Trade, Review of Economics and Statistics 82(4), Wolf,.C. (1997). Patterns of Intra- and Inter-State Trade. NBER Working Paper National Bureau of Economic Research, Cambridge, Mass. Woudsma, C. (1999), NAFTA and Canada-US Crossborder Freight Transportation. Journal of Transport Geography 7, pp Yilmazkuday. (2012): Understanding Interstate Trade Patterns. Journal of International Economics 86 (2012)

46 Tables Table 1: Transport statistics used to estimate Spanish interregional trade MODE ROAD TRAIN SIP AIRCRAFT DESCRIPTION AND MAIN FEATURES Permanent Survey on Goods Transport by Road (Encuesta Permanente de Mercancías por Carretera) Source: Ministerio de Fomento (Spanish Ministry of Public Works) Product Disaggregation: 160 products (class. NSTR-3 digits) Availability: 1995 present Remarks: Permanent survey on the activity of a large sample of trucks in Spain: each trip includes origin-destination, product type, volume, distance (km), etc. Survey may also include international transit flows from ports/airports to final destinations. It should be noted that figures obtained from truck surveys can be inconsistent with figures on production/purchases from firms/household surveys. RENFE Statistics on Complete Wagon and Container Flows Source: Statistics Department of RENFE Product Disaggregation: approx. 40 categories (RENFE classification) Remarks: Every domestic flow recorded: high quality, low product detail. No information on products transported by container (30% of rail flows). Statistics from Spanish Ports (Puertos del Estado) Indirect estimation of interregional flow matrices using optimization procedure based on: a) Tons loaded/unloaded per Spanish port, flow type, product type. Source: Statistical Yearbook. Puertos del Estado. Data: Annual. 27 Spanish ports. Product Disaggregation: 40 products (Puertos del Estado classification) b) Set of Spanish domestic flow matrices with Ports of Origin and Destination Source: Domestic maritime flows by Origin and Destination Puertos del Estado. Data: Annual. 38 largest Spanish ports (at the time). Product Disaggregation: 52 products (CSTE) O/D matrices of domestic flows of goods by airport of origin and destination AENA. Source: AENA & Ministerio de Fomento. Data: Annual. Main Spanish Airports. Product Disaggregation: None Remarks: No information on sectoral disaggregation of domestic flows by air. 46

47 Table 2: Variables Name (variables for importing regions are reported in parentheses) F ; F ; F ; F kr kt ks ka Pop i (Pop j ) GDP i (GDP j ) GO k i Description Inter-provincial flows in Spain by transport mode (R, T, S and A) and sector (k = all sectors, or product specific flows for 15 sectors) Population of province i (Population of province j) Gross Domestic Product of province i (Gross Domestic Product of province j) Dummy variable that takes value 1 if i = j and 0 otherwise. d For distance within the Iberian Peninsula: actual distance traveled by trucks on their deliveries (EPTMC). For distance between provinces within the Iberian Peninsula and islands: actual distance to main ports + actual distance from ports to islands by ship. Intra Dummy variable that takes value 1 if i = j and Adjac Wholesales pc i (Wholesales pc j ) Logistic Infraest i (Logistic Infraest j ) Island i (Island j ) Imp. int. all i (Exp. int. all i ) Imp. int.all k i (Exp. int. all k j). Border int. core EU i (Border int. core EU j ) Border int. other i (Border int. other j ) 0 otherwise. Dummy variable that takes value 1 if i and j share a common border within Spain and 0 otherwise. Ratio between number of wholesale activities in province i (j) and population of same province. Variable expressed in logs and used as alternative to the platform infrastructure. Number of logistic infrastructures located in province i (j). Expressed in levels. Dummy variable that takes value 1 if exporting region i (importing province j) is an island province: Islas Baleares; Las Palmas and Santa Cruz de Tenerife. International imports of all products in province i regardless of transport mode (Idem for exports in province j) International imports of product k in province i regardless of transport mode (Idem for exports in province j) Dummy variable that takes value 1 for Spanish provinces bordering on France or Andorra and 0 otherwise. (Idem for province j) Dummy variable that takes value 1 for Spanish provinces bordering on Portugal or Africa and 0 otherwise. (Idem for province j) Source Several sources reported in Table 1. Prepared for the C- intereg project. INE. Regional Accounts. INE. Regional Accounts. INE. Industrial Survey and National Accounts. Prepared for the C-intereg project. EPTMC. Ministerio de Fomento. Puertos del Estado. Authors Authors La Caixa, Anuario Económico, Borrowed from Alama et al. (2013b) Authors Official international trade data. AEAT Official international trade data. AEAT Authors Authors 47

48 Table 3. Ordinary least squares. M1 M2 M3 M4 M5 M6 M7 M8 M9 Dependent variable All (aggreg.) Road Ln(F R ) Train Ln(F T ) Ship Ln(F S ) Aircraft Ln(F A ) All (aggreg.) Road Ln(F R ) All (aggreg.) Road Ln(F R ) Rbar-squared Sigma Ln(GDP i ) 1.876*** 1.729*** 2.395*** 1.611*** 0.962*** 1.362*** 0.997*** 1.119*** 0.62** (15.5) (13.556) (18.793) (14.776) (15.941) (5.749) (3.983) (4.569) (2.391) Logist. infrastructure i 0.298*** 0.202** (3.348) (2.107) Ln(wholesales pc i ) 2.445*** 2.455*** (4.843) (4.643) Island i -2.45*** *** -3.65*** 6.791*** 5.858*** *** *** -3.08*** *** (-4.994) ( ) (-7.066) (15.371) (23.952) (-5.397) ( ) (-6.134) ( ) Border int. core EU i (0.289) (0.815) (-1.479) (-0.706) Border int. other i *** *** ** ** (-2.937) (-2.749) (-2.048) (-2.027) Ln(imp. int. all i ) (0.426) (0.177) (1.552) (1.159) Ln(GDP j ) 2.233*** 2.224*** 2.292*** 1.557*** 0.933*** 1.928*** 1.772*** 2.223*** 1.524*** (18.436) (17.433) (17.968) (14.272) (15.44) (8.517) (7.246) (10.874) (6.631) Logist. infrastructure j 0.277*** (2.877) (0.949) Ln(wholesales pc j ) *** (1.317) (2.824) Island j 0.903* *** *** 9.532*** 7.965*** *** *** (1.841) (-6.426) (-5.205) (21.585) (32.582) (0.485) (-7.787) (-0.355) (-8.264) Border int. core EU j * (-1.333) (-0.399) (-1.839) (-1.198) Border int. other j * (-1.291) (-1.965) (-0.615) (-1.621) Ln(Exp. int. all j ) (-0.701) (-0.762) (-1.315) (-0.111) Ln(d ) *** *** -1.31*** 1.102*** 0.712*** *** *** *** *** ( ) ( ) (-6.415) (6.307) (7.362) (-8.846) ( ) (-8.559) ( ) Intra 1.508* *** 0.729* 3.072*** 2.257** 3.227*** 2.26** (1.778) (1.404) (1.625) (6.92) (1.725) (3.461) (2.395) (3.643) (2.409) Adjac 2.62*** 1.99*** 2.691*** 1.988*** (5.738) (4.118) (5.908) (4.133) Ln(F T ) 0.055*** 0.06*** (2.635) (2.976) Ln(F S ) (0.189) (0.546) Ln(F A ) 0.248*** 0.248*** (5.622) (5.691) Constant *** *** *** *** ** *** *** (-2.988) (-0.414) ( ) ( ) ( ) (-1.251) (2.5) (-5.917) (-2.71) *** p<0.01, ** p<0.05, * p<0.1; all variables except the dummies in log form. 48

49 Table 4. Spatial autocorrelation tests. Model 6-9. M6 M7 W i W j W=W i +W j W i W j W=W i +W j Moran s I - test for spatial autocorrelation in the residuals Moran s I Moran s I-statistic Marginal probability LM error test for spatial autocorrelation in the residuals LM value Marginal probability LM error test for spatial autocorrelation in the dependent variable LM value Marginal probability Robust LM error test LM value Marginal probability Robust LM lag test LM value Marginal probability Combined LM lag and LM error test LM value Marginal probability M8 M9 W i W j W=W i +W j W i W j W=W i +W j Moran s I - test for spatial autocorrelation in the residuals Moran s I Moran s I-statistic Marginal probability LM error test for spatial autocorrelation in the residuals LM value Marginal probability LM error test for spatial autocorrelation in the dependent variable LM value Marginal probability Robust LM error test LM value Marginal probability Robust LM lag test LM value Marginal probability Combined LM lag and LM error test LM value Marginal probability

50 Flow type Table 5. Spatial autoregressive models. M10 M11 M12 M13 M14 M15 M16 Road Train Ship Aircraft Road Road Road Ln(F R ) Ln(F T ) Ln(F S ) Ln(F A ) Ln(F R ) Ln(F R ) Ln(F R ) W type W=W i +W j W=W i +W j W=W i +W j W=W i +W j W W W Rbar-squared sigma log-likelihood Ln(GDP i ) 0.564** 1.162*** *** 0.291*** 0.25* 0.293*** (2.265) (4.723) (-0.018) (5.828) (3.759) (1.934) (3.762) Ln(wholesales pc i ) 1.942*** 0.967* ** *** (3.814) (1.915) (-0.819) (2.035) (1.588) (9.285) (1.639) Island i *** *** 1.074*** 2.723*** *** *** -1.73*** ( ) (-7.932) (2.924) (11.636) ( ) ( ) ( ) Border int. core EU i ** *** *** *** (-1.346) (-2.413) (-4.976) (-2.778) (0.444) (-3.689) (0.42) Border int. other i * 0.332* *** (-1.179) (-1.949) (1.765) (0.297) (-1.586) (-2.758) (-1.599) Ln(imp. int. all i ) *** * (1.33) (1.219) (6.484) (-1.516) (0.702) (1.931) (0.712) Ln(GDP j ) 1.319*** 1.975*** 0.558*** 0.836*** 0.561*** 0.956*** 0.567*** (5.951) (9.014) (4.097) (8.956) (8.545) (8.802) (8.568) Ln(wholesales pc j ) 1.333** *** 0.705** 0.467*** (2.509) (-1.464) (-0.993) (-0.101) (2.785) (2.496) (2.787) Island j *** *** 1.6*** 2.56*** *** *** -1.26*** (-6.763) (-8.906) (3.765) (9.252) (-6.617) (-6.266) (-6.634) Border int. core EU j * * ** ** (-1.459) (-1.843) (-1.759) (-2.067) (-1.351) (-2.042) (-1.356) Border int. other j ** * * (-0.91) (1.628) (2.57) (0.945) (-1.698) (-1.151) (-1.697) Ln(exp. int. all j ) *** 0.112** *** (-0.272) (-3.552) (2.181) (-6.269) (-1.315) (-1.423) (-1.32) Ln(distance) *** ** *** 0.176* -0.7*** *** -0.71*** (-7.713) (-2.205) (-2.673) (1.702) (-9.474) ( ) (-9.503) Intra 2.379*** *** *** 1.363*** 1.158*** (2.642) (1.431) (4.086) (-0.453) (4.141) (2.915) (4.133) Adjac 1.408*** 1.252*** *** 0.668*** 0.644*** (3.038) (2.743) (-1.143) (-0.227) (4.447) (2.779) (4.46) Ln(F R ) 0.052*** *** (2.708) (-0.84) (3.179) Ln(F T ) 0.052*** 0.048*** 0.037*** (2.683) (4.04) (4.455) Ln(F S ) *** 0.037*** (-0.159) (2.791) (3.775) Ln(F A ) 0.227*** 0.219*** 0.109*** (5.43) (5.29) (4.239) Constant *** *** *** *** *** *** -3.38*** (-4.199) (-6.028) (-4.872) (-6.818) (-2.905) (-7.462) (-2.939) rho 0.387*** 0.445*** 0.873*** 0.574*** 0.737*** 0.565*** 0.734*** (13.206) (14.528) (54.755) ( ) ( ) ( ) ( ) *** p<0.01, ** p<0.05, * p<0.1; all variables except dummies in log form. 50

51 Table 6. Sector-specific analysis. Spatial autoregressive model. W = ub2 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14 Rbar-squared Sigma Log-likelihood Ln(GO k i) 1.223*** 0.09*** 0.794*** *** 0.061*** 0.075*** 0.172* 0.175** 0.401*** (15.559) (5.213) (8.467) (0.424) (-0.399) (-2.893) (3.756) (4.565) (1.778) (2.248) (4.789) (0.767) (0.902) (-0.449) Ln(wholesale pc i ) 1.145*** 0.646*** 1.019*** * 0.749*** 1.971*** 0.896*** 1.428*** 1.727*** 1.206*** 0.47*** 1.078*** 0.781** (5.379) (2.97) (3.821) (-0.806) (1.683) (4.657) (9.825) (3.519) (6.734) (7.437) (4.403) (2.666) (4.005) (2.551) Island i ** *** * ** *** ** * (0.115) (-2.484) (-4.252) (-0.305) (1.011) (-1.684) (-0.098) (-0.114) (-2.396) (-1.262) (-1.343) (6.638) (-2.02) (-1.876) Border int. core EU i * *** -0.68*** *** *** *** *** *** (-1.651) (-4.403) (-3.788) (-0.541) (-3.45) (-3.604) (-1.534) (-4.094) (-1.362) (-3.767) (-4.071) (-1.457) (-1.444) (-1.556) Border int. other i 0.411*** ** * *** *** *** (2.697) (-2.559) (-0.151) (-0.454) (-1.297) (-1.873) (-3.669) (0.696) (-4.252) (-4.287) (0.713) (1.271) (-1.283) (-0.634) Ln(imp. int.all k i) *** 2.253*** 1.916** 2.965*** 2.788*** 3.32*** 2.411*** 5.261*** 3.446*** *** 2.154*** 4.424*** *** (-5.079) (2.98) (2.211) (5.511) (4.914) (4.687) (3.046) (5.057) (4.063) (1.098) (3.924) (2.915) (7.218) (11.269) Ln(GDP j ) 0.853*** 0.738*** 1.494*** 0.378*** 0.348*** 0.444*** 0.961*** 0.633*** 0.754*** 0.622*** 1.165*** 0.302*** 0.957*** 1.072*** (8.339) (8.374) (14.521) (6.178) (5.592) (6.276) (10.468) (6.717) (8.257) (6.748) (11.702) (4.298) (9.182) (9.787) Ln(wholesale pc j ) (-0.548) (-0.266) (-1.177) (0.723) (1.65) (0.934) (1.591) (1.223) (0.433) (1.324) (-0.196) (1.089) (-0.36) (0.967) Island j * *** ** *** 0.473** ** *** (1.169) (-1.741) (-2.67) (-0.532) (0.994) (-0.956) (-1.101) (-1.996) (-0.744) (-1.589) (-3.916) (2.32) (-2.467) (-3.228) Border int. core EU j ** * -0.55*** ** *** -0.44*** *** *** ** (-2.26) (-1.901) (-2.989) (-2.214) (-3.653) (-3.518) (-0.37) (-0.429) (-1.466) (-4.109) (-0.733) (-1.452) (-3.796) (-2.233) Border int. other j *** ** *** -0.23* (-0.461) (-0.999) (-0.31) (0.224) (-1.418) (-3.893) (-2.075) (-2.72) (-1.73) (-0.549) (-1.357) (0.543) (-0.41) (-0.304) Ln(exp. int.all k j) ** *** *** *** *** (-0.706) (-2.221) (-4.21) (-0.151) (0.003) (-0.467) (-1.522) (-0.902) (-0.06) (-0.347) (-3.782) (-2.917) (-0.925) (-3.022) Ln(distance) *** *** *** *** *** *** -1.18*** *** *** *** *** *** *** ( ) ( ) ( ) (-4.697) (-3.833) (-7.37) ( ) (-9.254) (-7.238) ( ) (-11.34) (-1.101) (-9.738) ( ) Intra 2.011*** 4.239*** 1.921*** 1.46*** 1.878*** 3.459*** 1.637*** 3.061*** 2.863*** 2.398*** 2.573*** 5.961*** 5.272*** 1.68*** (4.573) (11.183) (4.428) (5.56) (7.036) (11.635) (4.201) (7.605) (7.394) (6.043) (6.172) (19.692) (11.866) (3.692) Adjac 1.86*** 2.502*** 1.559*** *** *** 1.008*** 1.469*** 1.467*** 1.175*** 2.388*** (8.224) (12.841) (6.972) (-0.081) (0.385) (9.66) (1.112) (6.854) (5.052) (7.193) (6.828) (7.533) (10.441) (0.288) Constant *** -8.01*** *** *** *** *** *** *** *** *** *** *** *** *** (-5.398) (-3.794) (-9.142) (-6.645) (-7.639) (-6.645) (-7.66) (-7.438) (-9.526) (-3.606) (-6.089) (-5.055) (-8.585) (-12.47) rho 0.652*** 0.661*** 0.664*** 0.646*** 0.622*** 0.734*** 0.647*** 0.676*** 0.679*** 0.658*** 0.674*** 0.669*** 0.604*** 0.633*** ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) *** p<0.01, ** p<0.05, * p<0.1; all variables except dummies in log form. 51

52 Table 7. Sector-specific analysis. Spatial autoregressive model. W = ub3 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14 Rbar-squared Sigma Log-likelihood Ln(GO k i) 0.989*** 0.131*** 1.1*** 0.031* *** 0.039** 0.065*** 0.685*** 0.319*** 0.392*** * (13.97) (6.332) (11.558) (1.852) (-0.581) (-5.786) (2.321) (3.583) (6.062) (3.949) (4.24) (-1.769) (-1.222) (-0.87) Ln(wholesale pc i ) ** *** ** *** *** * 0.704*** (-2.559) (-3.471) (-2.18) (-3.131) (-4.082) (-1.71) (3.329) (-1.524) (-0.117) (0.56) (1.189) (-0.779) (0.452) (-1.176) Island i ** * *** (1.297) (-0.71) (-1.615) (1.044) (2.588) (0.326) (1.911) (1.459) (0.179) (1.259) (-0.11) (7.274) (-0.639) (-0.308) Border int. core EU i 0.361** * ** * (2.194) (-1.037) (-0.624) (1.078) (-0.887) (-0.535) (1.702) (-0.838) (2.472) (-0.396) (-1.186) (1.662) (1.285) (1.341) Border int. other i 0.415*** ** *** ** -0.48*** (3.032) (-2.074) (0.748) (-0.348) (-0.936) (-1.225) (-2.757) (1.007) (-1.97) (-3.356) (0.994) (1.319) (-0.816) (-0.014) Ln(imp. int.all k i) *** *** 5.425*** 8.665*** 4.158*** 5.961*** *** 4.217*** 4.29*** *** (-1.64) (4.095) (0.754) (3.421) (7.265) (7.771) (5.3) (5.04) (0.42) (0.429) (2.863) (3.916) (6.511) (10.099) Ln(GDP j ) 0.752*** 0.848*** 1.439*** 0.507*** 0.412*** 0.588*** 0.976*** 0.659*** 0.803*** 0.619*** 1.254*** 0.361*** 0.979*** 1.135*** (8.224) (8.111) (14.206) (5.547) (5.126) (5.611) (10.237) (6.355) (7.865) (6.413) (11.245) (3.529) (8.834) (9.586) Ln(wholesale pc j ) (-0.524) (-0.455) (-1.205) (0.463) (1.503) (0.827) (1.596) (1.163) (0.41) (1.25) (-0.241) (0.886) (-0.349) (0.956) Island j * *** ** * -1.27*** 0.677** ** *** (1.053) (-1.841) (-2.802) (-0.578) (0.991) (-0.997) (-1.25) (-2.255) (-0.922) (-1.731) (-3.968) (2.277) (-2.585) (-3.396) Border int. core EU j ** *** ** *** *** (-1.514) (-0.913) (-2.193) (-0.918) (-2.777) (-2.549) (0.324) (0.345) (-0.68) (-3.292) (0.086) (-0.281) (-3.103) (-1.458) Border int. other j * *** ** *** -0.29* (-0.76) (-1.24) (-0.602) (-0.108) (-1.725) (-4.039) (-2.356) (-2.992) (-1.923) (-0.855) (-1.511) (0.286) (-0.676) (-0.554) Ln(exp. int.all k j) ** *** *** ** *** (-0.632) (-2.159) (-4.13) (-0.13) (0.185) (-0.06) (-1.426) (-0.82) (-0.008) (-0.235) (-3.593) (-2.476) (-0.806) (-2.975) Ln(distance) *** *** -1.17*** *** *** -0.76*** -1.16*** -0.97*** *** *** *** *** *** ( ) ( ) ( ) (-3.873) (-3.218) (-6.425) ( ) (-8.297) (-6.343) ( ) ( ) (-0.494) (-9.013) ( ) Intra 2.025*** 5.148*** 2.056*** 2.142*** 2.512*** 5.016*** 1.863*** 3.58*** 3.335*** 2.667*** 3.045*** 8.517*** 5.826*** 2.032*** (5.131) (11.408) (4.768) (5.438) (7.273) (11.202) (4.588) (8.043) (7.622) (6.387) (6.407) (19.297) (12.264) (4.069) Adjac 2.064*** 3.446*** 1.882*** 0.478** 0.46** 2.671*** 0.562*** 1.993*** 1.519*** 1.93*** 2.069*** 2.161*** 2.968*** 0.535** (10.172) (14.856) (8.474) (2.358) (2.583) (11.584) (2.688) (8.701) (6.746) (8.984) (8.459) (9.508) (12.138) (2.078) Constant ** * *** -6.66*** *** *** *** *** *** *** *** *** *** (-2.429) (-1.817) (-6.267) (-3.088) (-7.557) (-7.648) (-7.409) (-5.36) (-4.982) (-0.851) (-4.167) (-4.141) (-5.85) ( ) rho 0.679*** 0.59*** 0.668*** 0.478*** 0.501*** 0.606*** 0.627*** 0.639*** 0.64*** 0.639*** 0.633*** 0.517*** 0.573*** 0.593*** ( ) ( ) ( ) ( ) (728.38) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) *** p<0.01, ** p<0.05, * p<0.1; all variables except dummies in log form. 52

53 Figures Figure 1: Domestic transport-mode competition scheme Road: kr F i Train: Ship: kt F ks F j Province i (producer = exporter) Aircraft: ka F Province j (consumer = importer) Figure 2: ub-spoke structure with national and international transit flows i RoW () ub j RoW kr F Intra-national transit flows: A i Spanish province receives product k from another Spanish province i, which is the true producer (i.e., exporter). The load is then re-exported from to j within the country. kr F International transit flow: a RoW Spanish province receives k products from abroad, which are then re-exported to another province j within the country. Province is the ub, where the logistics, warehouse, wholesale facilities are. A ub province could also be one with a port, airport or large train station. kr F Intra-national transit j flows: a Spanish province delivers product k to final destination j, which is the true consumer (i.e., importer). The load has been produced not in but in i. kr F International transit RoW flow: a Spanish province delivers product k to final destination j. This load has been produced not in but abroad. 53

54 Figure 3: Scheme describing domestic multi-modal transport flows i (N) Node j kr F ; in kt F ; in ks F ; in ka F in N (multimodal node). Province i (producer = exporter) Province N receives products from other provinces within the country (we assume international flows are excluded) kr F ; Nj kt F ; Nj ks F ; Nj ka F Nj Province j (consumer = importer) 54

55 Figure 4. Total Trade Flows (Exports) by Provinces (2007). Euros. Figure 5. IMRI for Road Freight Flows (2007). 55

56 Figure 6. IMRI for Train Freight Flows (2007). Figure 7. IMRI for Ship Freight Flows (2007). 56

57 Figure 7. IMRI for Aircraft Freight Flows (2007). Figure 8. IMDI by Province. 57

58 Figure 9. Bivariate Local Moran s I between IMDI and Per-Capita GDP. 58