WPS8226 Policy Research Working Paper 8226 Roads and the Geography of Economic Activities in Mexico Brian Blankespoor Théophile Bougna Rafael Garduno-Rivera Harris Selod Development Research Group Environment and Energy Team & Development Data Group October 2017 Policy Research Working Paper 8226 Abstract This paper estimates the impacts of road improvements markets, the analysis finds significant and positive causal on local employment and specialization in Mexico for effects of improved domestic accessibility on employ- 1986–2014, through changes in access to domestic markets ment and specialization. It also finds that employment and travel costs to ports and the U.S. border. Instrument- is stimulated by lower transport costs to the U.S. border, ing for road placement endogeneity and addressing the but harmed by lower transport costs to ports. Hetero- recursion problem in regressions that involve access to geneous effects are found across sectors and regions. This paper is a product of the Environment and Energy Team, Development Research Group and the Development Data Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http:// econ.worldbank.org. The authors may be contacted at bblankespoor@worldbank.org, tbougna@worldbank.org, rafael. garduno@cide.edu, and hselod@worldbank.org (corresponding author). The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Roads and the Geography of Economic Activities in Mexico * Brian Blankespoor† Théophile Bougna‡ Rafael Garduno-Rivera§ Harris Selod ¶ Keywords: Mexico, Roads, Market Access, Market Potential, Specialization, Industrial Lo- calization. JEL classification: R12; R30; C23. * We thank Luis Rodrigo Gonzalez Vizuet, Shiyan Zhang, and Lucie Letrouit for excellent research assistance. We are indebted to Mike Bess for the many inputs he provided regarding the history of roads in Mexico, and to Matt Turner, Uwe Deichmann, Maria Marta Ferreyra and Mark Roberts, who provided very useful comments on the preliminary versions of this paper. We also thank participants to the authors’ workshop for the World Bank study on ‘Cities, Productivity and Growth in Latin America and the Caribbean’, and at the Toulouse Development, Labor and Public Policy Seminar. Funding from the World Bank and from dfid under the World Bank Strategic Research Program (srp) ‘Transport for Sustainable and Inclusive Growth’ is gratefully acknowledged. † Development Data Group, The World Bank, E-mail: bblankespoor@worldbank.org ‡ Development Research Group, The World Bank, E-mail: tbougna@worldbank.org § Centro de Investigación y Docencia Económicas, Mexico. E-mail: rafael.garduno@cide.edu ¶ Corresponding author. Development Research Group, The World Bank. E-mail: hselod@worldbank.org 1 Introduction There are many benefits associated with investments in transport infrastructure which can stimulate growth through trade, structural transformation, agglomeration and productivity (see Redding and Turner, 2015, and Berg el al., 2017 for surveys). The impacts can be large: Calderón, Moral-Benito and Servén (2009) estimate that a 10 percent improvement in the stock of infrastructure (including transport infrastructure) increases per capita gdp by 0.7 to 1 per- cent. In Latin America, in particular, the potential to stimulate growth through transport investments could be especially significant given the large infrastructure gap in most Latin American countries (see Fay and Morrison, 2007; Gonzalez, Guasch and Serebrisky, 2007). Perrotti and Sánchez (2011) estimate that ’closing the infrastructure gap’ in Latin American countries would require investments in the amount of up to 6.2 percent of annual GDP for the period 2012- 2020. In theory, benefits could then materialize in the form of a reduction in transportation costs (or travel time), more intense competition, and deeper economic integra- tion. Locations where access to markets is improved would become more attractive for firms, potentially triggering a process of concentration and specialization of local economic activity whereby innovation, the sharing of input markets, and improved matching on the labor market would increase productivity, and in turn lead to local economic development and an increase in living standards. In this paper, we focus on the first steps in this potential chain of events, in the case of Mex- ico, a country which experienced relatively large investments in roads over the past decades and from which lessons could be derived for the rest of the continent. In Mexico, road trans- port accounts for 57% of freight transport and remains a vital component of the economy. The Mexican road network links the country from north to south and between its two oceanic coastlines; and some of the most important road connections link the capital city with border crossings to the US. We investigate the extent to which these road improvements increased access to markets and, in turn, affected the location of economic activities. Our paper contributes to a growing literature that looks at the effect of transport infrastruc- ture on local economic development outcomes (Donaldson and Hornbeck, 2015; Duranton and Turner, 2012; Michaels, 2008; Baum-Snow, 2007; Baum-Snow et al., 2017; Jedwab and Storey- gard, 2016; Alder, 2016; Alder, Robert and Tewari, 2017; Straub and Bird, 2015, Ghani et al., 2015; Faber, 2014; Rothenberg, 2015, Fyre, 2015; Galasso and Oettl, 2017). In view of this lit- erature, our paper looks at the effects of roads on the geographical concentration of economic activities and the specialization of localities. There are two challenges to estimate the causal impact of roads on local economic out- comes: the non-random placement of roads, and the recursion problem inherent in regression of economic outcomes on market access measures. To address these two problems, we adopt three separate identification strategies. First, we use the so-called ’doughnut’ strategy, which 1 allow to instrument access to markets excluding zones that are more likely to be prone to an endogeneity bias due to non-random road placement. Second, in order to avoid the recursion problem, we substitute market access indicators with measures of access to infrastructure that do not involve population or income. These measures include the number of roads intersecting a circle of a given radius around the centroid of each locality, or the number of kilometers of roads within a given distance to that centroid. Third, we use these access to infrastructure variables as instruments for market access indicators. Our paper makes three principal contributions. First, we construct a new panel of geo- referenced roads data in Mexico over three decades (for the period 1986-2014). Second, we also take advantage of the panel nature of our data to estimate time-varying measures of market access for more than 2,000 localities (whereas previous studies, either just used a few dates or had a sample with a limited number of observations). Third, to the best of our knowledge, our paper is the first to study the effect of market access on specialization. Our key findings are that transportation improvements over the last three decades increased access to markets in Mexico—which we measure in concurrent ways—and that these improve- ments in accessibility had a positive impact on local employment and economic specialization. The effects are large and heterogeneous across sectors. A 10 percent increase in market access results in a 1.6 to 2.1 percent increase in employment and a large increase in locality specializa- tion. The effect of our market potential measure is slightly larger. A 10 percent increase results in a 2.9 to 6.5 percent increase in employment, and a 13.0 percent increase in output special- ization. We also find heterogeneous effects across sectors with employment in commerce and services benefiting more than manufacturing from road improvements. The rest of the paper is structured as follows. Section 2 presents the Mexican context. Section 3 describes the data used, explains the construction of our market access and market potential measures, and provides information on the trends in industrial concentration and specialization in Mexico. Section 4 presents our empirical methodology and discusses iden- tification strategies. Section 5 discusses our empirical results regarding the effects of market access and market potential on employment and specialization, and Section 6 concludes. 2 The Mexican context: Lessons from the literature 2.1 The concentration of industries and specialization of localities Economic activity in Mexico is highly concentrated in the Mexico City Metropolitan Area, which in 2010, contributed to a quarter of the national Gross Value Added even though it cov- ers less than 0.3 percent of the national territory. In the wake of Michael Porter’s work (Porter, 1998), there has been increasing policy interest in Mexico to facilitate the industrial concen- tration and specialization of localities and the development of industries, so as to ultimately 2 achieve higher economic growth. However, economic studies that shed light on these issues are scarce and only a few case studies focus on industrial concentration and specialization. A first set of such studies measures the level of concentration of various industries and its increase over time (Unger, 2003) and tries to identify the determinants of industrial clustering, concluding that, in Mexico, labor force skill play a very important role to attract clusters, much more than wage differentials (Unger and Chico, 2004). Complementing these observations, several authors identify the locations that have become more specialized over time, focusing on different spatial scales. Pérez and Palacio (2009) find that specialization increased during the 1994-2004 period. Focusing on Mexican cities, Kim and Zangerling (2016) find that, between 1990 and 2010, specialization only increased in the Mexico City Metropolitan Area, but did not in other cities which remained very much diversified. A second set of case studies points at the potential benefits of clustering in Mexico. Monge (2012) suggests that clustering of the tequila industry in the state of Jalisco (in Western Mexico) reduced transaction costs and that entrepreneurs in that sector benefited from labor special- ization. Dávila (2008), who studied the economic performance and the commercial integration with Texas of industrial clusters in northeastern Mexico1 during the 1993-2003 period, suggests that economic clusters played a role to foster productive innovations and stimulate bilateral trade. 2.2 Road investments: A historical perspective In Mexico, large investments in roads started during the Spanish Colony (1521-1810), but these roads were mainly focused on connecting natural resources (especially silver and gold) with the port of Veracruz, to ship them to Spain. During the presidency of Porfirio Días (between 1884 and 1911), while railways flourished throughout Mexico, the road network, which essentially dated back to the colonial period, received little financial support from the government and consequently deteriorated (Bess, 2016b). The Mexican Revolution, between 1910 and 1920, did not improve the situation. In 1918, municipal surveyors in Mexico City evaluated that the conflict had damaged around 4,000 km of roads serving the federal capital (Bess, 2016b). In the 1920s and 1930s, during the presidency of Alvaro Obregon and Plutarco Elias Calles, national and state political leaders engaged in the reconstruction and enlargement of the road network, mobilizing, to this aim, the private sector, public school teachers and rural communities all over the country (Bess, 2016a). Thousands of kilometers of highways were constructed. The 1940s and 1950s were characterized by a great hope in the promises of industrialization and a generalized drive toward economic modernization. Road construction was perceived as necessary to allow market growth and improve the accessibility of regions (Bess, 2014). This 1 i.e. Chihuahua, Coahuila, Nuevo Leon and Tamaulipas 3 conception led to the building of hundreds of kilometers of new roads in Mexico by state road-building agencies, mobilizing large public spending and private domestic investments as well as foreign ones. The U.S. invested millions of dollars of direct investments in Mexican transportation industry and infrastructure. During President Miguel Aleman’s term (1946-1952), the first freeway (from Mexico City to Acapulco) was opened and became a model for the construction of future freeways. During this period, road building played a key role in the modernization of the Mexican economy and the development of major commercial industries (Bess, 2014). In the 1960s, roads were built to respond to the needs of private firms and also to serve the national and state governments’ objective to build strategic relationships with rural commu- nities. Under President Adolfo Lopez Mateos (in office between 1958 and 1964), an unprece- dented amount of 300 million pesos was raised in bonds for the building of new highways (Bess, 2017). In addition, a government owned company was founded to build more than a thousand kilometers of toll roads in the center of Mexico. The presidency of Carlos Salinas de Gortari (1988-1994) constituted a significant milestone in the history of road construction in Mexico with the launch of a very ambitious program, the Programa Nacional de Solidaridad, which led to the construction of 5,800 kilometers of privately financed highways at a cost of $15 billion (Foote, 1997). At the same time, cutting its own spending on road infrastructure, the government privatized toll road operations, a lucrative business for road-building firms. This allowed Mexico to build, in just six years, ’what had taken two decades to achieve in western European countries’ (Foote, 1997). However, the extremely high tolls ended up deterring trucks from using these new roads. In response, the federal government announced, in January 1997, the mobilization of $3.3 billion over 30 years to restructure the highway network. This outlay was added to a $ 1.7 billion toll-road rescue plan to help the state-owned Mexican banks that had financed road building on non-market terms (Foote 1997). After 2000, the opposition party, the Partido de Acción Nacional, eventually gained power. Road building policies were nevertheless pursued in continuity with past policies, which con- sidered the construction of roads as a symbol of modernization. President Vicente Fox (in of- fice between 2000 and 2006) mobilized hundreds of millions of pesos for road building through a program for basic infrastructure, and his successor, President Felipe Calderón (in office be- tween 2006 and 2012), constructed and renovated more than 23,000 km of roads as part of a pro- gram to address rural poverty. Eventually, in April 29, 2014, the federal government launched the National Infrastructure Program2 2014-2018 (nip), projecting a substantial increase in in- vestment compared to the last twenty years (Pérez-Cervantes and Sandoval-Hernández, 2015). The most ambitious part of this new program focuses on the south of the country, while no large project is planned in the north and while the center of Mexico, more populated and richer, 2 Programa Nacional de Infraestructura. 4 receives smaller projects. 2.3 Roads and economic activities To our knowledge, the impact of roads on the location of economic activities in Mexico has not been studied yet. A set of studies, however, have investigated the link between infrastructure and productivity. For instance, Becerril-Torres et al. (2010) study the effect of total infrastruc- ture (measured with an index that includes roads, ports, airports, and telecommunications) on the convergence across states in technical efficiency. They find that their infrastructure index was associated with greater regional productivity, but mainly during the Import Sub- stitution Industrialization (isi) period (1970-1985), before Mexico entered the North American Free Trade Agreement (nafta) in 1994. Similarly, Brock and German-Soto (2013) find that lower levels of infrastructure investments had a lower effect on regional productivity during the nafta period, and conclude that continued investment in transportation will be necessary to boost industrial sector growth. Focusing specifically on the effect of road infrastructure on productivity growth in the manufacturing sector in Mexico, Duran-Fernandez and Santos (2014a and 2014b) conclude that road infrastructure had positive effects on productivity and on the average product of labor. 3 Data and descriptives statistics 3.1 Firms and employment We use panel data for employment and firm locations from two different sources published by the Mexican National Institute of Statistics (inegi).3 The first source is the Economic Census for the years 1986, 1989, 1994, 1999, 2004, 2009, and 2014, which provides employment figures and breakdown by broad sector of activity, as well as total income at the level of municipalities and for formal businesses. The second source of data is the Directory of Economic Units (denue)4 for the years 2004, 2009 and 2014. It is an exhaustive dataset which contains detailed micro- geographic information on all formal establishments, including the 6-digit naics classification, the number of employees of the establishment, turnover, the municipality identifier, and exact geographic coordinates. For the year 2014, the database contains information on approximately 3,000,000 establishments throughout the country. Note that although the denue database has finer geographic identifiers than the Economic Census (actual coordinates vs. municipality of location) and a finer industry identifier (6-digit naisc vs. broad sectors of activity), it covers less years than the Economic Census. In the analysis, we use either one, as appropriate. 3 Instituto Nacional de Estadísticas y Geografía. 4 Directorio Estadístico Nacional de Unidades Económicas. 5 Table 1 presents changes in the distribution of formal jobs in Mexico between 1999 and 2014 according to Economic Censuses. As can be seen, the bulk of jobs are in services, commerce and manufacturing. Although employment has increased in all sectors over the period, the greatest increases have occurred for commerce and services jobs, which relative shares have gone up while the relative shares of all the other sectors have gone down. The decrease in the relative share of manufacturing employment from 30 t0 24 percent is particularly noticeable. In 2014, more people were actually employed in the commerce sector than in manufacturing, showing that Mexico is following a trend of tertiarization similar to that of developed economies in previous decades. Table 1: Sectoral distribution of formal jobs in Mexico (1999 – 2014) Sector 1999 share 2009 share 2014 share Agriculture 174,127 1.27 180,083 0.90 188,566 0.87 Mining 113,189 0.83 14,2325 0.71 166,548 0.77 Manufacturing 4,1754,00 30.47 4,661,062 23.17 5,073,432 23.51 Commerce 3,792,466 27.68 6134758 30.50 6,389,648 29.61 Services 5,257,100 38.37 8,762,918 43.56 9,537,235 44.20 Other 190,033 1.39 235,688 1.17 220,929 1.02 Total 13,702,315 100.00 20,116,834 100.00 21,576,358 100.00 Source: Economic Censuces (inegi) 1999, 2009 and 2014. Figure 1 plots the location of establishments in the denue database in 2004 and 2014 for all firms (top panel) and for manufacturing firms (bottom panel) respectively.5 It reveals a high concentration of economic activity in the center of the country and in the surroundings of Mexico City, with a greater concentration of manufacturing firms than for the other sectors, but no visually noticeable change over the past decade. 3.2 Roads In this paper, we focus on the later waves of road construction and improvement that oc- curred since the mid-eighties. To have a consistent measure of road extent and road types over time, we constructed a geo-referenced database for the period 1985-2016. This was done by importing historical road type and road extent information from the American Automobile Association (aaa) paper maps into the 2014 road geometry published by DeLorme. Because the DeLorme road geometry is network-enabled and topologically correct, it allows for clean travel time computations needed to measure access to markets (see below). Using this fixed geometry, we then imported information on road type and extent for the reference years or 5 OnFigure 1, establishment locations are overlayed on roads. See the following subsection for details regarding the roads data construction. 6 Figure 1: The spatial distribution of firms in Mexico (2004 and 2014) Sources: DeLorme, aaa and denue (inegi). Note: Universe of all formal firm in 2004 (top left panel) and 2014 (top right panel). Manufacturing in 2004 (bottom left panel) and 2014 (bottom right panel). the nearest available year before which we have firm data in the denue and in the Economic Censuses (1986, 1994, 1999, 2004, 2009 and 2014), leaving us with a road panel for the years 1985, 1993, 1999, 2004, 2008 and 2016.6 We grouped the aaa road classes into four categories, which we label ‘Multilane divided’, ‘Two lanes or Divided’, ‘Pavement’, and ‘Gravel and Earth Road’. Figure 2 represents the evolution of the road network for the years 1985, 1999, 2008, and 2016. As can be seen, major road improvements took place in the middle of the studied period, between 1999 and 2008. This is confirmed by table 13 in Appendix A, which provides road length by road category for all six years during the 1985-2016 period.7 6 We use the roads information for 2016 as the aaa map for recent years before or in 2015 were not available to us. 7 Observe that official statistics provided by inegi for Mexico provide significantly greater total road lengths 7 Figure 2: The road network in Mexico (1985 – 2016) Source: DeLorme (2014) and authors’ calculations. Note: These maps represent a cross-sectional road geometry derived from DeLorme (2014) and updated by the authors using road category information from aaa maps (1985, 1999, 2008, 2016). 3.3 Geographic unit of analysis Because our focus is on accessibility and the location of economic activities, the natural geo- graphic unit of analysis is the urban or metropolitan area. However, because the data we use is produced at at a finer geographic scale—the municipality level—it is necessary to merge some municipalities into metropolitan areas. As inegi provides data for 2,377 municipalities8 and identifies 316 metropolitan areas—either as isolated municipalities or as groups of two municipalities—we are able to reconstruct a sample of 2,094 localities along a mixed classifi- cation that consists of the 316 metropolitan areas identified by inegi and the remaining 1,778 isolated municipalities that are not classified by inegi as metropolitan areas.9 In what follows, we use the term locality to refer to either an isolated municipality or a metropolitan area in our than what can be inferred from the aaa maps. This is because inegi data accounts for all types of roads, including minor segments, whereas the aaa maps focus on the main roads, which are likely to be more relevant for trading. 8 Since the 1989 census, new municipalities have been created by splitting some of the old municipalities. In order to work with a definition of localities that is stable across time, we merged these new municipalities back to their 1988 boundaries using the list provided by inegi (2006). 9 We consider here the 1990 definition of metropolitan areas. 8 sample. As a robustness check, all the regressions that we run on our mixed sample of 2,094 municipalities and metropolitan areas are also run on the sample of 2,377 municipalities. 3.4 Measures of accessibility Market access Each locality is characterized by its accessibility to markets and we resort to several acces- sibility measures. The first measure of accessibility for a locality i, which we refer to as market access (ma), is given by the following formula: M Ai,t = ∑ Pj ,tτij −θ ,t (1) j =i where Pj ,t is the population of locality j at time t (which proxies for the size of the local market in j ), τij ,t is the time required to travel between locality i and j given the state of the road network at time t, and θ is a measure of trade elasticity. From formula (1), it is easy to see that the market access indicator is the discounted sum of the populations of all the localities j that surround locality i, where the discount factor is inversely related to travel time. Travel times τij ,t are calculated on the reconstructed country- wide road network assuming that speed is a function of road type.10 As for the trade elasticity parameter, in the absence of a specific study for Mexico, we use the same value suggested by Donaldson (2016) in the case of India (θ = 3.8).11 Market access, which reflects the size of domestic markets accessible from location i, is frequently used in the literature in the absence of information on local incomes. Market potential The second measure of accessibility for a locality i, which we refer to as market potential (mp), is given by the following formula: Yj , t M Pi,t = ∑ T C σ −1 (2) j =i ij ,t 10 We use the following speed assumptions: Multilane divided: 90 km/h; Two lanes or Divided: 80 km/h; Pavement: 70km/h; Gravel or Earth Road: 40 km/h; and unknown category (i.e., not shown on aaa paper maps but present in the 2014 DeLorme geometry): 5 km/h. These speed assumptions are consistent with the travel times published by inegi for their own road geometry. 11 In the regressions presented in the next section, we also perform robustness checks by using alternative values of the market access indicator constructed with other values for θ, the maximum value encountered in the literature being θ = 8.2 (see Pérez Cervantes and Sandoval Hernández, 2015) 9 where Yj ,t is locality j ’s total income (in real terms)12 at time t, T Cij ,t is a transport cost function between localities i and j , and σ is an elasticity term. Since the market poten- tial formula has already been calibrated for Mexico, we use the same iceberg transport cost specification as the one estimated by Pérez Cervantes and Sandoval Hernández (2015), with T C ij ,t = e(.0557+.0024τij ,t ) for j = i and σ = 9. Similarly to formula (1), formula (2) measures the potential demand for goods traded from location i but in terms of income. Observe that in both formulas (1) and (2), we exclude ’own locality’ to reduce endogeneity concerns.13 Table 2 below details the mean of the market access and market potential indicators at the locality level over the study period. As one can see, there were large increases in both market access and market potential over the studied period. Table 2: Mean and median market access and market potential of localities (1986 – 2014) Market access Market potential Year Mean Median sd Mean Median sd 1986 24.41 0.00 389.89 – – – 1994 23.11 0.00 376.21 16,713.55 1,444.03 25,204.14 1999 24.90 0.00 393.58 445,492.00 55,543.83 617,044.60 2004 48.57 0.01 1,000.78 753,410.80 111,578.70 1,010,617.00 2009 53.09 0.01 1,093.79 1,271,577.00 220,128.80 1,680,713.00 2014 56.19 0.01 1,152.08 1,364,670.00 257,035.20 1,814,512.00 Sources: DeLorme, aaa and Economic Censuses (inegi). Notes: Mean and median market access and market potential values are calculated over the sample of localities in Mexico. On Figure 3, we represent market access and market potential at the locality level (over- layed on the road network) in 1986 and 2016. For comparison purposes, on both maps, the four categories correspond to the 1986 municipality quartiles of market access or market po- tential respectively. Although both maps show improvements along our two indicators, the contrast is starker with the market potential measure. This is understandably due to larger increases in real per capita income than in population. 12 We use the real gross production provided by the economic census at the municipality level to proxy for local income. 13 Also note that to increase accuracy of our measures, for metropolitan areas, we calculate market access and market potential by first calculating municipality level formulas (1) and (2) excluding all localities j that belong to the same metropolitan area as municipality i. We then compute a metropolitan area weighted average of these municipality indexes using municipality area as the weight. 10 Figure 3: Market access and market potential in Mexico Sources: DeLorme, aaa and Economic Censuses (inegi). Note: Market access in 1986 (top left panel) and 2014 (top right panel). Market potential in 1994 (bottom left panel) and 2014 (bottom right panel). Counts of road intersections and efficient road length In addition to market access and market potential measures, we also construct local in- dicators of road availability that do not depend on population or income but only on the characteristics of the roads network. We do this in two ways as in Baum-Snow et al. (2017). The first measure is the number of roads intersecting a circle of a given radius centered on a locality’s centroid. Another measure is the weighted length of roads within a circle—where the length of each road type is weighted by the corresponding speed—but excluding the roads in the locality itself.14 Table 12 in Appendix A shows the mean and median average efficient 14 The reason for this exclusion parallels the ’own locality’ exclusion in the market access and market potential 11 kilometers of roads within a circle of 200 km radius across localities for the 1986-2014 period. It can be seen that accessibility according to this variable has increased on average by 19.1 percent. Access to external markets Finally, to account for access to external markets, we also construct the minimum travel time and minimum travel cost (using the Pérez Cervantes and Sandoval Hernández, 2015 formula) to one among six major Mexican ports (documented by inegi) and to one among forty-four entry ports to the u.s. (documented by the u.s Department of Transportation, see us dot, 2014). Table 12 in Appendix A provides the mean and median minimum travel time and minimum travel cost to these ports and u.s. border entry ports. Over the 1986-2014 period, access to international markets slighlty improved on average as we measure a decrease in the average minimum travel time (by 8 percent) and in the average minimum travel cost (by 3 percent) to a u.s. border entry port. The same is true for the average minimum travel time and travel cost to a major port (which respectively decreased by 6.1 and 1 percent). 3.5 Trends in industrial concentration and specialization 3.5.1 Trends in industrial concentration: 2004 – 2014 We examine how industrial concentration evolved between 2004 and 2014. Although several measures are available in the literature, we resort to the Ellison and Glaeser concentration index (see Ellison and Glaeser, 1997), henceforth denoted eg. Using locality-level plant data and both the 4-digit and 6-digit naics industrial classifications, we calculate eg indexes for all industries and for manufacturing only, for the years 2004, 2009, and 2014 (see Appendix B for the formula and details). Table 3 reports our results, which leads to the three following comments. First, industries have become more geographically concentrated over the 2004-2014 period. This can be seen in the observed increase in the mean value of the eg index from 0.381 in 2004 to 0.430 in 2014 (a 10.3 percent increase). Second, manufacturing industries are on average more concentrated than overall industries in Mexico. The mean value of the manufacturing eg index is on average 12 to 15 percent higher than the mean value of the eg index calculated for all industries taken together. Third, applying commonly agreed upon thresholds in the literature, about 98-100 percent of industries (for overall industries and for manufacturing) can be said to be concen- trated. The fraction of concentrated manufacturing industries (98 percent) is greater than the one reported for Canada (75 percent) in Behrens and Bougna (2015) and is roughly similar to the one reported for the u.s. (97 percent), France (95 percent), and the u.k. (94 percent) formulas and is made to reduce endogeneity issues (see the econometrics section below). 12 in Ellison and Glaeser (1997), Maurel and Sédillot (1999), and Duranton and Overman (2005) respectively. Fourth, the average level of concentration increases as one moves from the 4 to the 6-digit industry classification (concentration is relatively more intense among more specific industry segments). This result is consistent with Rosenthal and Strange (2003) for the u.s. case and with Behrens and Bougna (2015) in the case of Canada. Table 3: Mean and median eg indices, naics 4- and 6-digit industries (2004 – 2014) 2004 2009 2014 Manufacturing Overall Manufacturing Overall Manufacturing Overall naics 6-digit Industries Mean 0.381 0.256 0.388 0.26 0.43 0.282 Median 0.284 0.146 0.297 0.156 0.319 0.154 Minimum -0.436 -0.739 0.045 -0.168 0.031 0.005 Maximum 1.001 1.006 1 1.006 1.353 1.013 Share < 0 1.23 0.95 – 0.27 – – Share ∈ (0, 0.05] 1.63 20.14 2.1 20 1.61 21.85 Share > 0.05 97.14 78.91 97.9 79.73 98.39 78.15 naics 4-digit Industries Mean 0.344 0.222 0.348 0.216 0.359 0.236 Median 0.264 0.116 0.273 0.12 0.285 0.118 Minimum -0.085 -0.225 0.033 0.004 0.02 0.005 Maximum 1 1 0.992 0.997 0.999 1 Share < 0 1.16 0.38 – – – – Share ∈ (0, 0.05] 2.33 25.95 3.49 24.23 3.49 25.84 Share > 0.05 96.51 73.66 96.51 75.77 96.51 74.16 Notes: Mean and median values for 248 (resp. 245 in 2004, and 238 in 2009) 4-digit and 6- digit naics classification of industries. Share< 0 means ‘not clustered’. Share∈ (0, 0.05] means ‘weakly clustered’. Share> 0.05 means ‘strongly clustered’. See Ellison and Glaeser (1997) for details. Figure 4 plots the distributions of the eg indices for the 6-digit industries in 2004, 2009, and 2014, for all industries (left panel) and for manufacturing only (right panel). As can be seen, these distributions are skewed towards 0.05, which shows that many industries are highly agglomerated, whereas only few of them are weakly agglomerated (the eg index is positive but smaller than 0.05). 3.5.2 Trends in specialization : 2004 – 2014 We now examine how specialization evolved between 2004 and 2014. For each localilty and for the years 2009 and 2014, we calculate the Krugman specialization index (henceforth denoted 13 Figure 4: Distribution of the eg Index (naics 6-digit) in 2004 (top panel), 2009 (middle panel), and 2014 (bottom panel). All industries (left panel) and manufacturing (right panel). EG index of all industries (6−digits) in 2004 EG index of manufacturing industries (6−digits) in 2004 200 60 150 Number of industries Number of industries 40 100 20 50 0 0 −1 −.5 0 .5 1 1.5 −1 −.5 0 .5 1 1.5 EG index EG index EG index of all industries (6−digits) in 2009 EG index of manufacturing industries (6−digits) in 2009 200 60 150 Number of industries Number of industries 40 100 20 50 0 0 −1 −.5 0 .5 1 1.5 −1 −.5 0 .5 1 1.5 EG index EG index EG index of all industries (6−digits) in 2014 EG index of manufacturing industries (6−digits) in 2014 200 60 150 Number of industries Number of industries 40 100 20 50 0 0 −1 −.5 0 .5 1 1.5 −1 −.5 0 .5 1 1.5 EG index EG index 14 KSI) at the 4-digit naics level, both for employment and output levels (see Appendix B for the formula and details). As standard in the literature, we consider as highly specialized localities for which the specialization index is greater than 0.75 and as not specialized, localities for which the specialization index is below 0.35. Our results, shown in Table 4, can be summarized as follows. First, localities are more specialized in terms of output than in terms of employment. The average output specialization index is on average 10 to 12 percent higher than the average employment specialization index. Second, there is an increasing trend toward specialization over the period, both in terms of output and employment specialization. Table 4: Mean and median specialization indices in Mexico : 2004 – 2014 2004 2009 2014 Employment Output Employment Output Employment Output Mean 0.117 0.126 0.155 0.174 0.199 0.209 Median 0.014 0.030 0.021 0.063 0.046 0.087 Minimum 0.000 0.000 0.000 0.000 0.000 0.000 Maximum 1.972 1.978 1.963 1.983 1.984 1.973 Share 0.35 91.750 90.72 88.90 88.20 85.20 84.73 Share ∈ (0.35, 0.75] 2.200 3.28 2.71 4.15 4.19 6.05 Share > 0.75 6.040 6.00 8.40 7.65 10.61 9.22 Notes: Mean and median values of the Krugman Specialization Index (ksi) for the 316 metropoli- tan area and 1,832 standalone municipalities in Mexico. ksi 0.35 means ‘not specialized’. ksi ∈ (0.35, 0.75] means ‘weakly specialized’. ksi > 0.75 means ‘highly specialized’. Figure 5 maps the employment and output specialization indices in 2004 and 2014. These maps show the trend towards increasing locality specialization as can be seen by the dots switching to yellow (weakly specialized) or red (highly specialized). To identify the factors that drive these changes in the employment and specialization of localities, we will resort to a multivariate analysis. 4 Econometric specification and identification issues 4.1 Econometric specification We take advantage of the panel nature of our data to estimate the effects of road improvements on the localization of economic activities and on the specialization of localities in Mexico. Our main empirical specification is the following ’market access’ model: int Ym,t = βM Mm,t + βC Mm,t Im,t + βC Cm,t + αt + µm + m,t (3) 15 Figure 5: Maps of employment and output locality specialization in Mexico (2004 – 2014) Sources: denue (inegi). Note: Employment specialization in 2004 (top left) and 2014 (top right). Output specialization in 2004 (bottom left) and 2014 (bottom right). where Y m, t is the dependent variable (employment or specialization index) for locality m, Mm,t is the market access or market potential of the same locality, Cm,t is a vector of time-varying locality characteristics (education, population, oil-reserves, and pre/post nafta period), Im,t is an interaction term (education and population, oil-reserves, pre/post nafta period, and capital city dummy, αt is a time dummy and µm is the location fixed effect, which absorbs time-invariant location characteristics. m,t is the error term. The coefficients of inter- est are βM and βC int , which account for the effect of domestic accessibility and of its interaction with controls. Note that because education (share of population with college, undergraduate, or graduate degree) and population size could be endogenous controls and interaction terms, we use instead dummy variables for localities that are above the initial median education and median population levels. 16 Specifically, we regress the log value of employment and/or of the specialization index of locality m and year t on the log of the market access (ma) or of the market potential (mp), so that our coefficients estimates can be interpreted as elasticities. Standard errors are clustered at the locality level to adjust for heteroskedasticity and within-locality correlation over time. The regression sample is a balanced panel of 2,094 localities with employment data for the years 1986, 1994, 1999, 2004, 2009, and 2014. The specialization and concentration indices are calculated at the establishment and locality levels. Since the latter micro-geographic data only span the 2004-2014 period, we will therefore restrict our analysis on specialization to this time period. In addition to the ’market access’ specification, we also estimate an ’infrastructure model’ where we substitute market access with our measure of efficient roads within a fixed radius surrounding a locality’s centroid. In that model, we also incluse a measure of access to inter- national markets (measured by the minimum time or travel cost to a port or port of entry into the u.s.).15 Table 5 presents the summary statistics for the key variables of our analysis and the overall, between and within standard deviations.16 4.2 Identification issues The panel structure of the data allows us to control for location-specific time-invariant fac- tors and general macroeconomic trends. However, we need to account for the three following problems: (i) omitted variables bias, (ii) the non-random road placement of roads, and (iii) the structural recursion problem with the use of market access regressors. Omitted variables bias. In our framework, the problem of omitted variables is mitigated by the panel structure of the data. We include locality fixed effects in the panel estimation, which absorb all time-invariant local characteristics such as initial wealth. Therefore, the estimated effects of market access and market potential cannot be attributed to time-invariant locality dif- ferences. This is an advantage over a cross-sectional framework, where unobservable location characteristics that explain outcomes could be correlated both with surrounding population or income and transport infrastructure that enter into the calculation of MA and MP indicators. Despite the use of panel data and the different controls included, our model is still po- tentially subject to problems of simultaneity bias or reverse causality. In this respect, the two main empirical challenges with our model specification are the non-random placement of roads 15 Itis not desirable to introduce our measures of access to external markets in the market access specification because of their potentially high correlation with our market access or market potential indicators. 16 The latter decomposition confirms the presence of time variation in our data (although most of the variation remains cross-sectional). 17 Table 5: Key variables and summary statistics Standard Deviation Variables Observations Mean Overall Between Within Total employment 12,564 6,812.303 90,112.890 85,792.800 27,619.900 Manufacturing employment 12,564 1,998.672 23,597.920 23,277.810 3,901.393 Commerce employment 12,564 2,149.258 26,322.070 25,274.430 7,369.452 Service employment 12,564 2,229.871 37,222.180 33,232.070 16,779.740 Mining employment 12,564 62.447 1,149.786 642.266 953.764 Market access 12,564 22.496 399.887 396.782 50.361 Market potential 10,470 365,276.400 573,321.300 430,809.700 378,379.700 Employment ksi 5,286 0.237 0.445 0.359 0.266 Output ksi 5,286 0.305 0.495 0.397 0.303 Education (literacy rate) 12,408 0.046 0.049 0.034 0.035 Population 12,564 47,149.370 445,092.100 442,553.500 48,283.380 Efficiency roads within 200 km 11394 12.440 0.419 0.411 0.081 Minimum travel time to port 12,544 6.885 0.551 0.544 0.090 Minimum travel time to border 12,544 6.300 0.543 0.531 0.113 Minimum travel cost to port 12,544 0.795 0.136 0.134 0.024 Minimum travel cost to border 12,544 0.648 0.121 0.119 0.020 Note: the above descriptive statistics are calculated on a sample of 12,564 observations covering 2,094 local- ities over 6 different dates. The standard deviation is decomposed into between and within components, which measure the cross sectional and the time series variation, respectively. and a recursion problem inherent with regressions of local employment or income variables on market access or market potential measures. We discuss below how we address these issues. Non-random road placement and the construction of the ‘doughnut’ IV. The expansion of the road network surrounding a locality can be endogenous and this may create a correla- tion between increases in local market access or local market potential and our left-hand side variable, for instance employment. This may occur if road construction occurs in localities that would otherwise have experienced relative increases or decreases in average employment (our review of road investments in Mexico suggests that roads may have been built to increase acces- sibility in lagging regions during the studied period). To deal with that problem, we adopt the so-called ’doughnut’ instrumentation strategy as proposed by Jedwab and Storeygard (2017). The basic idea behind this approach is to instrument the market access or market potential with similar measures excluding surrounding areas within a fixed radius (in addition to ‘own locality’ exclusion). Variations in this instrument are more likely to be exogenous. Following the identification strategy of Jedwab and Storeygard (2016) also implemented by Blankespoor et al. (2017), we construct the counterfactual measures of market access and market potential 18 excluding all localities j located within a 25, 50 or 75 km radius of locality i, hence the reference to a ‘doughnut’.17 The ’recursion problem’ or the structural endogeneity of market access and market po- tential. When considering an explained variable which is correlated with other variables that enter the calculation of the accessibility regressor in formulas (1) or (2)—for instance when regressing total employment on market access in Equation (3)—a recursion problem is bound to emerge, even when excluding ‘own locality’ from the measure of accessibility. To see this, notice that locality i’s market access is a function of locality j ’s population, which in turn is correlated with locality j ’s employment, which is a function of locality j ’s market access, and thus of locality i’s population. Locality i’s employment is then structurally correlated with its own market access in the absence of any causal effect of market access on employment. Fol- lowing Baum-Snow et al. (2017), we address this problem by calculating either counts of road intersections or a ‘doughnut’ of efficient road lengths in a given radius (in a similar way to the market access and market potential ‘doughnuts’).18 We then use the road count or efficient road variable instead of the market access or market potential indicator, or use it to instrument market access or market potential. Sources of variations in ‘accessibility’ are then only due to variations in roads, thus avoiding the recursion problem. The validity of our three instruments (both for market access and market potential) is tested by means of the Kleibergen-Paap statistics, which is robust in the presence of heteroskedas- ticity. After performing this test, we reject the hypothesis of weak instruments as the rule of thumb suggested by Staiger and Stock (1997) that the statistic must be larger than 10 is satis- fied. As for the exogeneity of the instruments, we rely on the Hansen-J statistics and strongly accept the hypothesis of exogenous instruments.19 The coefficients on market access and mar- ket potential are positive and statistically significant in all second stage regressions. To differentiate the role of access to domestic and external markets, we also estimate the infrastructure model as previously discussed. In this model, domestic market access is proxied 17 Because of the mixed nature of our sample, for metropolitan areas, we first compute these counterfactual measures at the municipality level, excluding both municipalities within a same metropolitan area and munici- palities within our distance threshold. We then define the metropolitan area ‘doughnut’ as the weighted average of these municipality indices, using municipality area as weights. 18 In this ‘doughnut’, the exclusion of surrounding areas addresses endogeneity of road placement. 19 See Baum et al. (2017) for a detailed explanation of test implementation in stata and for references. It is worthwhile noticing that, besides being efficient, our estimation results are also consistent with respect to the absence of heteroskedasticity or autocorrelation because of the Stock-Yogo weak identification employed for the estimation. In all our iv regressions, the first stage f-statistic is large, suggesting that the instrument is strong (Stock and Yogo, 2005). 19 by efficient road length20 and access to external markets by the minimum travel cost or travel time to a major port and to the u.s. border. One key advantage of this approach is that the infrastructure measure (efficient road length) does not have a structural dependence on population (like the market access) or on income (like the market potential). However, this infrastructure measure may still suffer from non-random placement. To deal with this problem, we use as an instrument the ‘doughnut’ road efficient length. As a robustness check, we also report the results from cross-section regressions where the efficient road length of the corresponding year is instrumented with the 1949 efficient road length.21 5 Estimation Results We exploit changes in the transportation network over the 1986 – 2014 period to estimate the effect of market access and market potential on localities employment and specialization in Mexico. 5.1 The effect of market access and market potential on employment Table 6 reports estimates from the regression of employment on market access, in which our market access measure is instrumented in three ways: using the doughnut market access, the road count within a 10 km circle, and the efficient road length within a 100km radius. The first column (ma only) reports our basic estimates without any controls. The second column reports the ols estimates of the full model, and the last three columns report result of the instrumented regression with each one of our three instruments. As can be seen from the in- strumented regressions, there is a positive and significant causal effect (iv) of market access on employment. Note that first stage coefficients (not shown) are positive and significant, which means that each market access measure is predicted by the appropriate instrument.22 Results indicate that between 1986 and 2014, a 10 percent increase in market access resulted in a 1.6 to 2.1 percent increase in employment. Note that we use as interaction terms, population and education variables (dummies equal to one if the locality has population or education above the median of the initial 1986 levels for the market access regression or above the median of the initial 1994 levels for the market potential regression). Also note that more urbanized locations benefited more from an increase 20 Efficient road length is calculated for different radii: 50, 75, 100, 150, 200, and 300 km. 21 1949 is the earliest date for which we could recover information on road extent and road types from paper maps. This historic measure can be argued be exogenous to economic outcomes observed more than 40 years later. 22 These first stage results are available upon request. 20 in market access, while areas with a more educated population benefited less from an increase in market potential. This suggests a stronger impact of accessibility in larger places with lower skilled workers. The estimates also provide evidence that the positive impact of improved domestic market access on employment was partially attenuated after nafta came into effect (the net impact is still positive though). One explanation of this result could be that under nafta, access to international markets started playing a relatively more important role, especially the northern localities given their proximity to the u.s.. Therefore, our result may point towards a substi- tution effect of international market access to domestic market access, an issue we explore in further detail below, in the section describing the infrastructure regressions. Finally, it is also noticeable that improvements in market access had large beneficial and statistically significant effects on localities with oil reserves (mostly located in the South). Table 6: The effect of market access on employment Total employment Variables ma only ols iv (Doughnut) iv (Road count) iv (Efficient roads) Market Access (ma) 0.176*** 0.149*** 0.208*** 0.163*** 0.170*** (0.0153) (0.0111) (0.0144) (0.0202) (0.0219) ma x Metropolitan dummy 0.0717*** 0.0528*** 0.0629*** 0.0574*** (0.0117) (0.0146) (0.0154) (0.0164) ma x Education -0.0468*** -0.0486*** -0.0474*** -0.0452*** (0.00334) (0.00403) (0.00329) (0.00339) ma x nafta dummy -0.114*** -0.119*** -0.115*** -0.117*** (0.00233) (0.00251) (0.00213) (0.00219) ma x Capital city -0.0225 -0.0294 -0.0228 -0.0237 (0.0226) (0.0390) (0.0250) (0.0251) ma x Oil dummy 0.839** 0.756* 0.832** 0.809** (0.385) (0.403) (0.399) (0.401) Above median education 0.155*** 0.124** 0.155*** 0.146*** (0.0581) (0.0550) (0.0522) (0.0508) Above median population 0.851*** 0.736*** 0.805*** 0.779*** (0.0803) (0.0823) (0.0918) (0.0959) Constant 6.836*** 5.790*** (0.0844) (0.0755) Observations 11379 11379 9778 11251 10293 Adj. R-squared 0.052 0.423 0.398 0.424 0.440 Note: * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. The doughnut iv is calculated by excluding all localities within a 25km circle. The road count IV is the number of roads intersecting a circle of 10km radius. The efficient roads doughnut iv is obtained by calculating the road efficiency units in a doughnut within two circles of 25 and 100 km radii. Table 7 reports estimates from the regression of employment on market potential, under ols and iv. Overall, we find a positive and significant causal effect of market potential on employment. The first column (mp only) reports our basic estimate without any controls and 21 shows a positive and statistically significant estimated coefficient: a larger market potential is associated with greater employment. Instrumentation shows that the effect remains large: A 10 percent increase in market potential is associated with a 2.9 to 6.5 percent increase in total employment. Interestingly, more urbanized locations benefited less from increases in market potential, whereas areas with more educated population benefited more. These results are at odds with what we found in the market access regression. This indicates that market access and market potential are not completely substitutable. One potential channel to explain the positive market potential impact on employement growth in more skilled localities may be that greater market potential (which is a function of surrounding incomes) increases the demand for goods which have stronger skill contents. Table 7: The effect of market potential on employment Total employment Variables mp only ols iv (Doughnut) iv (Road count) iv (Efficient roads) Market Potential (mp) 0.192*** 0.351*** 0.322*** 0.292** 0.645*** (0.00324) (0.0401) (0.0484) (0.148) (0.0342) mp x Metropolitan dummy -0.0180*** -0.00188 -0.00954 -0.0680*** (0.00652) (0.00822) (0.0232) (0.0144) mp x Education 0.0168*** 0.0188*** 0.0195*** 0.000723 (0.00207) (0.00286) (0.00705) (0.00270) mp x nafta dummy -0.0591*** -0.0557*** -0.0424 -0.141*** (0.0127) (0.0140) (0.0423) (0.00928) mp x Capital city -0.00822* -0.0113** -0.0108** -0.0109* (0.00430) (0.00450) (0.00427) (0.00646) mp x Oil dummy 0.0227 0.0168 0.0248 0.0135 (0.0157) (0.0190) (0.0172) (0.0229) Above median education 0.187*** 0.151*** 0.185*** 0.203*** (0.0417) (0.0404) (0.0374) (0.0359) Above median population 0.406*** 0.219** 0.302 0.973*** (0.0872) (0.101) (0.267) (0.171) Constant 3.749*** 2.134*** (0.0382) (0.334) Observations 10229 10229 8168 10098 9248 Adj. R-squared 0.463 0.503 0.511 0.503 0.471 Note: * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. The doughnut iv is calculated by excluding all localities within a 25km circle. The road count IV is the number of roads intersecting a circle of 10km radius. The efficient roads doughnut iv is obtained by calculating the road efficiency units in a doughnut within two circles of 25 and 100 km radii. 5.2 Spatial and sectoral heterogeneous effects Market access and market potential have heterogeneous effects on employment in different sectors and different economic regions. We report the estimated results for the three main sectors, namely services, manufacturing, and commerce. The results from the iv estimations 22 are reported in Table 8. The first three columns show the estimated coefficients for market potential and the last three columns for market access. Compared to other sectors, it is the services sector that benefits most from an increase in market access or market potential.23 A 10 percent increase in market potential is associated with 1.2 percent, 3.0 percent, and 4.1 percent increases in employment in the manufacturing, commerce, and services sectors respectively. Similarly, a 10 percent increase in market access is associated with 1.6 percent, 1.8 percent, and 2.5 percent increases in employment in manufacturing, commerce, and services respectively. Results also show that more urbanized or more educated localities benefit more from an increase in market potential, and less from an increase in market access.24 Finally, localities seem to benefit less from an increase in domestic market access or domestic market potential during the nafta period. As mentioned earlier, the net effect is positive, and the result could be explained by the increasingly important role of international market access at the expense of domestic market access under trade openness. Table 8: Sectoral heterogeneous effects of market access and market potential Market Potential: Employment Market Access: Employment Variables Manufacturing Commerce Services Manufacturing Commerce Services Market Potential or Market Access 0.119* 0.303*** 0.413*** 0.158*** 0.176*** 0.246*** (0.0623) (0.0418) (0.0807) (0.0184) (0.0125) (0.0807) mp or ma x Metropolitan dummy 0.0332** 0.00657 -0.0148 0.0363* 0.0711*** -0.0148 (0.0139) (0.00681) (0.0161) (0.0193) (0.0133) (0.0161) mp or ma x Education 0.0266*** 0.0195*** 0.0217*** -0.0399*** -0.0407*** 0.0217*** (0.00451) (0.00241) (0.00518) (0.00590) (0.00342) (0.00518) mp or ma x nafta dummy -0.0146 -0.0596*** -0.0744*** -0.0850*** -0.101*** -0.0744*** (0.0173) (0.0122) (0.0221) (0.00356) (0.00206) (0.0221) mp or ma x Capital city -0.00819* -0.0117*** -0.00924 -0.0470** -0.0121 -0.00924 (0.00427) (0.00379) (0.00568) (0.0228) (0.0363) (0.00568) mp or ma x Oil dummy -0.0200 0.0163 -0.000259 -0.391 0.683*** -0.000259 (0.0237) (0.0253) (0.0276) (0.485) (0.205) (0.0276) Above median education 0.197*** 0.125*** 0.152*** 0.234*** 0.0795* 0.152*** (0.0648) (0.0295) (0.0492) (0.0714) (0.0441) (0.0492) Above median population -0.299* 0.166* 0.438** 0.513*** 0.829*** 0.438** (0.171) (0.0863) (0.192) (0.102) (0.0779) (0.192) Observations 7839 8144 7726 9165 9747 7726 Adj. R-squared 0.168 0.534 0.412 0.162 0.403 0.412 Note: * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. The estimates are from the instrument regression with the doughnut iv calculated by excluding all localities within a 25km circle. Results are robust to others iv strategies (road count and efficient road units). Table 14 and Table 15 in Appendix C show the results for the instrumented regressions of market access and market potential on total employment for different subsamples stratified by region (border, capital, central, North and South). As can be seen, all things else equal, the North and the capital states benefit from greater impact of improved market access or im- 23 The services sector accounts for about 60 percent of Mexico gdp 24 This result is consistent with the use of a different human capital proxy i.e., locality literacy. 23 proved market potential. Interestingly, there is no significant effect of an increase in the market potential in the North, while the effect is important in the South (which is less developed).25 This result thus sheds light on the potential benefit to enhance economic opportunities through investment in transport infrastructure in lagging regions. It is also in line with the evidence that Southern Mexico has a greater potential to generate positive spatial spillovers (Deichmann et al., 2004, Baylis et al., 2012 and Alvarez et al., 2017). 5.3 The effect of market access and market potential on specialization Table 9 presents the causal effect of market potential and market access on locality specializa- tion. The first four columns present the impact on specialization in terms of output, whereas the last four columns present the impact on specialization in terms of employment. A 10 percent increase in market access causes a 7 percent increase in output specialization and a 3.4 percent increase in employment specialization. In contrast, we find no significant impact of market potential on specialization, neither in terms of employment nor in terms of output. More urbanized locations were not differently affected than less urbanized locations. Table 9: The effect of market access and market potential on specialization Krugman Specialization index: Output Krugman Specialization index: Employment Market Potential Market Access Market Potential Market Access ols iv (Doughnuts) ols iv (Doughnuts) ols iv (Doughnuts) ols iv (Doughnuts) Market Potential or Market Access 2.913*** 0.154 0.455*** 0.704*** 1.884*** 0.984 0.321** 0.339** (0.494) (1.863) (0.176) (0.231) (0.436) (0.815) (0.128) (0.160) mp or ma x Metropolitan dummy -2.245*** 0.464 -0.127 -0.286 -1.433*** -0.590 -0.0410 -0.0833 (0.507) (1.772) (0.146) (0.187) (0.441) (0.820) (0.0942) (0.0995) mp or ma x Education 0.0694 0.172** -0.259*** -0.291*** 0.0752* 0.122** -0.269*** -0.214** (0.0448) (0.0774) (0.0768) (0.0919) (0.0432) (0.0528) (0.0798) (0.105) mp or ma x Capital city 0.00864 0.00877 0.0458* 0.0837*** -0.00180 -0.00183 0.0955*** 0.102** (0.00871) (0.00988) (0.0239) (0.0304) (0.0138) (0.0135) (0.0170) (0.0439) mp or ma x Oil dummy -0.423 -0.577 1.009 0.950 -0.243 -0.233 1.785 1.756 (0.494) (0.470) (2.109) (1.945) (0.458) (0.538) (1.968) (2.478) Above median education 1.313*** 1.264*** 1.214*** 1.286*** 0.492 0.414 0.231 0.508 (0.353) (0.431) (0.390) (0.418) (0.403) (0.424) (0.423) (0.390) Above median population 29.22*** -4.382 1.503 0.994 18.32*** 7.713 0.780* 0.476 (6.422) (22.33) (1.017) (0.966) (5.581) (10.35) (0.472) (0.447) Constant -42.77*** -4.412*** -29.06*** -4.017*** (6.165) (0.985) (5.454) (0.504) Observations 4599 3628 4303 3628 4491 4149 4203 3556 Adj. R-squared 0.0649 0.0498 0.0234 0.0233 0.0385 0.0381 0.0200 0.0106 Note: * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. The estimates are from the instrument regression with the doughnut iv calculated by excluding all localities within a 25km circle. Results are robust to others iv strategies (road count and efficient road units). 25 Despiteimprovements, the per capita income in the northern states is two or three times higher than in the southern states, and the disparities in terms of other social and infrastructure indicators are even more dramatic. 24 5.4 The effect of access to infrastructure on employment and specialization We now turn to the estimation of the direct impact of local road improvements using measures of access to infrastructure that do not involve population or income. The focus on infrastruc- ture only allows us to avoid the structural endogeneity problem of market access and market potential regressors. In these regressions, we also introduce measures of access to international markets to contrast the effects of local and international accessibility. Following Baum-Snow et al. (2017) and Emran and Hou (2013), we regress our dependent variable—employment or specialization—on both measures of accessibility. For the local access variable, we consider our measure of road efficiency units. For access to external markets, we calculate two comple- mentary measures: the minimum travel time or the minimum travel cost to a major port and to the u.s. border.26 To correct for the non-random placement of roads, we also instrument local accessibility by applying the doughnut approach (i.e., we instrument our road efficiency measure with a doughnut of efficient roads). Finally, we also run a cross-section regression where our infrastructure metric is instrumented by the 1949 efficient road length. As in the previous section, first stage estimated coefficients are positive and significant, which means that our efficient road length measure is predicted by the appropriate instru- ment. Table 10 reports the estimates in the regression of employment on road length efficiency, where efficient road length is instrumented using the doughnut approach. Better local access to roads has a positive and statistically significant effect on employment. These effects are large: a 10 percent increase in efficient road length leads to a 6.7 percent increase in total employment. There are also heterogeneous effects across sectors: the increase in employment in the commerce sector (7.1 percent) is greater than in the services sector (6.0 percent) or in manufacturing (4.3 percent). As for access to external markets, a decrease in travel costs to the nearest u.s. border is pos- itively associated with an increase in employment in each sector. We find that a 1 percent reduction in travel costs to the nearest u.s. border leads to a 1.9 percent increase in manufac- turing employment, a 1.2 percent increase in commerce, and a 0.7 increase in services. These results are supported by the fact that the United States is, by far, Mexico’s leading partner in merchandise trade, and most of the Mexico-u.s. trade is through trucking. Note, however, that the impact of a reduction in travel costs to the nearest port is not positively associated with an increase in employment but with a decrease (for manufacturing in particular). An explanation could be that reducing travel costs to the u.s. facilitates exports and attracts em- ployment, whereas reducing travel costs to ports exposes industries to more competition and has a depressing impact on employment, especially in the manufacturing sector. Interestingly, when interacting with the nafta dummy (see Table 16 in Appendix C), we find that these two 26 Minimum travel cost is calculated using the same calibrated trade cost function as in our market potential formula. 25 opposite effects are magnified, consistently with the idea that trade openness results in local benefits for places that export but negative impacts for places that are exposed to competition with imported goods. Table 10: The effect of road infrastructure on employment Total employment Manufacturing Services Commerce Variables ols iv ols iv ols iv ols iv Road Efficiency 5.501*** 6.633*** 3.570*** 4.301*** 5.007*** 6.044*** 5.833*** 7.061*** (0.0941) (0.0911) (0.144) (0.141) (0.0793) (0.0719) (0.110) (0.104) Min travel cost to border -3.271*** -0.404 -3.808*** -1.887*** -3.354*** -0.730** -4.242*** -1.222*** (0.431) (0.404) (0.669) (0.634) (0.363) (0.318) (0.470) (0.437) Min travel cost to port 1.522*** 0.493 2.242*** 1.509** 1.683*** 0.743** 1.567*** 0.573 (0.467) (0.406) (0.737) (0.660) (0.383) (0.333) (0.527) (0.468) Above median education 0.224*** 0.238*** 0.339*** 0.348*** 0.159*** 0.171*** 0.130*** 0.145*** (0.0460) (0.0416) (0.0681) (0.0630) (0.0363) (0.0325) (0.0475) (0.0438) Above median population 0.427*** 0.404*** 0.314*** 0.304*** 0.415*** 0.394*** 0.420*** 0.396*** (0.0448) (0.0405) (0.0782) (0.0666) (0.0419) (0.0363) (0.0575) (0.0510) Constant -61.14*** -38.67*** -55.82*** -65.98*** (1.341) (1.978) (1.139) (1.609) Observations 11200 11200 10619 10614 11184 11184 10583 10575 Adj. R-squared 0.549 0.537 0.210 0.206 0.607 0.594 0.497 0.486 Note: * denotes significance at the 10% level, ** denotes significant at the 5% level, and *** denotes significance at the 1% level. The efficient roads doughnut iv is obtained by calculating the road efficiency units in a doughnut within two circles of 150 and 300 km radii. Finally, looking at the effects of road infrastructure on specialization, Table 11 shows that better local accessibility has a positive and statistically significant causal effect on specializa- tion. A 1 percent increase in efficient road length leads to a 5.9 percent increase in the locality’s output specialization, and a 4.5 percent increase in the locality’s employment specialization. Interestingly, reducing travel costs to u.s. borders or ports has not significant impact on spe- cialization, neither in terms of output nor in terms of employment. 26 Table 11: The effect of road infrastructure on specialization Output Specialization Employment Specialization Variables ols iv ols iv Road Efficiency 4.092*** 5.935*** 2.735*** 4.468*** (0.906) (1.261) (0.829) (1.187) Minimum travel cost to border -2.089 0.340 -2.332 -0.0469 (3.075) (3.217) (2.399) (2.694) Minimum travel cost to port 6.172* 5.003 4.489* 3.382 (3.201) (3.354) (2.531) (2.707) Above median education 0.907** 0.917** 0.471** 0.480** (0.415) (0.394) (0.228) (0.234) Above median population 0.977** 0.953** 1.033** 1.012** (0.415) (0.395) (0.476) (0.432) Constant -57.43*** -39.02*** (11.79) (10.84) Observations 4671 4517 4578 4410 Adj. R-squared 0.012 0.011 0.011 0.001 Note: * denotes significance at the 10% level, ** denotes significant at the 5% level, and *** denotes significance at the 1% level. The efficient roads doughnut iv is obtained by calculating the road efficiency units in a doughnut within two circles of 150 and 300 km radii. 6 Conclusion Transport investments have the potential to stimulate growth through trade, structural trans- formation, agglomeration, and productivity. This paper uses extensive micro-geographic data and geo-referenced digitized maps of the transport network in Mexico over three decades to provide empirical evidence of the causal effect of improved access to markets following road improvements on local employment and specialization. In addition to shedding light on the effects of transport infrastructure (roads) on local economic development, this paper also pro- vides evidence of the respective roles of domestic and external market access on the geography of economic activities. To estimate the causal effect of roads on localities’ economic activity and their level of specialization, this paper addresses two important endogeneity issues, namely the potentially non-random placement of roads and the structural endogeneity of market access and market potential regressors, also called the recursion problem. For this, we adopt three separate iden- tification strategies. We first use the so-called ’doughnut’ iv strategy to address the common non-random placement of roads problem, by excluding surrounding localities up to 25 km. Second, we address the structural endogeneity of market access and market potential by us- ing two raw measures of access to infrastructure (road count and road efficient length) that do not involve population or income (either as substitutes for market access indicators or as instruments). Our empirical exercise sheds light on the causal effect of road improvements on the increas- 27 ing trends in specialization and geographical concentration of employment in Mexico, and confirms that transport investments can have large effects on local economic activity through improved accessibility. The quantitative effects are large. A 10 percent increase in market ac- cess results in a 1.6 to 2.1 percent increase in total employment, a 7 percent increase in output specialization, and a 3.4 percent increase in employment specialization. In a country where the share of the manufacturing sector has been declining, we find heterogeneity in sectoral effects, with employment in commerce and services benefiting more from road improvements than employment in manufacturing. We also provide evidence of the potential benefits of the investments in transport infrastructures in lagging regions. There is no significant effect of an increase in the market potential in the North, while the effect is large and important in the South. Results also show that localities in the South are the most specialized and the large ef- fort of the Mexican government to increase accessibility in that region has played an important role. Finally, we show that access to external markets matter but in an heterogeneous way. Re- ducing travel costs to u.s. borders stimulates employment, whereas reducing travel costs to ports depresses employment, especially since nafta came into effect. This is consistent with both theories of exports to the u.s. stimulating local development, and trade openness exposes localities to competition from imports. Further investigation, however, will be needed to understand the possible mechanisms (ter- tiarisation of the economy, expansion of trade, or foreign direct investment) for the increase in specialization and employment caused by Mexico’s road improvement. For example, we do not know whether improved connectivity (through new or better roads) plays a role through improved accessibility to output markets, to intermediate inputs, to the pool of workers, or simply through reduction in travel times. We leave these important issues for further research. 28 References [1] Abrams, A., Lozano-Gracia, N., and C. Oram. 2016. Deblluring Nightime Lights. Unpub- lished manuscript. [2] Addison, D. and B. Stewart. 2015. Nighttime Lights Revisited: the Use of Nighttime Lights Data as a Proxy for Economic Variables. World Bank Policy Research Working Paper 7496. [3] Alder, S. 2017. Chinese Roads in India: The Effects of Transport Infrastructure on Eco- nomic Development. Unpublished manuscript. [4] Alder, S., Roberts, M. and M. Teewari. 2017. The Effect of Transport Infrastructure on India’s Urban and Rural Development. Unpublished manuscript. [5] Alvarez, A., Garduno-Rivera, R., and Nunez, H. M. 2017. Mexico’s North-South divide: The regional distribution of state inefficiency 1988 – 2008. Papers in Regional Science. [6] Anas, A. and K. Xiong. 2003. Intercity Trade and the Industrial Diversification of Cities, Journal of Urban Economics, 54, 258-276., [7] Asher, S. and P. Novosad. 2016. Market Access and Structural Transformation: Evidence from Rural Roads in India. Unpublished manuscript. [8] Becerril-Torres, O., Àlvarez-Ayuso, I. and L. Del moral-Barrera. 2010. Do Infrastructures Influence the Convergence of Efficiency in México? Journal of Policy Modeling, 32(1), 120- 137. [9] Baum-Snow, N., 2007. Did Highways Cause Suburbanization? The Quarterly Journal of Economics, 122(2), pp.775-805. [10] Baum-Snow, N., 2017. Urban Transport Expansions, Employment Decentralization, and the Spatial Scope of Agglomeration Economies. Unpublished paper. [11] Baum-Snow, N., Henderson, J.V., Turner, M.A., Zhang, Q., and L. Brandt. 2017. Roads, Railroads, and Decentralization of Chinese Cities. Review of Economics and Statistics 99(3), 435-448. [12] Baylis KR, Garduno-Rivera R, Piras G. 2012. The distributional effects of nafta in Mexico: Evidence from a panel of municipalities. Regional Science and Urban Economics. 42: 286 – 302. [13] Bazant, J. 1977. A Concise History of Mexico from Hidalgo to Cardenas 1805-1940. New York: Cambridge University Press. 29 [14] Becerril-Torres, O. , Álvarez-Ayuso, I. and L. del Moral-Barrera. 2010. Do Infrastructures Influence the Convergence of Efficiency in México?, Journal of Policy Modeling, 32(1), 120- 137. [15] Behrens, K., Bougna, T. and M.W. Brown. 2015. The World is not yet Flat: Transportation Costs Matter!’ CEPR Discussion Papers, DP 10356. [16] Behrens, K. and T. Bougna. 2015. An Anatomy of the Geographical Concentration of Cana- dian Manufacturing Industries, Regional Science and Urban Economics, 51, 47-69. [17] Berg, C., Deichmann, U., Liu, Y. and H. Selod. 2017. Transport policies and development, Journal of Development Studies, 53(4), 465-480, 2017. [18] Bess, M. 2014. Routes of Conflict : Building Roads and Shaping the Nation of Mexico, 1941-1952, The Journal of Transport History, 35(1), 78-96. [19] Bess, M. 2016a. ’Neither Motorists nor Pedestrians Obey the Rules : Transit Law, Public Safety, and the Policing of Northern Mexico’s Roads, 1920s-1950s." The Journal of Transport History, 37(2), 155-174. [20] Bess, M. 2016b. Revolutionary Paths : Motor Roads, Economic Development, and National Sovereignty in 1920s and 1930s Mexico. Mexican Studies / Estudios Mexicanos, 32(1), 56- 82. [21] Bess, M. 2017. Routes of Compromise: Building Roads and Shaping the Nation of Mexico, 1917- 1952. University of Nebraska Press, 246 pages. [22] Bird, J. and S. Straub. 2014. The Brasilia Experiment: Road Access and the Spatial Pattern of Long-Term Local Development in Brazil. World Bank Policy Research Paper, WPS 6964. [23] Blankespoor, B., Mesplé-Somps, S., Selod, S. and G. Spielvogel. 2017. Roads and Structural Transformation in Mali. Unpublished paper. [24] Bosker, M., Deichman, U. and M. Roberts. 2015. Hukou and Highways: The Impact of China’s Spatial Development Policies on Urbanization and Regional Inequality. World Bank Policy Research Working Paper WPS7350. [25] Bougna, Théophile. 2016. The Determinants of Localization: A Conditional Distance- Based Approach. Unpublished paper. [26] Brock, G. and B. German-Soto. 2013. Regional Industrial Growth in Mexico: Do Human Capital and Infrastructure Matter? Journal of Policy Modeling, 35(2), 228-242. 30 [27] Brulhart, M. and F. Sbergami. 2009. Agglomeration and Growth: Cross-Country Evidence, Journal of Urban Economics, 65, 48-63. [28] Calderón, C., Moral-Benito, E., and L. Servén. 2015. Is Infrastructure Capital Productive? A Dynamic Heterogeneous Approach, Journal of Applied Econometrics, 30(2), 177-198. [29] Calderón, C. and L. Servén. 2010. Infrastructure and Economic Development in Sub- Saharan Africa, Journal of African Economies, 19, aerc Supplement 1, i13-i87. [30] Ciccone, A. 2002. Agglomeration Effects in Europe, European Economic Review, 46(2): 213- 227. [31] Dávila A. 2008. Los clústeres industriales del noreste de Mèxico (1993-2003). Perspecti- vas de desarrollo en el marco de una mayor integración económica con Texas?, Región y Sociedad, 20(41), 57-88. [32] Deichmann, U., Marianne, Fay, Jun, Koo, Somik V. Lall. 2004. "Economic structure, pro- ductivity, and infrastructure quality in Southern Mexico", The Annals of Regional Science, 38:361?385. [33] Donaldson, D. 2016. Railroads of the Raj: Estimating the Impact of Transportation Infras- tructure, American Economic Review, forthcoming. [34] Donaldson, D. and R. Hornbeck. 2016. Railroads and American Economic Growth: A Market Access Approach, Quarterly Journal of Economics, 131(2), 799-858. [35] Duran-Fernandez, R., and G. Santos. 2014a. Road Infrastructure Spillovers on the Manu- facturing Sector in Mexico, Research in Transportation Economics, 46, 17-29. [36] Duran-Fernandez, R., and G. Santos, G. 2014b. Regional Convergence, Road Infrastruc- ture, and Industrial Diversity in Mexico, Research in Transportation Economics, 46, 103-110. [37] Duranton, G. and H. Overman. 2008. Exploring the Detailed Location Patterns of U.K. Manufacturing Industries Using Micro-Geographic Data,Journal of Regional Science, 48(1), 213-243. [38] Duranton, G. and H. Overman. 2005. Testing for Localization Using Micro-Geographic Data, Review of Economic Studies, 72(4), 1077-1106. [39] Duranton, G. and M. Turner. 2012. Urban Growth and Transportation, Review of Economic Studies, 79(4), 1407-1440. [40] Duranton, G., Morrow, P. and M. Turner. 2014. Roads and Trade: Evidence from U.S. Cities, Review of Economic Studies, 81(2), 681-724. 31 [41] Ellison, G. and E. Glaeser. 1997. Geographic Concentration in U.S. Manufacturing Indus- tries: A Dartboard Approach, Journal of Political Economy, 105(5), 889-927. [42] Ellison, G., Glaeser, E. and W. Kerr. 2010. What Causes Industry Agglomeration? Evidence from coagglomeration patterns, American Economic Review, 100(3), 1195-1213. [43] Emran, M.S. and Z. Hou. 2013. Access to Markets and Rural Poverty: Evidence from Household Consumption in China, Review of Economics and Statistics 95(2), 682-697. [44] Faber, B. 2014 Trade Integration, Market Size, and Industrialization: Evidence from China’s National Trunk Highway System, Review of Economic Studies, 81(3), 1046-1070. [45] Faggio, G., Silva, O. and W. Strange. 2014. Heterogeneous Agglomeration. serc Discussion Papers, serc DP0152. Spatial Economics Research Center, London School of Economics and Political Science, London, U.K. [46] Fay, M., and M. Morrison. 2007. "Infrastructure in Latin America and the Caribbean. Re- cent Developments and key Challenges. Directions in Development. Infrastructure." The World Bank. Washington, D.C. [47] Foote, W. 1997. Mexico’s Troubled Toll Roads. ICWA Letters, February 3. [48] Frye, D. 2015. Transportation Networks and the Geographic Concentration of Industries. Unpublished manuscript. [49] Galasso, A. A., and A. Oettl. 2017. Roads and Innovation. Review of Economics and Statistics 99(3), 417-434. [50] Ghani, E., Goswami, A., and W. Kerr. 2015. Highway to Success: The impact of the Golden Quadrilateral Project for the Location and Performance of Indian Manufacturing", The Economic Journal, 126(591), 317-357. [51] Gibbons, S., Lyytikainen, T., Overman, H. and R. Sanchis-Guarner. 2016. New Road In- frastructure: The Effects on Firms. cepr Discussion Paper DP11239. [52] Guasch, J.L., M. Fay. 2003. "Economic Activity, Agglomerations and Logistics in the Mex- ican Southern States". In Mexico: Southern States Development Strategy. The World Bank. Washington, D.C. [53] Jedwab, R. and A. Storeygard. 2016. The Heterogeneous Effects of Transportation Invest- ments: Evidence from sub-Saharan Africa 1960-2010. Unpublished Manuscript. [54] Kim, Y. and B. Zangerling. 2016. Mexico Urbanization Review. World Bank Publications. 32 [55] Monge, M. 2012 Analisis de la cadena productiva de Tequila: El caso de Jalisco. License dissertation. Departamento de Economía. UAM-Azcapotzalco. 69 pages. [56] Mun, S. 1997. Transport Network and System of Cities, Journal of Urban Economics, 42(3), 205-221. [57] Rothenberg, A. 2013. Transport Infrastructure and Firm Location Choice in Equilibrium: Evidence from Indonesia’s highways. Unpublished manuscript. [58] Pérez, S. ands M. Palacio. 2009. Desarrollo regional y concentración industrial: Impacto en el Empleo (1994-2004). In Observatorio de la Economía Latinoamericana, Número 117. [59] Pérez Cervantes, F. and A. Sandoval Hernández. 2015. Estimating the Short-Run Effect on Market-Access of the Construction of Better Transportation Infrastructure in Mexico. Banco de México Working Papers, N.2015-15. [60] Perrotti, Daniel and Ricardo J. Sànchez. 2011. La brecha de infraestructura en América Latina y el Caribe. Recursos naturales e Infraestructura Series No. 153, United Nations pub- lication, Santiago, Chile, July. [61] Porter, M. 1998. Clusters and the New Economics of Competition, Harvard Business Review, 76(6), 77-90. [62] Redding, S. and M. Turner. 2015. Transportation Costs and the Spatial Organization of Economic Activity. Handbook or Regional and Urban Economics 5b, Chapter 20, Elsevier. [63] Rosenthal, S. and W. Strange. 2003. Geography, industrial Organization, and Agglomera- tion, Review of Economics and Statistics 85(2), 377–393. [64] Staiger, D. and J. Stock. Instrumental Variables Regression with Weak Instruments. Econo- metrica, 65(3):557-586, 1997. [65] Roberts, M., Deichmann, U., Fingleton, B. and and T. Shi. 2012. Evaluating China’s road to prosperity: A new economic geography approach, Regional Science and Urban Economics, 42(4), 580-594. [66] Rothenberg, A. 2013. Transport Infrastructure and Firm Location Choice in Equilibrium: Evidence from Indonesia’s Highways. Unpublished manuscript. [67] Schaffer, M. 2010. xtivreg2: Stata module to perform extended IV/2SLS, GMM and AC/HAC, LIML and k-class regression for panel data models. http://ideas.repec.org/c/boc/bocode/s456501.html 33 [68] Storeygard, A. 2016. Farther Down on the Road: Transport Costs, Trade and Urban Growth In Sub-Saharan Africa, Review of Economic Studies, 83(3), 1263-1295. [69] Unger, K. and R. Chico. 2004. La Industria Automotriz en tres Regiones de México. Un análisis de clùsteres, El Trimestre Económico, 71(284), 909-941. [70] Unger, K. 2003. Los Clústeres Industriales en Mèxico: Especializaciones Regionales y la Política industrial, cide, División de Economía. Documento de trabajo número 278. [71] Volpe Martincus, C. and J. Blyde. 2013. Shaky Roads and Trembling Exports: Assessing the Trade Effects of Domestic Infrastructure Using a Natural Experiment, Journal of Inter- national Economics, 90(1), 148-161. [72] Volpe Martincus, C., Carballo, J. and A. Cusolito. 2014. Routes, Exports, and Employment in Developing Countries: Following the Trace of the Inca Roads. Unpublished manuscript. 34 Appendix A: Descriptives Statistics Summary Statistics of roads, travel time and travel cost data Table 12: Mean and median minimum travel cost & time, and efficient road length (1986 – 2014) Statistics Time to border Cost to border Time to port Cost to port Road efficiency 1986 Mean 1,150.14 2.28 663.12 1.94 244,535.80 Median 1,143.72 2.25 551.22 1.86 247,934.80 sd 449.24 0.34 398.62 0.30 83,307.17 1994 Mean 1,122.96 2.26 649.33 1.94 252,067.40 Median 1,110.02 2.23 542.41 1.86 255,266.10 sd 443.76 0.33 387.05 0.29 86,187.57 1999 Mean 1,111.30 2.25 637.15 1.93 265,420.80 Median 1,107.47 2.23 531.42 1.85 276,233.40 sd 427.93 0.32 374.74 0.28 89,278.03 2004 Mean 1,092.18 2.23 625.16 1.92 283,430.80 Median 1,080.79 2.21 518.52 1.84 289,366.60 sd 423.49 0.31 364.70 0.27 99,398.15 2009 Mean 1,007.39 2.17 583.87 1.89 297,425.70 Median 997.12 2.15 480.98 1.82 303,141.90 sd 388.01 0.28 363.45 0.26 99,650.99 2014 Mean 1,059.40 2.21 622.91 1.92 291,227.70 Median 1,053.18 2.19 519.05 1.84 296,489.10 sd 407.02 0.30 369.76 0.27 97,588.48 Note: Mean and median minimum travel time (in minutes) and travel cost (iceberg costs) to us border entry point and ports are calculated over the sample of 2,094 localities in Mexico. Efficient kilometers of roads reported are within a 200 km radius. 35 Table 13: Road network improvements in Mexico (1985 – 2016, in km) Road type 1985 1993 1999 2004 2008 2016 Multilane divided 0 2,299 4,885 6,239 16,233 16,036 Two lanes or Divided 1,888 3,221 4,013 4,706 6,315 6,591 Pavement 45,933 43,426 42,410 42,843 32,688 31,629 Gravel or Earth road 11,506 12,054 10,865 10,458 9,964 9,247 Total 59,328 60,999 62,173 64,246 65,200 63,504 Note: Authors’ calculations based on aaa map information for the corre- sponding years. Appendix B: Concentration and specialization Indices Krugman index The Krugman Specialization index (ksi) is a widely-used specialization measure. It measures the standard error of industry shares, by computing the share of employment which would have to be relocated to achieve an industry structure equivalent to the average structure of the reference group. The Index can take values in between zero (identical territorial struc- tures) and two (totally different structures). In our case, M = 2,094 localities, i.e., the number of metropolitan area and standalone municipalities in Mexico (consistent across years). We measure the locality specialization in terms of output and employment. I KSIm = ∑ |Sm,i − S i | (4) i=1 Where Sm,i is the output or employment share of industry i in locality m, S i and is the av- erage share of industry i in the total output or employment across all localities in Mexico, and I is the number of industries. If the relative specialization measure is zero, then the economic structure of a locality is identical to the economic structure of the overall economy. The higher the index, the more the economic structure of the locality deviates from the overall economy (reference group) and the more that locality is specialized. Ellison and Glaeser geographic concentration index The Ellison-Glaeser index (Ellison and Glaeser, 1997) defines concentration as agglomeration above and beyond what we would observe if plants simply chose locations randomly (as op- 36 posed to a uniform spatial distribution). This measure provides an unbiased estimate of ag- glomerative forces independently of their source. It can be interpreted as the probability that a firm choosing its location follows the prior firm rather than locating randomly. The Ellison- Glaeser index is given by the following formula: Gi − (1 − ∑r x2 r )Hi γi ≡ , ( 5) (1 − ∑r xr )(1 − Hi ) 2 where: • Gi ≡ ∑r (sri − xr )2 is the spatial Gini coefficient of industry i; • xr is the share of total employment in each locality r; • sri is the share of employment of locality r in industry i; 2 is the Herfindahl index of the plant size distribution of industry i; • Hi ≡ ∑j zji • zji represent the employment share of a particular firm j in industry i. Following Ellison and Glaeser (1997), an industry is strongly concentrated if γi > 0.05, weakly concentrated if γi ∈ (0, 0.05], and not concentrated if γi 0. Appendix C: Additional Results 37 Table 14: Spatial heterogeneous effects of market access Total employment Variables Border Capital Center North South Market Access (ma ) 0.154*** 0.328 0.188*** 0.336*** 0.221*** (0.0348) (0.274) (0.0248) (0.106) (0.0216) ma x Metropolitan dummy 0.0756 -0.0251 0.0613** -0.118 0.0949*** (0.0473) (0.123) (0.0242) (0.0999) (0.0247) ma x Education -0.0499*** -0.150 -0.0428*** -0.0173 -0.0495*** (0.0130) (0.212) (0.00744) (0.0151) (0.00508) ma x Capital city -0.0799*** -0.186*** -0.137*** -0.107*** -0.123*** (0.00637) (0.0232) (0.00423) (0.00587) (0.00413) ma x Oil dummy -0.0124 0.111 0.0246 -0.0785* (0.0319) (0.0896) (0.0234) (0.0453) Above median education 0.671* (0.406) Above median population -0.141 0.249 0.0165 0.250 0.225** (0.120) (0.774) (0.0775) (0.181) (0.106) Constant 0.862** 0.608 0.746*** -0.471 1.006*** (0.341) (0.449) (0.125) (0.417) (0.128) Observations 1117 206 3749 896 3798 Adj. R-squared 0.311 0.381 0.420 0.395 0.422 Note: * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. The estimates are from the instrument regression with the doughnut iv calculated by excluding all localities within a 25km circle. Results are robust to others iv strategies (road count and efficient road units). Table 15: Spatial heterogeneous effects of market potential Total employment Variables Border Capital Center North South Market Potential (mp ) -0.0257 0.690*** 0.381*** 0.0595 0.424*** (0.0951) (0.0757) (0.0791) (0.118) (0.0893) mp x Metropolitan dummy 0.0284 0.0338 0.0454*** 0.0737** -0.0223 (0.0294) (0.0400) (0.0144) (0.0339) (0.0154) mp x Education 0.0218** 0.0141** 0.0284** 0.0189*** (0.00920) (0.00584) (0.0133) (0.00398) mp x Capital city 0.0339 -0.166*** -0.0846*** -0.00115 -0.0791*** (0.0263) (0.0219) (0.0206) (0.0279) (0.0273) mp x Oil dummy 0.00265 -0.00494 -0.0132** -0.00822 (0.00984) (0.00426) (0.00523) (0.00866) Above median education -0.00480 (0.0207) Above median population 0.146 0.142** 0.0980* -0.0134 0.157** (0.104) (0.0674) (0.0560) (0.144) (0.0689) Constant -0.0609 -0.722** -0.373** -0.748* 0.470** (0.352) (0.368) (0.174) (0.396) (0.195) Observations 933 172 3119 744 3191 Adj. R-squared 0.214 0.750 0.634 0.489 0.534 Note: * denotes significance at the 10% level, ** denotes significance at the 5% level, and *** denotes significance at the 1% level. The estimates are from the instrument regression with the doughnut iv calculated by excluding all localities within a 25km circle. Results are robust to others iv strategies (road count and efficient road units). 38 Table 16: The effect of road infrastructure on employment: post-NAFTA Total employment Manufacturing Services Commerce Variables ols iv ols iv ols iv ols iv Road Efficiency 3.096*** 4.250*** 1.808*** 2.606*** 3.578*** 4.809*** 3.515*** 4.671*** (0.0979) (0.110) (0.164) (0.190) (0.0920) (0.100) (0.137) (0.158) Min travel cost to border -5.338*** -2.920*** -5.115*** -3.411*** -4.543*** -1.963*** -8.068*** -5.608*** (0.375) (0.388) (0.591) (0.621) (0.349) (0.366) (0.570) (0.573) Min travel cost to port 3.375*** 2.145*** 3.223*** 2.348*** 2.720*** 1.408*** 5.138*** 3.865*** (0.411) (0.382) (0.589) (0.605) (0.360) (0.353) (0.608) (0.580) Above median education 0.0928** 0.0948** 0.222*** 0.223*** 0.0783*** 0.0804*** 0.0621 0.0642 (0.0472) (0.0415) (0.0654) (0.0606) (0.0302) (0.0291) (0.0543) (0.0490) Above median population 0.194*** 0.193*** 0.159* 0.158** 0.185*** 0.183*** 0.250*** 0.248*** (0.0427) (0.0391) (0.0828) (0.0717) (0.0527) (0.0419) (0.0724) (0.0584) Constant -30.39*** -16.08*** -37.48*** -36.19*** (1.342) (2.224) (1.262) (1.914) Constant 7479 7479 7267 7257 7471 7471 7288 7276 Adj. R-squared 0.355 0.337 0.0894 0.0851 0.452 0.430 0.280 0.271 Note: * denotes significance at the 10% level, ** denotes significant at the 5% level, and *** denotes significance at the 1% level. The efficient roads doughnut iv is obtained by calculating the road efficiency units in a doughnut within two circles of 150 and 300 km radii. 39