The effects of road infrastructure on GDP per capita
Master Thesis International Economics and Business (IE&B) Groningen, July 2010
Author: Supervisor:
Joost van Uhm dr. T.M. Stelder
Student number: 1348574 Faculty of Economics and Business J.M.van.Uhm@student.rug.nl T.M.Stelder@rug.nl
Co-assessor:
prof. dr. J.H. Garretsen
Abstract
There have been numerous studies that show that road infrastructure is an important determinant of economic growth and income inequality. On the other hand, several studies conclude that a relationship between road infrastructure and economic growth can not be proven. This thesis tries to find a relationship between regional road infrastructure and regional gross domestic product (GDP) per capita. This is done by answering three different research questions, regarding the relative regional lengths of roads in 1980, the increase in the length of highways between 1980 and 2006, GDP per capita in 1997 and 2006 and its increase between 1980 and 2006. Hence, the initial road infrastructure endowments in 1980 as well as the growth in the length of highways between 1980 and 1997 and between 1980 and 2006 will be related to the GDP per capita in several years and the growth in GDP per capita over the period from 1980 to 2006. A distinction between highways and major roads is made in order to find out if the two types of road have had different influences on GDP per capita. The results show that the length of highways in 1980 has had a positive impact on GDP per capita in 1997 and 2006. Evidence for the relationship with GDP per capita growth only exists for highways on the NUTS 3 level. No relationship between the growth in the relative regional length of highways and GDP per capita growth has been found.
Table of contents
1. Introduction and research questions……… 4
1.1 Introduction………... 4
1.2 Facts on road transport……….. 6
1.3 Research questions……..……….. 8
2. Literature review and hypotheses………. 10
2.1 A definition of infrastructure………. 10
2.2 Characteristics of infrastructure……….………… 10
2.3 The role of infrastructure in regional economic development ……….. 13
2.4 Different economic effects of infrastructure improvements ……….… 15
2.5 Evaluation of the economic effects of road infrastructure ……… 17
2.6 Hypotheses………. 18
3. Data description...………... 19
3.1 NUTS regions……… 19
3.2 The road network………. . 20
5. Estimation results……… 33
6. Conclusions and limitations………... 35
6.1.1 Conclusions on the first research question………. 35
6.1.2 Conclusions on the second research question……… 36
6.1.3 Conclusions on the third research question……… 36
6.2 Limitations and future research………. 37
References……… 39
Appendices………... 45
Appendix A: Data overview and descriptive statistics……… 45
Appendix B: Diagnostic checks………... 49
1. Introduction and research questions
1.1 IntroductionTransport, communications and accessibility are important determinants for the economic development of regions. Without these essential components, economic development is virtually impossible for most sectors of the economy, because the transport infrastructure shapes the accessibility of places which in turn shapes the economic activity. In this thesis, the effects of one part of the transport infrastructure, regional road infrastructure, on regional gross domestic product (GDP) per capita in selected Western European countries are investigated. This is done by answering three different research questions, regarding the relative regional lengths of roads in 1980, the increase in the length of highways between 1980 and 2006, GDP per capita in 1997 and 2006 and its increase between 1980 and 2006. Hence, the initial road infrastructure endowments in 1980 as well as the growth in the length of highways between 1980 and 2006 will be related to the GDP per capita in several years and the growth in GDP per capita over the period from 1980 to 2006.
Two of the components named above, transport and accessibility, are important parts of the European Union’s policy. This is clear when the Territorial Agenda of the European Union is read: “Mobility and accessibility are key prerequisites for economic development of all regions of the EU.” In practice this means that regions having a high accessibility to raw materials, suppliers and markets are in general economically successful regions and improve their competitive position in the global market. If so, transport infrastructure improvement might be an important policy instrument to promote regional economic development (ESPON, 2006).
High accessibility is a prerequisite but not a guarantee for above average economic performance, as can be seen in figure 1. Although this map is not only based on the accessibility by road, but is a weighted average of the accessibility by rail, road and airplane, it gives a good impression of the importance of accessibility for economic performance. About two thirds of the NUTS 3 regions perform in line with the assumption that GDP is significantly linked to accessibility. This means that one third of the regions are either economically underperforming or overperforming relative to their accessibility.
reasons why the study on convergence and divergence is interesting. The first reason is simply the desire to test theory. More important, several regional and national governments pursue policies that are dedicated to income convergence. Sala-i-Martin states that exactly when the European governments promoted convergence in the early 1990’s, the incomes were subject to divergence. This contradiction is the second reason to study the subject. The final reason is the welfare of people around the world. This is illustrated by the fact that the ratio of the lowest GDP per capita to the highest GDP per capita in the world is still increasing. This is seen as a major failure of the world economic system.
Figure 1: GDP per capita versus potential accessibility of NUTS 3 regions, 2006.
1.2 Facts on road transport
This paragraph will discuss the main characteristics of road transport in the European Union.
Goods transport is often measured in tonne-kilometres, which is the equivalent of the transport of one tonne of freight over one kilometre. The total size of European freight transport by road in 2006 was about 1855 billion tonne-kilometres (ERF – IRF BPC, 2009). The reason to select only road infrastructure in this thesis is that road transport remains responsible for the biggest part of freight and person transport in Europe. Out of the five different transport modes: road, rail, inland waterways, sea and air, road transport was the fastest growing transport mode between 1995 and 2007. Both the share and size of road freight transport in total transport is expected to be even higher in the future (European Communities, 2008).
As can be seen in table 1, road freight transport is the biggest freight transport mode in the EU-27, Iceland, Norway and Switzerland, both by tons of freight transported and by number of employees in the transport sector (European Communities, 2009).
Table 1: Total freight transport per mode in the EU-27, Iceland, Norway and Switzerland.
The share of road freight transport in total freight transport is higher for national transport than for total international freight transport. Road freight transport is less important in terms of total value in external trade within the EU-27. Then, total value of external trade by road is just 17.3% of the total value. The main transport mode of external trade within the EU-27 in terms of value was sea transport, which accounted for 48.5% in 2007. When external trade within the EU-27 is measured in terms of weight, road freight transport is even less important, measuring 6.2% of the total weight.
The employment figures in table 2 show that road freight transport and road passenger transport are the biggest modes of transport in terms of employees. The figures only refer to employment in the 1.14 million companies in the EU-27 whose main activity is transport and only to freight transported within the EU-27, so the actual figures on employees in the transport sector are even larger. The total turnover of the 600.000 road freight transport enterprises in 2006 was 280 billion euro (European Communities, 2009).
Table 2: Total employment in the EU-27 by mode of transport in 2006.
Source: European Communities (2009).
1.3 Research questions
In this thesis, the role of transport infrastructure, in the form of major roads and highways, on the (regional) economy in selected countries in Western Europe is investigated. The research will consist of three parts in which a different research question will be answered. The first research question will be regarded as an explorative research to see what the influence of the stock of road infrastructure on a certain moment is on the GDP per capita after several years:
What is the influence of the relative regional length of major roads in 1980 on the level of the regional gross domestic product per capita in 1997 and 2006?
This question is answered by making a comparison between the relative regional length of the major road network in 1980 and the regional economic situation in 1997 and 2006, measured by GDP per capita. The reason to exclude the regional economic situation in 1980 from the analysis is a lack of data for the independent variables. Data on the research level of this thesis, see paragraph 3.1, are scarce, especially before the year 1993. Therefore, the employment per 1000 inhabitants in 1980 could not be calculated. The regional length of highways is recalculated to the length per 1000 km2. This is done to avoid a scale effect because bigger regions usually have more roads than smaller regions. This way, the lengths are relative and can be compared per region.
The second part of the research will deal with the possible relationship between the relative regional length of highways in 1980 versus the growth in GDP per capita over the period from 1980 to 2006. The corresponding research question is:
What is the influence of the relative regional length of major roads in 1980 on the relative regional growth in gross domestic product per capita in the periods from 1980 to 1997 and from 1980 to 2006?
was not available, this research question can not be answered for the NUTS 2 level from 1980 to 1997.
The third and final part of this thesis concerns the relationship between the relative regional growth in the length of highways versus the relative growth in GDP per capita in the same period. In this way, it is examined if regions with a relative high growth in the length of highways had a high relatively growth in GDP per capita as well. The corresponding research question is:
Is there a relationship between the growth in the relative regional length of highways between 1980 and 2006 and the relative regional growth in gross domestic product per capita in the same period?
There can be a endogeneity problem here because the growth of the relative regional length of highways is likely to be influenced by the relative regional growth in GDP per capita and the other way around. The growth in the relative regional length of highways will be calculated by comparing two handmade road maps. The reason to select only highways in the final research question is a practical one and will be explained in paragraph 3.2.
Because this research is conducted on two different regional levels, NUTS 2 and NUTS 3, a distinction between the effects on both levels can be made. It can be interesting to see whether some effects only occur on a certain regional level or if contrary effects occur on the two levels.
2. Literature review and hypotheses
The relationship between infrastructure and regional economic development is a popular topic among researchers since the 1930’s. The first people to investigate this relationship were Christaller (1933), Lösch (1940) and Isard (1956). Before a literature overview of the relation between transport infrastructure and regional economic development is given, it is essential to know what is meant by ‘infrastructure’.
2.1 A definition of infrastructure
There are several different definitions of infrastructure. In military, infrastructure is used for “all building and permanent installations necessary for the support, redeployment, and military forces operations” (Department of Defense, 2001). In engineering and construction, infrastructure describes the fixed assets that are in the form of a large network (Association of Local Government Engineers New Zealand, 1998). The World Bank (1994) defines infrastructure as “public services (electric energy, water facilities), public works (roads) and other transportation (harbours and airports)”. Most of the definitions of infrastructure have two things in common. First, infrastructure is a capital good for which users do not pay the market price. Second, infrastructure leads to very high costs for the first user, and small marginal costs for every next user (Rietveld, 1989).
2.2 Characteristics of infrastructure
Another characteristic of infrastructure is the existence of the ‘border effect’ (Puga, 2008, Rietveld and Vickerman, 2004). This effect means that trade across international borders is often much lower than might be expected, based on the size and distance of the other country. One reason for the existence of the border-effect is that national infrastructure networks are often designed for the national market and economic activity is often not evenly distributed across a country but gets less dense near the border. This is why the natural tendency of national governments is to under-invest in cross-border infrastructure. Another reason for the existence of the border-effect is that, apart from the physical crossing, a national border can often be understood as a non-physical barrier with an economic, political, cultural or language dimension. In road transport, the border is a barrier due to both non-physical factors such as language differentials, and physical factors such as the low density of roads near borders (Bruinsma and Rietveld, 1993).
In their study on the role of transport infrastructure endowment and investment in regional growth, Crescenzi and Rodriguez-Pose (2008) find that a good infrastructure endowment is a precondition for economic development because regions with adequate initial highway networks tend to perform better than regions lacking this type of basic infrastructure. However, they also find that the effects of infrastructure endowment and investment differ greatly per region, depending for instance on the timing of the infrastructure investment, the human capital and innovation endowments.
The Organisation for Economic Co-operation and Development (OECD) argues that “the availability of transport infrastructure and services is essential to most – if not all – sectors of society and the economy” (OECD, 2008). Sugolev et al. (2003) conclude as well that infrastructure is a necessary ingredient for economic growth, but it cannot create economic growth on its own and “the efficient supply of the right kind of infrastructure (material and institutional) in the right place is more important than the amount of money disbursed or the pure quantitative infrastructure capacities created”. Canning and Pedroni (1999) show that a two-causality between infrastructure and economic growth is common: infrastructure affects productivity and output while economic growth affects the demand for infrastructure.
economic growth. The World Bank finds a large range of empirical results on the importance of infrastructure for economic growth, ranging from no effect to rates of return in excess of 100% per annum (Canning and Pedroni, 1999). These examples make it clear that the effects of (transport) infrastructure on (regional) economic development are heavily disputed.
2.3 The role of infrastructure in regional economic development
There have been numerous studies that show that infrastructure is an important determinant of economic growth and income inequality. The exact impact may depend, however, on the type of infrastructure (Ottaviano, 2008) and the sensitivity of type of industry to transportation costs (McCann and Shefer, 2004). In industries where the nature and characteristics of the transactions have not changed, for instance in agriculture and raw materials, transportation costs have fallen steadily over time. However, in industries where the demand lead-times have fallen dramatically, or in industries where the variety and complexity has increased, transaction costs have not fallen and have sometimes even increased (McCann and Shefer, 2004). As the European Union continued to expand in the 1980’s, the position of cities in the European road network became increasingly important (Dupuy and Stransky, 1996). For Eastern Europe, one important obstacle to the integration of these countries in the world economy is the very poor state of their infrastructure (Martin and Rogers, 1995).
As Ottaviano (2008) shows, infrastructural investments affect the geographical distribution of economic activities in various and sometimes unexpected ways. This can be either between locations (‘external geography’) or within locations (‘internal geography’) (Martin and Rogers, 1995, McCann and Shefer, 2004). Concerning the external geography, improving the infrastructure between a developed and a less developed location has several consequences. The first, somewhat unexpected, consequence is that economic activity will migrate to the developed location, as shown by Boarnet (1998): “Street-and-highway capital is associated with higher output within the same county and with lower output in counties with similar population density, income, or employment shares”. This effect is known as the ‘straw effect’ because economic activities are ‘sucked up by a straw’. Crescenzi and Rodriguez-Pose (2008) have proven that regions benefit from neighbouring regions which are well endowed, but that additional highways in a neighbouring region might lure firms away from their own region towards that neighbouring region. For these reasons, the countries in the sample are countries that do border each other. The second effect is called the ‘shadow effect’ and shows that improved local infrastructure does not always make a location more attractive (Ottaviano, 2008). This effect can best be explained in a three city model where city B is located between city A and C and there is one road: from A to C via B. When the local infrastructure of city B is improved, not only the travel times between city B and city A and city B and city C will be diminished, but travel times between A and C via B will also be lowered. In this way, the shipments between A and C via B will ‘cast a shadow’ on city B’s attractiveness. Similar influences of transport on infrastructure distinguished by Button (1998) can be found in figure 2.
Figure 2: Influences of transport infrastructure on a region
Vickerman (1994) creates three broad categories of regional impacts of transport infrastructure: core-periphery impacts, corridor effects and shadow effects. Core-periphery impacts consider the relative impacts of new infrastructure on regions which form part of the current economic core of an economy and those who are more peripheral. Corridor effects consider the way that the new infrastructure affects relative accessibility between and within regions by looking at accessibility to the network and hence whether a particular region is likely to suffer from the creation of a corridor. By a corridor, Vickerman (1994) means the effect of channelling traffic through a region, without bringing substantial potential benefits to the region. The third and last category consists of shadow effects, which are caused by the diversion of traffic away from traditional roads and modes. These effects can occur both nearby and far away from the new infrastructure.
Crescenzi and Rodriguez-Pose (2008) claim a negative effect of the improvement of road infrastructure: “the direct impact of further infrastructure development (…) may even be negative where the additional infrastructure increases the exposure to external competition”. This is exactly what Puga (2008) makes clear with the analyses of cross-border infrastructure projects. On the one side, these infrastructure projects connect lagging regions with key markets, making it easier for firms in the lagging regions to reach new customers. On the other side, improvements in infrastructure also make firms in lagging regions subject to stronger competition from firms in more developed areas. Under certain conditions, the latter effect dominates the former effect, meaning that improvements in (cross-border) infrastructure can encourage firm location in regions with better initial conditions rather than facilitating a more even distribution of economic activities. Puga (2008) claims that this has happened in Europe in the late 1970’s when income convergence came to a halt, despite the vast spending in infrastructure. Claims that these expenditures in infrastructure prevented even stronger divergence are not justified by any evidence. Hence, one characteristic of cross-border infrastructure improvements is not that investment is split among several regions, but that it mainly affects the costs of moving goods and people across regions rather than within regions.
2.4 Different economic effects of infrastructure improvements
these industries. Lakshmanan and Chatterjee (2005) state that the economic outcomes of transport improvements are dependent on the context in which the improvements are made: the state of the pre-existing transportation network, the state of economic development and the nature of competition in the regions.
There are differences in the economic impact of the improvement of road infrastructure in the short run and the long run. As Rietveld (1989) points out, direct effects will occur in the short run in the construction sector which benefits from the actual road building, indirect effects in the short run occur in all other sectors via intermediate deliveries towards the construction sector. In the longer run, operations and maintenance effects occur. But these effects are relatively small compared to spin-off effects like changes in income and employment and investments due to the improvement or extension of the infrastructure.
Vickerman et al. (1997) argues that “transport improvements have strong and positive impacts on regional development only where they result in removing a bottleneck”. The removal of bottlenecks and the completion of critical missing links have been recognised as a vital thing for the European transport network (Vickerman, 1994). Sanchez-Robles (1998) states that infrastructure may have a different degree of impact according to its relative size. In types of infrastructure that follow distribution networks, the payoff of the investment is related to the size and the configuration of the network, usually being smaller in the case of larger networks.
Rietveld (1989) adapted the scheme in figure 3 below from Pluym and Roosma (1984) to show that the possible effects of the improvements of transport infrastructure are hard to predict, especially when multiple sectors of the economy are considered. The two effects that can occur are contradictory.
Figure 3: The economic effects of improvements of transport infrastructure.
Source: Rietveld (1989), adapted.
2.5 Evaluation of the economic effects of road infrastructure
transportation infrastructure itself”. They state that a realistic analysis (of infrastructure) requires a consideration of the nature of, and changes in, spatial transaction costs, as well as firm behaviour and organisational arrangements. According to Ottaviano (2008) infrastructure investments generate externalities that may diffuse far across the economy. Hence, regions need to coordinate in both interregional and intraregional infrastructure projects. Furthermore, transport infrastructure investments may exert an influence on economic activity which is not limited by regional boundaries (Crescenzi and Rodriguez-Pose, 2008). Therefore, the economic effects on infrastructure should not only be analysed for the region in which the investment takes place, but also for the surrounding regions which will benefit from the investment via spillover effects.
These examples show that the effects of transport infrastructure improvements on the regional economy are found to be hard to calculate.
2.6 Hypotheses
From the literature review it becomes clear that the relationship between infrastructure and economic development is controversial. On the one hand there are studies that find that there is a positive relationship, while on the other hand there are studies that question this relationship. Also, researchers have found that different relationships can occur on different geographical levels. Based on this debate, three research questions have been formulated, which are represented by the following hypotheses:
H1: There is a positive relationship between the relative regional length of highways in 1980 and regional GDP per capita in 1997 and 2006.
H2: There is a positive relationship between the relative regional length of major N-roads in 1980 and regional GDP per capita in 1997 and 2006.
3. Data description
In this thesis an important part of the transport infrastructure, the major road network, in Western Europe will be analysed. The goal of this thesis is to find out if there is a relationship between the regional endowments in road infrastructure in and GDP per capita. The two regional levels that this research is conducted on and the variables that are used to investigate this relationship will now be introduced.
3.1 NUTS regions
The two regional levels that are used in this thesis are part of the NUTS classification system. The abbreviation ‘NUTS’ stands for ‘Nomenclature des Unités Territoriales Statistique’ (Eurostat, 2009). More than 30 years ago, the European Union introduced the NUTS classification system in order to provide a single uniform breakdown of territorial units for the production of regional statistics for the European Union. This was useful for users of statistics in order to have comparable statistical data across the European Union (European Union, 2003). At a national level, four hierarchical levels of detail exist in the NUTS classification system, based on existing national administrative subdivisions. In Belgium for instance, the four NUTS levels are: NUTS 0: the country itself, NUTS 1: regions, NUTS 2: Provinces and Brussels capital region, NUTS 3: Arrondissements. In the NUTS classification system, guidelines are applied which are based on population data. For instance, the regions on the NUTS 2 level have a minimum of 800.000 and a maximum of 3.000.000 inhabitants. This makes the NUTS classification system suitable for comparisons.
The reason for not including more countries in the sample is that the data for the independent variables in this research are not available for most other nations which have joined the European Union after 1986. Not all countries which were not a member of the European Union by 1986 do provide population data on the NUTS level, which goes back to 1980. Hence, long term effects on any NUTS level could not be analysed for those countries in this thesis. An overview of the number of NUTS regions per country in this research can be found in table A1 in appendix A.
3.2 The road network
The road network in the selected countries for 1980 was not available in a digital format so it had to be made entirely by hand. This was done by taking Bartholomew’s ‘Road atlas Europe’ from 1980 to see which roads where present at that time, and in which form: a highway or a major road (‘N-road’). Then, the roads that were present in 1980 were drawn as a raster layer on an empty NUTS map of Europe in ArcGIS 9.3 to create a digital replica of the map from the road atlas. In order to make a realistic calculation on the relative growth of the road network, the map of the road network in 2006 had to be in the same format as the one from 1980. Hence, the 2006 road map was made by hand in ArcGIS as well. In order to be able to investigate if highways and major N-roads have different effects on GDP per capita, these two types of roads were created as separate raster layers. These layers can be combined to obtain the full major roads map in both years. Figure A1 in appendix A is an overview of the handmade road map of 1980 in which both layers are visible.
The lengths of the highways and major N-roads per region have been calculated by using the ‘zonal statistics’ option in ArcGIS, which counts the number of raster cells per region. In order to make the numbers suited for comparison, the numbers have been recalculated to cells per 1000 km2. This is done to avoid a scale effect because bigger regions usually have more roads than smaller regions.
N-roads could not be calculated. This problem does not occur for highways since a highway is not an arbitrary definition.
There are some drawbacks of using a raster map. The first drawback is that a raster map is generally less accurate than a polyline map, which is a map consisting of lines instead of raster cells. On a polyline map, lines can be drawn more fluent and hence more accurate than raster cells. However, since every road is drawn the same way this problem is levelled out. A second drawback is that the position of a road on the map in a hand created map can vary from its actual location because of the relatively big cell size. However, I have tried my best to make the maps as accurate as possible and used a raster cell size that has a good balance between accuracy and workability. The reason to work with raster layers is that raster layers are easy to draw, its length is easy to calculate and the drawbacks mentioned above do not result in big problems for this research.
3.3 Data description
The database consists of panel data of thirteen countries and covers the same NUTS regions for the years 1980, 1997 and 2006. Panel data consist of a group of cross-sectional units, also called ‘individuals’, which are observed over time. According to Carter Hill et al. (2008), panel data sets can be ‘long and narrow’, ‘short and wide’ or ‘long and wide’. The panel data set in this research, consisting of three years, 1017 NUTS 3 regions and 183 NUTS 2 regions, fits in the category ‘short and wide’.
An overview of the countries in this research can be found in table A1 in appendix A. All the selected countries were a member of the EU 12, except Austria and Switzerland. These two countries have been added to the sample, simply because the data were available and to avoid an area of missing data on the map between Germany and Italy. Greece was also a member of the EU 12 but has been removed from the sample because of a lack of data. East-Germany has been added to the sample because it became a part of the EU after the reunification with West-Germany in 1990. Therefore, West-Germany includes the Federal Republic of West-Germany, better known as West Germany, and the German Democratic Republic, or East Germany.
select these countries is that the endowments of road infrastructure have not only regional and national effects, but also international effects. In order to be able to analyse these international effects, the analysed countries must border each other.
A detailed description of the variables and its source(-s) used in this thesis can be found in table A2 in appendix A.
3.3.1 Dependent variables
Gross domestic product per capita (LOGGDPcap). This variable is the dependent variable in research question one, which investigates the relationship between the relative regional lengths of highways in 1980 and the level of the GDP per capita in both 1997 and 2006. The reason to select GDP per capita instead of total GDP per region is twofold. First, a simple scale effect takes place: a larger region in terms of area is likely to have a larger total population and a larger total GDP. Furthermore, larger regions in terms of area are likely to have more roads than a smaller region. Second, changes in the regional population size over the years will have an effect on total GDP which can not be attributed to the relative regional length of highways. These two problems do not exist by taking GDP per capita as the dependent variable. The natural logarithm of the GDP per capita is used in order to minimise the influence of extreme values.
The second reason to select the period from 1997 to 2006 is the relatively good availability of data on the regional level in this thesis, the NUTS level. Although the NUTS classification system was introduced to create a large database in order to compare European regions, there are unfortunately not many data available.
Growth percentage of the GDP per capita (GDPcapGROWTH). This variable is the dependent variable in the economic models that are used to find an answer on research questions two and three. The variable is calculated for the period from 1980 to 2006 using the corresponding figures of GDP per capita. Because there is a lack of data of the GDP per capita in 1980, this variable could only be calculated for Belgium, France, Germany and Spain. Hence, the second and third research question, which have this variable as the dependent variable, only account for these countries.
3.3.2 Independent variables
The influence of the relative regional length of highways on the regional GDP per capita in 1997 and 2006 is aimed to be described by the infrastructure situation in 1980. However, the GDP per capita and GDP per capita growth can not be solely accredited to the infrastructure situation in 1980 and its growth. Although it is virtually impossible to pool all the possible determinants of regional GDP per capita and growth of GDP per capita in one model, the following explanatory variables have been selected.
Relative regional length of highways in 1980 (LOH80KM2). This is one of the two main explanatory variables of the GDP per capita in 1997 and 2006 in this thesis. The other one is the relative regional length of major N-roads in 1980. In their research, Crescenzi and Rodriguez-Pose (2008) use the regional length of highways as a proxy for the quality of regional transport infrastructure, but they multiply this by the square of the regional population.
and other major roads before shifting the emphasis to high-speed trains (European Communities, 2008).
Relative regional length of major N-roads in 1980 (LON80KM2). The relative regional length of major N-roads in 1980 is the second main explanatory. Just like highways, N-roads are major roads but their maximum speed is lower and the number of lanes on N-roads is usually less than the number of lanes on highways. Despite the lower transport capacity of these roads, N-roads are still important for the regional transport and hence the regional economy. In regions where there are no highways, the major N-roads are the most important roads. Not all N-roads are ‘major’ in the sense that some N-roads are more important for regional transport than others. Bartholomew’s Road Atlas 1980 made a distinction between major and minor N-roads, and only the N-roads that have been indicated as a ‘main N-road’ have been drawn on the 1980 road map in ArcGIS. The number of raster cells per region has been calculated as an index for an area of 1000 km2 to avoid
a scaleeffect, for the same reason as is the case with the highways.
Relative regional length of highways and major N-roads in 1980 (LOHLON80KM2). This independent variable is the sum of the length of highways and major N-roads. Both types of road are viewed as being equally important. This variable is also recalculated as an index for an area of 1000 km2 and only used in regression (3).
Population density (LOGPOPDENS). The population density of a region is an indication of the level of urbanisation. It is a better indicator than the population size alone, because a scale effect occurs when only the population size is analysed: bigger regions in size are expected to have a bigger population. Therefore, the population density per square kilometre is used. The natural logarithm of this variable is calculated in order to minimize the effects of the extreme values.
or she is employed. When this is the case, the person is counted in the population size in another region than where this person is employed. In this variable, there is no difference made between jobs in the different economic sectors. The natural logarithm of this variable is taken in order to minimise the effects of extreme values, just like the variables GDP per capita and population density.
Airports (AIRPORT). This is a binary variable which accounts for the presence of a major airport in the region. Generally, the presence of a large airport in a region is an indicator of the economic importance of that region. Several studies have included the presence of a major airport in the region as a variable. In a study by the Netherlands Economic Institute (NEI, 1993) on location decisions by firms, it turned out that the presence of a major airport in the region is quite often a prerequisite for migrating international firms.
The data on airport traffic are taken from the World Airport Traffic Report 2006 (Airports Council International, 2007). The selected airports are located in the European countries analysed and are either in the top 200 list of biggest world airports by passenger movements or in the top 200 list of biggest world airports by cargo handled in metric tonnes, or in both these lists. The total number of NUTS 3 regions under analysis with at least one major airport which meet the criteria mentioned above is 66. For NUTS 2 regions, the number of regions with at least one major airport is 56. All the NUTS regions that had at least one major airport in 2006 also had at least one in 1980. Therefore, this independent variable is also used in answering the research questions which deal with the growth in GDP and the relative regional length of highways from 1980 to 2006. The presence of multiple airports in one NUTS region is not accounted for.
In some cases the airport serving a city is not located in the same NUTS 3 region as the city itself. This is for instance the case for Paris and London. For London, all major airports are in another NUTS 2 region as the city itself since the ‘Inner London’ region only covers the city itself. One drawback of the list is that there is no difference made between passengers who begin or end their journey at the airport, and transit passengers who just use the airport to get on another airplane and have no or little influence on regional GDP and hence regional GDP per capita. Another drawback of this type of measurement is that the economic benefits of the presence of an airport in a region are totally assigned to the region in which the airport is located. When an airport is located at the very edge of a region and close to another region, which probably has no airport, the real economic effect is likely to be distributed to the neighbouring regions without an airport as well. This distributing effect to neighbouring regions is not accounted for.
Seaports (SEAPORT). This is a binary variable which account for the presence of a major seaport. Just like airports, seaports may be a good indicator of the economic importance of a region. In the same research by NEI (1993) mentioned above, the presence of a seaport is also a quite important factor for companies, but to a lesser extent than the presence of a major airport.
In 2004, over 90% of Europe’s trade with the rest of the world was shipped through its seaports, as well as 43% of intra-European trade (ESPO, 2004). A seaport is a place where various types of transport interconnect. The presence of a seaport in a NUTS region depends more on physical geography than the presence of airports in a region, simply because seaports are located at shores and airports can be located virtually anywhere.
The data on seaports are obtained from the ESPO Rapid data Exchange System 2007, which is an organisation which monitors 50 ports in Western Europe. Those ports are located in 40 different NUTS 3 regions and 35 different NUTS 2 regions in this analysis. Although the figures are from 2007, the NUTS regions which have a seaport in 2007 had at least one seaport in 1980 which makes it possible to use this variable in all the regressions. Just like with the airports, since this is a binary variable, the presence of multiple seaports in one region is not accounted for. The main shortcomings for this variable are the same as those for airports mentioned above. Hence, the distributing effect of seaports to other regions is not accounted for and the presence of multiple seaports in one region has no effect.
Growth percentage of the relative regional length of highways (LOHGROWTH). This variable is used to answer the final main research question, which investigates the relationship between the growth in the relative regional length of highways between 1980 and 2006 and the relative regional growth in gross domestic product per capita. The values are based on the two handmade roadmaps. Because this variable represents the increase in the regional length of highways in percentages, it could only be calculated for regions in which highways were already present in 1980.
4. Empirical research
Now, the economic models used to estimate the relationship between the lengths of highways in 1980 and the change in GDP per capita between 1997 and 2006 will be described.
4.1 Economic models
Initially, kernel density plots showed that the dependent and independent variables that were selected for the regressions were skewed. One common solution for this problem is the use of natural logarithms of the variables, which is exactly what has been done. In the kernel density plots, the skewness approved greatly although not all the deviations from linearity were solved. The variables that are not taken the natural logarithm of are relative regional length of highways in 1980 per 1000 km2, the relative regional length of major N-roads in 1980 per 1000 km2, the
relative regional length of highways and major N-roads in 1980 per 1000 km2, the presence of an airport and the presence of a seaport. The reason for this is that some of the values of these variables are zero. In that case, taking the natural logarithm would result in missing data.
In order to find an answer on the first main research question, hence to find out what the relationship is between regional infrastructure endowment in 1980 and GDP per capita in the years 1997 and 2006 on both the NUTS 3 and NUTS 2 level, four economic models are used. The differences between these four models are that four different measurements of the relative regional length of the road network are used: highways, major N-roads, those two combined and those two as separate variables.
LOGGDPcapN ,t = ά + β1*LOH80KM2N + β2*LOGPOPDENS N,t + β3*LOGEMPLOY N,t +
β4*AIRPORTS N,t + β5*SEAPORTS N,t + ε (1)
Where LOGGDPcapN,t is the natural logarithm of GDP per capita on a NUTS level at time t, ά is
a constant or intercept, LOH80KM2N is the length of highways in NUTS regions in 1980 per
1000 km2,LOGPOPDENS
N,t is the natural logarithm of the population density on the NUTS
level at time t, LOGEMPLOY N,t is the natural logarithm of the number of jobs in persons per
1000 inhabitants on the NUTS level at time t, AIRPORTS N,t is a binary variable for the presence
of a major airport in the NUTS region at time t, SEAPORTS N,t is a binary variable for the
LOGGDPcapN ,t = ά + β1*LON80KM2N + β2*LOGPOPDENS N,t + β3*LOGEMPLOY N,t + β4*AIRPORTS N,t + β5*SEAPORTS N,t + ε (2)
Where LON80KM2N is the length of major N-roads in NUTS regions in 1980 per 1000 km2.
LOGGDPcapN ,t = ά + β1*LOHLON80KM2N + β2*LOGPOPDENS N,t + β3*LOGEMPLOY
N,t + β4*AIRPORTS N,t + β5*SEAPORTS N,t + ε (3)
Where LOHLON80KM2N is the combined relative regional length of highways and major
N-roads in NUTS regions in 1980 per 1000 km2. The fourth model is more or less the same as the
third but instead of taking the total relative regional length of the highways and the major N-roads together it makes a distinction between these two types:
LOGGDPcapN,t = ά + β1*LOH80KM2N + β2*LON80KM2N + β3*LOGPOPDENS N,t + β4*LOGEMPLOY N,t + β5*AIRPORTS N,t + β6*SEAPORTS N,t + ε (4)
The second main research question, concerning the relationship between the relative regional length of the road network in 1980 versus the growth in GDP per capita over the period from 1980 to 1997 and from 1980 to 2006, will be answered using the following economic models, which are quite analogous to models (1) to (4), except the fact that the dependent variable is different and that due to lack of data the increase in the employed persons per 1000 inhabitants could not be calculated.
GDPcapGROWTHN = ά + β1*LOH80KM2N + β2* POPDENSGROWTHN,t +
Β3*AIRPORTS N,2006 + β4*SEAPORTS N,2006 + ε (5)
Where GDPcapGROWTHN is the total GDP per capita growth percentage on a NUTS level in the
period from 1980 to 2006, POPDENSGROWTHN,t is the growth percentage of the population
density per km2 on a NUTS level in the period from 1980 to 1997 or the period from 1980 to
2006.
GDPcapGROWTHN = ά + β1*LOHLON80KM2N + β2* POPDENSGROWTHN,t + β3*AIRPORTS N,2006 + β4*SEAPORTS N,2006 + ε (7)
GDPcapGROWTHN = ά + β1*LOH80KM2N + β2*LON80KM2N + β3* POPDENSGROWTHN,t + β4*AIRPORTS N,2006 + β5*SEAPORTS N,2006 + ε (8)
The economic model that helps to answer the third and final research question concerns the relationship between the relative regional growth in the length of highways from 1980 to 2006 versus the relative growth in GDP per capita in the same period. There is only one model since only the relative regional growth in the length of highways could be calculated.
GDPcapGROWTHN = ά + β1*LOHGROWTHN + β2*POPDENSGROWTHN + β3*AIRPORTS N,2006 + β4*SEAPORTS N,2006 + ε (9)
Where LOHGROWTHN is the total growth percentage for the length of highways in the period
from 1980 to 2006.
4.2 Empirical methodology
It is virtually impossible to pool all the possible determinants of regional GDP per capita in one model. This is not only the case because there are a lot of variables which influence the GDP per capita, but also because there are not many variables which are available on the NUTS level.
The panel data set consist of data for the years 1980, 1997 and 2006. Since it is not the aim of the first research question to find a relation between the GDP per capita in 1997 and GDP per capita in 2006, the data of 1997 and 2006 are regressed separately. Because the data are regressed separately, they can be seen as being from a single point in time (non-recurrent) and therefore the regressions are an example of cross section analysis. This is also the case for regressions (5) up to and including (9). Therefore, simple ordinary least squares (OLS) regressions are used to find evidence for the hypotheses.
4.3 Descriptive statistics
standard deviations are relatively high. High standard deviations mean a high spread of the values in the dataset. However, after taking the natural logarithm of the variables the standard deviations were relatively lower.
4.4 Diagnostic checks
Diagnostic checks are necessary in order to check for relationships between variables and the validity of the model (Nau, 2010). The checks that have been performed will now be discussed.
4.4.1 Normality
To test for the normality of distribution a Jarque-Bera test is conducted. This test is based on two measures: skewness and kurtosis. Skewness refers to how symmetric the residuals are around zero while kurtosis refers to the ‘peakedness’ of the distribution. Perfectly symmetric residuals have a skewness of zero, while a normal distribution will have a kurtosis value of 3 (Carter-Hill et al, 2008). The Jarque-Bera test scores and the corresponding P-values are found in table A3 in appendix A. The high Jarque-Bera test scores in the dataset are generally caused by the high kurtosis values. This means that a relatively large part of the variance is caused by extreme peak values. This problem is partly solved by taking the log of several variables. One extreme peak value in the GDP per capita that is present in the 1997 and 2006 observations on the NUTS 3 level is Inner London West. This region has the highest GDP per capita of all regions by far and has a relatively low value for highways per 1000 km2. One reason for this is that there are a lot of
financial centres in this region which do not rely much on transport infrastructure for their business. Another reason is that the region is in the very heart of London where there are obviously not many highways; the only highway in this region is the M4. The relative regional length of major N-roads is relatively high in this region.
4.4.2 Linearity
which is an augmented partial residual plot of a random chosen variable, the (green) smoothed line is relatively close to the (blue) ordinary regression line, so this particular variable is not suffering from non-linearity. Off course, this check is performed for every single variable variable and nonlinearity was not detected.
Figure B2 in appendix B shows the relationships between all the variables in regression (4) on the NUTS 3 level for 1997. There are not many relationships between the variables which do not look linear in some way. The best linear relationship can be found between the log of GDP per capita and the number of employed people per 1000 inhabitants. The relationship between the relative regional length of highways per 1000 km2 in 1980 and the relative regional length of
major N-road per 1000 km2 in 1980 looks negative linear. These two variables are not very
related to each other. Similar relationships occur for 2006.
In a kernel density estimate, like in figure B3 in appendix B, skewness can be detected. The figure shows that for instance the variable LOGPOPDENS97 is indeed skewed, since the blue line deviates from the red line. However, the skewness is relatively low and not problematic. Furthermore, the skewness was much higher at first but was greatly reduced after taking the natural logarithm of this variable.
4.4.3 Multicollinearity
Multicollinearity occurs when two or more independent variables have a linear relationship. The higher the degree of multicollinearity, the higher the instability of the regression model estimates of the coefficients. There are several ways to detect if multicollinearity occurs in the dataset. One way to detect if multicollinearity occurs is to compute the variance inflation factor (VIF). When the value of the VIF is higher than 10, hence the value of 1/VIF is lower than 0.1, the variable could be considered as a linear combination of other independent variables and multicollinearity exists (UCLA, 2010). As can be seen in table B1 in appendix B, which shows the VIF’s for regression (4) on the NUTS 3 level for 1997, all values of the VIF are substantially lower than 10. All other regressions have been checked as well and showed similar results. Hence multicollinearity does not exist according to the VIF in this data set.
relationship (-1), to no relationship (0), to a positive relationship (+1). When values are lower than -0.7 or higher than 0.7, multicollinearity does exist (Terpstra, 2009).
As can be seen in table B2 in appendix B, multicollinearity does only exist in this example between the variables LOGGDPcap and LOGEMPLOY on the NUTS 3 level in 1997. Tests on the NUTS 3 level for 2006 gave the same result. On the NUTS 2 level, the values in the correlation matrix are higher than 0.7 for LOGPOPDENS and LOH80KM2. However, the values are only slightly higher than 0.7 and are therefore not worrying.
4.4.4 Heteroskedasticity
Heteroskedasticity exists when the variances for all observations are not the same (Carter-Hill et al, 2008). Both the White test and the Breusch-Pagan test are used to detect heteroskedasticity. These tests share the same null hypothesis stating that the variance of the residuals is homogeneous (UCLA, 2010). The results of these two tests show that the null hypothesis that the variance is homogeneous is rejected.
The only independent variable that gets the result that it is suffering from heteroskedasticity as a single dependent variable for LOGGDPcap97 is SEAPORT. For LOGGDPcap06 there is no independent variable that suffers from heteroskedasticity.
4.4.5 Normality of residuals
Multiple regressions do not necessarily require normality, before drawing final conclusions, to review the distributions of the major variables of interest (StatSoft, 2010). Three ways of checking the normality of the residuals have been performed: the kernel density estimate, the standardised normal probability plot (P-P plot) and a plot of the quantiles of the variables against the quantiles of a normal distribution. A kernel density estimate was already used above to test for linearity. An example of a P-P plot can be found in figure B4 in appendix B. The standardised normal probability plot is sensitive to non-normality in the middle range of the data while the plot of the quantiles of the variables is sensitive to non-normality near the tails. This P-P plot shows some signs of non-normality, but not to a very worrying extent.
5. Estimation results
Now that the dataset is checked for normality, linearity, multicollinearity, heteroskedasticity and normality of residuals, regressions can be performed. All the regressions have been performed using STATA 10.0. Results of the main regressions (4), (8) and (9) can be found in table C1 in appendix C. Results of regressions (1), (2), (3), (5), (6) and (7) are discussed here but are not included in the results table since these regressions are explorative and not regarded as the main regressions of this thesis.
First, the results regarding the first main research question, concerning the relationship between the relative regional endowments of roads in 1980 versus the GDP per capita in 1997 and 2006, will be discussed. When the relative regional length of highways per 1000 km2 in 1980 is the
independent variable regarding the road network for GDP per capita, as is the case in regression (1), LOH80KM2 is significant at the 1% significance level for both years and both NUTS levels. However, on both NUTS levels, the coefficients for 2006 are lower than the coefficients for 1997.
In the case where the relative regional length of major N-roads per 1000 km2 in 1980 is the only independent variable for GDP per capita, as is the case in regression (2), LON80KM2 is significant at the 1% significance level for the four different versions concerning the NUTS level and year. Unlike the coefficients regarding the highways, the coefficients on both NUTS levels for the major N-roads do not diminish over time as much as those on the NUTS 3 level but stay almost equal.
Regression (3), in which the combined relative regional length of highways and major N-roads per 1000 km2 in 1980 is the only independent variable, LOHLON80KM2 is significant at the 1%
significance level for all four versions of the regression. The coefficients diminish over time, just like the regression with the relative regional length of highways per 1000 km2 in 1980.
The result of the regression (4), where the control variables are added, can be found in table C1 in appendix C. LON80KM2 is not significant anymore in three out of four cases. However LOH80KM2 is still significant at the 1% significance level for all four versions of the regression. The other independent variables show mixed results, except SEAPORT which is never significant.
generally lower coefficients of the independent variables in 2006 compared to those of 1997. This is likely to be true, since there has been more road building between 1997 and 2006 which has its own effects on the GDP per capita in 2006.
The adjusted R2 values, the values that describe the part of the variation in GDP per capita that is
explained by the independent variables, are quite low for regressions (1), (2) and (3), although the ones on the NUTS 2 level are slightly higher than those on the NUTS 3 level. The fact that the adjusted R2 values are low when the relative regional lengths of roads are the only independent variables is proof of the statement by Ottaviano (2008) that it is hard, if not impossible, to disentangle the economic effects of infrastructure from other concurrent effects.
Indeed, when other independent variables are added, the adjusted R2 values rise. For regression (4), the main regression of this thesis, the adjusted R2 values are much higher than in the first
three regressions. For 1997 on the NUTS 3 level, 60,85% of the variation in GDP per capita explained by the independent variables, 52,67% for 1997 on the NUTS 2 level, 56,75% for 2006 on the NUTS 3 level and 26,74% for 2006 on the NUTS 2 level.
Every significant independent variable has a positive regression coefficient which indicates a positive impact on the independent variable, GDP per capita. The relative regional length of highways in 1980 per 1000 km2 is significant at the 1% significance level in all 8 regressions in
which this variable is included. The relative regional length of major N-roads in 1980 per 1000 km2 is significant at the 1% significance level in all four versions of regression (2). For regression (4) however, only two out of four regressions are significant: at the 1% significance level on the NUTS 3 levels in 2006 and on the 10% significance level on the NUTS 2 level in 2006. The combined length of highways and major N-roads in 1980 is significant at the 1% significance level in all four versions of regression (3).
6. Conclusions and limitations
6.1 ConclusionsNow that the results of the three research questions regarding the relationship between the relative regional road network length and GDP per capita have been discussed, conclusions can be drawn. The conclusions for each of the three research questions are discussed separately.
6.1.1 Conclusions on the first research question The first research question is:
What is the influence of the relative regional length of major roads in 1980 on the level of the regional gross domestic product per capita in 1997 and 2006?
This question is answered by using four regressions which have different independent variables measuring the relative regional endowment of roads in 1980: the relative regional lengths of the highways, the major N-roads, the combined length of these two and the two lengths separately in one regression. Regressions (1) to (4) have all been regressed four times due to the fact that there are two NUTS levels and two years.
The fact that the coefficients of LOH80KM2 are significant for regression (1) and (4) in both years and on both NUTS levels and are lower for 2006 than for 1997 on both NUTS levels indicates that the influence of the relative regional length of the highway network in 1980 on GDP per capita diminishes over the years. This seems to be a fair conclusion since more highways have been built between 1997 and 2006 so the highway endowment in 1980 is of less importance for the relative regional highway length and GDP per capita the further we move from 1980. In regression (2), where LON80KM2 is the independent variable regarding the road network, the coefficients of 1997 are not significant on both NUTS levels. The coefficients for 2006 are significant but very small and almost similar for 1997 and 2006 which indicates that the relative regional length of major N-roads does not diminish in importance for GDP per capita over time. Regression (3), in which LOHLON80KM2 is the independent variable concerning the road network, similar mixed results as regression (2) show up.
6.1.2 Conclusions on the second research question The second research question is as follows:
What is the influence of the relative regional length of major roads in 1980 on the relative regional growth in gross domestic product per capita in the period from 1980 to 2006?
The relative regional endowment of the road network is still the main independent variable in this question but the dependent variable measures the growth in GDP per capita in the period from 1980 to 2006 instead of the GDP per capita level at a moment in time.
Regression (5), (6) and (7) all show significant results for the independent variables measuring the relative regional length of the road network on GDP per capita growth on the NUTS 3 level. On the NUTS 2 level however, these results are not significant. The main regression concerning the second research question is regression (8). The results of this regression can be found in table C1 in appendix C. The coefficients of LOH80KM2 on both the NUTS 2 and the NUTS 3 level are somewhat surprising; they show that GDP per capita growth will diminish when the relative regional length of highways increases. The coefficients associated with the relative regional length of major N-roads in 1980 are not significant.
Based on these finding it can be concluded that only the relative regional length of highways in 1980 has had a significant relationship with the growth of GDP per capita over the period from 1980 to 2006. The fact that this relationship is negative is somewhat unexpected.
6.1.3 Conclusions on the third research question
The third and final research question of this thesis concerns the relationship between the growth of GDP per capita and the growth in the relative regional length of highways instead of the relative regional length of the road network in 1980. More specific:
The results of regression (9) in table C1 show that there is no evidence in this research found that supports the relationship as described in research question three. Hence, no evidence is found on convergence or divergence on the two different NUTS levels.
In summary, the findings on the research questions are twofold. The effects of the relative regional length of highways on the regional economy in a later year are undoubtedly positive. However, this effect diminishes when time goes by. A relationship between the relative regional length of major N-roads in 1980 and GDP per capita in 1997 and 2006 has not been found. There is also evidence found that the endowment of roads in 1980 has had a positive effect on the growth in GDP per capita in the following 26 years, but only for highways on the NUTS 3 level: NUTS 3 regions that had a relatively high length of highways in 1980 were subject to a relatively large growth in their GDP per capita. For the NUTS 2 level, this relationship is not found. Therefore, nothing can be said about divergence and convergence and the two different regional leves.
6.2 Limitations and future research
The most important limitation of this research is the lack of independent variables which determine the value of GDP per capita. In the first place, the fact that GDP per capita is influenced by so many different variables makes it virtually impossible to make a comprehensive model of independent variables that determine GDP per capita. Although I think that the independent that were selected are good indicators of GDP per capita, more independent variables would have been added if these were available on the NUTS level. The lack of availability of data on the NUTS level is the main reason for not having more independent variables. Ideally, independent variables like the education level, foreign direct investments and working population per economic sector would have been added. However, these data are not available on the NUTS level. The reason to select NUTS regions in this thesis is that these regions are relatively easy to compare since there are guidelines regarding the number of inhabitants per region. However, at the moment there are not many variables publically available.
References
Airports Council International (2007). World airport traffic report 2006, updated. Geneva: ACI world headquarters.
Association of Local Government Engineers New Zealand (1998). Infrastructure Asset Management Manual. Edition 1.1.
Baier, S.L. and J.H. Bergstrand (2001). The growth of world trade: tariffs, transport costs, and income similarity. Journal of International Economics 53, pp. 1-27.
BART (2010). A history of BART. Bay Area Rapid Transit. Internet: www.bart.gov/about/history/index.aspx. Visited on February 24, 2010.
Bartholomew (1980). Road atlas Europe. Edinburgh: Bartholomew.
Biehl, D. (1991). The role of infrastructure in regional development. In Vickerman, R.W. (ed.), Infrastructure and regional development. European research in Regional Science, Vol. 1. London: Pion.
Boarnet, M.G. (1998). Spillovers and the locational effects of public infrastructure. Journal of Regional Science, Vol. 39, No. 3, pp. 381-400.
Brouwer, H.J. (2010). Vastgoedontwikkeling, Infra en de Amsterdamse Zuidas. Class at the university of Groningen, february 17, 2010.
Brown, D., M. Fay, S.V. Lall, H.G. Wang and J. Felkner (2008). Death of distance? Economic implications of infrastructure improvement in Russia. In: Infrastructure investment, growth and cohesion. The economics of regional transport investment. European Investment Bank Papers, Vol. 13, No. 2, pp. 126-147.
Button, K. (1998). Infrastructure investment, endogenous growth and economic convergence. Annals of Regional Science, Vol. 32, No. 1, pp. 145-162.
Canning, D. and P. Pedroni (1999). Infrastructure and long run economic growth. Paper presented at the 1999 Econometric Society Summer Meeting, Madison, Wisconsin.
Carter Hill, R., W.E. Griffiths and G.C. Lim (2008). Principles of econometrics. Third edition. Hoboken, NJ: John Wiley & Sons, Inc.
Chandra, A. and E. Thompson (2000). Doe s public infrastructure affect economic activity? Evidence from the rural interstate highway system. Regional Science and Urban Economics 30, pp. 457-490.
Christaller W. (1933). Central places in southern Germany. Jena: Fischer. English Translation by C.W. Baskin (Prentice Hall, London, 1966).
Crescenzi, R. and A. Rodríguez-Pose (2008). Infrastructure endowment and investment as determinants of regional growth in the European Union In: Infrastructure investment, growth and cohesion. The economics of regional transport investment. European Investment Bank Papers, Vol. 13, No. 2, pp. 62-101.
Department of Defense (2001). Dictionary of Military and Associated Terms. (Revised, 2009).
Diamond, D. and N.A. Spence (1989). Infrastructure and industrial costs in British industry. Department of Trade and Industry, HMSO, London.
Dupuy, G. and V. Stransky (1996). Cities and highway networks in Europe. Journal of transport geography, Vol. 4, No. 2, pp. 107-121.
Esfahani, H.S. and M.T. Ramírez (1999). Institutions, infrastructure, and economic growth. Journal of Development Economics, Vol. 70, No. 2, pp. 443-447.
