The Quantification of the Effects from

The Quantification of the Effects from

Transportation Infrastructure on Economic Growth

Name: Roald Camphuijsen Student Number: 1220683

Date: 27 August 2006 Supervisors: Prof. dr. H.H. van Ark

Drs. J. Bolt

Study Programme: International Economics and Business

Institution: Rijksuniversiteit Groningen

Table of Contents

Acknowledgements 3

Abstract 4

Introduction 5

Literature Review 7

Infrastructure Modalities 12

Model 13

Data 16

Testing the Model 25

Results 27

Conclusion 36

Bibliography 41

Appendix 45

Tables and Figures 47

Acknowledgements

I would like to thank C.R. Burggraaff for his practical suggestions and recommendations in this investigation, prof. dr. H.H. van Ark and drs. J. Bolt for their constructive comments during the writing of this paper, drs. G. Ypma for making available part of the data and dr. J.P.A.M.

Jacobs for giving his experienced advise on problems with multicollinearity.

Deleted: certain helpful

Abstract

There has been a lot of discussion on the importance of infrastructure on economic growth

and development. This thesis focuses on the individual impacts of different infrastructure

modalities on economic growth using cross-country data. Furthermore, the effects of

interactions between these infrastructure modalities have been assessed to examine if

economic growth is further enhanced. The main result of this paper is that especially the

presence of ports is important for economic growth. The other infrastructure modalities

become increasingly important when they interact with ports, while with other modalities no

significant effect has been found.

1. Introduction

In September 2000, the UN established a roadmap with set of objectives to be achieved within the next 15 years. These objectives were set with the aim of promoting sustainable worldwide development both socially, economically and environmentally. These objectives were baptized the Millennium Development Goals (MDGs). Governments putting forward the accomplishment of these goals, would be given special assistance and support.

The development of poor countries receives special attention when it comes to the

implementation of the MDGs for the reason that it is these countries with an urgent need of development. Governments should focus on the promotion of education, train teachers, nurses and engineers, and build infrastructure so that by this means the goals such as eradicating extreme poverty, promoting universal primary education and gender equality can be achieved.

Indeed, infrastructure is presumed to promote these objectives to a large extent and among others, several World Bank researchers dedicated investigations to the economic and social effects of infrastructure on the lives of people in poor countries. For instance, Gannon and Liu (1997) assess the effect of transportation possibilities on poverty reduction finding that reliable access to schools and health services contributes directly to the accumulation of human capital, which is a key factor in sustainable poverty alleviation. But also transportation investment projects may foster economic growth, target the transportation need of the poor, or directly generate employment opportunities for the poor.

In a study on the effect of infrastructure on poverty alleviation in Morocco, Levy (2004) finds a strong effect of road building with respect to four main areas of transport activity. In transport, improved roads lead to better connections to markets and services, reduced costs, and increased quality of services. In agriculture, road improvement leads to increased agricultural activity and higher value crops. In addition, in health and education, better roads resulted in a doubling in school enrolment in primary education and better healthcare facilities. Furthermore, improved roads had a significant effect on gender equality.

Enrolment in primary education for girls increased significantly more than for boys and women gained the most from increased visits to healthcare services. Moreover, the welfare of women increased because of the introduction of butane gas for cooking and employment opportunities made possible by better roads.

Deleted: and g

Deleted: research

Deleted: an

Deleted: ed

Deleted: of

Deleted: from

These examples illustrate to what extent infrastructure can have an effect on the lives and development of people on a microeconomic level. But also on a macroeconomic level improvements in transportation infrastructure have consequences for the rise in economic activity and development of regions or countries. However, economists do not yet agree upon their extent and direction.

There has been a lot of public interest in the linkage between infrastructure and economic growth. This thesis focuses on one of the most commonly used measures to estimate economic growth: income growth. In particular the effects of infrastructure on real GDP per capita growth will be examined. There are studies on the effect of infrastructure on economic growth or productivity, but this literature uses aggregate measures of infrastructure.

Furthermore, none of the previous literature has yet elaborated on possible interaction effects between different infrastructure types. This thesis extends the existing literature not only by dealing with different types of infrastructure that have distinct effects on economic growth but also by investigating the effect of an interaction between certain infrastructure modes on economic growth. The study is limited to the extent that the effects of infrastructure are only examined at the macroeconomic level, without taking into consideration additional effects on the social aspects of people’s lives that might be generated by infrastructure development, such as network externalities mentioned by Escobal and Ponce (2002).

The remainder of this thesis is organized as follows: in the following section I will give a review of the literature on the linkage between infrastructure capital and economic growth. It includes both theoretical and empirical studies with the purpose of illustrating the multiplicity of research there has been done in this area. In section three the research question of this study is elaborated upon in more detail and the hypotheses that will be tested are described. Section four elaborates on the model for the estimation of the impact of infrastructure on economic growth. The inferences made with respect to the econometric model are explained as well. Section five describes the data and gives some limitations with respect to the chosen measurement of certain variables. In section six the model is tested.

Both the impact of the various infrastructure types and the interaction of these types on economic growth will be quantified. In section seven the results and interpretations will be explained. In the last section I will conclude with a brief overview of the results and their implications.

Deleted: activities

Deleted: Despite this disagreement, t Deleted: is

Deleted: For that reason, t

Deleted: Literature has already been written

Deleted: Nevertheless, this approach

Deleted: impacts Deleted: ’

Deleted: paper

Deleted: written

Deleted: work

2. Existing Literature on the Importance of Infrastructure

It was not until the late 1980s that economists started to develop quantitative measures of the linkage between transportation infrastructure and economic productivity. Aschauer (1989) was the first to build a macroeconomic model in attempting to quantify this linkage. In a series of controversial studies he investigates the relationship between public investment and economic growth for the US employing aggregate time series data and a Cobb-Douglas function that is expanded by (nonmilitary) public capital. In his investigation he finds a strong positive relationship between public capital and output with an estimated elasticity of 0.39.

Furthermore, he explores the effects of components of nonmilitary public capital stock on productivity. He finds that only the so called ‘core’ infrastructure (highways, airports,

electrical and gas facilities, water, etc.) has a significant effect on private business output with an estimated elasticity of 0.24. Aschauer argues that these findings provide evidence for the fact that the slowdown in the productivity growth in developed countries in the 1970s was caused by the deficiency of public capital.

Aschauers’ investigation provoked a lot of debate about both the interpretation of the results and the appropriate empirical methodology. As a reaction, many other studies were carried out on the relationship between infrastructure and economic growth. Some of them have outcomes that are in line with Aschauers’ results. For instance, Munnell (1990a) also used aggregate time series data and a Cobb-Douglas production function but a different measure for labour productivity as dependent variable. She reached the same conclusion and found coefficients of 0.33 to 0.39 which are similar to those found by Aschauer.

Nevertheless, many others studies criticize these findings disagreeing on several points in both Aschauer’s and Munnell’s investigation. One point of criticism with respect to the time-series data used in both investigations, is put forward by Gramlich (1994). He argues that the estimated elasticities are unrealistically high and overstate the impact of public infrastructure investment on private sector output and productivity growth. Even when pooling the time-series data as in Munnell’s (1990b) investigation, which allows for more variance in the independent variables, the impact of infrastructure can still be overstated by confounding intrinsic state

productivity differences with variation in infrastructure capital.

Aaron (1990) came up with a similar point of criticism stating that time-series data are not very useful for the assessment of the impact of public infrastructure on economic growth

1

At the supra-national level in the US.

Deleted: only

Deleted: to have

Deleted: reasons

Deleted: irrationally

because there is insufficient variation in the data. Aaron re-estimates a model similar to Aschauer’s but does not find significant estimates for infrastructure. The lack of varation in time-series data is also addressed in Jiang (2001) who gives an overview of studies in this field of interest. Furthermore, he points out that time series data often show the characteristics of non-stationarity which means that data series tend to move in similar directions over time.

If no adjustment is made for non-stationarity in the data, as in Aschauer’s studies, then the estimated effects are the result of spurious correlation.

A further comment on Aschauer’s work is made by Eisner (1991) who raises the issue of causality between the independent and dependent variables. Eisner claims that the

estimated effects may be running the other way and that the demand for public infrastructure capital is raised by increased private output. If this is true, then increases in private output may increase the demand for public infrastructure capital meaning that the causal relation runs the other way around.

Even though Aschauer’s work has received a lot of criticism, it initiated the attention of macroeconomists on the importance of the relationship between public infrastructure capital and productivity.

A more recent study is done by Anderson and Lakshmanan (2004) who elaborate on the underlying mechanisms between infrastructure and productivity. They separate the macro- effects of transportation infrastructure into three mechanisms. The first one is gains from trade caused by specialization. As Adam Smith (1776) observed, the division of labor is limited by the extent of the market and transportation infrastructure can extend the size of the market.

The productivity benefits of specialization can occur through comparative advantage in which the resources are used where they are most efficient. There is also a benefit via scale

economies that are realized when markets for producers grow (see also Canning and Benathan, 2000). Transportation infrastructure is important because specialization and trade will only occur if the efficiency gains from trade are larger than the transportation costs required. A second mechanism mentioned by Anderson and Lakshmanan (2004) is the ‘big push’ effect to overcome coordination failures. These failures are caused by the fact that no single sector can support transportation or other infrastructure for its industrialization. But if all sectors industrialize they can jointly support this infrastructure. Therefore, public infrastructure can be seen as a way to overcome this coordination failure and may provide a signal for firms in different sectors to foresee widespread industrialization and act

accordingly. The third mechanism explained Anderson and Lakshmanan is the shift to better

Deleted: insufficiency

Deleted: mentioned Deleted: done

Deleted: flow

Deleted: an

Deleted: But also, t Deleted: get larger

Deleted: The fact that makes t Deleted: here is that

Deleted: called

Deleted: in this article

technology in freight handling which has an impact on productivity to the extent that a fundamental transformation of production technologies brought about specialization and large-scale production.

The benefits of specialization and economies of scale linked with transportation costs are also described in a pioneering paper by Krugman (1991). This paper led the way to the so- called new economic geography. Krugman develops a simple model that shows how and when, as a consequence of a fall in transportation costs, manufacturing firms become concentrated in some regions or countries leaving others relatively undeveloped. The central finding is that a decrease in transportation cost (as a result of an improvement in the

transportation infrastructure) may cause a concentration of economic activities in certain regions that have better access to the large markets even if they do not have the lowest production costs. This spatial concentration is beneficial because firms can take advantage of economies of scale by limiting production to locations with a large market and, due to decreased transportation costs, serve other locations from the centrally located sites. This advantage is made possible by commercial integration (pooled market for workers, support the production of non-tradable specialized inputs, informational spillovers).

In addition to the benefit of specialization and economies of scale in Anderson and Lakshmanan’s work, Krugman (1991) also illustrates a negative effect of an improvement in infrastructure capital and the resulting decrease in transportation costs. Since a region with a relatively large non-rural population is not only attractive because of its market potential but also because of the availability of goods and services, this region will attract ever larger populations when transportation costs decrease at the expense of other regions. This means that if one region has slightly more population than another, the former ends up gaining population at the other’s expense when transportation costs fall. To put it differently, a decrease in transportation costs does not only enhance specialization and economies of scale, it also distracts human resources from other regions, leaving those regions relatively

undeveloped. Therefore higher transportation costs militate against regional divergence.

Tabuchi (1998) takes Krugman’s model a step further and combines it with a model developed by Henderson

. Henderson extended the single city model to a model of a system of cities. He incorporates commuting costs and housing space consumption in the model

2

Henderson, J.V. 1974. “The Sizes and Types of Cities”. American Economic Review, Vol. 64.

Deleted: made possible by Deleted: improvements Deleted: is not further elaborated on and

Deleted: as mentioned by Anderson and Lakshmanan (2004),

Deleted: of this literature

Deleted: encouraging to

Deleted: realizing

Deleted: He suggests that it depends on initial conditions which regions end up with the population and concentration of commercial activities.

Deleted: even more

Deleted: when transportation costs fall

Deleted: that region

Deleted: ,

Deleted: put up Deleted: the

which generate agglomeration diseconomies of congestion. On the other hand, there are also localization economies when firms in the same industries agglomerate. In equilibrium, each city specializes in producing an export good and trade these goods between cities.

Tabuchi recognizes that both Krugman’s and Henderson’s model are incomplete. The former ignores intra-urban commuting costs while the latter ignores interregional

transportation costs. Tabuchi combines these two models assuming both positive inter-urban costs and intra-urban costs. In this unified model possible causes for concentration and dispersion of firms and workers between regions are examined. The results show that dispersion of firms and workers takes place in situations with small or large transportation costs, whereas concentration occurs with intermediate levels of transportation costs.

Therefore, when transportation costs decrease monotonically, a transition from dispersion to agglomeration takes place, whereas a further decrease in transportation costs causes a situation of re-dispersion. Tabuchi also found that the welfare level in the dispersed state is usually lower than in the agglomerated state.

Following Krugman’s model in which commercial activity is concentrated in one region while the other region are left underdeveloped, Luo (2004) investigates the role of the infrastructure investment in China and the way in which different regions react to a 10 percent increase in transportation network density. She states that an improvement in transportation facilities is essential to reduce relative geographic disadvantages and to support the catch-up of the western part of China. The underdevelopment of western China limits the potential for domestic market enlargement and hinders the possibility of the relocation process of

industries to these regions. Furthermore, due to their remoteness, the western provinces have limited market accessibility because of high transport costs. Better transportation

infrastructure lowers transportation costs and makes the unfavourable geographic position less important. However, it may also add to a situation in which different regions compete for resources as described in Krugman (1991). The key question Luo investigates is how to efficiently locate infrastructure investment to optimize the effects on regional development.

By simulating the impact of a 10 percent increase in transportation network density in each province, she shows that infrastructure investments in transportation hubs in central China are more effective to improve western development than transportation hubs in the western regions. Hubs in central China lower the effective remoteness of the western region by reducing transport costs between West and East on the one hand and by creating regional

economic centers that lower the peripherical degree of the West on the other hand.

Deleted: Similar to Deleted: is Deleted: somewhat Deleted: location Deleted: improving Deleted: the Deleted: the Deleted: western

Deleted: Improving Deleted: would

Deleted: although

Deleted: The reason for this is that h

Deleted: on the one hand Deleted: on the other hand

The impact of transportation costs on economic growth is assessed on a much larger scale by Sachs (1998,1999). In his papers geography and transportation costs play a crucial role in the economic development of a country. Using geography, Sachs makes a separation between countries in the tropical and sub-tropical zone, and countries that are not. He states that being a tropical country lowers the effect on a countries’ economic growth potential. One of the causes for low economic growth is bad infrastructure, especially for the regions located far from the sea. Sachs states that so-called ‘land-locked countries’ are in an even worse position because they have no connection to the sea. This characteristic is very important for a country’s growth capability because it affects its trade potential. A coastal country has the advantage of a better connection to world markets, which of course affects total transportation costs. In addition, Sachs states that it enables the country access to sea-based trade, which is with no exception the least expensive form of transport. Therefore, a country with a

connection to the sea has serious transportation cost advantages compared to a land-locked country. Cost disadvantages for many land-locked countries can be double the trade costs of maritime countries (Faye, McArthur, Sachs & Snow 2004).

Limão and Venables (1999) extend Sachs’ research by investigating not only the importance of geography on transportation costs, but also by assessing the importance of the levels of infrastructure in a country. They show that infrastructure, both own and

neighbouring countries’ transit routes, is a significant and quantitatively important

determinant of transportation costs and of bilateral trade flows. They measure infrastructure as an index of road density, the paved road network, the rail network, and the number of

telephones per person. Limão and Venables confirm that being landlocked raises

transportation costs by around 50%, but improving the infrastructure from the median for landlocked economies to the 25

highest percentile reduces this disadvantage by 12 percentage points. Improving the infrastructure of the transit economy by the same amount reduces transportation costs by another 7 percentage points. At the same time they compute that the elasticity of trade to transportation costs is at around -2.5, which means that the median landlocked country has only 30% of the trade volume of the median coastal economy.

Furthermore, this percentage could be increased to over 40% improving infrastructure to the highest 25

percentile. Halving transportation costs would increase the volume of trade by a factor of 5.

Deleted: a much larger scale

Deleted: he

Deleted: al

Deleted: He Deleted: the Deleted:

Deleted: Following Sachs, t Deleted: countries’

Deleted: having Deleted: the

Deleted: paper

Deleted: trying to

Deleted: between

Deleted: also

Deleted: and i

Deleted: improving

infrastructure to the 25^th percentile

3. Infrastructure Modalities

In the previous section I have introduced the most important literature on the impact of transportation infrastructure investment and transportation costs on trade and and economic growth and wealth in general. Although this literature on transportation infrastructure

overviews the most important effects on economic growth, it is still deficient in the sense that it uses only aggregate measures for transportation infrastructure to assess the impact on economic growth. The aggregate analysis does not explain much about the individual contributions of different types of infrastructure to economic growth. Limão and Venables (1999), for instance, use an index combining road, rail and telecommunications density as an aggregate measure for infrastructure. But this index only says something about the combined density of infrastructure, not about the individual contributions of the infrastructures the index is made of. One of the main criticisms on Aschauer’s work

also centers around the use of an aggregated instead of disaggregated measure.

Therefore, in this thesis emphasis will be put on the assessment of the individual effects of the different infrastructure modes on economic growth. These infrastructure modes are air transport, ports, roads, railways and telecommunications.In addition to the individual impacts of infrastructure modes, I will also consider the influence of an interaction between the different infrastructure modes on economic growth.

The rationale for assessing the impact of individual infrastructure modes is that the different characteristics of each infrastructure mode give rise to the use of different kinds of infrastructure in distinct circumstances. As Bennathan, Fraser and Thompson (1992) mentioned, railways are most efficient when used for long-haul transport of heavy and raw materials, while road infrastructure offers the best characteristics for transportation and distribution of refined materials and goods and over a shorter distance than a railway network can. For example, in a country like Brazil where materials, goods and persons have to travel over large distances, roads are very inefficient to use and sometimes even non-existent and therefore air transportation might be the best way to move things to another place. Air

transportation is especially efficient when it concerns high value with respect to volume goods and when large distances have to be covered. This is, for example, true for the transportation

5

As mentioned in Gramlich (1994).

8

The use of time-series or a pooled time-series model, such as in Aschauer (1989) and criticized by Aaron (1990), are therefore not feasible.

Deleted: S

Deleted: illustrated Deleted: written Deleted: s

Deleted: This literature can be distinguished in two categories.

The first one is characterized by the empirical literature that assesses the impact of transportation infrastructure investment³ on economic growth or productivity. The second category can be explained by the literature that models the impact of changes in transportation costs⁴ on the economic geography and welfare. Aschauer’s papers belong to the most important literature of the former. His research initiated the idea that there was a linkage between public infrastructure and economic growth while before that time macro-economists had not paid much attention to this area.

Krugman’s pioneering paper (1991), as most important example for the second category, gave rise to new literature that elaborated on the changes in the economic geography and welfare of different regions resulting from an improvement in transportation infrastructure. ¶

¶

Deleted: An

Deleted: effect of infrastructure on economic growth

Deleted: Moreover, o Deleted: ⁶

Deleted: This is done to make a clearer assessment of the linkage than is done in previous literature that only focused on aggregate measures. Furthermore, i Deleted: assess another effect that has never been considered in previous literature:

Deleted: the Deleted: less

Deleted: Another example is Deleted: . In this case Deleted: would be Deleted: overcome Deleted: T

of diamonds in Botswana (Faye et al., 2004). As a consequence, every infrastructure mode has a different impact on economic activities because they are employed in distinctive situations. Therefore, the primary focus of the investigation will be based on the assessment of the hypothesis that different infrastructures have a different individual influence on economic growth.

Due to their dissimilarities, infrastructure modalities are no perfect substitutes for each other. These modalities have distinct characteristics and applications and therefore likely to have different impacts on economic growth. Even if some types of infrastructure are substitutable to a certain extent, like roads and railways, they are never perfect substitutes.

It can be beneficial to have certain infrastructure modalities working together in a complementary way, meaning that if two or more types of infrastructure in a country are present, they interact and strengthen each other. For instance, in a country where only ports are present, there is no way of efficiently distributing the goods. Evidently, it is not realistic to hypothesize that a country only has one type of infrastructure, nevertheless some countries have a very small infrastructure stock which constrains the growth potential of the economy.

When in addition to ports also railways or roads are widely present, more efficient distribution of imported and exported commodities will benefit the economy, because the roads and railways support the other infrastructure variable and the other way around. The same reasoning can be applied to the communication network that makes distribution and

transportation more efficient. This phenomenon will be elaborated on in the second part of the thesis and is the main novelty relative to the presented knowledge on infrastructure and economic growth. The corresponding hypothesis that will be tested is that due to an increase in the efficiency of use , additional economic growth can be observed from the interaction of two complementary infrastructure types.

4. The Model

In this section I will elaborate on the econometric model used for the estimation of the impact of infrastructure on economic growth.

A cross-country study seemed the best suitable way to obtain accurate estimates for the infrastructure variables in determining economic growth, as the time series on

transportation infrastructure contain very little variance around their mean

. In order to have

Deleted: could serve as an illustration for that utilization

Deleted: first

Deleted: will Deleted: impacts

Deleted: In addition, d Deleted: the Deleted: variables Deleted: , and it can be beneficial to have certain types working together

Deleted: If they were substitutable for each other then the effect on economic growth would be the same if a country had a comparable amount of rails, roads, or any other type of infrastructure. Obviously this situation is not very realistic, and e Deleted: This can also be concluded from the fact that they have distinct characteristics and applications and will therefore have different impacts on economic growth.

Deleted: On the contrary, some types of infrastructures may very well be complements, Deleted: available

Deleted: investigation Deleted: addition of this thesis

Deleted: appropriate

Deleted: Due to the fact that the data on transportation

infrastructure contain very little variance around their mean across time, the possibility of using time- series or a pooled time-series model⁷ was excluded. Therefore, a

as much variation in the data as possible, which allows for a better estimation of the model’s coefficients, a large number of different countries were included in the sample. The list of the total number of 74 countries is presented in appendix 1. Since especially data on railways are scarce, the number of countries to be included in the sample was highly dependent on the availability of these railway data. The few countries that do not report data on one of the included variables at any time over the period calculated, are assumed to have a value of zero for that observation

. This is done in order to preserve the sample size.

A common way to estimate the effects of different variables on economic growth is to use a model estimating Barro-style regressions (Barro and Sala-i-Martin, 1995). In this model all variables are included that can possibly have an effect on economic growth and it is determined by an Ordinary Least Squares (OLS) regression stated as:

i i i

i

c X Z

Y = + β

+ β

+ ε (1)

where Y

represents the economic growth in country i, X

a vector of control variables relating to country i and Z

a vector of infrastructure variables relating to country i.

Barro and Sala-i-Martin (1995) included explanatory variables such as initial GDP, educational attainment, life expectancy, investment ratio, public spending on education, political instability, terms of trade, quality of institutions, and several others, to explain GDP per capita growth. While a multitude of possible variables have been suggested in growth literature (see for example Barro, Sala-i-Martin, 1995 or Acemoglu et al 2001), only some of the most commonly accepted variables will be included in the present model to explain the average economic growth rate in addition to the variables measuring transportation infrastructure. These variables will be included because they are the most important basic explanatory variables in Barro and Sala-i-Martin (1995) as well. It concerns the following three variables: the initial GDP per capita, the education level, and the average investment rates. I also include the country’s area, population density and urbanization index because they correlate with the infrastructure variables I introduce in the model and leaving them out may introduce omitted variable bias.

9

Countries for which the observations are assumed to have a value of zero due to unavailability of data are:

Burma (Myanmar) and Cuba with respect to the initial level of GDP per capita and GDP per capita growth. For Sudan it has been done with respect to its investment rate.

11

See also the list of definitions in appendix 2.

Deleted: many

Deleted: In total 74 countries have been included¹⁰.

Deleted: many

Deleted: these

Deleted: this

Introducing these control variables together with the infrastructure variables, the model is stated as follows:

ε β

β β

β β

β β

β

β + + + + + + + + + +

=

Δ Y

X

X

X

X

X

X

....

X

X

(2)

where:

Y = average annual growth of GDP per capita.

X1 to X6 = the control variables as indicated on the previous page

. X7 = the amount of transported passengers by air (in millions).

X8 = ports or navigable rivers (as a percentage of land area).

X9 = fixed line and mobile phone subscribers (per 1,000 people).

X10 = the total kilometers of railway track.

X11 = paved roads (as a percentage of total roads).

ε = error term.

It is assumed that the model’s error term is both normally distributed and has a zero mean ( E [ e ] = 0 ). This assumption implies that the average of all omitted variables and

other errors made when estimating the model is zero. Another assumption of the multiple regression model is homoskedasticity. This means that the variance or the uncertainty of the model is the same for each observation (var [ e ] =

σ

). If this assumption does not hold, then a model contains heteroskedasticity which means that some observations are inherently less influenced by unmeasured factors. This is not an unusual situation with cross-sectional data (Hill, Griffiths and Judge, 2001)

. Also confidence intervals and hypothesis tests may be misleading because they use incorrect standard errors due to the heteroskedasticity. Therefore, a test for the presence of heteroskedasticity will be performed, which can be done with a scatter plot of the least squares residuals against the dependent variable. When no patterns of any kind can be detected, then the model exhibits homoskedasticity satisfying the assumption.

An additional assumption is made with respect to exact collinearity. A further explanation of collinearity can be found in section six where the model is tested. An

12

When heteroskedasticity is present, then the least squares estimator is no longer the best least squares estimator (the so-called BLUE-estimator - Best Linear Unbiased Estimator) although it is still unbiased.

Deleted: where i

Deleted: Some errors will be positive and some negative and will therefore average out to zero.

Deleted: If this assumption holds, then the model is on average correct.

Deleted: ¶

¶

Deleted: . When

heteroskedasticity is present, then the least squares estimator is no longer the best least squares estimator

Deleted: although it is still unbiased

Deleted: This means that none of the explanatory variables can be an exact linear function of any of the other explanatory variables.

Deleted: T Deleted: (

assessment of collinearity will be made with the aim of removing redundant variables from the regression. In this way, a better estimation of the model’s parameters can be made.

Finally, a note has to be made about causality between some of the explanatory variables and the dependent variable. There is no evidence about the precise direction of causality between infrastructure variables and economic growth (Button, 2002, Eisner, 1991).

Although it has been shown that infrastructure stocks affects the economic growth rate

, it seems also very plausible that higher economic growth leads to more infrastructure stock accumulation to satisfy the needs for better and more transportation (Canning, 1998). The latter direction of causality poses some problems to the model. Although reverse causality is likely to be present and to some degree inevitable, lagged values of the independent

variables

have been taken as instruments to minimize this effect.

5. The Data

In this section the infrastructure variables will be discussed, as well as the way data were obtained and the limitations of certain data measures.

The key variable that has to be explained by the infrastructure variables, is economic growth. As in most earlier studies, it is approximated by the increase in real GDP per capita.

Data were obtained from the World Development Indicators and were very complete for all countries in the sample. I decided to take the average linear growth of real GDP per capita over the years 1997-2001. It is calculated by taking the logarithm after dividing the linear growth between 1997 and 2001 by the total number of years. This measure is therefore less affected by volatility of the annual growth rate.

The collection of data for the infrastructure variables was a substantial task. As a starting point, data on roads, railways and phone lines were taken from the World Development Indicators, on air transport from the statistical yearbook of the International Civil Aviation Organization

(ICAO), and the statistics on ports and navigable rivers from

13

See section 2.

15

See measurement description on the next pages.

16

I am grateful to Gerard Ypma (Groningen Growth and Development Centre, University of Groningen) for making this dataset available.

Deleted: the direction of

Deleted: Up until today, t

Deleted: . Whether economic growth influences the stock of infrastructure or the other way around is not clear

Deleted: It Deleted: to think

Deleted: On the contrary, it has been shown that infrastructure stocks also affect the economic growth rate¹⁴. The latter will again be tested for in the following section

Deleted: , but t Deleted: former

Deleted: This problem causality in the direction from the dependent to the independent variables makes it more difficult to obtain and isolate the individual influence of the infrastructure variables on the dependent variable.

Deleted: earlier Deleted: used

Deleted: , which is 5 in this case Deleted: This way of calculation is a more effective way to reflect economic growth during a certain time, because t

Deleted: not

Deleted: rather problematic Deleted: In principle Deleted: most of the Deleted: . The data Deleted: were obtained Deleted: are

Deleted: Thanks t Deleted: .

the Gallup et al. dataset (1998)

. The last two sources were very complete and useful for the dataset, but the World Development Indicators were deficient in data provision on roads and railways. Especially for African and other developing countries data were often incomplete, but the same also applied to some OECD countries like France. Since infrastructure stocks usually change very slowly and cross-country data are used instead of annual observations, there is a scope for interpolating and extrapolating data series containing some gaps. In particular roads and railways showed some gaps in the dataset. Therefore, in some cases I averaged these variables over several years

. In order to check the consistency of road infrastructure data provided by the World Development Indicators, I compared them with the data presented by the World Road Statistics, which was obtained from the International Road Federation. The World Bank obtains its data from this source, but the IRF shows figures of earlier years as well. Since the data on railway statistics for Latin American countries showed some strange movements in the railway stock, I checked them with the Statistical Yearbooks of Latin America (CEPAL, 1999, 2000, 2001, 2002, 2003, 2004). The figures shown in these yearbooks were more consistent but showed on average a quantity of railway stocks

somewhat above those shown in the World Development Indicators. For the reason of greater consistency, I replaced the railway figures of Latin American countries by the figures shown in the Statistical Yearbooks of Latin America.

Another problem arises due to the fact that differences in infrastructural quality across countries are not accounted for in the data base. Hulten (1996) finds that management and efficient use of infrastructure may be more important than the quantity of stock. However, the data used in this investigation do not account for differences in quality across countries as will be explained in the following paragraphs.

Road infrastructure is measured as the average percentage of paved roads to total roads over the period 1993-1996. I use this measure instead of total kilometers of paved roads to express the intensity of road use. Obviously, it is very likely that larger countries have more paved roads in total length. But taken as a percentage of the total road network, paved roads could turn out to be only a small percentage of total roads in small and large countries alike.

For that reason it is a fairly good estimator for the quality of road infrastructure stock in a

17

Thanks to my second supervisor, Jutta Bolt, for obtaining these data for me.

18

Road statistics on the following countries have either been interpolated or extrapolated: Bangladesh, Croatia, Philippines, Romania, Turkey, Uganda. And for railway statistics the following: Burma (Myanmar), Ethiopia, Ghana, Nigeria, Sudan, South Africa.

Deleted: on the other hand

Deleted: sometimes Deleted: although Deleted: problem Deleted: the Deleted: -,

Deleted: used this method to have a suitable

Deleted: for Deleted: a database of

Deleted: Therefore, checking upon the correctness of the World Development Indicators has been done by comparing these figures with earlier figures of the World Road Statistics.

Deleted: the ones Deleted: higher

Deleted: with the method of measuring the infrastructure capital because although figures on the quantity of stock per country are shown,

Deleted: As

Deleted: in a study on the effect of quality and quantity of infrastructure on economic growth,

Deleted: chose to Deleted: taking the Deleted: in the first place because the figures of paved roads as a percentage of total roads are readily available in the World Development Indicators.

Furthermore, using this measure takes into account the differences in total road length between countries and therefore makes them more easily comparable Deleted: .

Deleted: sending them

country as compared to other countries. However, a shortcoming of this measure is that it does not take into account differences in the quality of road surface, or differences in road width. Nor is it a fully reliable measure because for instance a hard-surfaced gravel road in Australia will not be counted as a paved road while in reality it may be of better use than a paved road in Africa.

Railway infrastructure has been calculated as average amount of total kilometers of railway track over the period 1993-1996

. The figures used in this dataset concern railways open to the public and do not accounts from some tracks that are owned by private companies for industrial use and are not open to the public, even though these rail lines may contribute to economic growth as well, like the sugar industry railways in Latin America (Canning, 1998).

Another limitation of this measure is that it does not take into account the quality of the railway track. Furthermore, the importance of having railway terminals to load and offload goods is disregarded here

.

The variable on communication infrastructure is described by the average amount of fixed and mobile phone subscribers per 1000 people over the period 1993-1996

. Various authors find an impact of telecommunications on economic growth (Canning, 1998, Limão and Venables, 1999, Sachs, 1998). In earlier years, fixed lines were the dominating communication technology, but in the most recent decade the mobile phone use has risen dramatically and in some cases it has almost replaced the fixed phone. This happens mainly because of further developments and lower cost of mobile phone use in countries where main lines are scarce.

For the approximation of water transport, the dataset of Gallup et al (1998) has been used

. They measure the proportion of a country’s total area within 100 km of the ocean or

19

The differences in total length between larger and smaller countries is controlled for by adding an extra explanatory variable to the equation measuring the total land area.

20

Another possibility was to take the amount of railway terminals as a measure for railway infrastructure.

However, I found this measure inferior to the total kilometers of railway track because the total amount of terminals does not say anything about the amount or quality of railway infrastructure. For instance, some countries may have four terminals on a distance of 100 km while others have only two. If the country with two terminals has them located more strategically and efficiently, then taking the amount of terminals as a measure for railway infrastructure can give erroneous results. In addition, the size of the railway terminals would not be taken into consideration while size may be an important proxy for railway infrastructure. Another option was to take the flow of transported goods or passengers as a measure for transportation infrastructure. But due to unavailability of figures for especially African countries, this option was taken out of the possibilities. Taking these arguments into consideration, the total kilometers of railway track was the best measure for estimating railway infrastructure.

21

Ideally, it would be best to distinguish between fixed and mobile users but only a combined measure was available in the World Development Indicators for all countries in the sample.

22

Due to unavailability of better data, this proxy for port infrastructure has been used even though it does not take into account the quality and size of a port. It is likely that size and quality of ports in developed countries is higher, but this has not been taken into account in this measure.

Deleted: A Deleted:

Deleted: e

Deleted: Nevertheless, this measure will be used due to the unavailability of better data on roads.

Deleted: take care of Deleted: rail lines

Deleted: with this measure Deleted: . Nevertheless, it seemed the best proxy for estimating railway infrastructure Deleted: total

Deleted: has been

Deleted: determinant Deleted: last

Deleted: it has become easier and cheaper to get a

ocean-navigable river, excluding coastline above the winter extent of sea ice and the rivers that flow to this coastline. As with road infrastructure, this measure only deals with the quantity of infrastructure but not with the quality. Clark, Dollar and Micco (2001) find that improving a port’s efficiency or quality from the lowest 25

percentile to the 75

percentile, reduces shipping costs by 12 percent. Therefore, differences in quality matter, but have not been accounted for in the dataset of this investigation.

The final infrastructure variable is air transportation. A logical measure of air infrastructure would be a measure of total airports in a country. But such a measure is inappropriate because countries with two large international airports would then indicate less air transport capacity than countries with several small airports.

To circumvent this problem, I proxy this infrastructural variable by instead the total flow of passengers using the air transport infrastructure stock in a country. Hence the air infrastructure variable is measured by the average total passengers carried (international and domestic) for the years 1996-1997.

Earlier figures on air transportation were not available.

Variables other than the measures for infrastructure have been obtained from the World Development Indicators. As mentioned above, the first variable to control for initial differences in the level of GDP per capita is measured by GDP per capita for the year 1993

as an instrument variable for the GDP per capita for the 1997 variable. As Barro and Sala-i- Martin (1995) indicate, this instrumental procedure lessens the tendency to overestimate the convergence rate because of temporary measurement error in GDP. Following the absolute convergence hypothesis, countries with a lower initial level of GDP per capita tend to have higher growth rates than countries with a higher initial level of GDP per capita and therefore GDP levels will converge over time. The log is taken so that the very large differences in GDP levels per capita in poor and rich countries are scaled down in order to obtain a better estimation.

The education level is approximated by gross secondary school enrolment for the year 1990. It measures enrolment at a given level of education, regardless of age, expressed as a percentage of the population in the theoretical school-age group corresponding to this level of

23

The CIA World Factbook provides figures about the amount of airports, paved and unpaved, but these figures are not useful because they come from 2006 data.

24

Measured in US dollars with 2000 as base year.

Deleted: last Deleted: rational Deleted: obviously not all airports have the same importance, neither in size or economically.

Therefore

Deleted: of a country’s total airports

Deleted: in such a case, Deleted: be of less importance Deleted: As a result, the problem arises of which airfields to enclose in the sample. How to make a distinction between different capacities or economic importance is a very cumbersome task when the data needed are often not available Deleted: decided that for approximating

Deleted: , not its physical stock should be measured, but Deleted: people Deleted: is

Deleted: Consequently, Deleted: over Deleted: mention

Deleted: because by doing that, Deleted: is

Deleted: have

education.

Since this measure does not include people already enrolled, I have chosen the year 1990 to approximate the level of enrolment for the population contributing to economic growth by the year 1997. Other measures for education such as years of education or the literacy rate were ruled out, because using them would limit the observations total in the dataset.

Another common variable in growth regressions is the investment ratio. In many studies it is said to be one of the main drivers of increases in productivity and growth. The investment ratio is measured by the gross fixed capital formation which according to the European System of Accounts (ESA), consists of “resident producers’ acquisitions, less disposals, of fixed assets during a given period”. It can be seen as the factor capital in a production function. This variable is measured as a percentage of a country’s real GDP. In this study, the average over the years 1990-1995 is taken in order to lessen the effect of measurement error in a single year or an occasional drop in the investment rate.

The other three variables to control for are the urbanization index, population density and a country’s land area. It is useful to control for these variables because indirectly they have an effect on the quantity of infrastructure in a country and the need thereof. In countries with a large land area, it is reasonable to think that more infrastructure is needed to achieve the same efficiency as countries with a smaller area. Therefore, this variable is included and measured by the total area in squared kilometres. When large shares of the population live in cities and just a small share of the population in rural areas, less infrastructure is needed to supply for a large share of the population because they are concentrated in smaller areas. The urbanization index reflects the portion of the total population living in cities. Some countries, like Ethiopia and Cambodia, have urbanization indexes of around 14%, while other countries like Belgium and Argentina have rates of around 90%. A similar variable is included to take into account the density of the population, measured by the total population divided by the total land area in square kilometres. Intuitively, the higher the population density, the more infrastructure is needed because more people make need to make use of a fixed amount of infrastructure supply. The data on these variables entering the regression are the averages of 1993-1996.

25

Therefore, it is possible to have numbers higher than 100% because people outside the theoretical age group participating in secondary education are included. See, for example, Austria, Finland and the Netherlands.

Formatted: English (U.S.) Formatted: English (U.S.) Deleted: Therefore, it is possible to have numbers higher than 100%

because people outside the theoretical age group participating in secondary education are included. See for example Austria, Finland and the Netherlands.

Enrolment at the secondary level is a more complete measure of a population’s degree of education, as the higher this number, the more educated the population is in general.

Deleted: it

Deleted: For simplicity, it can be seen as the factor capital in a production function.

Deleted: to control for the differences in infrastructure provision due to differences in land area and is

Deleted: Nevertheless Deleted: , w Deleted: then

Deleted: To control for these differences, the urbanization index is included in the regression.

In a paper reviewing literature on the effects of public investment on economic growth from the theoretical and empirical literature, De Haan and Romp (2005) provided two

important points of criticism. In the first place they state that the role of outliers has largely been ignored by the empirical growth literature. Secondly, unusual observations are hardly ever controlled for. These issues have been addressed in this study. In the first place, the dataset has been checked for data that are not in line with the rest of the sample and show extreme figures compared to the other data in the sample. Some infrastructure data show large jumps over the time period covered. This might indicate unusual observations which I

checked with figures shown in the CIA World Factbook to ensure the correctness of the data.

In addition, the dataset has been checked for outliers which can cause the regression estimation to be less accurate and should therefore be deleted from the sample. Only two outliers have been excluded. Sweden has been erased from the sample because its figure of fixed and mobile phone subscribers was extremely high which affects the precision of the estimation. OLS measures an average correlation and outliers disturb the slope of the estimation. The figure for Sweden was 867 subscribers per 1000 people while the sample mean was 191 and the median 112. The other outlier was Bangladesh, which showed a population density of 915 people per square kilometre, while the sample mean and the median were 104 and 76 people per square kilometre, respectively. These figures have also been checked with the CIA World Factbook.

In table 1, a summary of the model’s most important descriptive statistics is given including the mean, the median and the maximum and minimum values of the individual samples. The standard deviation, skewness and kurtosis figures measure the normality of the distribution. Since the last three measures show very high numbers for some variables and therefore are likely not to be normally distributed, there was a need to normalize these variables.

A common way to normalize distributions is by taking the logarithm of the variable. For that reason, I added the descriptive statistics of all variables in log-form as well.

In this way, improvements in normality can easily be detected and in case of an improvement, the original measure in linear form was substituted by its measure in log-form.

The average growth rate of GDP per capita over the period 1997-2001 is 1.5%. The sample standard deviation is 2.06. While the median is 1.8%, there are countries that achieved

26

In a normal distribution skewness is around zero and kurtosis around three. Therefore, the samples of the model’s variables should have values close to zero and three, respectively.

Formatted: English (U.S.) Formatted: English (U.S.) Deleted: on

Deleted: , De Haan and Romp (2005)

Deleted: give Deleted: an overview of Deleted: . In addition, they criticize the existing literature on Deleted: ; i

Deleted: Also Deleted: investigation Deleted: a situation of an Deleted: and to deal with it Deleted: the specific data Deleted: In this way, the problem of unusual observations has been dealt with.

Deleted: because outliers Deleted: erased

Deleted: erased Deleted: with

Deleted: can be found

Deleted: in

Deleted: for some variables Deleted:

Deleted: are Deleted: In a normal distribution skewness is around zero and kurtosis around three.

Therefore, the samples of the model’s variables should have values close to zero and three, respectively.

a much higher per capita growth rate such as Albania with an average growth rate of 7.1% or Armenia with an average growth rate of 5.9%. Some other countries show a negative average growth rate, such as several Latin American and African countries. Zimbabwe has the largest negative growth rate of -3.8%. None of the OECD countries show negative figures in the period under consideration.

Large differences can also be found with respect to the level of GDP per capita, since the sample includes not only developing countries but also the rich developed countries. The sample’s poorest country is Tanzania which has a GDP per capita level of US $498. Other countries with a very low level of GDP per capita are Mali, Ethiopia and Malawi with US

$545, US $575 and US $670, respectively. It is very clear that the sub-Saharan African countries have the lowest GDP levels of the sample. While the sample mean is US $7,642, some OECD countries have GDP levels per capita of over US $20,000, such as Austria, France, Italy and Japan. Denmark has the highest level of GDP per capita, almost 50 times higher than the poorest country Tanzania: US $24,322. Since these differences are very large, the logarithm of these figures will be taken in estimating the regression to obtain a scale with less divergence in GDP levels, and a smaller standard deviation.

In order to control for differences in land surface, the total area in square kilometres was included in the regression. As can be seen in table 1, there are large differences in country size. The largest country in the sample is India with an area of 2,973,190 square kilometres, while the smallest country is Slovenia with only 20,151 square kilometres. The sample mean and median are 647,288 and 304,420 square kilometres, respectively. With a very large standard deviation of 1,122,428 there is scope for taking the logarithm of this variable for the regression estimation. In addition, the distribution of this variable is very skewed and shows large kurtosis which means that the values of this variable are peaked and asymmetrically distributed around the mean compared to a normal distribution. Therefore this variable is normalized by taking the logarithm which, as can be seen from the next column in table 1, shows better normality measures.

The last two variables with fairly large divergence in their maximum and minimum values are the population density and the amount of mobile and fixed phone subscribers. The highest population density of the sample is found in the Netherlands with an average density of 455 people per square kilometre. The Republic of Korea is a close follower with a value of 454 people per square kilometre. In general, most Asian countries have high population densities, usually above 100 people per square kilometre. Low population densities are to be found in Africa where Namibia that has the lowest density of only 2 people per square

Deleted: But s

Deleted:

Deleted: calculated

Deleted: exist

Deleted: s Deleted: that has

Deleted: will be

Deleted: , Deleted: in Deleted: following

kilometre, compared to a sample mean of 95 people. Also, the distribution is skewed and peaked, indicating that data are not normally distributed. Instead its distribution is shown when the logarithm is taken. These figures show a rather normal distribution and therefore in the estimation of the regression, population density will be used in log-form.

With respect to mobile and fixed phone subscribers it is very evident that

Scandinavian countries have by far the most subscribers per 1,000 people. Denmark has the highest rate of 748, followed by Finland 728 (Sweden was already excluded from the sample with a rate of 867 subscribers per 1,000 people, because such a high figure compared to the average sample distorts the estimation). As could be expected, most African countries hardly have any subscribers, ranging from 2 subscribers per 1,000 people (Uganda) to 57 subscribers in Tunisia. South Africa is an exception showing a figure of 112, while most sub-Saharan countries only achieve rates of around 10 subscribers per 1,000 people. The sample is fairly normal distributed, its mean is 184 and the standard deviation amounts to 203.

Regarding the other variables much less divergence has been found. Average

investment rates range from 6% of GDP (Cuba) to 41% (Thailand), while most countries have levels of around 20% of GDP, and in some cases around 30% (several Northern African, South-East Asian and Eastern European countries). The sample mean shows an average investment rate of 21% of GDP and the standard deviation is only 6. Also, this variable shows a normal distribution.

The average gross secondary school enrolment rate shows a sample mean of 61%.

With an average rate of 4.7%, Tanzania has the lowest secondary school enrolment rate of the sample. With the exception of South Africa (66.3%), all sub-Saharan countries rate below the sample mean. Northern African countries score better, such as Egypt with a rate of 70.8%.

Many Latin American and Asian countries have enrolment rates below the sample mean, and OECD countries have relatively high rates, with the Netherlands achieving the highest rate of 119.5%. The sample standard deviation is 31.3. The sample distribution is a little flat, which can be seen by its low kurtosis measure, nevertheless it is better distributed than the

logarithmic measure.

Figures on urbanization show that among the countries of this study, 56% of the people lives in urbanized areas. Uganda and Malawi do not even come close to this figure with rates of 12.4% and 12.9%, respectively. In Europe and South-America much higher ratios can be found. Belgium Uruguay and the Netherlands have urbanization ratios of 96.9%, 90.4% and 89%, respectively.

Deleted: ,

Deleted: in the following column,

Deleted: the

Deleted: -

Deleted: - Deleted: mainly Deleted: -

For the transportation infrastructure, the range between maxima and minima has been made much smaller, since all variables have been normalized by either total roads, total land area, or total population. The sample mean for paved roads as a percentage of total roads is 52%. Many OECD countries have reached figures of 90-100%, while African countries sometimes not even attain levels of 10%, such as Tanzania (4.2%) or Uganda (6%). But also some Latin-American countries as Bolivia or Brazil have a paved road stock of less than 10 percent (4.9% and 9.7% respectively). With respect to its distribution, the linear form does a better job as compared to the log-form. Even though the linear form of this variable is a little flatly distributed, the skewness measure is sufficient and normalization assumption is not violated. Therefore the linear form will be used in the estimation.

A similar spread of values can be seen while looking at the figures showing the percentage of country area close to the ocean or ocean-navigable river. Nevertheless, these values are geographically determined and therefore high values are to be seen both for developed and less developed countries. Some countries like the Netherlands, Costa Rica and the Philippines are completely within the reach of the ocean or ocean-navigable river. Less fortunate countries like Mali, Sudan or Zimbabwe have no connection at all and are therefore landlocked. The rest of the sample shows figures ranging from hardly any coastal connection (Ethiopia, 2%) to almost completely connected to the coast (Belgium, 99%). The sample mean for this variable is 49% and standard deviation 35.

The values for the railway data show the total length of railways in a country. On average countries have 6,453 kilometres of railway track with a standard deviation of 10,122 kilometres, which is very high. The country with the largest railway network is India with 62,577 kilometres of railway. With only 468 kilometres of railway, the Philippines has the shortest railway network of all included countries. The sample distribution of this linear measure is both skewed and peaked. However, the measure of total kilometres of railways normalized by the logarithm clearly shows an important improvement in the compliance of the normality assumption and will therefore be used in the regression estimation.

As for the air transportation data, the minimum of the sample is equal to zero because this value has been used to fill the gap for Cambodia. Since there were no values reported for Cambodia, I assumed a zero value. On the other hand, some countries have transported more than double their own population, such as New Zealand (2.6) and Ireland (2.3). However, by far most countries show values between 0 and 1, indicating that compared to their total population only a smaller amount of passengers has been transported by air (including foreigners). The sample mean and standard deviation are 0.3 and 0.47. Similar to the railway

Deleted: -

Deleted: ,

Deleted: ,

Deleted: have