Agglomeration of economic activity and income inequality between regions in the European Union

(1)

Master’s Thesis

International Economics and Globalisation

Agglomeration of Economic Activity and

Income Inequality between Regions in the

European Union

Rosalie Dieleman (11136480)

rosalie.dieleman@student.uva.nl

Supervisor: Dr. Kees Haasnoot

(2)

Abstract

This thesis researches the e↵ect of agglomeration of economic activity on inequality of per capita GDP between regions in the EU. NUTS2 regional data for 262 regions in 28 EU member states is used, over the time period 2006-2016. The results show that, whereas in the EU15, regional di↵erences of agglomeration have gone up along with between-region income inequality, in the new member states income inequality between regions has gone down, and there was no clear pattern visible in the development of regional agglomeration. Moreover, the new member states experienced a much higher growth of GDP per capita. In the regression analysis, a statistically and economically significant relationship between employment density and GDP per capita was found for all regions. However, statistically significant spatial spillovers of agglomeration on per capita GDP were only found for the EU15, and not for new member states. From the results it is concluded that agglomeration of economic activity contributed to interregional income inequality in the EU15.

(3)

(4)

Introduction

Since the late 1980s, the area within the European Union curving from the United Kingdom to northern Italy has been nicknamed “the blue banana”. As this area was the heart of the EU manufacturing industry, it was visible from space at night as a band of light stretching across Europe. This patterns is also visible in figure 1.1 - which displays regional GDP in 2017 - where there is a darker blue area stretching from the London area to northern Italy displays areas with a relatively high GDP. This graph also reveals a clear core-periphery pattern within the EU, as the southern and eastern European regions have a much lower GDP than the western European regions.

Figure 1.1: GDP per NUTS2 region (2017)

Although the heart of EU economic activity has shifted eastward since the 1980s, there remains to be a divide between the core and peripheral regions. Moreover, this divide may have worsened in the past years as the sovereign debt crisis contributed to disconnecting Greece, Spain and Portugal further from the economic core of the EU (Taylor, 2015). These regional di↵erences in economic activity may also have an

(7)

e↵ect on income inequality between regions, as an article in “The Economist”(2016) suggests that since the global financial crisis “the gap between poor and rich regions in Europe is widening”. Economic inequality between regions has always been high on the EU agenda. Almost a third of the EU budget for the period 2014-2020 - 351.8 billion Euro - has been set aside for “Cohesion Policy”(European Commission, 2018). Cohesion Policy is aimed at “reducing disparities between the various regions and the backwardness of the least-favoured regions”. In the recent year, the inequality in eco-nomic development between regions in the EU has become increasingly topical and its relevance increasingly evident. Both large economic di↵erences between member states as well as regional di↵erences within member states have caused unrest in na-tional and EU politics. The economic di↵erences between northern European and peripheral member states such as Greece, Spain, and member states that acceded to the EU in recent years, are evident. The same holds for regional di↵erences within member states, notably Catalu˜na versus poorer regions of Spain, as well as northern regions of Italy versus its southern regions. Aside from the claim to independence of Catalu˜na, a very prominent example of this recent political unrest is Brexit, as eco-nomic researchers argue that income inequality and poverty boosted the “leave”votes in the Brexit referendum (Darvas & Wol↵, 2016).

It is well established in economic theory that increased international trade, or in-creased economic cooperation such as in the EU, can increase overall welfare. However, this is no guarantee that this increase in welfare is divided evenly over the population. Therefore, it is relevant to question what will happen to peripheral regions as eco-nomic integration in the EU continues. Generally, the presumption has been that manufacturing will tend to move towards the peripheral regions when access to these regions improves and trade costs decrease, as wages in these regions tend to be lower (Krugman, 1991a). However, as Krugman, one of the economists behind “New Eco-nomic Geography”(NEG), argues: “improved access might actually hurt, not help, peripheral industry”(Krugman, 1991a). With improved access, companies might still prefer to localize in agglomerated areas if positive externalities from agglomeration - such as being located close to consumers and intermediate good producers, result-ing to lower transport costs - might outweigh the higher costs of locatresult-ing in highly concentrated area’s.

Since the 1990s, the relatively new field of spatial economics has combined interna-tional trade theories with location theory in order to explain the regional distribution of production. New Trade Theory (NTT) explains specialization in production and trade by economies of scale in production (Amiti, 1998). By benefiting from scale economies, production can take place at lower average cost when the scale of production increases. When trade is possible, production of a certain good is predicted to take place in the country that has the largest market for the good in order to minimize trade costs such

(8)

CHAPTER 1. INTRODUCTION

as the costs of transport of final products. NEG extends NTT by including endoge-nous demand e↵ects. Endogeendoge-nous demand e↵ects occur due to mobility of workers or mobility of firms which demand intermediate goods, as more and more workers and firms move towards areas where production is taking place (Neary, 2001). Increased economic integration, such as in the EU, leads to lower trade costs. NEG suggests that this has the potential to cause a circular process of increased agglomeration of economic activity in certain regions, creating core and peripheral regions (Amiti, 1998).

Some of the stylized facts of these theories have been confirmed in empirical anal-yses (Amiti, 1998). The most geographically concentrated industries in the EU are characterized by economies of scale, and use a large share of intermediate goods in production. Additionally, production takes place in countries with close access to large markets, such as Germany (Amiti, 1998). However, the evidence for ongoing increasing agglomeration of economic activity within the EU is mixed, and depends on the sec-tors that are included in the analysis and the measure of economic agglomeration used (Br¨ulhart & Traeger, 2005). Research tends to focus on sectoral agglomeration and specialization as opposed to agglomeration of the EU economy as a whole,1 _{and recent}

studies on this topic are few. Additionally, there is a lack of research that incorporates data from a larger sample of EU countries. Moreover, research concerning the e↵ect of agglomeration on economic inequality between regions in the EU is limited. Taking Krugman’s suggestion that “improved access might [...] hurt peripheral regions”, the e↵ect of economic agglomeration on regional income inequality seems to be a relevant area of research. When considering that the EU spends a third of its budget with the aims of “reducing disparities between [...] regions”and “reducing the backwardness of the least favoured regions”, it seems crucial to study the possible existence of a process that might be hampering or even o↵setting the e↵ects of this policy.

In order to research agglomeration of economic activity in the European Union and its e↵ects on income inequality between regions, this thesis will focus on the following research question: what is the e↵ect of agglomeration of economic activity on inequality of per capita GDP between regions in the EU? This research will be executed following a two-stage approach. The first stage will focus on assessing whether agglomeration of economic activity and between-region income inequality has increased within the European Union. This will be done by calculating Gini indices for the agglomeration and income variables, and comparing the results over the years in the sample. The Gini index is a measure of inequality, which allows us to see the evolution of inequality over time. The second stage consists of a panel data regression analysis with per capita GDP per region as the dependent variable, and the number of jobs per square kilometre as a measure of agglomeration activity as the relevant independent variable. In order to take account of the spillovers of agglomeration of surrounding regions, “spatial lags”of

(9)

the number of jobs per square kilometre of neighbouring regions are introduced into the regression analysis. The regression analysis will allow us to assess whether geographic concentration of economic activity has a statistically and economically significant e↵ect on regional per capita GDP. For the econometric analysis, regional data from Eurostat concerning 262 sub-national regions of 28 EU member states over the period 2006-2016 is used (Eurostat, 2018).

The remainder of this thesis will be structured in the following manner. The next chapter will give an overview of the literature concerning New Trade Theory and New Economic Geography. Chapter 3 will then discuss the dataset that is used for the analysis. This chapter will address the regions included, the variables of interest and the control variables. Chapter 4 will subsequently discuss patterns of agglomeration and income inequality in the EU found in the literature, and presents and discusses the Gini indices of agglomeration and GDP per capita calculated with the dataset. Chapter 5 then introduces the methodology for the regression analysis, after which 6 presents and discusses the results of the regression analysis. Lastly, chapter 7 will summarize the results, conclude and touch upon policy implications.

(10)

Chapter 2

Literature Review

This chapter will give an overview of the relevant literature concerning the topics and theories introduced in the previous chapter. Section 2.1 will discuss literature concerning New Trade Theory and New Economic Geography. Subsequently, section 2.2 elaborates on the literature relating to the di↵erent mechanisms through which agglomeration has an e↵ect on income and income inequality.

2.1 New Trade Theory and New Economic

Geog-raphy

The main traditional theories of international trade explain trade and specialization on the basis of country or regional characteristics such as abundant factor endowments (Krugman, 1991b). However, these theories, that focus on comparative advantage, were unable to explain spatial concentration of economic activity and the fact that most in-ternational trade takes place between countries with similar characteristics (Ottaviano & Puga, 1998). NTT focusses instead on internal economies of scale in production and product di↵erentiation leading to monopolistic competition. Leading formalisations of NTT are those based the model of Dixit & Stiglitz (1977), and the model by Krug-man (1980), which incorporate scale economies. Internal scale economies encourage production on a large scale by one or very few producers in order to produce at lower cost. The existence of trade costs - such as transport costs (Krugman, 1980) and tari↵ and non-tari↵ barriers (Amiti, 1998) - then induces firms to locate in the country that has a larger market for their goods. Countries therefore specialize in the industries in which domestic demand is relatively large.

NTT did not account for the phenomenon of agglomeration of firms, nor did it account for di↵erences between countries or regions. NEG is an extension of NTT that addresses these issues (Ottaviano & Puga, 1998). NEG formalizes backward and for-ward linkages in the firm’s production process into cumulative causation mechanisms, which ensure that demand di↵erences are endogenous to the model. These linkages are brought about by mobility of firms and mobility of the labour force. Backward link-ages refer to the fact that final-goods producing firms locate near large final-consumer markets, and intermediate goods producers tend to locate close to their markets: end-goods producers. This backward linkage causes vertically linked producers to locate in

(11)

2.2. AGGLOMERATION EFFECTS ON INCOME AND INCOME DISTRIBUTION

each other’s facinity. Forward linkages refer to the trade cost saving of final-goods pro-ducers when they locate near intermediate good propro-ducers (Ottaviano & Puga, 1998). Forward linkages can also encourage consumers to move towards areas of greater eco-nomic activity (Krugman, 1991b). These linkages both have a “centrifugal e↵ect”, that is, they tend to increase agglomeration of firms and labour force. Backward and for-ward linkages have the potential to reinforce each other to cause clustering of vertically linked industries. This causation mechanism allows regions that initially had similar underlying structures to di↵erentiate into an economic core and peripheral regions.

There are, however, also factors that induce a move away from agglomeration. One such e↵ect is the “price index e↵ect”; more firms agglomerating in one area induces increased competition, which lowers the price level and thereby lowers the profits that can be made by firms. This can discourage firms from setting up shop in an area where there is already a lot of economic activity. (Neary, 2001). Another factor that one can think of that discourages agglomeration of firms are rising wages in areas where there is a lot of economic activity. As labour is not perfectly mobile - especially in the EU due to language constraints - wages tend to go up in areas where demand for labour is high. This can a↵ect the localization decision of firms to produce in an area where the cost of labour is lower (Krugman, 1991a).

Other NEG models show that agglomeration tends to take place when trade costs decrease, as the benefits of internal economies of scale can be exploited by producing in one location and transportation at low cost to markets further away. However, suf-ficiently low values of trade costs might o↵set this centrifugal e↵ect again (Krugman & Venables, 1995). Although the lowering of trade costs is expected to raise overall wel-fare, Krugman & Venables (1995) point out that if agglomeration into a smaller number of industrial districts is to take place, this can require a painful adjustment process in which regional unemployment might increase and regional real wages decrease in the peripheral regions. Income inequality between core and peripheral regions is therefore expected to rise during this adjustment process. The following section will discuss more mechanisms though which agglomeration has an e↵ect on income income inequality.

2.2 Agglomeration E↵ects on Income and Income

Distribution

Aside from the e↵ects of increasing or decreasing labour demand on wages, there are other mechanisms linked to agglomeration that a↵ect income. On the one hand, ag-glomeration can a↵ect income and income distribution through agag-glomeration driven externalities. On the other hand, spatial income distribution can be a↵ected through the types of companies and workers that benefit from agglomeration. An empirical

(12)

CHAPTER 2. LITERATURE REVIEW

study using regional data from a sample of European countries, concludes that the e↵ects of agglomeration on productivity are substantial, and estimate the e↵ect of a doubling of employment density to be an increase in productivity of about 5 per-cent (Ciccone, 2002). 19th per-century economist Alfred Marshal, already studied “the advantages which people following the same skilled trade get from near neighbor-hood to one another”. He divided these advantages - now referred to as “Marshallian externalities”- into three types of agglomeration externalities: the benefits of a large pool of skilled labour, local knowledge spillovers and easy access to local customers or suppliers (Diodato, Ne↵ke, & OClery, 2018). The latter has already been addressed as it is a central element of the NEG theory. This section will briefly discuss the other externalities to agglomeration.

One of these Marshallian externalities is recognized as the existence of “thick labour markets”. Labour markets are characterized by high costs for finding a suitable em-ployee, often referred to as “search and matching”costs. Agglomerated areas have access to a large pool of workers with specialized skills, which reduces the cost of search and matching, and can increase the flexibility of a firm to hire and fire accord-ing to output fluctuations (Storper & Venables, 2004). Therefore, labour markets in agglomerated areas are characterized by a high employee turnover with an efficient and rapid search and rehire mechanism. A reduction of these costs increases productivity and allows wages of workers in agglomerated areas to rise. On the other hand, this context of high labour market turnover can also contribute to increased learning of em-ployees and companies, as emem-ployees learn di↵erent skill-sets, and can transfer these skills between companies.

Another mechanism through which Marshallian externalities occur, is through local knowledge spillovers and technological innovation. Although hard to model or measure empirically, the notion that spatial proximity is in some way conducive to information flows which stimulate innovation, is widely recognized among economists (Diodato et al., 2018). A prominent example of the existence of such spillovers is the success of Silicon Valley, as Saxenian (1996) claims that the knowledge spillovers facilitated by its strong social networks is among the key factors causing its innovative success . Empirical research concerning agglomeration e↵ects for various types of jobs in the US, suggests that agglomeration increases income for occupations that are innovation- and creative-based (Gabe & Abel, 2011). Examples of these jobs are engineers, financial executives and information technology workers. They find, however, that personal service providers and low skilled workers do not benefit from agglomeration.

Agglomeration externalities thus contribute to di↵erences in productivity and dif-ferences in income between agglomerated and rural areas. The fact that especially high-skilled sectors such as engineering, information technology and the financial sec-tor benefit from agglomeration externalities, suggests that this can potentially increase

(13)

2.2. AGGLOMERATION EFFECTS ON INCOME AND INCOME DISTRIBUTION

the divide between earnings of high-skilled an low-skilled, as agglomeration causes productivity of the former category to rise faster than the latter. Moreover, as it is precisely these types of companies and high-skilled workers that tend to gain from ag-glomeration, a self-selection of highly skilled workers into agglomerated areas is likely to be part of the e↵ect of agglomeration on income.

(14)

Chapter 3

Data

This chapter introduces the data that are used for the calculation of the Gini indices and the regression analysis. Section 3.1 discusses the regions that are included in the sample. Secondly, section 3.2 discusses the agglomeration parameter and spillover e↵ects. Section 3.3 discusses the control variables that are be included in the regression analysis. Lastly, section 3.4 provides some descriptive statistics for the variables in the dataset.

3.1 Regions

For the empirical analysis, data from the Eurostat regional database is used (Eurostat, 2018). This database contains data of di↵erent classifications of regional levels, which are referred to as NUTS levels (Nomenclature of Territorial Units for Statistics). There are three levels in this classification, of which this research will use the smallest terri-torial unit for which the required data is available, the NUTS2 regions. NUTS regions generally correspond to existing administrative units within the member state. The main criterion for a NUTS region is a population criterion, which for the NUTS2 region is a population between 800.000 and 3 million (Eurostat, 2015). There are, however, exceptions for particular regions such as islands with a smaller population and regions such as the north of Sweden where the land area is vast and the population low. The regions in this dataset thus do not have the same size, which will be compensated for if necessary by dividing the relevant variables by the territory of the region, such that a uniform measure per square kilometre is used.

Appendix figure A.1 presents a map including all the 276 NUTS2 regions currently existing in the EU. As visible on this map, there are a number of overseas territories included in this classification of regions. For the purpose of this research, the regions that are not geographically situated within the European continent, will be discarded from the sample. Notably, these regions are: the French overseas territories Martinique, R´eunion, Mayotte, Guyane and Guadeloupe; the Spanish Canary Islands and Spain’s overseas territories in Morocco, Ciudad Autnoma de Melilla and Ciudad Autnoma de Ceuta; and the Portuguese islands Azores and Madeira. Appendix table A.1 contains the countries included in the dataset and the number of NUTS2 regions for each of the countries.

(15)

3.2. AGGLOMERATION PARAMETER

made. The London area was split up in 2013 from 2 into 5 NUTS2 regions. Therefore, the data for London over the time period 2006-2016 consists of many missing data points, as before 2013 the data is either available for the two ”old” NUTS regions or the whole London area, and after 2013, the data is available for the 5 new NUTS2 regions. Therefore, with the available data, the values for the entire London area were calculated, such that the 5 NUTS2 regions for London were reduced to one region, for which there is no missing data. Besides reducing the missing data issue for an important economic area in the EU, this measure also reduced the outlier value for the London area. Especially the London area ”Inner London West”, which includes Westminister and the City of London, was a big outlier with values for GDP per capita and jobs per square kilometre of twice the maximum of the current dataset.1_{. These}

extreme values compared to other regions in the sample can partially be ascribed to the fact that none of the other NUTS2 regions consist solely of an inner-city banking and business hub. After these modifications to the sample, 262 regions in 28 member states remain.

3.2 Agglomeration Parameter

The variable that will be used to measure agglomeration of economic activity is the number of jobs per square kilometre, as used in Br¨ulhart & Traeger (2005). As the regions in the dataset are unequal in size, writing the variable in terms of units per square kilometre attempts to remove most of the e↵ect of unequal territorial units. Furthermore, in order to be able to include agglomeration in a broader sense and to measure regional spillovers of agglomeration, it is necessary to also take account of agglomeration in the neighbouring regions. As information concerning mutual borders of regions is only available as visual information in the form of geographical maps, it was necessary to create a matrix that provides information concerning the borders between regions within the EU. For this purpose, neighbouring regions are considered to be only those regions that have land borders in common. This information was coded in a matrix in the form of dummies, 1 for bordering regions, 0 for regions that do not share a common border. Within this dataset, a distinction has been made between regional within-country borders, and regional between-country borders, as it is interesting to observe if there are di↵erences in geographical spillovers between regions within a country and between countries.

In the regression analysis, the estimator for within region agglomeration takes ac-count of missing observations. For the calculation of the parameter for agglomeration in surrounding regions - as explained in more detail in section 5.1 - this is not possible.

1_{For example, in UKI3 in 2016, ”Inner Londen - West”, the number of jobs per square kilometre}

(16)

CHAPTER 3. DATA

Therefore, the jobs per square kilometre for the 37 missing observations was extrap-olated using the available observations for those regions and the percentage change in national employment data. The extrapolated data-points are only used for the calculation of the spillover e↵ects.

3.3 Control variables

In order to control in the regression analysis for factors other than agglomeration that have an e↵ect on GDP per capita, various other variables will be introduced in the regression. In the literature, many factors are found to have a significant influence on GDP. The control variables will be limited to those factors that are found to have the largest e↵ect on per capita GDP, and are limited by the availability of data.

A cross-country study by Barro (1996) has found that higher levels of schooling, higher life expectancy, lower fertility rates, lower government consumption, better maintenance of rule of law, lower inflation, and improvements in terms of trade have a positive e↵ect on GDP. Technology is also a driving factor behind economic growth, and therefore a higher technology level can contribute to higher levels of GDP per capita (Solow, 1957) (Ketteni, 2011). Although (Ramrez, 2002) finds that the contribution of infrastructure services to GDP is substantial, this variable will not be included as a control variable as the regional indicators concerning infrastructure have many missing data, as well as it is likely to be closely linked to both GDP as agglomeration, which might cause issues of multicollinearity.

If available, data for the control variables were obtained on the NUTS2 regional level, if this was not available, country level data was used. Appendix table A.2 contains a description of the variables that are used, whether they are regional or country level, and their source.

3.4 Descriptive Statistics

Appendix table A.3 displays descriptive statistics with regards to the variables used in the statistical inference. First of all, it is visible that the amount of missing data is very limited, to a maximum of 1.3 percent per variable. Another noticeable outcome is the fact that there are relatively large divergences between the minimum and maximum values for all variables except for the fertility rate.

For the dependent variable, GDP per capita, the 200 lowest observations all consist of regions in Poland, Hungary, Romania, Bulgaria, and Slovakia, which are relatively “young”EU member states. The minimum value is 6100 Euro, which is the annual GDP per capita in “Nord-Est”Romania (NUTS2 region RO21) in 2006. This region has experienced a growth of around 70 percent over the time period 2006-2016: from

(17)

3.4. DESCRIPTIVE STATISTICS

6100 in 2006 to 10400 in 2016. Similarly, average GDP per capita in Bulgarian region Severozapaden (BG31) grew by around 32 percent, Polish region Podlaskie (PL34) by 55 percent, and Hungarian region ´Eszak-Magyarorsz´ag (HU31) by 33 percent. The growth over the same period in the upper side of the sample is much more modest, for instance around 16 percent for the richest region in the sample, Luxembourg, 15 percent for the London area, 24 percent for Ile de France (FR10), 13 percent for Hamburg, 16 percent for Noord-Holland (NL32) and 3.6 percent for Brussels. This is also visible in figure 3.1, in which the average growth rates of regions in the countries that joined the EU after 1995 eastern European member states with lower level of GDP per capita -have a higher average growth rate over the entire time period.

Figure 3.1: Average annual growth of regional GDP per capita divided into member states that acceded to EU before and after 1995

For the agglomeration parameter, number of jobs per square kilometre, there is a large spread between the observations, ranging from 1.38 in the most northern region of Sweden (Ovre Norrland) to 2747.77 in London. The mean of 142.29 for this variable is, compared to the maximum value, relatively low. Figure 3.2 displays the country averages of regional jobs per square kilometre in 2016. For most countries, this average lies below 200 jobs per square kilometre. When comparing the countries with relatively high averages, these are mostly countries that fit within the “blue banana”pattern as visible in figure 1.1, except for Malta and Czech Republic. When looking at the rela-tionship between the natural logarithms of JPSK and GDP per capita in a scatterplot, depicted in 3.3, we can see that there is a clear positive relationship between the two variables.2

2_{The natural logarithms are used as the range of observation for both variables is large, and the}

majority of the observations at the lower end of the distribution, which makes the scatterplot of the level variables less informative

(18)

CHAPTER 3. DATA

Figure 3.2: Country averages (unweighted) of the number of jobs per square kilometre in 2016

Figure 3.3: Scatterplot of the natural logarithms of GDP per capita and jobs per square kilometre in 2016

When looking at the descriptive statistics of the human capital parameters, there are large di↵erences between the observations in the sample. For tertiary education this ranges between 6.8 of the economically active population in Severoz´apad (Czech Republic) in 2008 to 57.15 percent in London in 2016. Life expectancy ranges from 70.6 (Latvia in 2006) and 85.2 (Madrid in 2016), whereas the standard deviation of 2.6

(19)

3.4. DESCRIPTIVE STATISTICS

is relatively small. It is remarkable that, compared to the spread in the other human capital variables, the di↵erence in fertility rates are rather small. Between all of the regions, the di↵erence between the lowest and highest fertility rate in the entire sample is 1.2, with a the standard deviation of 0.26. When looking at figure 3.4, it is visible that the highest country averages of regional fertility rates in the sample belong to western EU member states, notably the UK, Sweden, Ireland, France and Denmark. The lowest values belong to eastern EU member states as well as peripheral member states, notably Croatia, Cyprus, Poland, but also Greece, Spain and Portugal.

Figure 3.4: Country averages (unweighted) of regional fertility rates, 2016

The governance related parameters also show a great variation in their values, where there is a relatively large di↵erence between the minimum and maximum values of government consumption, but a relatively low standard error. The lowest value for the rule of law index is -0.12 - the observation for Romania in 2006 - which is relatively low on a scale of -2.5 to 2.5. The negative values for ROL in the sample concern observations from Bulgaria and Romania. Another extreme observation can be seen for the variable of inflation, for which the maximum value is of 15.6 percent, which is the observation for Latvia in 2008, at the onset of the financial crisis. It must be noted however, that before the crisis, in 2006, the inflation for Latvia was already relatively high, by around 7.5 percent.

(20)

Chapter 4

Patterns in Agglomeration and

Income Inequality in the EU

Before going on to present the outcomes of the Gini analysis and discuss the patterns of regional agglomeration and income in the EU over 2006-2016, this chapter will first discuss the patterns of agglomeration of economic activity in the EU found in earlier empirical research, in section 4.1, followed by the patterns of regional income inequality in the EU that were found in the data thus far, in section 4.2. Lastly, section 4.3 will discuss the patterns of regional agglomeration and income inequality found in the data by an analysis of Gini coefficients for both variables.

4.1 Agglomeration of Economic Activity in the EU

As mentioned in the introduction, the evidence concerning increasing economic agglom-eration in the EU is mixed. There is support for some of the predictions of NTT and NEG. Amiti (1998) argues that the patterns of geographical concentration and special-ization within the EU seem consistent with these theories. The industries that are most geographically concentrated, are those which are characterized by internal economies of scale, and use a large proportion of intermediate goods in end-goods production. Furthermore, localization of production is found to be consistent with the model’s pre-dictions, and takes place in countries with close access to large markets, such as core EU member states like Germany (Amiti, 1998). Brülhart (1998) also finds that industries characerized by internal economies of scale are localized in the core region of the EU. However, with regards to the trend in agglomeration of economic activity in the EU, research outcomes do not always point in the same direction. Brülhart (1998) based a study on national production data of manufacturing industries, and found that there was statistical evidence for increased localization in the period 1980-1990. However, Brülhart & Traeger (2005), using regional data of 11 EU member states, detect no statistically significant change in geographic concentration of aggregate employment in the EU over the time period 1975-2000. In contradiction to Brülhart (1998), this re-search used regional data at the NUTS2 level. More recent rere-search using country-level data of 14 EU member states, concludes that overall agglomeration of industries in the EU has increased by about 25 percent in the time period from 1970 to 2005 (Krenz & Rübel, 2010). Combes & Overman (2014) point out that changes over time show

(21)

4.2. REGIONAL INCOME INEQUALITY IN THE EU

a mixed pattern: about one third of the EU industries becomes more concentrated, whereas the rest becomes more dispersed for the time period 1970-1990.

The di↵erences in outcomes can be ascribed to di↵erent reasons, such as the fact that di↵erent levels of data were used, there were di↵erent time periods and coun-tries included, there was a di↵erent focus on induscoun-tries, and the fact that di↵erent instruments were used. For instance, Brülhart (1998) and Krenz & Rübel (2010) use “locational Gini indices”, whereas Brülhart & Traeger (2005) use entropy indices to evaluate the level of agglomeration.

4.2 Regional Income Inequality in the EU

The allocation of the EU cohesion budget, aimed at reducing disparities between regions within the EU, is targeted towards the regions that have an average income per capita corrected for purchasing power that is less than 75 percent of the EU average (Eurostat, 2017). Figure 4.1 below gives a clear picture of the relative di↵erences in average PPS income per capita on a regional level, in which the dark red regions are the regions targeted by EU cohesion spending. From this figure, which shows a similar pattern as the figure 1.1 in the introductory chapter, it is clear that regional average GDP per capita in the EU also displays a clear core-periphery pattern. When it comes to developments in this pattern, Bouvet (2010) finds in a panel of 197 European regions over the years 1977-2003, that overall interregional income inequality has decreased since 1977. However, he concludes that this decrease in inequality can be ascribed to a decrease in between-country inequality, whereas regional income inequality within poorer EU member states had actually increased over the same period. Bouvet finds that since the existence of the EU regional inequality in poorer EU member states had increased, while it has not significantly a↵ected regional inequality within richer member states.

4.3 Regional Inequality analysis

After briefly discussing the patterns of agglomeration of economic activity and regional income inequality in the EU found in earlier research, this section will present an anal-ysis of the patterns of agglomeration and regional income inequality in the EU through the discussion of Gini coefficients. Therefore, Gini coefficients will be calculated from the main variables of interest over the 11 years in the sample, that is, GDP per capita, and the number of jobs per square kilometre.

In order to be able to compare income inequality over time, or between regions, it is useful to display inequality as an index. For this purpose, in economics, the Gini

(22)

CHAPTER 4. PATTERNS IN AGGLOMERATION AND INCOME INEQUALITY IN THE EU

Figure 4.1: NUTS2 regional GDP per capita in PPS in relation to the EU-28 average (2015). Source: Eurostat, 2017

index is most widely used. The coefficient is based on the Lorenz curve, which is a cumulative frequency curve that compares the distribution of a variable with a uniform distribution. Therefore, this index ranges from 0 to 1, where 0 stands for equity, as there is no di↵erence between the Lorenz curve and the uniform distribution and 1 is the highest level of inequality. Figure 4.2 represents this concept visually. If the x-axis of this figure would be divided into N equal intervals, the formula for the Gini coefficient would amount to (World Bank, 2014):

Gini = 1 1 N N X i=1 (yi+ yi 1). (4.1)

A lot of research and policy discussions concerning income inequality centres around within-country inequality. Therefore, most Gini coefficients presented in papers and reports concern income inequality for the population within member states, or weighted averages thereof for the entire EU. There is generally fewer attention for between-region inequality. This research will, however, focus on the latter and will assess the pattern of between-region income inequality. Within-region inequality will therefore not be treated in this analysis, due to lack of data and the delimitation of this research. This

(23)

4.3. REGIONAL INEQUALITY ANALYSIS

Figure 4.2: The blue line represents uniform distribution, the black line is the Lorenz curve, which represents the cumulative distribution. Horizontal axis represents the cumulative pop-ulation of the sample, vertical axis cumulative distribution of the variable of interest. The Gini coefficient = A / (A+ B)

section will therefore present between-region income inequality Gini coefficient, in order to assess if, according to this measure, income inequality between regions in the EU has increased, decreased or remained relatively stable. Additionally, in the same manner, Gini coefficients will be calculated from the regional jobs per square kilometre in order to assess whether the EU economy has become more agglomerated.

Column 1 of table 4.1 presents the evolution of the Gini coefficient calculated with average regional GDP per capita, calculated for all regions in the sample. It is visible that, in the year preceding the crisis, the level of income inequality between regions drops slightly, after which it increases again during the recessional years of the crisis. From 2013 until 2016, however, income inequality between regions drops to a lower level than in 2006. In section 3.4, however, it was clear that the members states that acceded to the EU in or after 2004, had a higher growth of GDP per capita compared to the EU15. It therefore could be the case that, similar to the findings of Bouvet (2010), the decrease in inequality - or at least part of it - can be ascribed to a decrease in between-country inequality. This can also be seen in Appendix B.1, which shows the growth in GDP per capita on the country level over the period 2006-2016. It is clear that the highest growth in GDP per capita over this time period was achieved by eastern European member states. This seems to point towards “conditional convergence”, the central idea of neoclassical growth models that poorer countries have a faster growth of GDP per capita when corrected for control variables such as government policy, initial levels of human capital, and so on (Barro, 1996). It is therefore interesting to assess whether the movements of between region income inequality are di↵erent within these two subgroups. Column 2 and 3 present the Gini coefficients for the EU15 and the new

(24)

CHAPTER 4. PATTERNS IN AGGLOMERATION AND INCOME INEQUALITY IN THE EU

Table 4.1: Gini coefficients for average GDP per capita between regions, and agglomeration of economic activity between regions (JPSK). Columns 1 and 4 are calculated with all the EU member states included, 2 and 5 with the member states that entered the EU in or before 1995, and 3 and 6 are calculated from the EU member states that entered in or after 2004.

(1) (2) (3) (4) (5) (6)

GDP GDP GDP JPSK JPSK JPSK

EU28 EU15 NEW EU28 EU15 NEW

2006 20.23 13.90 23.54 65.46 65.06 59.31 2007 20.02 14.25 23.42 65.11 64.72 58.74 2008 19.73 14.56 22.73 65.17 64.81 58.89 2009 19.51 14.51 22.30 65.46 65.08 59.47 2010 19.78 15.11 22.03 65.70 65.21 60.21 2011 20.02 15.82 21.71 65.61 65.05 60.34 2012 20.08 16.20 21.01 65.93 65.40 60.53 2013 20.11 16.32 21.04 66.14 65.64 60.59 2014 19.88 16.29 20.81 66.33 65.91 60.43 2015 19.86 16.35 21.02 66.29 65.97 59.52 2016 19.66 16.29 20.76 66.57 66.23 59.92 Regions 262 204 58 262 204 58

member states. It is visible that the income inequality between regions in EU15 has increased over the time period 2006-2016, whereas income inequality between regions within the subgroup of new member states has decreased. The slight decrease in the Gini coefficient for the EU28 therefore seems to be coming from a decrease in inequality between regions in the new EU member states, as well as a decrease in regional income inequality between the EU15 and new member states. Whereas within the EU15, regional income inequality has increased since 2006.

Column 4 of table 4.1 presents the Gini coefficient for agglomeration of economic activity for all regions in the sample. It is clear that overall, the regional di↵erences in the number of jobs per square kilometre between region has increased over the time period. For the overall sample and for the EU15 (columns 4 and 5), there is a clear upward trend after the economic crisis. With regards to the new member states, column 6 shows a more ambiguous pattern. In the crisis years regional di↵erences employment density between regions seem to rise, however, in 2015 it drops again. Therefore, similar to the findings for between-region income inequality, the overall rise in the Gini coefficient for agglomeration seems to originate from the rise in regional inequality of agglomeration within the EU15.

(25)

agglom-4.3. REGIONAL INEQUALITY ANALYSIS

eration as displayed in the Gini coefficients, two distinct patterns are visible within the data. For the EU15 a rise of between-region income inequality moves along with a rise in between-region di↵erences in agglomeration. For the new member states, no clear trend in regional agglomeration is visible, whereas between-region inequality is going down.

(26)

Chapter 5

Methodology Regression Analysis

This chapter will discuss the methodology for the econometric regression analysis in section 5.1. Subsequently, section 5.2 will discuss the various robustness checks that will be introduced in order to observe if the results of the regression analysis are sensitive to the types of variables used or to specific observations included in the sample.

5.1 Regression Analysis

In order to investigate the e↵ect of agglomeration of economic activity on the average regional GDP per capita, a panel data regression analysis will be executed. The baseline model that will be used as the start for this regression analysis is based on the model that Barro (1996) uses for assessing the determinants of GDP per capita between countries. The regression model will include the following variables, as introduced in section 3.3:

GDPit= 0+ 1EDUit+ 2LIF Eit+ 3F ERit+ 4GOVit+ 5ROLit

+ 6T OTit+ 7IN F Lit+ 8T ECHit+ uit

(5.1)

Where GDP stands for purchasing power standards adjusted GDP per capita, EDU for education, LIF E for life expectancy, F ER for the fertility rate, GOV for government consumption as a percentage of GDP, ROL for an index that measures the rule of law, T OT for the change in the terms of trade, and IN F L for the yearly inflation rate. Appendix A.2 gives a more detailed description of all the variables used. In addition to the parameters included by Barro, a measure of technology is added to the equation. Barro also includes an interaction term between education and GDP and a democracy index. As Barro indicates that these terms only have a small e↵ect on GDP per capita, these variables are left out of the model. This model will serve as a baseline model.

In order to research the e↵ect of agglomeration of economic activity on GDP, it is necessary to consider the economic activity in the region itself, but also to take account of agglomeration of economic activity in the surrounding regions. This is important as we can expect spillovers from heightened economic activity in one region to the surrounding regions. When employment increases in a certain region, surrounding regions might profit from this as people can commute to the other region for jobs.

(27)

5.1. REGRESSION ANALYSIS

Moreover, this increased economic activity could in turn create additional jobs in the surrounding regions. In order to take these possible spillovers into account in the regression analysis, a spatial element will be added to the regression analysis in the form of neighbour e↵ects (Overman & Puga, 2002). The econometric model that is used will therefore essentially be a SLX model, a panel regression model which includes a “spatial lag”of an explanatory variable (Gibbons, 2012). In order to test for the e↵ect of agglomeration of economic activity on GDP per capita, the baseline model will be extended with the following variables, where JP SK stands for number of jobs per square kilometre.

JP SKit+ WJP SKit (5.2)

Where W denotes a binary contiguity matrix, as used by as used by Overman & Puga (2002). This matrix captures the spatial interdependence between the regions by a dummy. This dummy takes the value of 1 for two bordering regions, and zero for regions which do not share a common border. In this matrix, all neighbouring regions are taken into account, that is, both borders between regions in the same country, as well as borders with a region of another EU member state.

Many papers that incorporate a spatial lag in the econometric analysis, compose the spatially lagged variable as a (weighted) average of the values of the spatially interde-pendent regions.1 _{However, as this research is focussing on the e↵ect of agglomeration,}

taking the average of the number of jobs per square kilometre of neighbouring regions might not accurately reflect the overall level of agglomeration. This can be illustrated by an example considering figure 5.1, where the darker blue regions signify a higher

Figure 5.1: Example of regional economic activity

number of jobs per square kilometre. When taking the average of the neighbouring re-gions of UKJ4 and the average of the bordering rere-gions of UKJ1, the values of spatially

(28)

CHAPTER 5. METHODOLOGY REGRESSION ANALYSIS

lagged variable will be very similar, however, UKJ4 is a more secluded region that is mostly surrounded by water, whereas UKJ1 is surrounded by many regions at the heart of agglomeration of economic activity. Therefore, a weighted average might not be the most accurate parameter of the agglomeration in surrounding regions, and the sum of the variable JP SK for surrounding regions is used in the regression analysis instead. However, as a robustness check, the regressions will also be executed with spatially lagged variables which are calculated using the average of of the bordering regions.

In order to be able to distinguish if there are di↵erences in spillover e↵ects between bordering regions of the same country and bordering regions between countries, two separate variables are introduced:

XJP SKit, ZJP SKit (5.3)

Where X denotes the contiguity matrix which gives within country regional interde-pendence, and Z the between country regional interdependence.

Lastly, in order to control for the e↵ect of the economic crisis on the level of GDP per capita, a dummy for recessional years of 2007-2012 will be introduced into the regression.

OLS estimation provides consistent coefficient estimates if the spatial correlation of the dependent variable Y only occurs through exogenous characteristics such as in the SLX model, meaning that the spatial correlation only occurs through the spatial lags of the independent variables included in the model (Gibbons, 2012). However, the standard errors of an SLX model will be inconsistently estimated with OLS. Therefore, heteroscedastic robust standard errors will be applied, following Overman & Puga (2002). Moreover, fixed e↵ects will be included in the model, as the Hausman test applied to both the baseline model and the model with control variables shows that the fixed e↵ects model is a more appropriate fit for the data.2 _{As is customary in}

analysis of economic data, the variables that are related to income are tested for the existence of a unit root. As the existence of a unit root in the data was rejected in a Fisher type test, no further adjustments are made to the data before conducting the regression analysis.3

5.2 Robustness Checks

After this regression analysis, various robustness checks will be introduced in order to assess if the results of the regression analyses are sensitive to various observations

2_{For both models,the null hypothesis that the random e↵ects model is the most appropriate}

spec-ification is rejected at p-value of 0.01.

(29)

5.2. ROBUSTNESS CHECKS

within the sample or to the types of variables used. First of all, as mentioned in section 5.1, a spatial lag for agglomeration of economic activity in neighbouring regions based on an average of the number of jobs per square kilometre (instead of the sum) is introduced into the regression analysis. Secondly, by dropping various observations from the sample, it can be assessed if the results are robust to particular observations or instead heavily dependent on several “extreme”observations in the sample. The observations that will be excluded from the sample are those that are considered to be outliers due to their values or particular geographical circumstances.

The first category of observations that will be excluded from the sample consists of two groups of regions that have no within-country borders. The first group that will be dropped from the sample consists of countries that are made up of one NUTS2 region, as is done by Overman & Puga (2002).4 _{For these countries, due to the delimitation of the}

NUTS2 area, the agglomeration parameter of within country bordering regions is zero, which might a↵ect the estimate of the within bordering regions coefficient. Secondly, regions which consist only of one or multiple islands, and thus are not connected to the mainland, will be excluded of the sample.5 _{As islands, these regions are most likely not}

a↵ected in the same way by the phenomenon of agglomeration of economic activity, nor do they benefit from spillover e↵ects. It is interesting to see if the coefficients are robust to the exclusion of these regions for which agglomeration and agglomeration spillovers are likely not to work according to the same mechanism.

The second category of observations that will be excluded are regions that have extreme values for the observations of the variables JP SK and GDP relative to the average country values, which are partially due to the relatively small delimitation of the NUTS2 region as compared to other regions in the same country. These concern highly urbanized regions such as London, Brussels, Vienna, Berlin and Prague. It is therefore interesting to see if the results of the regression analysis are robust to these outliers in the sample, or instead might be dependent on these regions. Ten regions fall within this in this category.6 _{However, in order to completely exclude}

these observations from the sample, it is also necessary to remove the spillover e↵ects. Therefore, another 24 neighbouring regions are excluded from the sample.

The last robustness check will consist of splitting up the sample between the EU15 and the new member states, as done before in the Gini analysis.7_{. As already discussed}

4_{Estonia, Latvia, Lithuania, Luxemburg, Malta, and Cyprus.}

5_{4 NUTS2 regions of Greece which are solely made up of islands (EL41, EL42, EL43, EL63); the}

Spanish Illes Balears (ES53); the French island Corse (FR83), the Italian islands Sardegna and Sicily (ITG1, ITG2), and the Finnish island Aland (FI20); as well as the EU member states Cyprus and Malta.

6_{London, West Midlands, Manchester, Brussels, Berlin, Hamburg, Bremen, Vienna, Prague and}

Bucharest.

(30)

CHAPTER 5. METHODOLOGY REGRESSION ANALYSIS

in the descriptive statistisc in section 3.4 and in the Gini analysis in 4.3, the economic di↵erences between the EU15 and the new member states are substantial. Therefore, it is interesting to see if the regression analysis provides similar results when separating these samples. Moreover, it is likely that the agglomeration e↵ects, as described in New Economic Geography, do not happen overnight, but instead require some adjustment time in which the economies of the di↵erent member states become more integrated. As the new member states have only joined the EU just before or during the time period in the sample - in 2004, 2006 or 2013 - it is possible that agglomeration e↵ects are not visible in the same magnitude as in the EU15.

(31)

(32)

Chapter 6

Results

6.1 Baseline Model

The results of the baseline model and the regressions in which agglomeration param-eters are included, are presented in table 2. Column 1 presents the estimates of the model as given in equation 5.1. First of all, it is visible that the R-squared - a measure of the explanatory power of the model - for all of the models is relatively high, and has a similar value as the R-squared in the model estimated by Barro (1996).

Secondly, it is noticeable that the only variable that is not significant is the tech-nology parameter. This can perhaps be explained by the fact that this parameter is measured in terms of human resources in science and technology as percentage of the economically active population. As there are also other human capital parameters in-cluded in the model, namely education and life expectancy, this might have rendered the technology parameter in the regression redundant. Because the technology param-eter was not statistically significant in any of the following regressions, the subsequent regressions were run without the inclusion of this parameter.

When considering the signs and sizes of the coefficients in column 1, an interesting outcome is the fact that the coefficient for fertility is positive, whereas the literature predicts that this would be negative. This e↵ect is assumed in theory as popula-tion growth reduces the amount of productive capital per worker, as well as it steers resources away from the productive economy towards childcare. This negative rela-tionship has also been found in empirical research (Barro, 1996). The positive relation between per capita GDP and fertility that was found in this regression, can potentially be related to the low levels of fertility rates, especially for the eastern European coun-tries. As the population replacement level of the total ertility rate in the EU is 2.1, all of the regions in the EU between 2006-2016 had a fertility rate below the replacement rate (Hoorens et al., 2011). At a total fertility rate below the replacement rate, the population is declining and ageing. The economy of an ageing population is charac-terized by a declining share of working population relative to total population. As this decreases the number of taxpayers, increases the cost of healthcare and increases pressure on the pension provision, this can have a negative e↵ect on economic growth. It seems from the results of this regression analysis that the negative relationship be-tween fertility and income does not hold in a similar way for fertility rates below the replacement rate. Is appears as though the negative e↵ects of higher fertility rates play less of a role at the interval of fertility rates below 2.1 and the negative e↵ects weigh

(33)

6.1. BASELINE MODEL Table 6.1: Results (1) (2) (3) (4) (5) (6) GDP GDP GDP GDP GDP GDP JPSK 15.273*** 13.152*** 13.162*** 13.066*** 13.257*** (4.80) (4.86) (4.86) (4.85) (4.69) WJPSK 19.160*** 18.946*** (all) (3.15) 3.08 XJPSK 19.069*** 18.970*** (within) (3.24) (3.22) ZJPSK 21.783* (between) (12.17) EDU 147.140*** 182.794*** 165.895*** 165.735*** 167.416*** 153.699*** (42.41) (24.36) (23.81) (23.67) (24.22) (23.50) FER 7423.019*** 7392.413*** 6135.106*** 6131.960*** 6204.078*** 6607.954*** (890.63) (881.99) (874.79) (873.10) (879.07) (892.57) LIFE 553.732*** 580.998*** 582.394*** 581.315*** 590.908*** 521.602*** (108.02) (108.70) (102.68) (103.56) (103.07) (102.50) GOV -535.518*** -539.487*** -519.731*** -519.737*** -520.760*** -473.378*** (110.12) (110.32) (109.97) (110.01) (109.81) (114.15) TOT 93.722*** 90.762*** 60.561*** 60.461*** 62.841*** 46.111*** (13.38) (13.84) (13.97) (14.01) (13.87) (14.24) INFL -85.385*** -84.442*** -86.102*** -86.564*** -81.807*** –42.784 (27.05) (26.95) (26.12) (26.17) (26.13) (24.12) ROL 4842.755*** 4757.511*** 4324.730*** 4324.959*** 4343.734*** 4139.187*** (572.79) (532.25) ( 526.09) (526.07) (526.56) (521.51) TECH 56.658 (35.05) Dcrisis -479.39*** (74.00) Obs. 2838 2840 2840 2840 2840 2840 R2 _0.452 _0.467 _0.496 _0.496 _0.494 _0.503

standard errors in parentheses * p < 0.1, ** p < 0.05, *** p < 0.01

(34)

CHAPTER 6. RESULTS

more heavily. Moreover, the coefficient of fertility is relatively large compared to the other coefficients. An increase of one child in the total fertility rate is associated with an increase in GDP per capita of around 7400 Euro, with a standard error of over 890. One should however, consider that fertility rates generally move slowly from year to year: the largest country average change of fertility rate within the sample is 0.36 over 10 years. Nonetheless, this estimated e↵ect is relatively large.

There is a lot of variation in the size of the coefficients between the variables, which is in many cases - at least partially - related to the di↵erent units of measurement of the variables. When looking at the other human capital variables included in the model, we can see that the di↵erent human capital variables are all statistically significant and positively associated with movements in GDP per capita. The coefficient for education seems small compared to the coefficient for fertility, but considering that an increase of 1 percentage point in tertiary education attainment of the working population is associated with an increase of overall GDP per capita of around 150, it is actually quite substantial. Especially considering returns to education are generally thought to be diminishing in years of schooling, and this concerns tertiary education, the positive e↵ect of education on GDP per capita is substantial. Furthermore, an increase in newborn life expectancy of a year, which can be seen as a measure of health, is associated with an increase of per capita GDP of around 550.

When considering the government and economy related coefficients, the signs do not show any surprises. However, the coefficient for the rule of law index is relatively large. An increase of one in the rule of law index (which ranges from -2.5 to 2.5) is estimated to have an e↵ect of around 4800 Euro on GDP per capita, with a relatively small standard error. This is a substantial e↵ect considering that the largest yearly change in this variable was around 0.36. Even though approximately 60 percent of the yearly changes was less than 0.05, the estimated e↵ect of such a change in the rule of law index on GDP per capita is still relatively large as it is estimated that this would a↵ect GDP per capita by around 240 Euro. We can furthermore see that the ratio of government consumption to GDP is estimated to have a large negative e↵ect on GDP per capita. As government consumption rises, taxation necessarily increases and resources are steered away from productive activities, and thereby hamper economic growth. This relationship is found in the regression results. The coefficient that seems relatively low is that of inflation, as a one percent change in inflation is estimated to have a negative impact on GDP per capita by 85 Euro.

6.2 Model including Agglomeration

When adding the agglomeration parameter for agglomeration within the region itself, the explanatory power of the model, measured by R-squared, increases slightly. The

(35)

6.2. MODEL INCLUDING AGGLOMERATION

inclusion of the variable JP SK is significant at a 99 percent confidence level, and has, as expected, a positive relation to GDP per capita. A unit increase of this variable is associated with an increase in GDP per capita of around 15 Euro. Compared to the coefficients of the other variables in the model this coefficient might seem small, however, considering the observations of the variable JP SK range from 1.38 to 2747.77, this coefficient is in fact relatively large. Although the standard error is relatively large, accounting for this, the variable JP SK still has an economically significant positive e↵ect on GDP per capita.

When introducing the spatial spillover variable WJP SK, which includes all neigh-bouring regions (within and between countries), it is clear that there is a statistically significant spatial spillover e↵ect of agglomeration, as WJP SK is significant at the 99 percent confidence level. Interestingly, it has an even higher coefficient than JSP K. When adding the spatial lag of the agglomeration parameter, the coefficient for JP SK drops slightly. It is in line with new economic geography to find economic spillovers from surrounding regions, for instance because more economic activity in surrounding regions increases the size of markets and can therefore increase internal economies of scale. However, it is a remarkable outcome that the spillover e↵ect of agglomeration is estimated to be larger than the e↵ect of agglomeration in the own region. Overman & Puga (2002) also find surprisingly strong spillover e↵ects of changes in employment in neighbouring regions , which also points towards significant spillover e↵ects of changes in economic activity in surrounding regions. It is, however, still remarkable that the spillover e↵ect is estimated to be higher than the e↵ect of local economic activity. A possible reason for this outcome can be related to the existence of diseconomies of agglomeration. This is to say that, along with the agglomeration e↵ects discussed in the theories in chapter 2, there might also be a downward pressure of agglomeration on GDP per capita. Possible issues that could create these diseconomies are (traffic) congestion, increased crime rates, and increased pollution (?, ?). Additionally, it is possible that with increasing agglomeration, an increasing amount of resources needs to be steered away from productive resources towards resolving these issues. As it could be the case that the cost of agglomeration - diseconomies of agglomeration - are local and thus have less spillover e↵ects than the benefits of agglomeration - economies of scale and Marhallian externalities -, this could be a possible explanation for the fact that spillovers of agglomeration are higher than local agglomeration e↵ects.

The fourth column introduces the regression estimated with two separate parame-ters for spatial spillovers, for within-country and between-country regional spillovers. When these e↵ects are separated, we can see that the estimate of the e↵ect of within-country spillovers is very close to the estimate of overall spillovers WJP SK. On the other hand, it is visible that the within-country spillover e↵ect has a higher statistical significance than the between-country spillover e↵ect. Although the coefficient value

(36)

CHAPTER 6. RESULTS

of the between-country spillovers resembles that of the within-country spillovers, the standard error for the between-country spillover is relatively large compared to the standard error of the other spillover estimates. This could be related to the fact that the number of between-country borders is limited compared to the number of within-country borders. As there are less observations, the estimate for this coefficient is likely to be more uncertain. Although more uncertain, the estimate still suggests that there are positive spillover e↵ects from neighbouring regions from two di↵erent member states.

Regression 6 introduces a dummy for the economic crisis. The dummy variable is significant at a 99 percent confidence level, has the expected negative sign, and its estimated e↵ect on GDP per capita of approximately minus 480 Euro per year. Moreover, the R-squared of the model increases slightly with the introduction of the dummy. The agglomeration variables are robust to the addition of this parameter as their values only change slightly. However, the variables that are estimated to have a negative e↵ect on GDP per capita, especially inflation, show quite a large change of the coefficient with the introduction of this dummy.

Overall, from this regression analysis, it can be concluded that agglomeration of economic activity within a region, as well as the spillover e↵ects from agglomeration of economic activity in surrounding regions, both within the same country as between countries, have a statistically and economically significant positive e↵ect on regional average GDP per capita. Interestingly, the regression results indicate that the spillover e↵ects of agglomeration in neighbouring regions are higher than the e↵ects of local (within-region) agglomeration.

6.3 Robustness Checks

6.3.1 Di↵erent Calculation Spatially Lagged Agglomeration

Parameter

Table 6.2 presents the results of the regression analysis which includes the spatial lags calculated as averages of agglomeration in surrounding regions, instead of the sum, which was used in the previous section. These regressions show very similar outcomes to those presented in table 6.1. The coefficient for the parameters of within-region agglomeration is robust compared to the regression results in table 1, and does not change considerably. The spatial lag of agglomeration in bordering regions of another country is again only statistically significant at a 90 percent confidence level, and the coefficients for WJP SK (overall) similar to that of XJP SK (within-country), as was the case in the regressions in table 6.1. However, the value of the coefficient for the parameters WJP SK and XJP SK have increased by around fourfold compared to the

(37)

6.3. ROBUSTNESS CHECKS

Table 6.2: Robustness check: spatial lags calculated as average of bordering regions (instead of the sum) (7) (8) (9) GDP GDP GDP JPSK 12.973*** 13.071*** 12.969*** (4.78) (4.61) (4.58) WJPSK 81.614*** 81.051*** (average) (14.14) (14.03) XJPSK 79.566*** (average) (12.19) ZJPSK 62.719** (average) (29.53) EDU 164.921*** 152.435*** 150.943*** (24.03) (23.68) (23.44) FER 6200.771*** 6676.038*** 6321.798*** (867.32) (883.39) (888.64) LIFE 617.321*** 555.196*** 550.521*** (103.30) (102.73) (102.19) GOV -516.560*** -469.302*** -469.849*** (109.57) (113.86) (113.54) TOT 64.910*** 50.033*** 45.637*** (13.74) (14.01) (13.90) INFL -89.699*** -45.580* -50.858** (26.21) (24.21) (24.26) ROL 4197.056*** 4007.089*** 4050.232*** (542.99) (538.36) (529.59) Dcrisis -488.059*** -482.848*** (74.07) (74.69) Observations 2840 2840 2840 R2 _0.490 _0.499 _0.504

Standard errors in parentheses * p < 0.1, ** p < 0.05, *** p < 0.01

regressions in table 6.1. It therefore seems, at first sight, that the parameters are not robust to the di↵erent methods of calculating the spatially lagged variable. How-ever, when considering that the average number of bordering regions in this sample is 4.28, the coefficients for WJP SK presented in table 6.2 comes very close to the coeffi-cients presented in table 6.1. When taking the coefficoeffi-cients produced by regression (8),

(38)

CHAPTER 6. RESULTS

on average, an increase of one job per square kilometre in a bordering region would have an e↵ect of 1

4.27 ⇤ 79 = 18.50. In regession (6) this coefficient was 18.95. The

average number of within-country borders is 3.32, and thus the average e↵ect would be a higher than the comparable result in 6.1, namely around 24. For between-country borders, the average number in the sample is 0.96, which means that the estimate for between-country spillovers is rather high. Overall, the results seem relatively ro-bust to the method of calculating the spatially lagged variable. For WJP SK and XJP SK, the estimated e↵ects on GDP per capita are close to those found in the earlier regressions. For ZJP SK the outcome is less robust, however, this is not sur-prising considering the smaller statistical significance and large standard error in the results in table 6.1. Nevertheless, estimating the same regressions with a di↵erently calculated spatial spillover parameter confirm a positive e↵ect of agglomeration and spatial spillovers of agglomeration on per capita GDP.

6.3.2 Excluding Regions of the Sample

Table 6.3 presents the outcome of the various robustness checks that consist of ex-cluding observations from the sample as discussed in section 5.2. Column 10 presents the results of the exclusion of member states that consist of one region and regions that consist solely of islands. When comparing regression 10 to regression 6, the re-sults seem relatively robust to the exclusion of the regions that do not have any within country bordering regions. The coefficient for spillover e↵ects hardly changes, the same holds when WJPSK is included in the regression as opposed to XJPSK and ZJPSK (regression not included in the table). The only clear change in the coefficients of the agglomeration variables is the fact that there is a slight drop in the coefficient of JPSK.

Column 11 presents the results of the regression in which small, highly urbanized regions where excluded from the sample. When eliminating these regions, the coef-ficient estimates of all agglomeration variables are a lot higher than those presented in table 6.1. The coefficient of JP SK has increased by almost sevenfold, whereas the coefficient for XJP SK - within-country neighbouring regions - has gone up more than half compared to the estimates in table 6.1. Moreover, the variable ZJP SK has gained in statistical significance, and the value of its coefficient has increased, but decreased relative to XJP SK. This seems a more intuitive result, as it is more in line with expectations that economic spillovers within a country are higher between countries. The results of the estimations in the first series of regressions in table 6.1 therefore do not seem to be robust to the exclusion of these highly agglomerated regions, which amounts to around 12 percent of the initial observations. However, this does not imply that the results found in previous sections are obsolete, as it is still evident from the results that agglomeration e↵ects and spillovers do exist, and are in fact estimated to

Agglomeration of economic activity and income inequality between regions in the European Union

Master’s Thesis

International Economics and Globalisation