i Master Thesis
University of Groningen Faculty of Economics and Business MSc International Economics & Business
31-08-2018
UNDERSTANDING THE RELATIONSHIP BETWEEN LIFELONG LEARNING AND REGIONAL COMPETITIVENESS – A STUDY ON THE EUROPEAN NUTS2
SCALE
ABSTRACT
Technological change demands stronger connections between education and employment. Lifelong learning is a way to enhance the employability, flexibility and resilience of the labor force which in turn strengthens a region’s competitive position. This thesis examines the relationship between lifelong learning and regional competitiveness for 266 NUTS2 regions in the European Union in the years 2010, 2013 and 2016. After employing pooled OLS and fixed effects regressions, this study finds moderate support for a positive influence of lifelong learning on regional competitiveness. This effect only materializes for higher levels of regional institutional quality.
Keywords: Regional Competitiveness, Lifelong Learning, Regional Labor Markets, Economic Policy.
Author: Florine Marjolein Zendijk Student number: 2348268
ii
TABLE OF CONTENTS
ABSTRACT ... i
LIST OF TABLES AND FIGURES ... iii
1. INTRODUCTION ... 1
2. LITERATURE REVIEW ... 3
2.1 Why does the objective of this study call for an analysis on the regional level?... 3
2.2 Regional competitiveness ... 4
2.3 The role of lifelong learning in developing human capital ... 6
2.4 Institutional quality ... 7 2.5 Population density ... 8 2.6 Summary ... 9 3. METHODOLOGY ... 10 3.1 Model ... 10 3.2 Endogeneity ... 11 3.3 Robustness check ... 12 4. DATA ... 13 4.1 Regional data ... 13 4.2 Regional competitiveness ... 13 4.3 Lifelong learning ... 16 4.4 Institutional quality ... 18 4.5 Population density ... 22 5. RESULTS ... 23
5.1 OLS regression assumptions ... 23
5.2 Pooled OLS estimations ... 25
iii
LIST OF TABLES AND FIGURES
Table 5.1: Pairwise correlations ... 24
Table 5.2: Regression results pooled OLS and fixed effects estimation ... 26
Table 5.3: Regression results pooled OLS with lagged (t-3) explanatory variables ... 30
Table A1: Changes in NUTS 2006 and 2010 towards 2013 ... 42
Table A2: Missing data ... 43
Table A3: Variance Inflation Factors ... 44
Table A4: Robustness check using log unemployment as dependent variable ... 44
Figure 4.1: Regional and national spread long-term unemployment in 2016 ... 15
Figure 4.2: Country-average long-term unemployment rates over time ... 16
Figure 4.3: Regional and national spread lifelong learning in 2016 ... 18
Figure 4.4: Country-average lifelong learning over time ... 18
Figure 4.5: Regional and national spread institutional quality in 2017 ... 20
Figure 4.6: Country-average institutional quality over time ... 21
Figure 5.1: Average Marginal Effects of Lifelong Learning in pooled OLS model ... 27
Figure 5.2: Average Marginal Effects of Lifelong Learning in fixed effects model ... 29
Figure A1: Scatterplot regional competitiveness and lifelong learning ... 40
Figure A2: Scatterplot log regional competitiveness and log lifelong learning ... 40
1
1. INTRODUCTION
‘The United Kingdom is a nation divided by a common economy’ (Parsons, 2018). Its economy faces the largest within-country inequality in Europe. Winners of globalization reside in high-tech office buildings in the City in London, while a four-hour drive northwards, North-Eastern regions experience a completely different development path and are subject to large-scale economic decay and high regional unemployment rates (Annoni, Dijkstra, & Gargano, 2017). Currently, enormous uncertainties exist about the pace and extent of technological change induced by globalization. Some even fear a future of mass unemployment in specific regions (Economist, 2017).
Human capital is essential in absorbing the type of shocks which formerly flourishing regions in the North-Eastern regions of the UK have faced. Guided by knowledge, the employability, flexibility and resilience of the labor force can be improved and labor market participation can be increased, which in turn substantially strengthens the competitive position of regions. However, differences between regions in lifelong learning rates, i.e. the percentage of the population between 25 and 64 years old taking part in education and training, are pronounced. Whereas the lifelong learning rate barely exceeds 10% in Northern Ireland, this percentage in Inner-London is 18.5% and in the Hampshire and Isle of Wright-region even 20%.
2 impact of lifelong learning on competitiveness, a novelty by which this study adds to the literature. In addition, this study will add regional nuance to the existing literature while focusing on the influence of education on various economic indicators. Within countries, marked regional differences in competitiveness and to a lesser extent in lifelong learning exist. Does the one influence the other? If so, under what circumstances and in which specific regions is the effect the most evident? Ideally, regional as well as national educational policy can be finetuned based on this knowledge.
Before we can proceed with our inquiry into these questions, at least one more general question must be answered. What makes regions relevant units of analysis vis-à-vis nations, to achieve the objective of this study? Firstly, people tend to relocate as well as look for jobs within their own region. Secondly, as shown above for the UK, within-country differences seem to be pronounced and analysis on the national level would ignore essential regional nuance. Thirdly, the political authority of regions in Europe has grown largely over the last decade: on average public expenditure managed by regional authorities accounts for a third of total public spending in the EU (Dijkstra, 2017).
As such, the increased global attention to regional competitiveness reflects growing consensus, in academia as well as in politics, that regions are the primary spatial units that compete to attract investments (OECD, 2018b). This focus on regional competitiveness has forced educational systems to become intimately linked to technological and economic development according to Stromquist & Monkman (2014), who recognize education as an ‘undisputable pathway to social mobility’. In the fast-changing and technology-driven era of globalization, lifelong access to up-to-date information and knowledge, combined with the skills to use these resources, is of paramount importance for the economic situation of individuals, regions and countries (European Commission, 2017).
The above clarifies why the analysis is conducted at the regional level. For this analysis, 2010, 2013 and 2016 versions of Eurostat regional data (European Commission, 2018) and the European Quality of Institutions Index (Charron & Lapuente, 2018) are used. Differences between European regions in lifelong learning and regional competitiveness will be extrapolated on a NUTS2 scale using pooled OLS and fixed effects regression analysis. The NUTS2 level is roughly comparable to provinces in most Member States and is the scale level which is regularly used by the EU when applying regional policies. NUTS2 regions have a population ranging from 800,000 to 3 million (Eurostat, 2018b).
3
2. LITERATURE REVIEW
What drives competitiveness? Many contemporary economic studies devote attention to the role of human capital in gaining competitiveness. However, drawing conclusions based on analyses on (supra) national levels can sometimes be oversimplified for labor market analyses. To shine a more regional light on competitiveness, this section summarizes and reviews the literature on economic analysis on lower spatial levels and assesses the role of human capital herein. First, the presumed positive relationship between lifelong learning and regional competitiveness will be explained. Afterwards, the way in which institutional quality and population density supposedly influence this relationship will be discussed.
2.1 Why does the objective of this study call for an analysis on the regional level?
The term ‘region’ is described in the Oxford English Dictionary (2018b) as ‘an area, especially a part of a country or the world having definable characteristics but not always fixed boundaries’. I interpret these ‘definable characteristics’ to be of physical geographical nature or a consequence of historical human impact created by the interaction of people with the environment.
Taking the above into account, why does the objective of this study call for an analysis on the regional level? What makes regions relevant units of analysis in this situation, vis-à-vis nations? Firstly, people tend to move within their own region. Data for the Netherlands show that 80% of all relocations of people are within their own NUTS2 region (Compendium for the Living Environment, 2016). Although the Netherlands is quite small compared to most other Member States, there is no reason to assume that this percentage is significantly lower elsewhere. This region-bound effect is even more evident for lower educated people, as Mincer (1991) finds that geographic mobility increases with education. Mincer shows that inter-regional migration is twice as frequent among workers with sixteen or more years of education, than among people who have received twelve years of education or less. Should I find evidence for a positive impact of lifelong learning on regional competitiveness, this ‘stickiness’ to a region gives a local economy more time to reap the benefits of the initial investment in lifelong learning. The second argument for a regional focus is provided by Porter (2003), who argues that studies focused on (drivers of) competitiveness often take nations as the unit of analysis, whereas many essential determinants of economic performance are to be found at the regional level. The latter is supported by Annoni & Dijkstra (2013) who state that on various indicators, the disparities within a country exceed the differences between countries. Examples of these determinants include unemployment rates, labor productivity and wages. An analysis on the national level can in some cases be too generic, which foregoes local economic particularities as they are absorbed in the national economic picture. Moreover, it can be challenging to compare Malta (420,000 inhabitants) with for example France and Germany with respectively 65 and 82 million inhabitants.
4 regions. Differences in regional competitiveness are rooted in persistent poverty and a lack of opportunities in ‘left behind’ regions. The discontent following this economic decay results in persistent patterns of uneven regional development, which may eventually lead to a major political breakdown (Gordon, 2018; Rodríguez-Pose, 2018). These ‘regions that don't matter’ are highly vulnerable to the threats of populism, as the cornerstones of contemporary economic growth (open markets, migration, economic integration and globalization) are fought in these regions via the ballot box (Horner, Schindler, Haberly, & Aoyama, 2018). Policy can have a fundamental impact on declining the inequality between regions.
Thirdly, the political authority of regions in Europe has grown significantly after the turn of the Millennium (Hooghe, Marks, & Schakel, 2010). More specifically, Charron, Dijkstra & Lapuente (2015) argue that in the last two decades, regions have become more important in managing EU funding. Currently, public expenditure managed by regional authorities accounts on average for a third of total public spending in the EU (Dijkstra, 2017). With these budgets, local authorities perform thorough regional labor market analyses and obtain an accurate overview of the specific human capital abundance and shortages in their region. This detailed picture of the labor market makes regional authorities effective entities to set up appropriate lifelong learning programs.
2.2 Regional competitiveness
Understanding various sources of competitiveness has always played an important role in economics, given the truism that regional competitiveness significantly contributes to the improvement of living standards in an area. In the wake of Smith and Ricardo, literature has developed conceptually as well as empirically, which has brought forth an anthology of factors affecting economic growth levels and rates (Schwab, 2015).
Despite the burgeoning importance of the examination of competitiveness on the regional scale, specific definitions of regional competitiveness are still subject to debate (Kitson, Martin, & Tyler, 2004). Kitson, Martin & Tyler argue (p. 992) in an export-oriented way that ‘at its simplest’, the definition of regional competitiveness is ‘the success with which regions compete with one another in some way’. A broader notion of the term is introduced by Dijkstra, Annoni & Kozovska (2011), who define regional competitiveness as the ‘ability to offer an attractive and sustainable environment for firms and residents to live and work’. Although this definition can very well be aligned with the aim of this study, it is still too broad. Besides, Dijkstra et al. explain later in their paper that their definition covers (regional) institutions as a part of the regional competitiveness, whereas my study explicitly assesses the effect of institutions on regional competitiveness.
5 impression that competitiveness is just a mean, instead of a goal in itself. Regional competitiveness is not something where taking part is more important than winning; regions need to perform nationally and internationally to obtain or sustain a certain standard of living for their inhabitants. Despite the availability of definitions mentioned above, this thesis refers to regional competitiveness as ‘a regional economy’s ability to optimize its labor force in order to compete and prosper in national and global markets and to adapt to change in these markets, ultimately to minimize long-term unemployment’.
Substantial variation in contemporary regional competitiveness can be noticed among formerly flourishing industrial regions in Europe. Some industrial regions used to be specialized in machinery and vehicle production and were, therefore, less dependent on coal resources and heavy industry. Regions taking this development path broadly managed to maintain their industrial wealth (Rosés & Wolf, 2018). Examples include Stuttgart and Lombardy with GDPs per capita in 2016 of 50,000 and 36,000 (Eurostat, 2018a), both among the highest in their respective countries.
Other regions have successfully made the transition from manufacturing towards a service-oriented economy. A striking example is the formerly heavily industrialized region North Rhine Westphalia in western Germany, which has managed to attract a diversified service sector in the last two decades (Tödtling-Schönhofer & Davies, 2013). Similarly, Bilbao in northern Spain has also made a successful transition from heavy industry to services. An extensive strategic plan for the region led to the reconversion of the port and former mining areas, and the installation of new infrastructure. Aided by the strategic plan, the flourishing service sector now forms the backbone of the economy which accounts for 75% of the economic value added (Tödtling-Schönhofer & Davies, 2013).
A dramatic decrease in wealth is the harsh reality for regions which proved unable to adapt their production structure and resource dependency. A textbook example of such a region is Hainaut in Belgium, one of the wealthiest regions in Europe a century ago and currently one of the poorest. Rosés and Wolf (2018) name Wallonia, Wales and Nord-Pas-de Calais as regions sharing this unfortunate fate. All these regions had a GDP per capita of around 25,000 in 2016, among the lowest in their respective countries.
How can these markedly different regional labor markets be explained? Part of the answer to this question can be found in Baldwin’s ‘second unbundling’ (2006), i.e., the decreased coordination and communication costs which made it possible to coordinate complex (manufacturing) activities over long distances. This second unbundling further nurtured de-industrialization in Europe and subsequently induced structural changes on the labor market, where jobs shifted away from formerly industrial regions. Unsurprisingly, these developments led to rising inter-regional inequality (Ezcurra & Rodríguez-Pose, 2013).
6 that efficient labor markets produce incentives for employers as well as for employees to increase their efficiency. This report argues why efficient labor markets can stimulate productivity (and hence competitiveness). The allocation of human capital to the most productive sector allows for an improvement in productivity of that sector, which means that economies must rely on labor market flexibility and human capital adjustment to respond to external shocks.
2.3 The role of lifelong learning in developing human capital
The connection between regional competitiveness and the development of human capital is mainly a result of resources gained because of the region’s competitive position vis-à-vis other regions. These gains can be reinvested to improve the development of human capital in this region. Cross-country regressions show a significant two-way relationship between human capital and economic growth (Mincer, 1996). This connection is expected to exist for regions as well, and allows for vicious or virtuous cycles, where good or bad performance in one region reinforces performance in that particular region and in the other.
The Oxford English Dictionary (2018a) defines human capital as the ‘the skills, knowledge and experience possessed by an individual or population, viewed in terms of their value or cost to an organization or country’. The definition implicitly includes investments in people, for example in the form of lifelong education and training, to increase the productivity of the labor force. Even though there is no panacea for achieving competitiveness, a region’s human capital stock is seen as an essential component of economic performance (Benhabib & Spiegel, 1994). Besides the macroeconomic changes as outlined earlier, the meaning as well as the significance of employment has also changed remarkably on the individual level (Alheit, 2009). Firstly, every younger generation spends a much smaller fraction of their time at work than their parents and grandparents ever did. Secondly, workers experience a wider gap in years between finishing their full-time education and retirement due to increased life expectancy in most European countries. Considering these two changes in combination with the rapidly changing nature of work in many sectors where knowledge is becoming more important than physical resources, it is fair to conclude that the labor market has become more complex.
7 using the following conceptual definition of lifelong learning: ‘All taught formal learning activities undertaken by people in the age group between 25 and 64 years and over which result in an improvement in knowledge, skills and competencies in an employment-related perspective’.
As lifelong learning can be distinguished in different classes, it also serves multiple purposes. Examples include the teaching of workers of necessary vocational skills, upgrading their skills for re-employment, obtaining new qualifications, achieving personal growth and pleasure in the learning activity itself (European Union, 2018). Via which mechanisms do these lifelong learning activities exert their influence on regional competitiveness? According to Alheit (2009), the answer to this question departs from the ‘trivial, overriding consensus that, in the wake of the technological innovations engendered by the post-industrial information society, knowledge has become the key resource of the future’. More specifically, van Leeuwen (1999) discerns two mechanisms through which lifelong learning can contribute to a (regional or national) economic situation. Firstly, lifelong learning generates a better educated and therefore more productive workforce. Thus, through lifelong learning, the effective use of the human capital in the region is increased. As technology advances and the daily tasks for many jobs adjust to these changes, the skills to use digital technologies are considered as being of vital importance to make an area economically competitive (Blake, 2014). A significant share of the labor force had finished its studies well before computers made their entrance in the offices. Lifelong learning is essential to increase the working efficiency on the individual job level, and subsequently to increase individual labor productivity and aggregate regional competitiveness. Secondly, as wages of people undertaking lifelong learning generally increase, this generates additional labor taxes and social security premiums. In a similar vein, should lifelong learning help someone to find a job, this will reduce the social security payments. These relative gains arising from the additionally generated taxes and the lower social security bill can be invested in other areas of the economy to permit sustained investments in the regional competitive position.
Thus, the significance of lifelong learning is beyond doubt. This is also recognized by the United Nations, which in 2015 adopted the ‘Agenda for Sustainable Development’ including a commitment to ‘ensure inclusive and equitable education and promote lifelong learning opportunities for all’ (UNESCO, 2016). Soldi et al. (2016) also acknowledge the improvement of job-specific and transversal skills, the facilitation of transition to employment and the maintaining and updating of skills of the labor force. Nonetheless, labor market forecasts conducted as part of the seventh EU Cohesion Report (2017) still indicate an upcoming shortage of people with vocational education and training qualifications in some Member States (Dijkstra, 2017).
2.4 Institutional quality
8 Quality of government matters substantially for economic development and is an important determinant of regional economic growth (Dijkstra, 2017).Similarly, Rodrik, Subramanian & Trebbi (2004) estimate the respective contributions of institutions, geography, and trade in determining income levels around the world and go as far as to conclude that the quality of institutions trumps every other indicator.
Institutional quality has also proven to significantly and positively influence educational quality (Sokoloff & Engerman, 2000). However, in what way and via which channel this relationship becomes apparent, depends heavily on (institutions related) contingencies of the area under study and the current phase of the relevant area's development. For instance, Well (2007) argues that skilled human capital has a stronger growth-enhancing effect in economies closer to the technological frontier.
As such, how exactly can institutional quality reinforce the presumed relationship between lifelong learning and regional competitiveness? Institutions set the outlines of the ecosystem connecting employment and education. As high quality (local) institutions are assumed to have an accurate overview of the specific skills-shortages in their region, they should be able to equip the labor force with a mix of relevant skills to make a positive impact on the local labor market. Thus, the institutional quality can enhance the effectiveness of lifelong learning by adapting the content of adult education to the needs of the labor market. More effective lifelong learning is, in turn, expected to have a relatively stronger effect on regional competitiveness than lifelong learning which is organized by lower quality institutions.
The general institutional approach departs from differences between locations and then analyses how place-specific institutions affect local economic development (Boschma & Frenken, 2006). Nonetheless, an anthology of definitions of ‘institutional quality’ exists. An all-encompassing, but for the purpose of this thesis too broad definition of institutions is formulated by North (1990): ‘Institutions are the rules of the game in a society or, more formally, are the humanly devised constraints that shape human interaction.’ The ‘rules of the game’ are also stressed by Levchenko (2007), in stating that the quality of institutions equals the ‘quality of contract enforcement, property rights and shareholder protection’. Even so, Levchenko’s (2007) definition is too narrow, as institutional quality has more aspects than solely the legal side. Charron et al. (2015) cover a broader spectrum of institutional quality, which is according to them ‘the extent to which states (regions, red.) perform their required activities and administer public services in an impartial and uncorrupt manner’. This definition aligns well with the purpose of this thesis and will, therefore, be used in the remainder.
2.5 Population density
9 More than century ago, Marshall (1890) was the first to devote academic attention to the clustering of firms (and subsequently people) in the same location, recognizing increasing returns to scale in those clusters. Marshall discerns three possible sources underlying these economies of scale effects: knowledge spillovers, non-traded local inputs and a local skilled labor pool. Knowledge spillovers occur when many firms or people are clustered together in a geographical area and arise from face-to-face contact with and easy access to employees from other firms. Furthermore, he asserts that non-traded local specialist inputs can be beneficial in situations where many firms of the same industry are put together. Lastly, he argues that a local skilled labor pool, a so-called ‘thick local labor market' creates agglomeration advantages by reducing labor acquisition costs, for employers as well as for employees, and through the set-up of specialist courses. Especially the knowledge spillover argument and the thick labor market argument have enjoyed widespread attention in the field of spatial economics.
Similarly, Combes, Duranton, Gobillon & Roux (2010) find a relationship between the level of spatial agglomeration of a region and this region's economic prosperity. The latter is also linked to the sectoral and specialization pattern of a region and the position the region occupies in global value chains. Employment on the higher end of a global value chains generally generates higher income than for example manufacturing employment in the early stages of a global value chain.
Nonetheless, no matter how flourishing a regional economy can be, and despite Marshall's productivity benefits and positive spillover effects, the spatially-boundedness of spillovers is clearly emphasized in the literature (Capello, 2009). Thus, there are limits to the extent to which an economically flourishing city or region can carry the economy of an entire country forward. In fact, the inverse is true. London's prosperity is as vulnerable to Brexit's repercussions as that of the North Eastern part of England. In a similar vein, Trump's policies affect the welfare of America well beyond the so-called ‘flyover and rustbelt’ states where a significant part of his electorate resides (Rodríguez-Pose, 2018).
2.6 Summary
10
3. METHODOLOGY
The methodology-section is subdivided into two parts. The first part explains the econometric model. The second part discusses the methodological challenge of endogeneity and introduces the robustness check.
3.1 Model
This study primarily aims to examine the relationship between lifelong learning and regional competitiveness. Aforementioned papers reason that the connection between regional competitiveness and the development of human capital is primarily a result of resources gained because of the region’s competitive position vis-à-vis other regions. Lifelong learning is, in turn, thought to increase this development of human capital. As lifelong learning increases human capital and as human capital increases regional competitiveness, I expect lifelong learning to positively influence regional competitiveness. Formally, the first hypothesis reads as follows: Lifelong learning has a positive influence on regional competitiveness.
Furthermore, the authors of the cited papers appear to agree that institutional quality is an essential determinant of regional competitiveness. Besides, institutional quality has proven to enhance educational quality. More specifically, institutional quality can improve the effectiveness of lifelong learning by adapting the content of adult education to the needs of the labor market. Consequently, more effective lifelong learning is expected to have a stronger effect on regional competitiveness than lifelong learning which is organized by lower quality institutions. Taking the above into account, the following second hypothesis can be formulated: Institutional quality reinforces the relationship between lifelong learning and regional competitiveness.
The first hypothesis implies that 𝛽1 is significantly larger than zero and the second hypothesis entails that 𝛽3 significantly exceeds the value of zero.
Following these hypotheses, the pooled OLS model takes the following form:
𝑅𝐶 = 𝛽0 + 𝛽1𝐿𝐿𝐿 + 𝛽2𝐼𝑄 + 𝛽3𝐿𝐿𝐿 ∗ 𝐼𝑄 + 𝛽4𝑃𝐷 + ε (1) In equation 1, regional competitiveness (RC) is a non-linear function of lifelong learning (LLL) institutional quality (IQ) and population density (PD). For reasons to be outlined further on, regional competitiveness and lifelong learning have been measured applying their natural logarithms.
To control for unobserved region and time heterogeneity, fixed effects models will be estimated. Fixed effects models remove unobserved regional characteristics which do not vary over time. Examples of factors which are absorbed by a fixed effects model include deeply rooted cultural factors or specific characteristics of the primary, secondary and higher education institutes in a country or region. Other time-invariant region-specific characteristics influencing both regional competitiveness and lifelong learning include wages and the sectoral division in the region. The fixed effects model is to be estimated as follows:
11 In equation 2, competitiveness in region i and year t is a function of the same variables which have been introduced in equation 1. In addition, the model also controls for region and year fixed effects (αit).
Firstly, pooled regression analyses will be performed. Afterward, fixed effects models will be employed. Lastly, pooled OLS models with lagged explanatory variables will be estimated. The rationale behind the lagged models will be explained later on in this chapter. All models are estimated including year dummies, which capture annual fluctuations in the dependent variable, regional competitiveness, that are not caused by a change in lifelong learning or any of the control variables.
3.2 Endogeneity
Generally, endogeneity can occur in two ways: reverse causality and omitted variable bias. Either form potentially causes unreliable regression results, which reflect merely correlations than causal links (Wooldridge, 2012). Reverse causality comes down to the following: some independent variables which significantly influence the dependent variable might also be in a causal relationship the other way around, which casts uncertainty about which variable is driving the result. In case of omitted variables, the effect of an independent variable influencing a dependent variable only works because of the impact of a third variable determining them both. This third variable can either be unknown or unavailable.
As explained before, the link between regional competitiveness and the development of human capital is primarily a result of resources gained because of the region’s competitive position, as these gains can be reinvested to improve the development of human capital. This vicious circle in estimating the economic effects of cognitive skills and education raises endogeneity concerns which are voiced by Hanushek & Woessmann (2012). These authors name cultural differences between countries and differences in economic institutions as factors which are correlated with the economic situation as well as with cognitive skills and the factors determining them. Subsequently, Hanushek & Woessmann (2012) argue that it is ‘virtually impossible’ to identify causality in the relationship between cognitive skills and economic growth, due to the limited observations underlying cross-country growth models and inaccurate measures of cognitive skills. As they cast doubt about the presence as well as the direction of causal links between the dependent and independent variable(s), the caveat posed by Hanushek and Woessmann (2012) concerns both forms of endogeneity.
Institutional indicators also form a ‘usual suspect’ to endogeneity problems (Eicher & Leukert, 2009). This could happen because more competitive regions tend to have had better institutions in the past, as shown in the seminal paper by Acemoglu, Johnson & Robinson (2001).
12 As the use of lagged explanatory variables is held to form a solution to endogeneity problems (see for example Green, Malpezzi & Mayo, 2005), I will also test some models using lagged explanatory variables. Furthermore, Delgado, Henderson and Parameter (2012) conclude that a change in the average level of schooling may be slow to manifest itself in improved economic growth rates. This conclusion, i.e. the existence of a potential delayed effect of lifelong learning on explanatory variables, is tested as well. In doing so, I suppose that a change in institutional quality can also be slowly to manifest itself. The empirical model using lagged explanatory variables takes the following form:
𝑅𝐶𝑡 = 𝛼𝑡−3 + 𝛽1𝐿𝐿𝐿𝑡−3+ 𝛽2𝐼𝑄𝑡−3+ 𝛽3𝐿𝐿𝐿𝑡−3∗ 𝐼𝑄𝑡−3+ 𝛽4𝑃𝐷𝑡−3+ 𝜀𝑡−3 (3) The symbols in equation 3 are the same as one ones introduced in the previous two equations. t-3 refers to a one-period lag, which encompasses three years as the moments of observation have intervals of three years.
3.3 Robustness check
13
4. DATA
Whereas the previous sections outlined the theoretical rationale and the methodological approach in respect of this study, this section explains how the variables have been measures and presents some descriptive statistics and their implications for the analysis.
4.1 Regional data
The unit of analysis of this study is a region at the NUTS2 level. NUTS is a coherent system to classify the EU territory in a way that allows for a statistical comparison between the areas. The NUTS2 level is roughly comparable to provinces in most Member States and is included in EU legislation. NUTS regions do not cover functional economic connections, such as local labor markets, but favor existing administrative units instead. Usually, every three to four years a new NUTS classification is published, following internal territorial changes in Member States.1 Countries encompassing only one single region because of their small sizes are Cyprus, Estonia, Latvia, Liechtenstein, Luxembourg and Malta.2
I focus on the NUTS2 scale as opposed to NUTS1 because in most Member States, the within-region differences in competitiveness in NUTS1 within-regions are still significant. Furthermore, as shown by the Compendium of the Living Environment (2016), people often migrate within a NUTS2 region and the majority of them will stick to the same regional labor market. An analysis on the NUTS1 level would therefore lose important nuance. The NUTS3 level, on the other hand, is too small and detailed for the purpose of this study, as it is hard to determine in which NUTS3 region people gather skills through lifelong learning and which NUTS3 region benefits from those skills. Contrarily, the skill-gathering and skill-using region on the NUTS2 level is more likely to be the same region.
This study maintains the 2013-classification of NUTS2 areas. As revisions in the NUTS classification hamper cross-sectional comparison and comparison over time, all former NUTS2 regions which have been merged or split in the years before or after 2013, have been reclassified into NUTS-2013, allowing to compare the regions over the years. Table A1 in the appendix provides an exact overview of the transformation of all regions into the 2013 classification. These adjustments resulted in an unbalanced panel3, including observations for all 266 NUTS2 regions specified in the NUTS-2013 classification and for all three measuring years (2010, 2013 and 2016). Data availability for both lifelong learning and institutional quality hinders an analysis over more years.
4.2 Regional competitiveness
As stressed in previous sections, this study focuses on the labor market determinants of regional competitiveness, and refers to regional competitiveness as ‘a regional economy’s ability to
1 Please consult http://ec.europa.eu/eurostat/web/nuts/background for an accurate overview of all regions distinguished in NUTS 2013.
2 Because of the pronounced economic differences with regions in Europe’s mainland, the overseas regions of Guadeloupe, Martinique, Guyane and Réunion (all belonging to France) and Ciudad Autónoma de Ceuta and Ciudad Autónoma de Melilla (belonging to Spain) have been omitted from further analysis.
14 optimize its labor force in order to compete and prosper in national and global markets and to adapt to change in these markets, ultimately to minimize long-term unemployment’. The social distress caused by the absence of work is, as stated before, held by some to be one of the fundamental roots of the recent election results in the UK and the US. Efficient and flexible labor markets are recognized as vital components for optimization of the labor force, which subsequently leads to regional competitiveness (Dijkstra et al., 2011). Dijkstra et al. further state that within the broad concept of ‘efficient labor markets’, (un)employment rates are a valuable source of information regarding the activity of the regional economy. Besides, Dijkstra et al. explain that specifically long-term unemployment rates highlight the presence of structural obstacles in a regional economy. Similarly, Wadsworth (2007) reports that unemployment rates are regularly used to express the limited capacity in the labor market or as a measure of social distress caused by the absence of work. In addition, high rates of long-term unemployment indicate that a regional labor market operates inefficiently (OECD, 2018a). Therefore, the regional long-term unemployment rate is used to operationalize regional competitiveness in this study. I consider the long-term unemployment rate to be more reflective of social distress than the overall unemployment rate. Nonetheless, the latter variable is used for robustness tests. The long-term unemployment rate is measured using regional statistics derived from the EU Labor Force Survey (European Commission, 2018). National statistical agencies gathered regional unemployment statistics and these datasets are subsequently harmonized by Eurostat. The long-term unemployment rate is operationalized as the ratio of people who have been unemployed for twelve months or more over all unemployed people.
Figures 4.1, 4.3 and 4.5 contain boxplots showing the within-country variance in long-term unemployment rates, lifelong learning and institutional quality. Furthermore, these figures show how these spreads per country relate to each other. All boxplots show data over 2016. The boxes show the observations that fall between the 25th and the 75th percentile, with a small line indicating the mean. The lines extending the boxes represent the upper and lower adjacent values. The adjacent values represent the upper and lower quartile, plus 1.5 times the distance between the upper and the lower quartile. The dots indicate the values which fall outside the range of the upper and lower adjacent values.
Long-term unemployment rates show wide variation across European regions. Long-term unemployment rates below 1% in 2016 are registered in the German regions Oberbayern, Freiburg and Unterfranken. Regions with low long-term unemployment rates are further located in the United Kingdom, all Scandinavian countries, Romania and the Czech Republic. The reverse side of the spectrum shows the long-term regional unemployment rates of above 21% in Greece (Dytiki Ellada and Kentriki Makedonia). Many Spanish regions also face high long-term unemployment rates.
15 east-west. All regions belonging to the former GDR in East Germany have the highest unemployment rates in the country. These regions are the outlying dots in Figure 4.1.
However, for other countries, the within-country differences in long-term unemployment rates appear to be on the smaller side, especially for smaller countries in which some regions show among the lowest long-term unemployment rates in Europe. In those countries (Scandinavian countries, Austria, the Netherlands) the difference between the lowest and the highest long-term unemployment rate is marginal.
Figure 4.1: Regional and national spread long-term unemployment in 2016
Source: Own calculations based on data regional compiled by the European Commission (2018).
Figure 4.2, 4.4 and 4.6 show respectively long-term unemployment rates, lifelong learning rates and institutional quality scores over time for the five largest EU-countries by population and Denmark, as this is the country with the highest lifelong learning rates. Please note that the lines per country represent an unweighted average of the regional values for that specific variable. Nonetheless, with three years of observations in a panel dataset, it is difficult to identify a clear trend in the data, as the time-span is short and the values for years in between have been estimated.
From Figure 4.2 can be derived that Spain has seen a sharp decrease in the mean long-term unemployment during the most recent financial crisis and in its aftermath. Long-term unemployment rates have decreased in Italy as well, albeit not as sharp as in Spain. The other four countries show a moderate increase (Germany and the United Kingdom) or decrease (France and Denmark) in mean long-term unemployment rates between 2010 and 2016. Thus, the long-term unemployment rate shows both within- and between country variation and is, except for Spain, relatively constant over time. The variation in the data over time hopefully contains valuable information for the regression analysis. However, the lack of substantial variation over time may induce problems for the fixed effects regression analysis, as fixed effects estimators cannot estimate the effects of variables that are individually time-invariant.
16 Figure 4.2: Country-average long-term unemployment rates over time
Source: Own calculations based on data regional compiled by the European Commission (2018).
Thus, regional long-term unemployment rates are used as a proxy for regional competitiveness. In sub-section 5.1 will be explained why the logarithm of the long-term unemployment rate will be used in the regression estimations. I have calculated the natural logarithm (ln) of long-term unemployment, i.e. the logarithm to the base of mathematical constant e (e ≈ 2.718). Furthermore, regions generally aim at long-term unemployment rates which are as low as possible, whereas they strive for high levels of competitiveness. To be able to interpret regional competitiveness as ‘the higher, the better’, I multiplied the term ln( long-term unemployment) with -1. The calculation of the regional competitiveness variable is also shown in equation 4. In the remainder, the variable regional competitiveness is consequently constituted this way.
𝑅𝐶𝑖𝑡 = −(ln (𝐿𝑜𝑛𝑔 𝑡𝑒𝑟𝑚 𝑢𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡𝑖𝑡)) (4)
4.3 Lifelong learning
The data on lifelong learning originates from the Eurostat Regional Education Statistics (European Commission, 2018). This dataset covers school-based general and vocational education and training. The data are collectively processed and harmonized by Eurostat. Lifelong learning has been operationalized as ‘participation of adults aged 25-64 in education and training’. A major downside of this dataset is the exclusion of both informal lifelong learning and based training in the statistics. However, as informal learning and work-based training are more difficult to influence by policymakers than formal learning and as this study aims at providing policy guidance regarding lifelong learning, I expect this data-limitation not to hamper the achievement of the objective of this study.
Figure 4.3 reveals high lifelong learning rates in Scandinavian countries. The Danish capital region Hovestaden outcompetes all other European regions in terms of lifelong learning in
0 5 1 0 1 5 2010 2012 2014 2016 Year Denmark France Germany Italy
17 2016, as 36.1% of the population in the age group 25-64 indicated their participation in education or training in this region. Without exceptions, all Scandinavian region are in the top-18 countries with the highest lifelong learning rates. Other countries with high performing regions on this variable are France, the Netherlands and the United Kingdom.
Many of the lower performing regions are located in Eastern, and to a lesser extent, Southern Europe. Romanian regions Vest and Nord-Vest have respectively 0.9% and 1% of their population participating in lifelong learning. These relatively low values are broadly similar to the ones in many Greek and Bulgarian regions.
Overall, EU15 countries tend to have higher lifelong learning rates than countries that joined the EU in 2004. This could be a result of the policies initiated as a part of EU lifelong learning guidelines that operated before new Member States acceded in 2004, 2007 and 2013 (European Commission, 2000).
Furthermore, Figure 4.3 shows the remarkably low differences in lifelong learning rates between regions in the same country. Recall that the boxes contain all lifelong learning rates between the 25th and 75th quartile. The difference between the 25th and 75th quartile barely exceeds 3% in many countries. The differences between countries tend to be more pronounced. Thus, lifelong learning shows to be much more determined on a national level than on a regional level, as I expected a priori. However, there is some variation in lifelong learning between regions in the same country. Considering this, in combination with the large variation in lifelong learning between countries, still enough variation in the lifelong learning rates is observed to perform the analysis.
18 Figure 4.3: Regional and national spread lifelong learning in 2016
Source: Own calculations based on data regional compiled by the European Commission (2018).
Figure 4.4: Country-average lifelong learning over time
Source: Own calculations based on data regional compiled by the European Commission (2018).
4.4 Institutional quality
The ‘European Quality of Government Index’ (hereinafter referred to as QoG) is used to measure institutional quality. The dataset is based on a project initiated by the Quality of Government Institute of the University of Gothenburg, Sweden and conducted by Nicholas Charron, Lewis Dijkstra and Victor Lapuente. Surveys measuring perceptions of institutional quality and corruption on the NUTS1 as well as NUTS2 level in the EU are conducted in 2010,
0 1 0 2 0 3 0 4 0 L if e lo n g L e a rn in g AT BE BG CY CZ DE DK EE EL ES FI FR HR HU IE IT LT LU LV MT NL PL PT RO SE SI SK UK 5 1 0 1 5 2 0 2 5 3 0 2010 2012 2014 2016 Year Denmark France Germany Italy
19 2013 and 2017. The QoG, funded by the EU and covering all 28 EU Member States, is the first and so far only dataset that enables between as well as a within-country comparison of institutional quality. The first wave in 2010 generated 34,000 responses, whereas the second wave in 2013 resulted in 85,000 returned surveys.
The impact of national-level institutions on regional-level institutions is evident and is exposed in, for example, the national legal system, trade policies and in security systems. Therefore, the QoG takes the following four items into account: control of corruption, the rule of law, government effectiveness and voice, and accountability to measure institutional quality on the national level. To capture regional variation, respondents were asked to rate regionally organized public services such as education, healthcare and law enforcement, based on their own experiences. By composing a regional index, Charron et al. (2015) aimed at adding the national context to regional institutional quality scores and have sought for a careful balance between institutional scores on these two spatial scales. In the resulting composite index, quality of institutions scores range from -2.84 to 2.64, whereby a higher score indicates better quality of institutions.
Despite the aforementioned advantages of using the QoG dataset, its use also has three limitations. Firstly, as admitted by Charron et al., the number of observations per region (only 200 for the 2010 wave) is relatively limited, which may result in representativity issues. Secondly, for five countries4, the QoG data are composed on the NUTS1 level. In Slovenia, data are aggregated to the country level. To be able to compare regional institutional quality among NUTS2 regions, these data had to be transformed to the NUTS2 level. Because the exact QoG cannot be identified for these regions on the NUTS2 level, I have assumed that if a NUTS1 region composes of x NUTS2 regions, those x NUTS2 regions all have the same institutional quality scores as the single NUTS1 region. Sweden is one of the countries with only a NUTS1 classification and NUTS2 regions SE11 and SE12 jointly form NUTS1 region SE1. So, if SE1 has an overall institutional quality score of 2.5, SE11 and SE12 are also assumed to have institutional quality scores of 2.5. This assumption slightly reduces preciseness of the true and unknown institutional quality scores on the NUTS2 level. However, as the regional differences in institutional quality between NUTS2 regions in the same NUTS1 regional are only marginal, I assume that this transformation does not impact the regression results to a significant extent. Thirdly, the QoG data comes in three waves: 2010, 2013 and 2017. The data for lifelong learning and regional competitiveness have been compiled over 2010, 2013 and 2016. In the fixed effects models, the 2017 institutional quality data has been treated as if it were data over 2016. Although this causes a bias in the estimation, the deviation will be only minor as institutional quality is relatively constant over time (see Figure 4.6 on the next page).
Even within Europe, where institutional quality is generally high, vast differences in institutional quality exist between countries. In 2017, the index showed reasonably predictable patterns on the country level as well as on the regional level. The first 38 regions in the ranking are solely located in Scandinavia, the Netherlands or Germany. Moreover, all but two regions (Estonia and Jihovýchod in the Czech Republic) not belonging to the EU15 are in the bottom
20 half of the ranking in 2016. Contrarily, EU15 countries with more than one region in the bottom half of the ranking are Greece (13), Spain (10), France (4), Italy (21) and Portugal (2).
The levels and spread in institutional quality between regions in the same country are shown in Figure 4.5. Countries with lower QoG scores tend to have more regional variation, as can be derived from this figure. The best performing countries, the three Scandinavian countries and the Netherlands, barely show regional variation in institutional quality. The Finnish region Åland outlies this rule with an exceptionally high QoG score.
Although Germany and Austria are federal countries, these countries have lower regional variation in QoG scores than for example Romania, Bulgaria or Czech Republic, which are far more centralized.
Overall, institutional quality on the national level is relatively stable. Figure 4.6 depicts institutional quality over time for six countries over time. Germany and the United Kingdom see a slight improvement in their institutional quality, whereas recent years have seen a fall in institutional quality in Italy. The latter can partially be explained by a noticeable decline in QoG scores in north-western Italian regions Valle d’Aosta and Piemonte, and Abruzzo, a region located halfway the eastern coast. Generally, among the regions showing an institutional quality score below the EU-average are many southern European regions.
As such, the regional variation in institutional quality seems to be sufficient to perform the regression analysis. However, the overall stability of regional institutional quality over time might hamper a fixed effects regression.
Figure 4.5: Regional and national spread institutional quality in 2017
Source: Own calculations based on Charron & Lapuente (2018).
21 Figure 4.6: Country-average institutional quality over time
Source: Own calculations based on Charron & Lapuente (2018).
Figure 4.7 shows the regional institutional quality scores for 2017. Basically, this figure shows the same information as Figure 4.5. Additionally, the map indicates per country where the higher and lower scoring regions are located. This map is intended to clarify the interpretation of the interaction term institutional quality in the next section.
Figure 4.7: Institutional quality score per region
Source: Charron & Lapuente (2018).
22 4.5 Population density
The data to measure the variable ‘population density’ is derived from Eurostat. Eurostat collects demography data from statistical institutes in each Member State. The surface area which is used to calculate the population density per region includes the total area, excluding inland and coastal water. The population density, the number of persons per square kilometer, is calculated as the ratio between the annual average population and the land area.
23
5. RESULTS
This section will present the empirical findings and their economic implications. The section starts with a discussion of the prerequisites for conducting regression analysis. Subsequently, a pooled ordinary least squared estimation is performed. Afterward, region as well as year fixed effects are introduced to address the issue of regions’ varying competitiveness based on unobserved regional characteristics. The results of both the pooled and fixed effects regression models will first be discussed and compared. Afterward, lagged effects will be implemented and a robustness check using the unemployment rate as a proxy for regional competitiveness will be performed. Lastly, the results of the analyses will be discussed.
5.1 OLS regression assumptions
In this sub-section, the assumptions of linear regression analysis will be tested. I will check for linearity, normality of the error term of the model, heteroscedasticity and multicollinearity. All supporting graphs and tables are to be found in the appendix.
Firstly, I will test for linearity between the dependent and independent variable. A scatterplot of the relationship between regional competitiveness and lifelong learning (Figure A1) shows a weak non-linear relationship between the two variables. Taking the natural logarithm of both the dependent variable and independent variables shows the best fit (see Figure A2). Lifelong learning and regional competitiveness will therefore be calculated using their natural logarithm. Secondly, I test whether the errors are normally distributed around their mean. A histogram of the residuals of the regression of lifelong learning on regional competitiveness does not show the desired bell-shaped form (see Figure A3). The distribution is skewed with a long tail on the left side of the mean. Additionally, a Jarque-Bera test for normality has been conducted. The null hypothesis of this test states that the errors are normally distributed. The Jarque-Bera test generates a p-value of 0.000. With this p-value, the null hypothesis is rejected, indicating that the errors are not normally distributed. Nonetheless, as the sample size is reasonably large, the non-normality of the errors is not likely to hamper the estimation of confidence intervals and related hypothesis tests according to the Central Limit Theorem.
Thirdly, OLS regression assumes the absence of multicollinearity, i.e. the case of multiple explanatory variables having a similar or identical variation which might distort OLS estimates. In other words, in the case of multicollinearity, one explanatory variable can explain another explanatory variable to a large extent. Multicollinearity causes a bias in the standard errors of the regression, thereby hampering the reliability of the estimates.
24 Table 5.1 Pairwise correlations Regional competitiveness Lifelong learning Institutional quality Population density Regional competitiveness 1.000 Lifelong learning 0.3612*** 1.000 Institutional quality 0.4851*** 0.6878*** 1.000 Population density -0.1112*** -0.0542 -0.1014*** 1.000
Notes: The correlations have been measured in the original form of the variables, no logarithms. *** Statistical significance at the 1% level ** Statistical significance at the 5% level * Statistical significance at the 10% level.
From Table 5.1 can be derived that all independent variables are significantly correlated with the dependent variable, regional competitiveness. The correlation between lifelong learning and institutional quality appears to be on the high side (ρ = 0.6878). This high correlation raises multicollinearity concerns, even though the VIF method does not detect multicollinearity between any pair of variables. Lifelong learning and institutional quality could (partially) explain the same part of the variance in regional competitiveness. Regression analysis cannot distinguish which part of the variance in regional competitiveness is caused by a change in lifelong learning, and which part of the variance is due to a change in institutional quality. Fourthly, to be able to perform OLS-regression analysis, the standard errors should not be heteroskedastic. Violation of this assumption causes efficiency loss as well as inconsistency in the estimates. To test whether heteroscedasticity is present, a scatterplot showing the data points as well as the fitted regression line for the impact of lifelong learning on regional competitiveness is generated. The scatterplot, as shown in Figure A1, reveals heteroscedasticity. For higher lifelong learning rates the variance in regional competitiveness decreases. Moreover, the additionally performed White test for homoscedasticity with a null hypothesis which predicts homoscedastic errors, generates a p-value of 0.0010. This p-value allows to reject the null hypothesis at the 1% significance level. With this information, I can assume the presence of heteroscedasticity in the data.
Cameron & Miller (2015) pose a caveat that failure to control for correlated errors can lead to ‘misleadingly small standard errors’, which in turn results in misleadingly narrow confidence intervals, large t-statistics and low p-values. To offset the effects of this heteroskedasticity, the use of White (robust) standard errors would be sufficient. However, as I suspect the variance to be homoscedastic within a region, I will use standard errors which are clustered on the regional level. This kind of standard errors is robust to heteroscedasticity effects by calculating average residuals for each region.
25 explain that larger and fewer clusters have less bias but more variability. As this study heavily relies on the assumption that within-country regional differences often exceed between-country regional differences, I cluster the errors on the regional level.
Lastly, OLS regression also assumes stationarity as well as the absence of serial correlation in the data. As my dataset encompasses only three years, I cannot test for the presence of either stationarity or serial correlation. However, the stability of especially lifelong learning and institutional quality over time (see Figures 4.4 and 4.6) gives at first glance no reason for concern about either of these problems.
5.2 Pooled OLS estimations
First of all, a pooled OLS regression of the logarithm of lifelong learning on the logarithm of regional competitiveness (hereinafter referred to as ‘log-log’) will be tested. One may recall equation 4 that regional competitiveness is measured as –(ln(long-term unemployment)). Both lifelong learning and regional competitiveness have been estimated using their logarithms, so the coefficients are to be interpreted as follows. If an explanatory variable changes by one percent, regional competitiveness is expected to change by βx%. Institutional quality and population density have not been measured using their logarithms.
Pooled models ignore the unique attributes of individuals and universal effects across time. Thereby, the errors are clustered on the regional level. The advantage of ‘ignoring’ time-effects in the pooled model is the additional data; each region has now three combined observations for 2010, 2013 and 2016.
The results of the pooled regression models are presented in Table 5.1. All models use regional competitiveness as dependent variable. Model I shows a significant and positive influence of lifelong learning on regional competitiveness (β = 0.45, p < 0.01). As model I is a log-log model, this beta can be interpreted as follows: If lifelong learning increases by 1%, regional competitiveness is expected to increase by 0.45%. The R2 of model I is 0.3064, indicating that 30.64% of the variation in regional competitiveness can be explained by the model.
26
Table 5.2
Regression results pooled OLS and fixed effects estimation
DV: log RC Pooled OLS Fixed effects
I II III IV V VI Log LLL .4461*** (.0434) .1784*** (.0521) .1712*** (.0515) -.1414** (.0565) -.0951* (.0507) -.0951* (.0513) IQ -.0858 (.0996) -.0985 (.0996) .2368** (.1117) .2496** (.1146) Log LLL*IQ .2297*** (.0405) .2365*** (.0411) .1207** (.0461) .1133** (.0484) PD -.0003* (.0002) -.0003* (.0002) Constant -.1.5473*** (.0964) -1.2112*** (.1103) -1.1084*** (.1172) -.3902*** (.1142) -.6237*** (.1027) -.5121*** (.1230) Year dummies
YES YES YES YES YES YES
Observations 779 779 779 779 779 779
F-test 89.40*** 87.98*** 75.71*** 66.72*** 50.45*** 42.52***
R2 within 0.3713 0.4240 0.4266
R2 between 0.1800 0.4036 0.4057
R2 overall 0.3064 0.4398 0.4439 0.0153 0.4011 0.4040
Notes: *** Statistical significance at the 1% level ** Statistical significance at the 5% level * Statistical significance at the 10% level. Region-level cluster robust standard errors are shown between brackets.
27 Figure 5.1: Average Marginal Effects of Lifelong Learning in pooled OLS model
Note: post-estimation corresponding to equation 1.
Population density seems to only moderately impact regional competitiveness. Based on the cited agglomeration literature, I would expect a more pronounced relationship. This might have to do with the way in which lifelong learning has been measured. Informal learning has been excluded from the statistics. However, this form of lifelong learning is reasonably common in daily practice. Following Marshall’s seminal agglomeration advantages theory (1890), ‘information is in the air’ in densely populated areas. Apart from following institutionalized education or training, people can learn informally by learning from a co-worker or Friday night pub conversations with friends.
In assessing the first regression results as outlined above, one should keep in mind that the pooled OLS model does not take the potential presence of unobserved heterogeneity into account. Unobserved heterogeneity would, in this case, imply that regions differ from each other in underlying characteristics which are not included in the model but which might correlate with the other independent variables. In addition, the unobserved heterogeneity can also express over time, in the form of unobserved characteristics which are similar for all regions, for example business cycles, can vary over time. To avoid missing out on individual effects that are unique to a region and cause biases in the regression estimates, fixed effects models are also estimated and their results are presented in the next section.
5.3 Fixed effects
The figures 4.2, 4.4 and 4.6 did not reveal large variation over time in lifelong learning, institutional quality and to a lesser extent in regional competitiveness. However, there is still some variation in the data, giving rise to estimation of fixed effects models in addition to the pooled OLS models. The results of the fixed effects models are shown in the right part of Table 5.1. Model IV suggests that an increase in lifelong learning of 1% is expected to induce a
-1 -. 5 0 .5 1 E ff e c ts o n L in e a r P re d ic ti o n -3 -2 -1 0 1 2 3 Institutional Quality
28 decrease in regional competitiveness of 0.14% (p < 0.05). This model has a remarkably low overall R2 of 0.0153, indicating a marginal explanatory power of the model. This low R2 seems to be caused by a large difference between the within and between R2 in Model IV. The R2 of the same estimation in the pooled model yields an R2 of 0.3064. What might be even more remarkable, is the difference in overall R2 between model IV on the one hand and the models V and VI on the other hand, as the latter models have R2’s of 0.4011 and 0.4040 respectively. Thus, lifelong in itself barely explains the variation in regional competitiveness in the fixed effects model and this variance is largely explained when institutional quality enters the regression. This might point at a misspecification in the fixed effects model IV.
The significant and negative coefficient for lifelong learning in model I seems counterintuitive and contradicts existing literature in the field. Nonetheless, there might be one situation in which lifelong learning could have a negative impact on regional competitiveness. The negative lifelong learning coefficient might reveal a pattern of lifelong learning inducing people who have gained new skills to leave the region in pursuit of happiness elsewhere. This human capital drain might be a cause of decreased regional competitiveness. However, as the study by the Compendium for the Living Environment (2016) shows, most movements take place within the region people already live in, so this is not likely to be the most plausible explanation for the negative coefficient.
To assess the effect of lifelong learning on regional competitiveness for varying values of institutional quality, Figure 5.2 has been created. Like Figure 5.1 corresponding to the pooled OLS model, Figure 5.2 shows a positive impact of lifelong learning on regional competitiveness for higher values of institutional quality. However, following the fixed effects model, this effect only materializes for values of institutional quality that exceed 0.9 whereas for the pooled OLS model, the positive impact of lifelong learning on regional competitiveness already becomes apparent for values of institutional quality exceeding -0.9.
29 Figure 5.2: Average Marginal Effects of Lifelong Learning in fixed effects model
Note: post-estimation corresponding to equation 2.
5.4 Lagged effects
As explained in sub-section 3.2, both lifelong learning and institutional quality could potentially have a delayed effect on regional competitiveness and the use of lagged explanatory variables is believed to form a solution to endogeneity problems. Furthermore, the existence of a potential delayed effect of lifelong learning on explanatory variables, is also tested. Therefore, a pooled OLS model with lagged explanatory variables has also been estimated. Please note that, due to the three-year intervals between the observations, a one period lag encompasses a period of three years.
Generally, the outcomes of the lagged model are similar to those of pooled OLS model without lag. However, the R2 is slightly higher in the complete models (III) without the lagged effects. That is, explanatory variables in the models without lagged explanatory variables explain a slightly more substantial part of the variance in regional competitiveness than the lagged explanatory variables do. This can be interpreted as such that the pooled OLS model without lag is a better specification of the model at hand. Again, the differences are marginal.
Figure 5.3 shows that lifelong learning positively influences regional competitiveness in the lagged model for values of institutional quality exceeding -1.20.
-1 -. 5 0 .5 E ff e c ts o n L in e a r P re d ic ti o n -3 -2 -1 0 1 2 3 Institutional Quality
30
Table 5.3
Regression results pooled OLS with lagged (t-3) explanatory variables
DV: log RC I II III Log LLL .4230*** (.0422) .2486*** (.0510) .2440*** (.0511) IQ -.1392 (.0999) -.1477 (.1000) Log LLL*IQ .2002*** (.0415) .2047*** (.0420) PD -.0002 (.0001) Constant -1.5646*** (.1026) -1.3659*** (.1155) -1.2991*** (.1253)
Year dummies YES YES YES
Observations 779 779 779
F-test 88.84*** 72.95*** 62.02***
R2 0.2874 0.3661 0.3679
Notes: *** Statistical significance at the 1% level ** Statistical significance at the 5% level * Statistical significance at the 10% level. Region-level cluster robust standard errors are shown between brackets.
Figure 5.3: Average Marginal Effects of Lifelong Learning in lagged OLS model
Note: post-estimation corresponding to equation 3.
5.5 Robustness check
The pooled OLS model and pooled OLS model with a one-period lag have also been estimated with the overall unemployment rate (as opposed to the long-term unemployment rate) as a proxy
-. 5 0 .5 1 1 .5 E ff e c ts o n L in e a r P re d ic ti o n -3 -2 -1 0 1 2 3 Institutional Quality
