The Impact of Global Value Chain Participation on Regional Inequality – An Empirical Analysis

(1)

Master Thesis

The Impact of Global Value Chain Participation on

Regional Inequality – An Empirical Analysis

Svea Müller

Student Number: S3975010 / 21800401

Email: s.muller.5@student.rug.nl

Double Degree Programme

M.Sc. Economic Development and Globalization

University of Groningen

M.Sc. International Economics

University of Göttingen

June 16, 2020

Supervisor:

Co-Assessor:

Prof. Dr. Bart Los

Prof. Dr. Krisztina Kis-Katos

(2)

Abstract

This paper empirically analyses the relationship between regional inequality and participation in global value chains, regarding the participation intensity, the functional specialisation and the development level of countries, for 35 high- and middle-income countries between 2000 and 2011 with static and dynamic panel data models. The global value chain revolution leads to the global spreading of unbundled economic activities and the specialisation of countries in tasks. New light is shed on the old question of how trade affects regional inequality. On average a higher participation intensity in global value chains seems to be associated with higher regional inequality. Functional specialisation is identified as a novel key driver of the effect as functions seem to be associated with differently strong agglomeration economies.

(3)

List of Abbreviations

CEEC Central-and Eastern European Countries

CPI Consumer Price Index

FDI Foreign Direct Investment

FE Fixed Effects

FS Functional Specialisation

GMM Generalized Methods of Moments

GVC Global Value Chain

GVCI Global Value Chain Intensity GVCS Global Value Chain Specialisation

ICT Information and Communications Technology

ISIC Rev. 3 International Standard Industrial Classification Revision 3 MNE Multinational Enterprise

NEG New Economic Geography

OECD Organisation for Economic Co-operation and Development

OLS Ordinary Least Squares

R&D Research and Development TiVA Trade in Value Added

TL Territorial Level

VA Value Added

WIOD World Input-Output Database

WIOT World Input-Output Table

(5)

II

List of Tables

Table 1: Summary Statistics ... 22

Table 2: Top 10 Countries - Values and Changes. ... 23

Table 3: Pairwise Correlation Coefficients. ... 25

Table 4: Regression Results of the Pooled OLS Model. ... 26

Table 5: Regression Results of the FE Model. ... 27

Table 6: Regression Results of the FE Model – Developed Countries. ... 28

Table 7: Regression Results of the FE Model – Emerging Countries. ... 28

Table 8: Regression Results of the Dynamic Model. ... 30

Table 9: Regression Results of the Dynamic Model – Lagged Main Regressors. ... 31

Table 10: Regression Results of the FE Model - Theil. ... 32

Table 11: Regression Results of the FE Model - Theil - Subsampling. ... 32

(6)

III

List of Figures

Figure 1: Developments in Hungary, China, and Germany, 2000 – 2011. ... 2

Figure 2: A Stylized Example of a Global Value Chain. ... 8

Figure 3: Value Added and Agglomeration Economies in Functions. ... 12

Figure 4: The Structure of the WIOT ... 19

(7)

1

1. Introduction

The German engineering company and automotive supplier Bosch closes its factory in Bremen, Germany. In the end of 2020, the production of steering columns will be offshored to Eger, Hungary. Only a development department with 30 people will remain in Bremen (Weser Kurier 2019). Bosch is a multinational enterprise (MNE) with presence in 60 countries, continuously opening new production facilities like in Wujin, China in April 2017 (Bosch 2017, p. 13). Cases like these are no rarity in a global world. Massive developments in information and communications technology (ICT) in the recent decades make the geographical spread of production activities possible. This development is known as “the second unbundling of globalization” (Baldwin 2006). Countries become interconnected and no longer depend on domestic resources and technology. Instead of setting up production from zero they can join existing supply chains. Global value chains (GVC) emerge. In the example, Germany adds value in product development, Hungary in assembly. Intermediate inputs from third nations, which also may import parts of the inputs, may be involved. The steering column is reimported by Germany or exported. The same applies to other car components. Trade in the final car implies trade in intermediates and embodies domestic and foreign value added (VA) (Feenstra 1998, p. 39). The decreasing ratio of VA to gross trade represents the GVC revolution (Johnsen und Noguera 2012) which is accompanied by a global cross-country convergence trend as it offers developing countries the possibility to become “emerging markets” (Baldwin 2012). But what happens within countries? How do Bremen, Eger and Wujin develop due to the GVC participation and how does that affect regional inequality within these countries? The existing literature tends to focus on gross trade or the impact of the second unbundling on inequality from the cross-country perspective. This thesis aims to address the missing link by analysing the relationship between regional inequality and GVC participation for 35 countries between 2000 and 2011. Figure 1 shows that Hungary (+86%) and China (+52%) both experienced a drastic increase in GVC participation between 2000 and 2011. Germany experienced a weaker increase of 25% but was already at the beginning of the period double as integrated in the global production network as Hungary and China. This reflects the opportunities of the GVC revolution especially for the relatively less developed countries. The rising integration was accompanied by different trends of regional inequality. In Hungary there was a parallel rise of regional inequality by 13% while it fell in China by 22%. Germany also experienced a decrease of 11%. This shows that GVC participation may affect regional inequality differently across, but also within, different country groups. This thesis aims to find reasons for – and patterns in – the different working mechanisms of GVC participation on regional inequality.

(8)

2

Note: Regional inequality is measured by a Gini index at Territorial Level 2 based on regional GDP per capita and population data from the OECD. GVC participation is measured by GVC income (Timmer et al. 2013) as share of GDP based on data from the WIOD. Detailed information on measurement and data follows in section 3. Source: Own calculations.

The observed parallel overall trend of rising regional inequality and rising GVC participation between 2000 and 2011 shows two things. First, rising regional inequality is a present phenomenon. The Europe 2020 strategy wanted to reach “inclusive growth” leading to “territorial cohesion” (European Commission 2010). Yet, between 2000 and 2011, and still today in 2020, territorial cohesion seems far from reality – not just in Europe. Regional divergence can have severe consequences. Rodríguez-Pose (2018) shows how the “revenge of the places that don’t matter” may look like: political populism and “ballot-box revolts”. In fact, the recent years have been characterised by events like the Brexit or presidential elections which showed the emergence of spatially concentrated populist power. Such developments can determine and endanger the faith of the whole nation.

Second, the trend of rising regional inequality is accompanied by a trend of rising GVC participation. International trade as a possible driver for regional inequality is nothing new. What is new, however, is that trade is no longer about final goods. This sheds new light on the old question of how trade affects regional inequality. When Paul Krugman (1991, p. 486), founding father of the New Economic Geography (NEG), asked where production activities will end up, he thought about a bundle of activities and where it will locate when it is freely mobile between regions within a country. Where will the car production end up within Germany? The GVC revolution extends the location possibilities as it unbundles activities and allows for their global spreading. Where will different activities of car production end up across Germany, Hungary, and China? The following first research question arises:

(9)

3

Does GVC participation impact regional inequality within participating countries? Specified in the light of the observed parallel trends between 2000 and 2011:

Does the increase in GVC participation lead to an increase in regional inequality between 2000 and 2011?

GVC participation implies GVC specialisation as countries carry out certain production stages associated with certain tasks. Hence, the second research question is:

Is the impact of GVC participation on regional inequality different for countries with a different specialisation (in terms of tasks carried out within the chain)?

From the example it becomes clear that the location decisions of MNEs influence which regions participate. Eger (Northern Hungary) is a relatively poor region, suggesting that regional inequality in Hungary decreases. However, there is evidence that the majority of inward Foreign Direct Investment (FDI) is attracted by the already better-off capital region. Pavlínek (2004, pp. 51f.) shows that in Hungary 56.5% of the total FDI stock was located in Budapest in 2000. The overall presence of Bosch in Hungary suggests that this pattern still holds today. Most wholly owned companies (4 out of 9) are located in Budapest (Bosch 2018, pp. 149-153). Although the strong increase in GVC participation may have benefitted Hungary’s development, it was (and still is) Budapest which reaped most of these benefits. An additional boost for the already better-off region increases regional disparities. The same picture of highly concentrated FDI inflows to better-off regions is also found for e.g. the Czech Republic, Slovakia or Poland (see e.g. Pavlínek and Smith 1998 and Pavlínek 2004). The FDI location decisions of MNEs from rather more to rather less developed countries, represent power asymmetries between countries within value chains which are linked to development levels. The following third question arises: Is the impact of GVC participation on regional inequality different for countries with a different level of development (in terms of GDP)?

(10)

4

2. Literature Review

GVC participation might impact regional inequality in a new way. However, when the second unbundling started, there were already immense regional inequalities present. To understand which forces resulted in this unequal spatial distribution of economic activity within countries is crucial to understand how GVC participation plays along with these forces and potentially moderates them in a new way. Therefore, this section reviews the theoretical and empirical literature on regional inequality and GVCs. The first section 2.1 explains the basic idea of why location matters and strikes the important role of trade in determining regional inequality. Afterwards, section 2.2 introduces the concept of global value chains. Connecting these two leads to the testable hypotheses of this thesis presented in section 2.3.

2.1 Regional Inequality

Generally, regional inequality implies that regions within countries are different. The questions why economic agents prefer some locations over others and why economic activity clusters in some regions, are the subjects of investigation in the urban economics and economic geography literature. As the basic underlying question, this literature asks why location matters. Some regions might be preferable over others due to exogenous factors. This reasoning is known as the “first nature” explanation in the literature. However, even if pure geography would be assumed absent, economic activity would not distribute equally across space. This is shown in the Von Thünen (1826) model. Originating from a “featureless plane”, economic agents face a trade-off between trade and housing costs and locate according to maximising net utility. This spatial equilibrium framework is still the core idea of modern urban economics. It is first transferred to the multi-city case by Henderson (1974).

It becomes clear that the decisions of economic agents determine the distribution of economic activity independent of exogenous geography. This is known as the “second nature” explanation. The underlying reason why location matters for firms are transport costs. But even with transport costs there must be economies of scale to make the concentration of economic activity beneficial (Krugman 1993, p. 131). The necessity of transport costs combined with scale economies is expressed by the “spatial impossibility theorem” which dates back to Starrett (1978). It states that without indivisibilities or costs of producing at small-scale, backyard capitalism would result in order to avoid transport costs (Brakman, Garretsen and van Marrewijk 2020, p. 173). Consequently, models which want to explain the spatial distribution of economic activity must include both features.

(11)

5

dispersion forces and consequently whether manufacturing concentrates in just a few regions or is distributed equally across space. Concentration implies regional inequality. The main agglomeration force is the “home market effect”. It implies that larger markets are beneficial because they imply closeness to consumers and demand and closeness to other producers. Firms want to locate where demand is high, and demand is high where markets are large. This is often called “circular causation” or “positive feedback loop” in the literature (ibid., pp. 485f.). Contrary, the main dispersion force is the “degree of competition effect” making location in small markets with less competition beneficial. Dominant agglomeration forces will occur whenever transport costs are relatively low, the share of manufacturing relatively high and economies of scale relatively strong (ibid., p. 497).

There are two types of increasing returns in the model. First, firms produce under increasing returns. They arise from the fact that production involves fixed and marginal costs (ibid., p. 488). Second, size matters also regarding the market, as firms benefit from larger markets. The first are internal-, the second external increasing returns. These external increasing returns to scale are agglomeration economies. They can be technological or pecuniary (Brakman, Garretsen and van Marrewijk 2020, pp. 194f.). In both cases the interaction between firms is beneficial which makes them locate close to each other. The difference is the type of interaction. Pecuniary agglomeration economies refer to benefits from market interactions which are transmitted by market mechanisms, i.e. price effects, while technological agglomeration economies arise from market transactions (Fujita and Thisse 1996, p. 345). Beneficial non-market interactions between firms are informational spillovers and knowledge exchanges. When firms have different knowledge pools, the spillover rents are higher in places where more firms locate. The need to locate close to each other, results from the fact that the quality of information decreases with increasing distance (ibid., p. 348).

The well-known Marshallian trilogy based on Alfred Marshall (1890) determines three sources of agglomeration economies: specialised suppliers of inputs, a pooled labour market with industry-specific skills, and informational spillovers (Krugman 1991, p. 484f.). They are a mixture of technological and pecuniary agglomeration economies (Fujita and Thiesse 1996, p. 345). The mechanisms of agglomeration economies at the micro-level are identified by Duranton and Puga (2004). These are sharing, matching, and learning. The sharing mechanism implies that concentration is beneficial to firms as they can share large indivisibilities in production and a variety of differentiated intermediate suppliers (ibid., pp. 2068ff.). Matching means that agglomeration implies a larger number of agents, which implies a better expected quality of matches between firms and workers with heterogeneous skills. This is important as mismatch is costly (ibid., pp. 2086f.). Lastly, a concentration of economic agents may also facilitate learning (ibid., p. 2098).

(12)

6

firms with similar Research and Development (R&D) activities are located than in Eger. For Eger one may assume that it is a region where already other MNEs may have located fabrication activities. Some may potentially produce intermediate inputs for steering columns. Then it may hypothetically be that Bosch chose Eger (next to saving on production costs) because of potential benefits from the sharing of differentiated intermediate suppliers.

As transport costs make location matter, it becomes clear why trade liberalisation is seen as an important determinant for regional inequality. The general logic is that regions benefit from better access to foreign markets. Spatial convergence emerges if the benefitting regions are the ones lagging behind before the trade liberalisation. If, however, the regions which were better off reap the benefits, there will be spatial divergence (Brülhart 2011, p. 78). Put differently, if trade liberalisation is linked to an increase in agglomeration forces, regional inequality will increase and vice versa. Although this sounds straightforward, the predictions of NEG models building on Krugman (1991) are far from uniform. This is noted by Brülhart (2011, p. 80) who reviews the literature and concludes that a-priori one does not know which model is most suitable as predictions are highly dependent on parameter choices. An overview over different model extensions and their predictions for the impact of trade liberalisation on the spatial distribution of economic activity is given by Brakman, Garretsen and van Marrewijk (2020). The variety of different predictions due to modelling choices is obviously a major drawback of the NEG framework. Further criticism aims at the highly stylised model assumptions, especially the “treatment of geography and […] history” (Garretsen and Martin 2010, p. 129). With geography being a fixed and abstract concept, critics see the explanatory power of NEG for the real world as limited. Additionally, NEG models would focus on equilibrium outcomes instead of helping to explain how geography developed over time (ibid., p. 129f.). Despite its drawbacks, the main aim of the NEG framework is to explain what drives economic agglomeration across space at different geographical levels. This includes explaining “the existence of strong regional disparities within the same country” (Fujita and Krugman 2004, p. 140), which is the purpose of this thesis. It appears to be a reasonable theoretical approach for addressing the spatial consequences of the second unbundling as the same is done by Baldwin (2016, pp.186-196). Building on the framework of Baldwin and Forslid (2000), which combines NEG with modern growth theory1_{, he explains why the second unbundling is the time of “The}

Great Convergence”. The drawback of this approach is that it ignores the within country perspective (Brakman, Garretsen and van Marrewijk 2020, p. 399). This is exactly the missing link this thesis approaches from an empirical perspective.

The mixed theoretical predictions make empirical studies inevitable. Brülhart (2011, pp. 68-72) reviews 11 empirical studies which analyse the relationship between the spatial concentration of economic activity within countries and trade in a cross-country setting. Out of these 11 studies six find no significant effect. Three find spatial convergence, one spatial divergence. In one study the result depends on the location of the largest city. Newer empirical studies suggest that trade openness leads to higher regional inequality (Ezcurra and Rodríguez-Pose 2013) at

1_{Modern growth theory, dating back to Romer (1986, 1990), endogenises technological change. Baldwin and}

(13)

7

least when combined with certain country characteristics and especially in developing countries (Rodríguez-Pose 2012). The impact on regional inequality in developing countries is also analysed by many country studies. The country cases reviewed by Brülhart (2011, pp. 75ff.) often focus on Mexico and find predominantly spatial divergence. Daumal (2013) finds the same result for India, however, the opposite result for Brazil. For China, the increase of regional inequality through globalisation is found e.g. by Wan, Lu and Chen (2007).

Regarding the empirical approach, the majority of cross-country studies reviewed by Brülhart (2011) apply pooled OLS and/or static OLS with country and/or year fixed effects. Three studies use a dynamic GMM approach. I will combine these frequently used approaches. After assessing the multivariate association between regional inequality and GVC participation in a pooled OLS model, I estimate a static OLS model with country and year fixed effects, as well as a dynamic panel model with Difference GMM and year fixed effects. This comes close to the empirical strategy applied by Rodríguez-Pose (2012). In line with that study, I will use a logged Gini index as the measure of regional inequality. The choice of control variables is inspired by Rodríguez-Pose (2012) and Ezcurra and Rodríguez-Pose (2013).

The theoretical and empirical findings of the relationship between trade and regional inequality are mixed. However, the patterns of international trade changed significantly in the last decades. The GVC revolution adds new aspects to the question where economic activity will locate and therefore may shed new light on the relationship between trade and regional inequality.

2.2 Global Value Chain Participation

The patterns of international trade changed over the last decades. In traditional trade theory, i.e. the Ricardian and Heckscher-Ohlin framework, countries trade different final goods produced under perfect competition and constant returns to scale. However, the observed trade patterns look different in three main aspects. First, there are high shares of intra-industry trade, implying that similar countries trade similar goods. This leads to the New Trade Theory which is based on Krugman (1979, 1980) and builds on the Dixit-Stiglitz (1977) monopolistic competition model. The introduction of imperfect competition, production under increasing returns and love-of-variety is the basis for the NEG model (Krugman 1991). Second, firms in trade are more productive than their non-trading counterparts (Bernard et al. 2007). Heterogeneous firm models dating back to Melitz (2003) evolve. Third, and central for this thesis, firms fragment their production into different stages and locate them where it is cost efficient. GVCs and global production networks emerge. In the second unbundling trade is no longer about final goods (see e.g. Baldwin 2016).

The concept of the “two great unbundlings” is introduced by Baldwin (2006). The first unbundling is induced by falling transport costs since the late 19th_{century. This allows splitting}

(14)

8

possibilities in globalisation arise as countries can join existing supply chains. The South industrialises while the North deindustrialised, their income converges (Baldwin 2012). From the NEG perspective Baldwin (2016, pp. 195f.) explains this “Great Convergence” with an increase in “the freeness of knowledge spillovers” which means that “moving ideas” becomes less costly due to the ICT improvements. It plays as a dispersion force.

GVC analysis emerges as a tool to identify the chances and risks resulting from the second unbundling. The basic idea of the GVC concept can be traced back to the value chain model introduced by Porter (1985) and the global commodity chains framework (see Gereffi and Korzeniewicz 1994 for an overview). According to Porter (1985, p. 26), the value chain “divides a firm into the discrete activities it performs in designing, producing, marketing, and distributing its product”. When talking about GVCs in the remainder of this thesis, the following definition provided by the global value chain initiative (globalvaluechains.org) holds:

A global value chain describes the full range of activities that firms and workers, divided among multiple firms and geographic spaces, do to bring a good from its conception to its end use. This includes activities such as design, production, marketing or distribution. This shows that GVCs are about different countries carrying out different activities. Grossman and Rossi-Hansberg (2008) introduce a framework which captures this shift from trade in goods to “trade in tasks”. Figure 2 visualizes a GVC in a stylised way, using the countries from the introductory example in a hypothetical scenario of car production.

Note: This is a hypothetical and fictional scenario which uses the introductory example as illustration, but which is in no way related to any actual production activities of Bosch.

(15)

9

Obviously, the final car not only embodies domestic value added from the country engaged in the last production stage, in this example Germany. The German car manufacturer imports intermediate goods, among others the steering column from Hungary. For the production of that steering column, Hungary may in turn import intermediate inputs e.g. certain steel parts from China. The final car embodies Chinese, Hungarian, and German value added. This also includes indirect domestic VA from other domestic intermediate suppliers or re-imported domestic VA. This second component includes the VA that Germany added in the product development of the steering column.

Such interconnection of countries are likely to impact the level of regional inequality in the participating countries through three main mechanisms. First, manufacturing activity which used to be bundled in a single country, Germany in this case, is now relocated unbundled and globally, e.g. to China and Hungary. This implies a change in the share of manufacturing in national income in the participating countries – a key determinant for regional inequality in the NEG model. Second, the participating countries specialise in specific activities. Germany might be most engaged in R&D, while China and Hungary might specialise in fabrication. These activities might be subject to differently strong agglomeration economies. Third, GVC participation is largely driven by the location decisions of MNEs. If MNEs target lagging behind regions, regional inequality in participating countries is likely to decrease. If they target predominantly the better-off regions, the opposite will happen.

Already the simple example of Figure 2 shows that trading final goods involves much more trade than just the transport to the final consumer (Hummels, Rapoport and Yi 1998, p. 80) and that value added can generate from multiple sources. Regarding domestic value added one needs to consider direct, indirect, and even re-imported domestic value added. In reality these networks become more complicated. Hungary may potentially export the steering column to a different car manufacturer, e.g. in France. This export from Hungary to France would still embody German value added. Also, the foreign value added may be contributed by far more than two countries. Traditional trade statistics based on gross trade are incapable to capture these complex interconnections. This is because gross export flows can no longer be reliably related to domestic value added and national income (OECD 2012, p. 1). The same holds for traditional specialisation indices based on gross exports like the revealed comparative advantage index based on Balassa (1965).

New measurement approaches based on new data sources are needed. Multi-country input-output databases allow for tracking the value added along value chains. Prominent examples are the Trade in Value Added (TiVA) database from the OECD and the World Trade Organization (WTO) or the WIOD (Timmer et al. 2015). Although input-output tables make it possible to trace value added, there are drawbacks of a rather broad sector aggregation and a limited country-coverage with a focus on developed countries (Baldwin and Lopez Gonzalez 2015, p. 1689). Nevertheless, this data combined with input-output analysis inspired by Leontief (1953) allows for constructing measures of GVC participation and resulted in an extent literature on measurement approaches.

(16)

10

(2001) present the famous vertical specialisation measure calculated as the amount of imported inputs used in exports. Newer approaches like Johnson and Noguera (2012) or Los, Timmer and de Vries (2015) extent and improve these measures. Koopman, Wang and Wei (2014) propose a measure of upstream and downstream participation which is similar to the measures of backward and forward participation calculated by the OECD. These indicators are used in studies about the impact of GVC participation on wage inequality (Lopez Gonzalez, Kowalski and Achard 2015) or on developing countries (Kowalski et al. 2015). ”Headquarter” economies are characterised by a relatively low share and “factory” economies by a relatively high share of intermediates in exports (ibid., p. 17). The high dependency of factory on headquarter economies is called “hub-and-spoke asymmetry” (Baldwin 2006, p. 8).

This thesis wants to answer whether GVC participation impacts regional inequality and if the impact differs across differently specialised countries. Consequently, a measure for the intensity and one for the specialisation of country i’s GVC participation in year t is needed. A reliable intensity measure captures the full extent of activities performed by a country, how much income the country generates from these activities and how high this income is compared to its overall economic size. The concept of “Global value chain income” proposed by Timmer et al. (2013) measures exactly this. As a share of GDP this seems to be a good measure for the GVC Participation Intensity (GVCI). Regarding specialisation, the measure of “Functional Specialisation” introduced by Timmer, Miroudot and de Vries (2019) appears a reliable measure for the GVC Specialisation (GVCS). It tracks domestic VA in exports and matches it with the type of workers, defined by occupation, involved in production (ibid., p. 9f.). Occupations are assigned to the four functions fabrication, R&D, marketing and management. While Hungary, China and Germany all increasingly derived income from GVCs between 2000 and 2011 (Figure 1), the way this income was derived differed between countries. During the whole period, Germany received most income from R&D activities and China from fabrication activities in line with Figure 2. Hungary was also most active in fabrication from 2000 to 2005, but derived income increasingly from R&D and marketing between 2006 and 2011. It may well be that differences in specialisation patterns are the reason why the increased GVC participation had different impacts on regional inequality in the three countries.

Different production activities are linked to a different level of value added by the “smile curve”2_{. The idea is that in early stages like R&D there is high VA involved. It goes down in}

fabrication and increases in later stages like marketing or distribution (Low 2013, p. 73). Kaplinsky (2004) states the concern that despite the new participation possibilities for relatively less developed countries, there is the risk that they specialise in the simple low VA tasks. Lower barriers to entry in production results in increased competition and falling terms of trade in manufacturing. Increased economic activity accompanied by falling returns is a situation Kaplinsky refers to as “immiserising growth” (ibid., p. 82f.).

How countries can achieve better positions within GVCs is a crucial subject in the GVC literature known under the concept of upgrading. Upgrading in value chains can be understood as “the process by which economic actors - nations, firms, and workers - move from low value

(17)

11

to relatively high value activities in global production networks” (Gereffi 2005, p. 171). This process can be differentiated into four types. Process upgrading means an enhanced efficiency when turning inputs into outputs while product upgrading implies moving into better product lines and functional upgrading implies engaging in functions with higher skill-levels. Inter-sectoral upgrading means that clusters of firms move into new value chains (Humphrey and Schmitz 2002, p. 1020). The change of functional specialisation from fabrication to R&D in the case of Hungary is an example for functional upgrading. Functions may be subject to differently strong agglomeration economies through which functional upgrading potentially has an impact on regional inequality. Relevant from the regional perspective may also be the inter-sectoral upgrading. If economic activity clusters regionally and the whole cluster is able to move up the value chain, this leads to a strong catch-up if the region is lagging behind. Contrary, if it is the better-off regions which participate, this upgrading will deepen regional disparities.

Upgrading possibilities are linked to the way a GVC is governed. While upgrading is the bottom-up perspective, governance is the top-down view (Gereffi and Fernandez-Stark 2011, p.4). Five governance modes can be distinguished. Next to the two extremes of markets and hierarchy, there are modular, relational, and captive chains. Which mode the lead firm chooses is determined by the complexity of the knowledge transfer, the extent to which it can be codified and the capabilities of suppliers (Gereffi, Humphrey and Sturgeon 2005, p. 85ff.). Pietrobelli and Rabellotti (2011) highlight special upgrading opportunities in modular chains. Modular chains arise when the three determinants of the mode choice are “high”, i.e. information is complex but can be codified and suppliers have high capabilities. The firms are obliged to comply with the quality standards of the global market but do not get direct support. The lead firm takes the role as “a crucial external stimulus for learning and innovation” (ibid., p. 1262). The authors suggest that these benefits could trickle down to other regions within the participating country through domestic production linkages.

2.3 Hypotheses

After having reviewed the literature on regional inequality and GVCs, there seems to not be much literature combining these phenomena to the best of my knowledge. In order to connect the two concepts and state expectations about their relationship, I transfer the idea of the NEG framework to the second unbundling but with the focus on the within-country view. To assess the impact of specialisation within GVCs, I rethink the idea of the smile curve in terms of agglomeration economies in functions.

(18)

12

It is proven by the literature that larger cities pay higher wages because firms located in bigger places are more productive on average (Combes et al. 2012). If offshoring firms were primarily concerned to save on labour costs, they would opt for the rather small (and poor) regions instead of the better-off big regions. Such a reasoning would be the equivalent to the global spreading argument at the regional level and would favour the expectation of less regional inequality. However, there is evidence that FDI tends to be located in core regions. For example, in China FDI is highly concentrated in the coast regions (Huang and Wei 2016). This is in line with the observations of highly concentrated FDI in already better-off regions in Central and Eastern European Countries (CEEC) made by Pavlínek. This adds an additional agglomeration force at the regional level, which makes increasing inequality the most probable outcome. As suggested by Figure 1 (and later confirmed by Table 2), especially the relatively less developed countries experienced a massive increase in GVC participation between 2000 and 2011. When, however, in these countries the already better-off regions benefit predominantly, it is likely that agglomeration forces dominate at the regional level. Hence, the following expectation is stated: H1: The intensity of GVC participation is expected to have a significant effect on regional inequality in participating countries. The overall impact is expected to be regional inequality increasing.

Regarding the GVC specialisation I adopt the idea of the smile curve but instead of value added I assign agglomeration economies to different activities. Drawing on the sharing, matching, learning mechanisms introduced by Duranton and Puga (2004), I assume an inverted U-shape relationship between activities and agglomeration economies - the mirror image of the original smile curve for value added. As Figure 3 visualizes, agglomeration economies are assumed to be highest for fabrication while they are assumed to be lower for R&D or marketing activities.

Source: Smile curve for value added based on Low (2013, p.73). Curve for agglomeration economies based on own assumptions.

I reach this assumption by identifying which micro-mechanisms could be associated with which function and then identifying the strength of agglomeration rents for each mechanism. In the case of fabrication, I see the sharing mechanism in the form of sharing of indivisible facilities and intermediate inputs as dominant. Duranton and Puga call it the “large indivisibility

(19)

13

argument” (ibid., p. 2068). It is important whenever there are large fixed costs creating internal increasing returns in production. This is the idea of a “factory town”. This factory town might in fact be a whole “factory region”. In the case of the high VA activities I see matching and learning mechanisms as dominant. Matching reduces the probability of costly mismatches and these costs are higher in high skill functions as the training is more complicated (ibid., p. 2097). Assigning an agglomeration strength to the mechanisms assigned to the functions, I assume the sharing mechanism in fabrication to be stronger than the matching and learning mechanisms in the high VA functions. This is in line with the reasoning of Baldwin (2016 pp. 194f.) that ideas become more mobile in the second unbundling. Arguments like the one by Fujita and Thisse (1996, p. 348) which claim that agglomeration is necessary because the quality of information decreases with distance, seem to become less powerful in the light of the second unbundling. Information and ideas are more important in the high VA functions. Fabrication still requires large factories, maintaining the relevance of sharing. This results in the following hypothesis: H2: The effect of GVC participation on regional inequality is expected to be different for countries with a different specialisation (in terms of tasks carried out within the chain). The effect is expected to be regional inequality increasing for countries specialising in low value added activities and regional inequality decreasing (or at least less inequality increasing) for countries specialising in high value added activities.

Keeping in mind that the location decision of lead firms is assumed to be a strong agglomeration force in relatively less developed countries and considering that especially rather less developed countries may specialise in the low value added, but strongly agglomerating, fabrication activities, I state the following hypothesis:

H3: The effect of GVC participation on regional inequality is expected to be different for countries with a different level of development (in terms of GDP). If there is an overall regional inequality increasing effect, it is expected to be stronger in developing countries.

3. Data and Methodology

(20)

14 3.1 The Empirical Model

3.1.1 Static Model

In order to investigate the association between GVC participation and specialisation, the static analysis starts with a simple pooled OLS regression. The following model is constructed:

ln(𝐼𝑛𝑒𝑞𝑢𝑎𝑙𝑖𝑡𝑦 ) = 𝛼 + 𝛼 ln(𝐺𝑉𝐶𝐼 ) + ∑ 𝛽 ln(𝐺𝑉𝐶𝑆 ) + ∑ 𝛾 ln 𝐺𝑉𝐶𝐼 ∗

𝐺𝑉𝐶𝑆 + 𝛼 𝑙𝑛(𝐺𝐷𝑃𝑐𝑎𝑝 ) + 𝛼 ln(𝑅𝑒𝑔𝑖𝑜𝑛 𝑛𝑢𝑚𝑏𝑒𝑟 ) + 𝛼 ln(𝑅𝑒𝑔𝑖𝑜𝑛 𝑠𝑖𝑧𝑒 ) +

𝛼 ln(𝐺𝑜𝑣𝑒𝑟𝑛𝑚𝑒𝑛𝑡 ) + 𝜀 . (1)

𝐼𝑛𝑒𝑞𝑢𝑎𝑙𝑖𝑡𝑦 is a measure for the regional inequality in country i in year t. It measures how different the single regions within a country are regarding their economic activity. This inequality is explained by two key regressors. 𝐺𝑉𝐶𝐼 is a measure for the intensity with which country i participates in GVCs in year t and accounts for the fact that some countries are more integrated in - and dependent on - the global production networks. The coefficient in front of this regressor measures the partial effect of the intensity of participation in GVCs on regional inequality and thus addresses Hypothesis 1. 𝐺𝑉𝐶𝑆 measures country i’s specialisation in function k in year t with K being the total number of different functions. Each function can be included separately or up to K-1 functions can be included jointly in the regression to avoid multicollinearity. The coefficient in front of this regressor gives the partial effect of the degree in specialisation in function k on regional inequality and hence addresses Hypothesis 2. The interaction term between the two is included to find out how intensity and specialisation in GVCs jointly influence regional inequality. Again, the number of interaction terms included can vary across specifications and maximum K-1 interaction terms can enter jointly. Next to the main regressors, I include for each country-year pair the national GDP per capita (𝐺𝐷𝑃𝑐𝑎𝑝 ), the number of regions (𝑅𝑒𝑔𝑖𝑜𝑛 𝑛𝑢𝑚𝑏𝑒𝑟 ), which is time invariant, and their average size (𝑅𝑒𝑔𝑖𝑜𝑛 𝑠𝑖𝑧𝑒 ) in terms of population, as well as the size of the public sector (𝐺𝑜𝑣𝑒𝑟𝑛𝑚𝑒𝑛𝑡 ).

Model (1) can explain the general association between the variables, however, a major drawback is that it does not control for individual heterogeneity. This “heterogeneity bias” (Wooldridge 2016, p. 413) needs to be addressed in order to make statements about causality. Hence, model (1) is next estimated as a static OLS model with country and year fixed effects.

ln(𝐼𝑛𝑒𝑞𝑢𝑎𝑙𝑖𝑡𝑦 ) = 𝛼 + 𝛼 ln(𝐺𝑉𝐶𝐼 ) + ∑ 𝛽 ln(𝐺𝑉𝐶𝑆 ) + ∑ 𝛾 ln 𝐺𝑉𝐶𝐼 ∗

𝐺𝑉𝐶𝑆 + 𝛼 𝑙𝑛(𝐺𝐷𝑃𝑐𝑎𝑝 ) + 𝛼 ln(𝑅𝑒𝑔𝑖𝑜𝑛 𝑛𝑢𝑚𝑏𝑒𝑟 ) + 𝛼 ln(𝑅𝑒𝑔𝑖𝑜𝑛 𝑠𝑖𝑧𝑒 ) +

𝛼 ln(𝐺𝑜𝑣𝑒𝑟𝑛𝑚𝑒𝑛𝑡 ) + 𝜂 + 𝜈 + 𝜀 . (2)

(21)

15

rejecting the null hypothesis of all 𝜂 being jointly zero at the 1% significance level. The null hypothesis of all year dummies being jointly zero is rejected at the 5% significance level. Lastly, 𝜀 is the error term, containing factors other than the independent variables affecting regional inequality. For model (1) I perform a White test for heteroskedasticity and for model (2) a modified Wald test for groupwise heteroskedasticity in FE models. Both tests have the null hypothesis of homoskedasticity which is rejected for the respective model across all specifications at conventional significance levels. Although heteroskedasticity does not bias the OLS estimator, it affects inference (ibid., p. 375). To account for that, I use heteroskedasticity robust standard errors which are clustered at the country level in the case of model (2).

The alternative to the FE model would be the random effects (RE) model. If consistent, RE is more efficient and allows for estimating the effect of time-invariant variables (Hill, Griffiths and Lim 2012, p. 557). However, RE requires 𝜂 to be uncorrelated with the explanatory variables (Wooldridge 2016, p. 444). A Hausman test for the equivalence of estimates of FE and RE under the null hypothesis of no significant differences, is rejected which indicates inconsistency of RE and the need to use FE.

In line with Rodríguez-Pose (2012) who estimates the effect of trade on regional inequality with a comparable empirical model, I decide on a log-log model transformation. The advantages of such a functional form are that it allows for interpretation independent of units of measurement, can reduce problems of heteroskedasticity and skewed distributions, and makes the OLS less sensitive to outliers (Wooldridge 2016, pp. 172f.). There seems to be a problem with heteroskedasticity in the data and some of the independent variables have a wide range with some extreme high outliers. The log transformation narrows this range and decreases the positive skewness remarkably. The same applies for the distribution of the dependent variable which becomes less skewed after the log transformation. These changes are visible in the summary statistics including measures for skewness and kurtosis provided at the end of this section. The underlying requirement for log transformations of positive values is fulfilled. 3.1.2 Dynamic Model

Applying economic reasoning makes it obvious why there might be a dynamic nature in the relationship proposed in (1) and (2). Regional inequality in a certain year is very likely determined by its level in the previous year, i.e. autocorrelated. This economic reasoning is confirmed by statistical tests. The Wooldridge test for autocorrelation in panels, with the null hypothesis of no autocorrelation, is rejected at the 1% level. The positive and significant correlation coefficient of 0.99 between regional inequality and its lag confirms this. Neglecting the dynamics may lead to an overestimation of the effect from the main regressors on regional inequality in the static models. Part of the effect would be in reality due to previous levels of regional inequality. In order to model the dynamic relationship, the lagged dependent variable enters as a regressor. The model in (2) becomes:

ln(𝐼𝑛𝑒𝑞𝑢𝑎𝑙𝑖𝑡𝑦 ) = 𝛼 + 𝜑 ln 𝐼𝑛𝑒𝑞𝑢𝑎𝑙𝑖𝑡𝑦, + 𝛼 ln(𝐺𝑉𝐶𝐼 ) + ∑ 𝛽 ln(𝐺𝑉𝐶𝑆 ) +

∑ 𝛾 ln 𝐺𝑉𝐶𝐼 ∗ 𝐺𝑉𝐶𝑆 + 𝛼 𝑋 + 𝜂 + 𝜈 + 𝜀 (3)

(22)

16

OLS is no longer unbiased and efficient in the presence of a lagged dependent variable as a regressor. An unbiased and efficient alternative are estimators based on GMM developed by Hansen (1982). In line with Rodríguez-Pose (2012), I use the Difference GMM which dates back to Arellano and Bond (1991). For the application in Stata I use the command “xtabond2” (see Roodman 2009) as it allows for necessary adjustments not available under the build-in command “xtabond”. I include one lag of the dependent variable which is assumed to be the only endogenous regressor in (3). Following the recommendations by Roodman (2009, p. 128), and in line with Rodríguez-Pose (2012), all other regressors enter the instrument matrix as strictly exogenous. Time dummies are included, and standard errors are heteroskedasticity robust (Windmeijer-corrected based on Windmeijer 2005). I use the option “collapse”. This prevents the number of instruments to exceed the number of countries and the Hansen test to report a perfect p-value of 1 (Roodman 2009, p. 98). Furthermore, I use the option “small” to account for my rather small sample size (ibid., p. 123).

In a second step I estimate the long-run coefficients of the dynamic model. This is important to find out how GVC participation and specialisation affect regional inequality in the medium- and long-term. I follow the approach by Wickens and Breusch (1988, p. 190) and calculate the long-run coefficient 𝜃 for a regressor 𝑥 using its short-run coefficient 𝛼 (or 𝛽 or 𝛾 ):

𝜃 = 𝜆 ∗ 𝛼 (4)

with 𝜆 = 1/(1 − 𝜑) where 𝜑 is the coefficient for the lagged dependent variable as in (3). Regional inequality may not be the only path-dependent variable in the model. The integration into global production networks is a continuous process. Countries with a high level of integration are likely to also be highly integrated in the next period and vice versa. The same reasoning applies to specialisation in functions. Functional upgrading is a time-taking process which does not happen out of sudden. This path-dependency is confirmed by a high correlation coefficient between the main regressors and their respective lags. To account for this persistency, I repeat the dynamic analysis including one lag of the main regressors. Next to allowing for the possibility that the effects of GVC participation on regional inequality may take time to materialise, this analysis also helps to rule out concerns about reversed causality. As a robustness test, I will treat the lagged main regressors as endogenous.

In order to address Hypothesis 3, the models (2) and (3) will be estimated separately for two different country groups. In the group of emerging economies are the 10 middle-income countries3_{plus the 6 Central- and Eastern European countries}4_{in the sample. They are included}

to obtain subsamples of equal size and because they show similar characteristics as the middle-income countries, i.e. a relatively high average level of regional inequality and a strong increase in GVC participation over the observed time period. The remaining 19 countries belong to the group of developed countries5_.

3_{Bulgaria, Brazil, China, India, Indonesia, Lithuania, Mexico, Romania, Russia, Turkey.} 4_{Czech Republic, Hungary, Malta, Poland, Slovakia, Slovenia.}

5_{Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Japan, Korea,}

(23)

17 3.2 Variables

3.2.1 Regional Inequality

The inequality between regions within countries in terms of varying degrees of economic activity is the object of study in this analysis. According to Brakman, Garretsen and van Marrewijk (2020, p. 68f.) there are three criteria that inequality measures must fulfil when they are applied to the spatial dimension. They are “symmetry”, “scale independence” and “region size independence”. They imply that the measure of inequality should not change when regional observations are swapped, activities are increased or decreased by the same amount in each region, or the number of regions changes, respectively. As the most popular measures satisfying the criteria the authors name the Gini coefficient and generalized entropy measures. In line with Brakman, Garretsen and van Marrewijk (2020) I calculate the Gini and Theil index, which belongs to the class of general entropy measures, for each country-year pair in my panel. Regarding the specific measurement I follow approaches used in empirical studies about the relationship of trade on regional inequality. Both measures are based on regional GDP per capita before any possible transfers.

The regional Gini index 𝐺 for country i in year t is calculated following the approach by Rodríguez-Pose (2012, p. 117):

𝐺 =

∑ ∑ , ∀ 𝑚, 𝑞 ∈ 𝑖 (5)

where 𝑦 and 𝑦 is the GDP per capita in region m and region q, respectively. The population shares of these regions are 𝑝 and 𝑝 , respectively, and 𝜇 = ∑ 𝑝 𝑦 . Consequently, the index 𝐺 represents the sum of all pairwise differences between regions in a country i and year t, weighted by their respective population shares and divided by GDP per capita of all regions. The regional Theil index 𝑇(0) follows Ezcurra and Rodríguez-Pose (2013, p. 94):

𝑇(0) = ∑ 𝑝 log ( )

,

∀ 𝑚 ∈ 𝑖 (6)

where 𝑦 and 𝑝 are respectively the GDP per capita and population share of region m in country i and 𝜇 = ∑ 𝑝 𝑦 . The index 𝑇(0) corresponds to “Theil`s second measure of inequality” also known as the “mean logarithmic deviation” (ibid., p. 94).

(24)

18

between component of large regions and the within component of smaller regions within the large regions. As the central purpose of this thesis is to analyse the general relationship between GVC trade and regional inequality and the data to perform a decomposition analysis is not available, I consider the advantages of the Gini as more relevant. Hence, I choose the Gini as the dependent variable in the main analysis while the Theil index will serve as a robustness test. The data to calculate (5) and (6) is obtained from the Regional Database provided by the OECD (OECD 2019). Regional GDP per capita is among the Regional Economy indicators. It is measured in PPP corrected millions USD in constant prices (base year 2015). Using constant prices and PPP allows to measure the developments in regional inequality across countries while being sure that these changes are not just driven by inflation or price differences between countries. The population measured in persons is obtained from the Regional Demography indicators. Population shares are obtained by dividing the regional by the national population. When measuring regional inequality, the choice of territorial units is a crucial question. The OECD uses the territorial classifications Territorial Level 2 (TL2) and Territorial Level 3 (TL3). The TL2 regions broadly correspond to states within the country, e.g. 51 states in the case of the USA. The TL3 classification disaggregates the regions into smaller administrative regions. To find out how crucial the choice of territorial level is, I calculate Gini and Theil indices at both levels when available. The difference in results is remarkable with TL3 indices being larger than at the TL2 level. The largest difference can be identified for Ireland in 2011 where the TL3 Gini is 0.21 while the TL2 Gini is 0.07. This indicates that a large part of the differences between regions is hidden in the TL2 classification and uncovered at the TL3 level. Consequently, regional inequality is different at different territorial levels. I would like to uncover as much as possible of the regional inequalities, however, the main problem with the usage of TL3 regions is data availability. Using TL3 regions would lead to the exclusion of 6 out of 10 middle-income countries including China, Mexico or India. For maintaining the variation in the data, the only reasonable choice is the TL2 classification.

3.2.2 Global Value Chain Participation

As noted in the previous section, measures based on gross trade are not representative in the context of GVC trade. Hence, the gross trade volume, often used as the key regressor in empirical studies about the impact of trade on regional inequality, is not suitable for the purpose of this thesis. In line with the definition of a GVC, the measures must instead cover all interconnections between countries and firms, i.e. all sources of value added. Hypothesis 1 asks for the impact of the degree of participation in GVCs on regional inequality. The concept of “GVC income” proposed by Timmer et al. (2013) appears to be a suitable base to construct a measure of participation intensity. GVC participation implies specialisation in tasks. Under Hypothesis 2 the impact of this specialisation is investigated. The measure of “Functional Specialisation” by Timmer, Miroudot and de Vries (2019) is chosen to capture varying degrees of specialisation in different functions. Both measures are based on the World Input-Output Database (Timmer et al. 2015), release 2013. To facilitate the understanding of the calculation of the measures, the concept and structure of the WIOD is introduced.

(25)

19

http://www.wiod.org. In order to construct the WIOTs, officially published input–output tables are merged with national accounts data and international trade statistics. “A WIOT provides a comprehensive summary of all transactions in the global economy between industries and final users across countries.” (Timmer et al. 2015, p. 577). Consequently, the WIOT is a suitable tool to analyse GVCs like the one in Figure 2 as it covers the whole production network. It allows to determine how much intermediate inputs from which country are embodied in - and how much value they add to - the final product. Figure 4 illustrates how a WIOT is structured.

Source: Based on Timmer et al. (2015, p. 577).

Each country produces output in 35 industries. This supply from country-industries can be used either as intermediate inputs in the production of country-industries or as final demand. Final demand is split into the five categories of final demand of households, non-profit organizations and the government as well as gross fixed capital formation and changes in inventories and valuables. Consequently, each row indicates the distribution of country-industry output over usage which is summed up to total use. The total use is equal to the total gross output. Value added is gross output minus intermediate inputs (Timmer et al. 2015).

The measures for GVC intensity and specialisation can be obtained by applying input-output analysis based on Leontief (1936, 1953) to the WIOT. This in turn requires to define the data structure in Figure 4 in matrix notation. The yellow part in Figure 4 is defined as y and represents the vector of output with dimension (1435 × 1)6_{. The blue part represents the final}

demand. Summing over the five categories and over the 41 countries results in a vector of final demand f with dimension (1435 × 1). As described above, gross output minus intermediate inputs results in value added. Hence, the red part becomes w, the (1435 × 1) vector of domestic industry value added. Lastly, the green part is known as the matrix of intermediate inputs Z with dimension (1435 × 1435). The elements of this matrix 𝑧 (s,u) represent the output of sector s in country i used as intermediate input in sector u by country j (Dietzenbacher 2012, p. 2). The intermediate input coefficient matrix A can be obtained by dividing the elements of matrix Z by the column vector y. The output of sector s in country i used as intermediate input in sector u by country j is divided by the total output of sector u in country j. Hence, the elements 𝑎 (s, u) represent the extra input from sector s in country i needed for one extra unit of output

6_{35 industries multiplied by 41 countries = 1435.}

(26)

20

in sector u in country j (ibid., p. 9). Going down along the column of country j and sector u gives the complete composition of the output in that country-industry. It is identified how much each row country-industry contributes to the final output of that column country-industry. Remembering Figure 2, Germany would be column country j and the car industry would be industry u. At the bottom of the column would be the output value of the final car. Going down the column, one would find the contributions from the row country-industries. The first row would be the domestic German car industry which assembles the final car and was involved in the development process. Hungary would appear as a first-tier supplier of the steering column. The row below might be the Chinese steel industry as second-tier supplier. Going down the column firstly shows the full list of country-industries involved and secondly, how much of their intermediate inputs are needed for the final German car. This is a stylised example and in reality the column of contributions is much longer.

With matrix A the well-known Leontief inverse (Leontief 1936) denoted by L can be calculated as L = (𝐈 − 𝐀) 𝟏_{where I is a (1435 × 1435) identity matrix. An element 𝑙 (s, u) of L indicates}

how much extra production of sector s in country i is needed in order to satisfy one extra unit of final demand for the good produced by sector u in country j (Dietzenbacher 2012, p. 10). The variable 𝐺𝑉𝐶𝐼 measures the Global Value Chain Intensity of country i’s participation in GVCs in year t. It is based on the concept of “Global value chain income” introduced by Timmer et al. (2013) and follows the authors calculations closely (ibid., p. 622-625). Global value chain income is calculated according to:

𝐯 = 𝐩(𝐈 − 𝐀) 𝟏_𝐟 ₍₇₎

where 𝐯 is the (1435 × 1) vector of value added levels and p is the (1435 × 1) vector of direct value added coefficients. The elements of p are obtained from dividing the elements in w by the respective elements in y. The hat indicates a diagonal matrix of (1435 × 1435) where the elements of p are on the diagonal. The vector f contains final demand for manufactured goods7_.

Equation (7) captures all activities from all countries that contribute to satisfy worldwide final demand. The complete composition of domestic and foreign intermediate inputs and the respective VA contributions are traced. If the elements in v are summed up for each country over its respective industries the result is a (41 × 1) vector which gives for each country i the total value added, i.e. income, derived from the participation in GVCs in year t in millions USD in current prices.

In order to adjust for price changes, Timmer et al. (2013, p. 636) calculate the real GVC income in constant 1995 prices using the US Consumer Price Index (CPI) as the deflator. I calculate constant 2015 prices using the US CPI available from the OECD (OECD 2020a) as the deflator. For the measure of GVC Intensity (𝐺𝑉𝐶𝐼 ), I divide the GVC income of a country i in year t

7_{This includes the following 14 WIOD sectors with their respective ISIC Rev. 3 code in brackets: Food, beverages}

(27)

21

in constant 2015 million USD by the country i’s GDP in year t in constant 2015 million USD. GDP data is obtained from the OECD (OECD 2020b).

Consequently, 𝐺𝑉𝐶𝐼 , is the share of GVC income in GDP of country i in year t:

𝐺𝑉𝐶𝐼 =

(8)

where a higher share implies a more intense participation in - and higher interconnectedness with – the global production network. I consider this measure a suitable transformation of the traditional trade openness measure (total trade volume/GDP) to GVC trade.

𝐺𝑉𝐶𝐼 gives the intensity of participation. However, the measure is silent about how this income is generated, i.e. which tasks the single countries perform. The example in Figure 2 suggests that the participating countries specialise in specific activities e.g. Hungary in fabrication and Germany in R&D. This specialisation in tasks is a crucial feature of GVC trade. This thesis wants to shed light on the consequences of this specialisation for regional inequality. The measure of “Functional Specialisation” introduced by Timmer, Miroudot and de Vries (2019) links the domestic value added in exports to the activities of workers involved in its production. The activities are aggregated into the four functions fabrication (FAB), R&D (RD), marketing (MAR) and management (MGT). How much income domestic workers derive from a function gives the country’s contribution to this function (ibid., p. 8). To trace the income of workers by occupation the authors construct the “Occupation Database” which is available at http://www.wiod.org. I follow the calculations of the authors (ibid., pp. 9-11) to obtain 𝐺𝑉𝐶𝑆 , the degree of specialisation in function k for country i in year t. First, the output needed in each country-industry in order to produce exports is calculated:

𝐲 = (𝐈 − 𝐀) 𝟏_𝐞 ₍₉₎

where e is the (1435 × 1) vector of the exports of each country-industry. Looking at Figure 4 these are the row sums of Z which represent the deliveries of row- to column country-industries. The deliveries of domestic to domestic country-industries are not included. Now d, the (1435 × 1) vector of domestic VA required for the production for exports is obtained:

𝐝 = 𝐩𝐲. (10)

A matrix B (4 × 1435) is defined which contains 𝑏 , the income of workers performing function k in industry s and country i as a share of the total value added in s. This income by function is linked to value added in exports by multiplying B with d. As the interest lies in each country’s contribution to the functions, one needs to perform this multiplication for each country separately, i.e. B becomes (4 × 35) and d (35 × 1) such that:

𝐠 = 𝐁𝐝 (11)

is of dimension (4 × 1) and 𝑔 is the value added by function k in country i’s exports, which is used to finally calculate the Functional Specialisation (FS) measure, i.e. 𝐺𝑉𝐶𝑆 :

𝐺𝑉𝐶𝑆 = 𝐹𝑆 = ( / ∑ )

(28)

22

which is the share of function k in overall income in country i’s exports in year t divided by the income share of function k for all countries in their exports. If the index is larger than one the country is specialised in function k (ibid., p. 8). As I want to analyse the impact of varying specialisation degrees in different functions, I have four variants of 𝐺𝑉𝐶𝑆 according to the four functions (K=4) which include also values below one.

3.2.3 Controls

GDPcap is the national GDP per capita and accounts for the overall standard of living. The famous Kuznets curve proposes that the relationship between economic growth and inequality has an inverted U-shape with inequality first increasing and later decreasing. Williamson (1965) demonstrated the importance of this relationship on the regional level. GDP per capita in USD constant prices and PPP (year 2015) is obtained from the OECD (OECD 2020c).

Region Number and Region Size account for the heterogeneity in territorial regions which is inherent when measuring regional inequality. Even when using the consistent TL2 classification, there is a lot of variation across countries in terms of how many regions are used and how large they are. Following the approach used in Ezcurra and Rodriguez-Pose (2013, p. 96) I include the number of regions and their average size in terms of population.

Government addresses the size of the public sector. The way the government is active in transfer payments between regions directly influences the regional inequality. I follow the approach by Ezcurra and Rodriguez-Pose (2013, p. 96) who use the final consumption expenditure of the government as a share of GDP. This also captures spending on infrastructure and education which are important determinants of regional inequality and GVC participation as pointed out by Kowalski et al. (2015). Military expenditures are excluded. The data is obtained from the World Development Indicators, World Bank (World Bank 2020b).

Table 1 provides summary statistics of all variables included in the main static (1) and (2) and dynamic (3) model before (A) and after (B) the log-transformation. A normal distribution has a skewness of 0 and a kurtosis of 3. Especially the region number and region size, but also the Gini, become more normal after the log transformation.

(29)

23

4. Analysis and Results

This section starts with preliminary descriptive evidence of the association between regional inequality and GVC participation and specialisation (4.1). Afterwards, the regression results from the static models (4.2) and the dynamic model (4.3), as introduced in the previous section, are presented. At the end, the robustness of the results is discussed (4.4).

4.1 Preliminary Evidence

Throughout the period of study 22 out of 35 countries experienced an increase in regional inequality, including middle- as well as high-income countries. This overall trend of increasing regional inequality across the world highlights the importance to better understand the sources of regional disparities within countries. One potential source could be the dramatic increase in GVC participation that most of the countries experienced during the analysed period. Table 2 highlights the top 10 countries which faced the most extreme levels of regional inequality and GVC participation on average and which experienced the most extreme changes compared to 2000. It is revealed by Table 2 that the average level of regional inequality was higher, and that of GVC participation lower, in less developed countries.

Note: Values in level, i.e. 0.38 for Russia in 1A is the average value of the Gini index over 2000 and 2011 and not the logged value which is used in the regression analysis.

Source: Own calculations.

Eight out of the top 10 unequal countries (1A) are middle-income countries and six of them are among the top 10 least participating countries (2B). In contrast, all top 10 equal countries (2A) are high-income countries and four of them are among the top 10 most participating countries

(30)

24

(1B). Regarding the changes, the pattern is less straightforward. Six countries that experienced the highest increase in GVC participation (3B) are among the countries which faced the highest increase in regional inequality (3A). While four are middle-income countries (Bulgaria, India, Lithuania and Romania), the other two are high-income countries (Australia and Slovakia). At the same time, two countries with the highest increase in GVC participation are among the countries which experienced the strongest decline in regional inequality (4A). These are both middle-income countries (Brazil and Russia). Consequently, emerging countries seemed to be more unequal and less integrated into the global production network between 2000 and 2011. Most countries that experienced a strong increase in GVC participation saw their level of regional inequality also increase strongly, independent of their development level.

Performing a simple linear regression of the logged Gini coefficients on the logged GVC participation intensity measure for each country shows a mixed pattern. The regression coefficients in Figure 5 suggest that in some countries there is a positive relationship while in other countries it is negative and close to zero in others. Each group contains middle-income countries which are indicated in red. The absence of a clear linear relationship between trade openness and regional inequality is no novelty in the literature (see e.g. Rodriguez-Pose 2012, p. 117). As suggested by Figure 5, this also holds for GVC trade.

Source: Own calculations.

Table 3 provides pairwise correlation coefficients. Against Hypothesis 1, the correlation coefficient between the Gini and the GVC intensity is negative (-0.52). However, without controlling for the development level, this is not surprising as Table 2 shows that there is a clear pattern of middle-income countries having high levels of regional inequality but low levels of GVC participation. This is supported by a high negative (positive) correlation coefficient between the Gini (GVCI) and GDP per capita. For the other hypotheses, the evidence is in line with the expectations. Regarding Hypothesis 2 there is evidence that specialisation in different functions affects regional inequality differently. As expected, the sign of the correlation between the Gini and specialisation in fabrication is positive and negative for all other functions. This supports my idea of the inverted smile-curve of agglomeration economies in functions. It directly links to Hypothesis 3 which states that the inequality increasing impact of

The Impact of Global Value Chain Participation on Regional Inequality – An Empirical Analysis

Master Thesis