An exploration of space in FDI

(1)

Master thesis

An exploration of space in FDI

Examining third-country effects and the main motivation behind FDI using

a sophisticated spatial econometric approach

July, 2012

Faculty of Economics and Business University of Groningen

Name: Martijn G.J. Regelink Student number: 1526359 Mail: m.g.j.regelink@student.rug.nl

(2)

1

Abstract

This thesis tests for the relevance of third-country effects and the main motivation behind the decision of MNEs to invest in a particular host country. We employ a sophisticated spatial econometric approach, including direct and indirect effects estimates recently proposed by LeSage and Pace (2009) and a test for the procedure that should be used to normalize the spatial weights matrix. We also run tests on sub-samples of horizontal, vertical and more complex forms of FDI, as well as tests using a more general spatial Durbin model specification and alternative spatial weights matrices. Using data on US outward FDI activity in 20 European states between 1999 and 2008, this study finds stronger relationships between formal FDI theory and results than previous research. Results are foremost indicative of export-platform FDI.

(4)

3

1. Introduction

The boom in foreign direct investment (FDI) by multinational enterprises (MNEs) since the 1980s has led to a substantial interest in the economics literature to empirically investigate the determinants that underpin FDI behavior. Most of the models and resulting empirical examinations on the determinants of FDI have been based on a partial equilibrium analysis; testing hypotheses within a two–country framework (see Blonigen (2005) for an overview of the literature). This assumes that decisions by an MNE in a home country to invest in a particular host country are independent of decisions to invest in any other country. A number of scholars have pointed out that such an assumption is deficient and that the impact of so-called third countries should be taken into account (Baltagi et al., 2007a; Blonigen et al., 2007; Garretsen and Peeters, et al., 2007). For example, a vertical FDI decision by an MNE involves picking the best low-cost host at the expense of alternative host countries (Blonigen, 2005). Likewise, an MNE that decides to use a host country as an export-platform takes the market potential of neighboring countries into account. Moreover, excluding third-country effects can lead to serious econometric problems such as biased coefficient estimates and too high R2 statistics (Anselin, 1988; Elhorst, 2010a).

In recent years, a handful of papers have taken third-country effects into account, among which Blonigen et al. (2007), Baltagi et al. (2007a), Poelhekke and Van der Ploeg (2009) and Uttama and Peridy (2009) for US outward FDI, Garretsen and Peeters (2009) for Dutch outward FDI, Martínez-Martín (2011) for Spanish outward FDI, Ledyaeva (2009) for Russian regional FDI inflows and Chou, Chen and Mai (2011) for Chinese outward FDI. By using information on distances between potential host countries these studies have tested for the impact of third-country effects. Following an approach popularized by Blonigen et al. (2007), particular attention has been paid to the impact of two spatial variables: a spatial lag on FDI which captures the distance weighted effect of FDI into countries surrounding a particular FDI host country, and a surrounding-market potential variable which captures the distance weighted impact of market sizes of countries surrounding the FDI host country. A specific objective of Blonigen et al. (2007), and subsequent research, has been to use the sign and significance of these third-country variables to identify the main motivation behind FDI.

(5)

4 exports or by affiliate sales (Markusen, 1984). Blonigen et al. (2007) argue that such an investment is not affected by third-country effects. Through vertical FDI a firm wants to profit from factor cost differences across regions or countries by shifting parts of the production process to the lowest cost location (Helpman (1984). This implies a competition effect between host-countries and thus Blonigen et al. (2007) predict that FDI in one country comes at the expense of other potential investment locations. In recent years more complex rationales for FDI behavior have been brought up, explicitly taking third-country effects into account. Export-platform FDI occurs when an MNE invests in another country with the purpose of using this country as a base to export final products to markets other than the home country of the MNE (Ekholm et al., 2007). Not only does this mean that an MNE takes the size of a neighboring market into account, it is likely that an investment into one country acts as a substitute for an investment into a proximate country. Complex vertical FDI or vertical specialization takes place when an MNE off-shores parts of its production chain over various regions in order to benefit from factor cost differences between various host countries (Bergstrand and Egger, 2007). An investment into one country now acts as a complement of an investment into another.

In line with Blonigen et al. (2007) and subsequent studies, this thesis tests for the relevance of third-country effects on FDI, as well as the relationship between third-country effects and the main motivation behind FDI. While spatial relations have shown up in the data, empirical tests have often produced results which are inconsistent with the above-mentioned basic theoretical models of FDI. Given the presumed importance of spatial linkages, we hope to offer new insights in the relevance of third-country effects and the main motivation of FDI by using a more sophisticated empirical approach than previous research.

(6)

5 more complex forms of FDI, in order to see if we can obtain more precise predictions on the relationships between third-country effects and FDI theory.

To let space enter our empirical analysis we follow Blonigen et al. (2007) by adding a spatial lag on FDI and a surrounding-market potential variable to a gravity model specification. In contrast to previous work, a number of recent insights from the spatial econometrics literature are introduced to come to a more sophisticated empirical approach. First, we use direct and indirect effects estimates recently proposed by LeSage and Pace (2009) rather than the point estimates of the independent variables to test whether spatial spillover effects exist. Second, we compare a wide array of models in which we not only test for the inclusion of time-period and spatial fixed effects, but also for the procedure that should be used to normalize the spatial weights matrix that describes the spatial arrangement of the countries in the sample. Moreover, we run robustness checks on two alternative distance weighting schemes and on what the spatial econometrics literature would describe as a more general spatial Durbin model specification in which the surrounding-market potential variable is measured differently. As far as we know, we are the first study on the impact of third-country effects on outbound FDI to take all of these methodological considerations into account. This study thus does not only offer insights into the relevance of spatial linkages and the main motivation behind FDI, but also offers an exploration of state-of-the-art applied spatial econometric techniques in FDI research.

Overall, this thesis seeks answers to a number of questions. First, the guiding research questions of this thesis are: what is the relevance of third-country effects for US outward FDI into Europe and to what extent can the main motivation of MNEs be derived from these effects? In addition, we ask two sub-questions related to our empirical approach. First, in line with previous literature on the subject we wonder whether or not the omission of spatial interactions leads to biased results on the determinants of US outward FDI into Europe? Second, we wonder to what extent our more sophisticated empirical approach will render biased research using a similar though less advanced approach. Hereto we will compare results obtained by our adjusted specifications with results obtained from model specifications similar to those used by Blonigen et al. (2007). A strong bias could offer insights into why previous research has not been able to find strong relationships between third-country effects and formal FDI theory.

(7)

(8)

7

2. Literature review

2.1 FDI and the importance of space

Much research has already been done on the factors that drive FDI. Apart from a large body of the literature that has focused on partial equilibrium models in which exogenous factors (e.g. market size, skills, trade costs, taxes, the exchange rate, etc.) affect the magnitude of foreign direct investments into a certain location, general equilibrium models have given us insight into the underlying rationale on why a firm wants to invest abroad (Blonigen, 2005; Navaretti and Venables, 2004). Early work on formal theory pertaining to these general equilibrium models makes a distinction between investments that are driven by direct access to a foreign market (horizontal FDI) or differences in factor prices (vertical FDI) (Helpman, 1984; Markusen, 1984). In both cases firms face a distinct trade-off. With horizontal FDI (HFDI), firms avoid the trade costs that come with serving a foreign market by exports. However, such benefits are met by foregoing economies of scale and disintegration costs due to the dispersion of plant activity. In the case of vertical FDI (VFDI) there is a trade-off between disintegration cost and the benefits of saving on factor costs.

Notwithstanding the useful insights that these general equilibrium models provide, as they, among others, enable researchers to tie back long run determinants of MNE activity to micro-economic decision-making (Blonigen, 2005), a possible shortcoming is that they are typically restricted to a bilateral country setting. Various authors have explained why such a setting is problematic. First, as Baltagi et al. (2007a) remark, the trade-offs pertaining HFDI and VFDI, seem to ignore the actuality that firms set-up foreign affiliates in a multi-country world. This fact provides room for the idea that the choice by an MNE to enter a particular target market also depends on the characteristics of certain other potential host countries. For example, when a vertical investment is made it is likely that the MNE not only compares prices between home and host country, but also takes the prices of inputs in third countries into account.

(9)

8 third-country effects into account can help to explain some of the basic stylized facts that are at odds with the theoretical models of HFDI and VFDI. For example, it is difficult to explain why horizontal FDI has surged within the EU, while trade costs have become significantly lower over the years (Neary, 2008). It is suggested that export-platform motivations by multinationals might account for this basic fact. Export-platform FDI occurs when an MNE invests in another country with the purpose of using that country as a base to export final products to markets other than the home country of the MNE. In such a setting a multinational is not so much motivated by the size of the host market, but rather by the size of surrounding markets (Ekholm et al., 2007).

Thirdly, Garretsen and Peeters (2009) explain that from a theoretical point of view bilateral models are at odds with suggestions from the new economic geography literature. The new economic geography literature puts emphasize on the impact of agglomeration effects on the economic landscape (Fujita et al., 1999; Head et al., 1995, 1999). Agglomeration effects arise, because firms want to be close to other firms in order to benefit from certain forward and backward linkages, knowledge spillovers and markets for specialized factors (Navaretti and Venables, 2004). Obviously such considerations by firms would indicate that the decision by one firm to choose a particular location is influenced by the location decisions of other firms. This assumption is at odds with the basic models on FDI, which take the spatial distribution of firms as a given. If agglomeration effects are indeed important, taking the spatial distribution as given would clearly be deficit.

2.2 FDI and the impact of space

(10)

9 thereby ignoring endogenous spatial relationships through the dependent variable. Moreover, it allows researchers only to examine the existence of FDI activity, not the magnitude of it (Blonigen et al., 2007).

A more flexible approach to account for the impact of third countries on FDI flows is offered by spatial econometric techniques, which allow spatial linkages to be modeled directly within a linear regression framework (Blonigen et al., 2007). Although a more thorough discussion on the use of spatial econometrics techniques is given in the next section, it is important to note that most of the empirical studies that use spatial econometrics to test for the impact of third-country effects either use a spatial lag model, or a spatial error model. The former model is able to account for third-country effects operating through an endogenous spatial interaction effect, i.e. a spatial lag between FDI in the target market and FDI in geographically proximate countries (Garretsen and Peeters, 2009). In contrast, a spatial error model uses a spatial autoregressive error term to control for stochastic shocks to FDI in neighboring countries that spillover into the host market. Both spatial lag and spatial error model specifications have been adjusted in most studies to incorporate exogenous spatial interaction effects as well, which arise from the effects of spatially weighted independent variables in neighboring markets on FDI in the target market. The abovementioned market potential variable is an example of an exogenous spatial interaction effect.1

Until now, only a handful of papers have used techniques from spatial econometrics to examine the impact of third-countries on FDI. Coughlin and Segev (2000) are the first paper to use spatial econometric techniques in FDI research. Using a spatial error model they try to explain the geographical dispersion of US FDI inflows across China. They identify positive spatial autocorrelation in their data (reflected in the error term) and conclude that increased FDI into a Chinese province has a positive effect on FDI into nearby provinces. Although the authors suggest that such autocorrelation can be attributed to agglomeration effects, they do not connect the spatial dependence to formal FDI theory. In a more complex approach, Baltagi et al. (2007a) adjust a basic three factor knowledge-capital model to include exogenous spatial dependencies that are linked to the various motivations that drive FDI. At the same time, they test for spatial autocorrelation in the error term. Using data on US outbound FDI between 1989 and 1999 across a variety of sectors and countries, they find a number of significant third-country effects, lending support to the existence of more complex

1_{The term ‘spatial interaction effect’ is commonly used in the spatial econometrics literature to denote that the}

(11)

10 forms of horizontal and vertical FDI. Building on the framework of Baltagi et al. (2007a), Uttama and Péridy (2009) examine the impact of third-country effects and regional integration on FDI flows to five Association of Southeast Asian Nations (ASEAN) member states. Their findings lend support to vertical and complex vertical modes of operation, as well as some support for export-platform FDI. Just as results obtained by Baltagi et al. (2007a), the findings by Uttama and Péridy (2009) are somewhat ambiguous though, as they cannot identify which type of FDI is more common.

A different approach is adopted by Blonigen et al. (2007). This paper is of particular importance to our work as we build on their framework. They use a standard gravity model, and adjust it with a spatial lag and a surrounding-market potential variable closely linked to the one used by Head and Mayer (2004). In line with Baltagi et al. (2007a), the work of Blonigen et al. (2007) is grounded upon FDI theory, making it possible to link third-country effects to horizontal FDI, vertical FDI, export-platform FDI and complex vertical FDI. These studies deviate from earlier research, that either limits the existence of space to one channel (Head and Rise, 2004), or is based upon a spatial error model that misses theoretical foundations (Coughlin and Segev, 2000). Compared with Baltagi et al. (2007a) the empirical specification used by Blonigen (2007) has the advantage that it is better suited to indicate which type of FDI dominates. Furthermore, in order to avoid the possibility of biased estimates, a spatial lag specification is often preferred over a spatial error model (Elhorst, 2010a). The later point will be explained in more detail in the method section, where we will give a more thorough explanation on why we follow the approach used by Blonigen et al. (2007) instead of the spatial error approach outlined by Baltagi et al. (2007a).

2.2.1 Third-country effects and FDI theory

(12)

11 on FDI into geographically proximate third countries (Blonigen et al., 2007).2 This variable is measured by the distance weighted sum of FDI into third countries surrounding a particular host country. Such a variable is endogenous, because it captures whether or not the dependent variable in one country depends on the behavior of the dependent variables in other countries. The surrounding-market potential variable captures the spatially weighted market size effect of countries surrounding an FDI host country and is measured by the distance-weighted sum of third-country GDP. Its effect on FDI is assumed to be exogenous, as its behavior depends on independent explanatory variables in third countries (Elhorst, 2010a). In line with predictions by Blonigen et al. (2007), as well as Garretsen and Peeters (2009), these two variables can be used to construct a variety of hypotheses related to four types of FDI.

As was explained, in the event of pure horizontal FDI a company from home country

d is motivated by market access to a particular host market i. Formal theory explains that such

an investment is governed by a trade-off between the possibility to avoid trade costs and the costs of setting up an extra plant (Navaretti and Venables, 2004). The proximity concentration trade-off explains that when trade costs are sufficiently low, a firm would choose to supply a foreign market by exports, rather than by supplying it directly through affiliate sales, in order to capitalize on economies of scale and save on setting up an extra plant. Subsequently, when trade costs are sufficiently high a firm may choose to supply a host market directly if the initial returns to scale and disintegration costs are relatively low. In such a case, the decision by an MNE to invest in host country i is not affected by the decision to invest in country j (j ≠ i), because it is taken independently of any type of agglomeration or demand effects in other markets. Basic theory simply predicts a trade-off that is related to the question whether a firm wants to serve that particular country by exports or by affiliate sales. In line with predictions in the literature, we thus hypothesize that HFDI in host country i is not affected by investments into host countries j, nor do the size of the markets of j affect HFDI in

i. This implies that the spatial lag coefficient and surrounding-market potential variable will

both have an insignificant effect on HFDI.

In the case of purely vertical FDI an MNE is motivated by cheaper access to factor inputs abroad.3 General theory explains that such an investment is facilitated by low trade

2_{Adverbs like ‘proximate’ and ‘nearby’ are used throughout the literature on third-country effects to indicate}

that the geographical distance between host country i and third countries j is a crucial element in determining the impact of third-country variables on FDI in host i.

3_{We note that access to natural resources is sometimes identified as a separate form of VFDI in the literature}

(13)

12 costs, while the size of surrounding markets (j ≠ i) has no impact on the volume of VFDI to host country i. An insignificant relation between VFDI and the surrounding-market potential variable is therefore expected. On the other hand, Blonigen et al. (2007) explain that an investment by a firm from parent country d to host country i is to the detriment of an investment into third country j, as firms try to minimize costs by choosing the lowest cost location for (part of) its production. For that reason Blonigen et al. (2007) predict a negative spatial lag coefficient.4

A gap in the theoretical framework of Blonigen et al. (2007), and related studies, is that they do not explain why the negative spatial lag in the case of VFDI operates through distance. Although a full substantiation on this topic is beyond the scope of this research, we expect that distance captures intensified competition over VFDI between proximate investment locations. Typically, the investment decision by an MNE involves a two-stage game (Oman, 2000; Charlton, 2003). First, firms draw up lists of potential low cost locations based on the economic and political fundamentals that drive VFDI. Second, they start a bidding procedure for the investment among the selected candidates. In this second stage, strong policy competition has been observed between the selected candidates competing over the same investment by offering generous incentives (tax breaks, grants, special facilities etc.) to the MNE.5 Although in theory these bidding procedures could be global, case study evidence suggests that strong policy competition between low cost locations for among others vertical investments is predominantly intra-regional, and shaped by intra-regional bidding procedures, and investment agencies who create incentives based on the behavior of neighboring countries (Oman, 2000; Charlton, 2003; Blomström and Kokko, 2003). The underlying reasons why these competition effects are mainly encountered intra-regional instead of global are rather unexplored in research, although the surge in regional investments in for example Central and Eastern Europe and Southeast Asia (Navaretti & Venables, 2004), suggests that the economic and political fundamentals that drive VFDI, as well as some other effects and formal theory, we will ignore the special case of access to natural resources and mainly relate VFDI to the idea of cheaper access to factor inputs by moving unskilled labor activity to economies were unskilled labor wages are low, moving R&D activities to countries were scientists are cheap, and so forth (Navaretti and Venables, 2004). For a spatial econometric study of the determinants of resource FDI we refer to Poelhekke and Van der Ploeg (2010). In case of our research it is important to note that their data observations show that resource FDI is probably not a main driver of US outbound FDI to Europe, as its magnitude is relatively small compared to that of other forms of FDI.

4_{We note that negative spatial autocorrelation is hardly encountered in spatial research, meaning that little is}

known about the mechanisms through which it takes place. Discussing some empirical examples, Griffith and Arbia (2010) indicate that negative spatial autocorrelation is shaped by competitive locational processes.

5

(14)

13 less observed country characteristics and peer influences, might on average be stronger correlated between proximate host countries than more distant countries.67

More complex rationales for FDI are export-platform motives and vertical specialization (Ekholm et al., 2007; Blonigen et al., 2007; Baltagi et al. 2007a; Garretsen and Peeters, 2009). Export-platform FDI is an investment by an MNE from home country d in host country i with the purpose of using this country as a base to export final products to other markets in countries j. When trade costs between markets i and j are sufficiently low in comparison with trade costs between home country d and markets j, the firm from country d could decide to use target market i to serve as an export-platform. Export-platform FDI has elements of both vertical and horizontal FDI (Ekholm et al., 2007). Like in HFDI, production is meant to serve a larger combined market to circumvent trade costs, while a specific location within such a combined market is chosen is on the basis of cost savings, resembling VFDI. The choice for a single production facility to serve a larger market implies that an investment by an MNE in host i will be at the expense of an investment in nearby country j. We therefore hypothesize that FDI from home country d to third countries proximate to i is negatively related to export-platform FDI from home country d to host i. Meanwhile, the size of nearby markets will determine the amount of exports the MNE can engage in. It is therefore hypothesized that the market size of third countries (j ≠ i) near country i is positively related to export-platform FDI from home country d to host i. In line with Baltagi et al. (2007a), we

6_{Often the generalization is made in spatial econometric research that countries (or regions) that share borders}

(or are close to each other) share on average more commonalities with each other than countries which are more distant (see for example Hall and Petroulas, 2008). From this perspective we make a few observations regarding the spatial distribution of cost conditions that drive VFDI. First, we make the rather straightforward observation that countries nearby each other share a relatively similar bilateral distance with the home country of the MNE. Although this similarity is not necessarily confined to proximate hosts, it does mean that any (dis)advantages spelling from low (high) trade costs are more likely to be shared by both. Second, particularly in the case of Europe there is some evidence available which suggests that the spatial distribution of among others income per capita, labour productivity and (low) wages follows a pattern in which on average countries (or regions) that share similarities across these characteristics are geographically closer together (see Combes and Overman (2003) for a literature overview; López-Bazo et al. (1999), Overman and Puga (2002), Melciciani (2006), and Ramajo et al. (2008) are examples of empirical research finding evidence of positive spatial autocorrelation).Third, as Hall and Petroulas (2008) note, neighbouring countries often share stronger institutional and cultural ties with each other than with more distant countries. Although not all these factors are modelled by standard FDI models, they are at some point likely to be taken into account by an MNE when comparing potential investment locations, increasing the likelihood that hosts nearby each other compete for the same investments. Fourth, similarities in investment climates are likely to be influenced by knowledge spill overs and peer influences between nearby countries. An example of strong regional peer influences was observed in the Czech Republic in 1998. After finding out that its hostile view of tax holidays for foreign firms made it lose out on low cost investments to Hungary and Poland, it changed its views and adopted the same policies as its neighbours (Charlton, 2003).

7_{Agglomeration effects contributing to clustering in high tech production, as well as demonstration effects}

(15)

14 assume that weighting these variables by distance is likely to capture trade costs and trade related costs, as well as some of the earlier described competition effects in case of the spatial lag. Keeping to the above-mentioned predictions, we expect the spatial lag coefficient to be negative, while the surrounding-market potential variable is expected to be positively related to export-platform FDI.

Complex vertical FDI (complex VFDI) or vertical specialization is defined as an investment whereby a firm off-shores parts of its production chain over a number of host countries (i and j), in order to benefit from factor cost differences between hosts (Bergstrand and Egger, 2007). Complex VFDI is driven by agglomeration effects, whereby having suppliers in proximate countries j makes an investment in country i more attractive (Blonigen et al., 2007). In such a case geographical proximity acts as a facilitator for intermediate trade in goods and services between complex VFDI hosts i and j. In addition to supplier-network effects, a number of other agglomeration effects, such as having large harbors or airports in surrounding markets, could make fragmentation in host country i more attractive (Poelhekke and Van der Ploeg, 2009). Blonigen et al. (2007) suggest that surrounding-market potential has no effect on complex VFDI. However, as Garretsen and Peeters (2009) explain, it could well be that the surrounding-market potential variable not only captures market demand, but also agglomeration effects. In the latter case we would expect that the size of surrounding markets has a positive effect on complex VFDI. We therefore expect the surrounding-market potential to either have a positive, or an insignificant effect on complex VFDI. In addition, the spatial lag coefficient is predicted to have a positive impact on complex VFDI.

Table 1 provides an overview of the hypothesized effects.8

8_{We have also investigated whether these hypotheses can be used to test for the impact of third-country effects}

at the more disaggregated services and goods level. The standard models outlined above focus on the determinants at the aggregate level, which are predominantly based on the manufacturing of goods. With a sharp increase in services FDI over the past two decades, researchers have asked the question whether or not theoretical frameworks should be adjusted when analyzing the services sector. Davies and Guillin (2011) explain that the key issue revolves around the use of physical distances as a proxy for trade costs in spatial, as well as non-spatial, variables. Whereas the role of physical distance in the goods sector as a proxy for trade costs is often rather straightforward (except for its use in the construction of the spatial lag in the model of VFDI), it is unclear in the case of intangible services, as characteristics of services often differ. Some services tend to be more tailored for local needs and take local culture and language into account (Ramasamy and Yeung, 2010). These services often require face-to-face contact and are less likely to be traded. By contrast, modern communication technologies have made other services easier to export. These characteristics could either diminish or increase the role of distance between potential host countries, and home and host countries.

(16)

15

Table 1 Hypothesized effects on four types of FDI

HFDI VFDI

Export-platform

Complex VFDI

Spatial lag on FDI 0 - - +

Surrounding-market

potential 0 0 + 0/+

Sources: Blonigen et al. (2007); Garretsen and Peeters (2009)

2.2.2 Related empirical observations

We now come back to the results obtained by Blonigen et al. (2007), as well as related studies, from which a number of interesting observations can be made. These observations serve as a benchmark to which we can compare our tests on the impact of third country effects and will help us to avoid pitfalls made by previous studies using a spatial econometric approach based on the framework set up by Blonigen et al. (2007). Blonigen et al. (2007) tests for hypotheses similar to those in Table 1 by using a sample of US outward FDI into 35 host countries for the period 1983-1998, as well as various sub-samples.9 In general, Blonigen et al. (2007) conclude that space matters. However, their results are sensitive to sample selection and the inclusion of country dummies. Results obtained on their full sample, as well as on a disaggregated European OECD sample, cannot be tied back to a particular model of FDI. Only when using a sub-sample of European OECD countries disaggregated by industry, they analogies from the trade literature in which researchers have found that when using the standard gravity model distance is just as, if not more, important in trade in services than in goods (Kimura and Lee, 2006).

The fact that the models outlined above are appropriate to use on tests of both services and goods FDI, does not mean that outcomes on the main motivation of FDI could not differ. Firm level characteristics like economies of scale and skill intensity are likely to deviate on average between the goods and services sectors and could translate into different motivations arising from the data as some papers have found. Garretsen and Peeters (2009) find evidence which suggests that Dutch service sector FDI to a particular host is a substitute for FDI to other hosts, while in manufacturing FDI to a particular host is a complement for FDI to other hosts. Using more disaggregate data across a range of industries, Blonigen et al. (2007) note that the main motivation of US MNEs varies across industries. No impact of third-country effects is found in the case of services FDI. We note though that a lack of industry level data for a number of countries could have biased these results. In contrast to observations by Garretsen and Peeters (2009) and Blonigen et al. (2007), Davies and Guillin (2011) for US outbound FDI and Ramasamy and Yeung (2010) for OECD inbound FDI mostly find similar motivations for investments in the goods and services sectors. For this thesis we have also run tests on sub-samples of FDI in goods and services. Results of these tests showed similar motivations for both sectors and did not deviate substantially from tests on aggregate FDI data. Due to the same data limitations that have plagued the study of Blonigen et al. (2007), we have refrained from an even more disaggregated industry analysis. Results of the sector analysis are not published in this master thesis, but are available upon request.

9_{In a related study, Blonigen et al. (2008) use the same framework as in Blonigen et al. (2007). However,}

(17)

16 are able to find some evidence of export platform FDI.10 In general though, Blonigen et al. (2007) conclude that most of the impact of the spatial interactions is picked up by country fixed effects. Moreover, they show that traditional determinants of FDI (such as trade costs, market size, etc.) are rather robust to the inclusion of spatially weighted variables, even though they do conclude that there is an omitted variable bias when a spatial lag and a surrounding-market potential variable are not included.

A number of other papers have used a similar approach to Blonigen et al. (2007). In general results are mixed. Analogous to Blonigen et al. (2007), albeit including a spatial error model as well, Garretsen and Peeters (2009) test for the impact of third-country effects on Dutch outward FDI into 18 host countries for the period 1984-2004. Their results mostly point out to the existence of complex VFDI. Weak evidence is found for export-platform motives when using a European sub-sample, although it should be noted that the spatial lag in this case has an impact close to zero. When dividing their sample by industry, they find that in the service sector Dutch FDI to a particular host is a substitute for FDI to other hosts, while in manufacturing, FDI to a particular host is a complement for FDI to other hosts. Garretsen and Peeters (2009) also incorporate a spatially weighted tax variable, which has a negative impact on all types of FDI. Despite these significant findings, it is important to note that Garretsen and Peeters (2009) confirm that spatial auto-regression is to a large extent picked up by country fixed effects. Similar results are obtained by Martínez-Martín (2011) for the case of Spanish FDI outflows to 50 host countries between 1993 and 2004. Her findings are suggestive of complex vertical FDI as well and she also finds that country dummies pick up a substantial part of the impact of spatial interactions. Again building on the framework of Blonigen et al. (2007), studies by Ledyaeva (2009) and Chou, Chen and Mai (2011) analyze the determinants of FDI for Russian regional FDI inflows and Chinese outward FDI, respectively. Both studies are not able to pick up any spatial interaction effects and confirm the observations by Blonigen et al. (2007), as well as Garretsen and Peeters (2009), that results are sensitive to sample selection. Ledyaeva (2009) concludes that FDI inflows into Russian regions between 1995 and 2005 are mainly horizontal in nature. Probably due to the small size of their sample group no fixed effects are employed. Chen and Mai (2011) run regression including and excluding time-period and spatial fixed effects on a sample of 61 host countries of Chinese outward FDI for the period 1993-2008. They are not able to find

10_{As indicated, the number of countries for which data is available in the industry analysis of Blonigen et al.}

(18)

17 support for the impact of third-country effects on Chinese FDI, as in most of their specifications (including a spatial error model) the surrounding-market potential and autoregressive parameters are insignificant.

Poelhekke and Van der Ploeg (2009), as well as Casi and Resmini (2011), adopt slightly different approaches compared to the aforementioned studies. Both argue that previous work has been too restrictive on the channels through which third-country effects operate. After re-estimating the specifications used by Blonigen et al. (2007) using similar samples, Poelhekke and Van der Ploeg (2009) try to unbundle the spatial lag by including a whole range of spatially weighted explanatory variables. This procedure is open to criticism, as it is rather bewildering to exclude a spatial lag a priori. By doing so, the possibility exists that they have missed out on effects which operate through the spatial lag. It is therefore that we attach less value to the results obtained by Poelhekke and Van der Ploeg (2009). The procedure adopted by Casi and Resmini (2011) is more refined. Casi and Resmini (2011) investigate what factors drive the uneven distribution of FDI inflows across European regions and whether regional performance is affected by national performance. Their dataset consists of 41 European countries, subdivided into 264 regions, for the period 2005-2007. On the whole, they find that spatial heterogeneity, measured with the use of dummies that help to identify national and regional effects, is more important than spatial auto-regression.11 Preferring a spatial error model over a spatial lag model, they argue that spatial interactions arise through other channels than FDI. Using a model specification including a number of exogenous spatial interaction effects and, in contrast to Poelhekke and Van der Ploeg (2009), excluding a spatial lag or a spatial error term only after both turned out to be insignificant, they identify the growth rate of GDP and labor costs in surrounding markets as channels through which spatial interactions arise. Nonetheless, for our research these findings are of less interest, as they are not obtained in a framework that enables researches to link results directly to general theory on FDI. Moreover, from an econometric perspective, one would like to include a spatial lag whenever significant in order to avoid biased estimates. We come back to this point in the next section.

Summing up the evidence, results obtained by using a framework based on the work of Blonigen et al. (2007) are mixed and often inconsistent with theoretical models that predict the main motivation of FDI. Still, these studies offer room for a number of important observations. First, the abovementioned papers show that sample selection is of utmost

11

(19)

18 importance in research into the spatial determinants of FDI. Although the majority of the discussed papers conclude that space matters, the impact of space is, in general, larger and more consistent with theoretical explanations, when samples are confined to a more coherent group of countries and/or aggregated by industry. Secondly, one of the more consistent findings is that spatial fixed effects seem to absorb a large chunk of third-country effects. Such an observation is in line with analogies from the trade literature which indicate that spatial fixed effects adequately control for the impact of third-countries in gravity model settings (Feenstra, 2002). As a third observation we note that there is a rather striking lack of conclusive evidence in support of export-platform FDI when a spatial econometric approach is applied.12 Above all in studies that use a European sub-sample one would expect to find stronger evidence in favor of export-platform FDI, as relatively high trade costs with the home country compared to trade costs between Europe’s relatively integrated markets are thought to facilitate such investments (Ekholm et al., 2007). Only very weak evidence of export-platform FDI is found by Garretsen and Peeters (2009). Blonigen et al. (2007) hypothesize that border costs might account for the absence of export-platform FDI evidence. The higher the border costs in the target region, the more firms are inclined to choose the larger market as a base from which products are spread, explaining a negative coefficient estimate for the surrounding-market potential variable. However, such a hypothesis is not only at odds with the standard theoretical model of export-platform FDI, it also contradicts with case study evidence which describes how among others small countries like Belgium, Ireland and the Netherlands have profited from a relatively large influx of export-platform FDI (Ekholm et al., 2007; Navaretti and Venables, 2004).

A fourth observation regards the lack of innovation in the methodological approaches used in previous studies, which might offer an alternative explanation for the often ambiguous findings by previous research. Apart from adding or subtracting variables to the initial model offered by Blonigen et al. (2007), which in some cases may have led to biased results (Poelhekke and Van der Ploeg (2009), none of the subsequent researches have questioned its methodological set-up. However, over the past few years the field of spatial econometrics has seen a number of important contributions to its rapidly expanding library, providing researchers with valuable new insights on how to model and interpret spatial spillover effects (Elhorst, 2010a). As indicated, the main novelties of our study are related to a more

12_{When we put our research in a broader perspective, we note that negative spatial autocorrelation, as one would}

(20)

(21)

20

3. Method

3.1 FDI and modeling space

In order to provide the reader with a better understanding of the method of this research and the novelties that are introduced, we start this section with a discussion on methods applied in prior studies. We proceed in two steps. We start with a discussion on the basic non-spatial linear regression models used in most FDI research and we will motivate which of these models will serve as a benchmark model in our research. After that we explain why we prefer a spatial lag model over a spatial error model.

3.3.1 Modified gravity model versus knowledge-capital model

Much of the previous work on FDI determinants is based on some sort of modified gravity model (Navaretti and Venables, 2004). These models consist of a number of host and parent country variables. To explain bilateral FDI flows some of the more common variables used are market size in the host and home country, and between country factors such as distance, sharing a common language, similarities in institutions, and sharing a border. Moreover these models are sometimes modified with variables on skill endowments, which allow making a distinction between HFDI and VFDI. When using aggregate data on FDI, such models have the disadvantage that they can only give some sort of aggregate effect with regard to various types of FDI (Navaretti and Venables, 2004). For example, in the case of HFDI trade costs are positively related to the inflow of FDI, while in the case of VFDI theory predicts that inflows increase when trade costs are low. Using a simple gravity model these conflicting predictions cannot both be shown in the data, as this model is only able to show a dominant form of FDI. Moreover, it could even be that these conflicting impacts can lead to none of the forms of FDI showing up in the data.

(22)

21 by using various interaction terms. However, in practice formal tests have shown that this model does not provide better results than purely horizontal models or models tested on a specific subgroup of countries (Navaretti and Venables, 2004). Moreover, this model is not able to indicate which motivation dominates. Thereby, our sample can be split in sub-samples for HFDI, VFDI, and complex forms of FDI, which makes the use of a model that incorporates multiple types of FDI less of an issue. We therefore follow the bulk of the literature by using a modified gravity model as our benchmark equation.

3.1.2 Spatial lag model versus spatial error model

In the literature section we explained that modified gravity models (Blonigen et al. 2007; Garretsen and Peeters, 2009), as well as the knowledge-capital models (Baltagi et al., 2007a), have been used to obtain baseline results, which are subsequently compared to models including spatial interaction effects. Spatial interaction effects have been modeled in two ways: with the use of a spatial lag model and with the use of a spatial error model. Blonigen et al. (2007) and related work, use a spatial lag model including both exogenous (e.g. surrounding-market potential) and endogenous (spatial lag) spatially weighted variables. In essence these model specifications are restricted forms of what the literature has labeled a spatial Durbin model (Elhorst, 2010a).13 An advantage of taking a spatial lag term into account, instead of a spatial error term, is that the former can be easily grounded upon FDI theory (Blonigen et al., 2007; Garretsen and Peeters, 2009).

Studies that have incorporated spatial interaction effects within a knowledge-capital model have used spatial error specifications to test for their hypotheses (Baltagi et al., 2007a; Uttama and Péridy, 2009). In a spatial error model spatial autoregressive errors account for stochastic shocks to FDI in neighboring countries that spillover into the host market. The advantage of taking spatial errors into account is that estimations become more efficient. However, a clear disadvantage is that no spatial lag term is taken into account. The literature explains that if such a term is relevant, excluding it would lead to biased and inconsistent results (Elhorst, 2010a).

13_{In a full spatial Durbin model specification all variables have a spatially weighted counterpart as well. A}

(23)

22 Ideally, a model that estimates spatial linkages would be able to include both exogenous (surrounding-market potential) and endogenous (spatial lag) interaction terms and correlation effects (spatial error terms). However, from the literature on spatial econometrics we know that this is not possible (LeSage &Pace, 2009; Elhorst, 2010a). To circumvent this problem, it is argued that the best option available is to exclude the spatial error term. This may lead to a less efficient estimator when the real data generation process is a spatial error model, but avoids estimations to be biased (Elhorst, 2010a).

The econometric implications and the observation that a spatial lag model can be easily grounded upon FDI theory form the reasons why in this thesis we follow Blonigen et al. (2007) by using a spatial lag model instead of a spatial error model to incorporate space.

3.2 Basic model

As indicated, to analyze the impact of the hypothesized third-country effects on bilateral FDI, we use a gravity model specification, which is still the most widely used model for empirical research on FDI (Blonigen et al., 2007). First, we estimate a basic model without third-country effects, which we can then compare to a model that includes spatial interactions. Since all variables are commonly measured in logs, the basic specification is as follows:

(1)

The dependent variable is an N×1 vector of outward US FDI in host country i at year t. It will be measured by US MNE affiliates sales. In the data section we will give a more detailed explanation on why we choose for affiliate sales data. For now, it is important to note that data on affiliate sales enable us to split aggregate data by motivation, as US affiliates sales data is available on local sales (proxy for HFDI), sales back to the US (proxy for VFDI, and sales to third countries (proxy for export-platform FDI and vertical specialization). By doing so, we try to circumvent some of the limitations of earlier studies that focused on aggregate effects.

(24)

23 model variables (GDP, population, trade costs, distance and institutional quality) complemented by a variable that measures the skill level or skill abundance of the host country. For reasons of comparability we try to keep this group of host variables as similar as possible to Blonigen et al. (2007). Below we will give a brief overview of the expected impact of these variables on FDI. Because these variables are used as control variables, we keep the discussion mainly limited to the effects of these variables on aggregate FDI flows. When the literature is clear on any deviations with regard to the impact of these variables on specific types of FDI, we will briefly mention it though. enters the estimation as the variable that measures the market size of host country i at time t. In line with general theory, we expect market size to be positively related to FDI (Navaretti and Venables, 2004). The size of the population in country i at time t ( ) is used to control for the tendency of FDI to move between wealthier markets (Blonigen et al., 2007). If population increases, GDP per capita reduces. In other words, this variable is expected to have a negative impact on FDI. The host trade costs variable ( ) is a proxy measure of national protectionism and

captures the barriers that might hamper trade between the home and host country (Carr et al., 2001). The literature explains that higher levels of trade costs are associated with an influx of HFDI, as exports are replaced by affiliate sales (Brainard, 1997). In the case of other forms of FDI, higher trade costs will lead to less FDI. The effect of trade costs on aggregate data thus depends on which form of FDI dominates. Distance between home country d and host country

i is a traditional gravity model variable used to control for trade costs related

to transportation costs (effect unclear) and the costs of management (negative effect) (Blonigen et al., 2007). Just as with the trade cost variable, the full effect is unclear and depends on which type of FDI dominates. In line with Car et al. (2001), a variable on the amount of skilled workers relative to the total workforce in a country is added In general, greater skills are thought to be positively related to larger influxes of FDI (Ekholm, 1997; Carr et al., 2001). However, when an MNE intends to invest in a country to profit from its low skilled labor abundance, it could also have a negative effect on FDI when vertical motives dominate the investment decision. To account for the impact of investment risk on FDI inflows Blonigen et al. (2007) include a variable based on a composite index of political, financial and other economic risk indicators. Unfortunately, this composite index is not publically available. As an alternative, we follow Garretsen and Peeters (2009) by adopting a variable on quality of government to serve as a proxy to account for the investment

(25)

24 related to all types of investments, as higher risk is associated with higher costs (Blonigen et al., 2007).

We discard parent characteristics (e.g. GDP, population, trade costs, etc.) from the estimation. It is no longer necessary to incorporate such characteristics since the home country is always the US. These characteristics will only vary over time. Blonigen et al. (2007), as well as Garretsen and Peeters (2009), have captured these variations with a time trend variable. Following Baltagi et al. (2007a), we add time dummies (denoted by the term) to our specifications to control for unobserved time variation in a more precise way. In addition, we also estimate a two-way fixed effects model (Baltagi, 2007a), including both time-period as well as spatial fixed effects, to control for spatial heterogeneity. In the models that include country dummies the variable will no longer be in the equation, since its values are fixed over time.

3.3 Spatial lag model

To test for spatial linkages, we will make use of an adjusted form of the spatial lag model, modifying our baseline equation with two spatial interactions: the exogenous surrounding-market potential variable and the endogenous spatial lag. All estimations will include time-period fixed effects, while two-way fixed effects models will include spatial fixed effects as well.

The adjusted spatial lag model is formulated as follows: 14

₍₂₎

The dependent variable now not only represents aggregate FDI outflows, but also FDI outflows by type. For notational purposes subscripts on type of FDI are left out of the

14_{An alternative, more formal formulation of Equation (2) would be:}

(26)

25 equation. The non-spatial variables are similar to the variables in Equation (1). The surrounding-market potential variable (SMP) is reflected by the interaction term . denotes an vector of GDP in all other countries other than host country i (j ≠ i) at time t. is an spatial weighting matrix that captures the spatial arrangement of the host countries in our sample. is equal for all time periods and is defined as:

(3)

The diagonal elements of W are set to zero, as no host country can be its own neighbor (Elhorst, 2010a). The off-diagonal elements of W are the spatial weights. They are calculated by taking the inverse distance between any pair of host countries i and j ( i j

i j ). 15

Subsequently, the SMP is calculated for each target market i at time t by the natural logarithm of the sum of weighted GDPs of all other host countries (j ≠ i) at time t in our sample The construction of this variable is closely linked to the market potential variable introduced by Harris (1954) and used by among others Head and Mayer (2004). In contrast to these studies, we follow Blonigen et al. (2007) by not including host country GDP, as their findings reject a common coefficient on host country GDP and surrounding-market potential.16 Furthermore, to keep our work comparable to Blonigen et al. (2007) it is important to note that we take the natural logarithm of , instead of multiplying the spatial weights matrix with the natural logarithm of , as is common in applied spatial econometrics when estimating an exogenous interaction effect. We come back to this point below, when we introduce a more general spatial Durbin model specification.

The spatial auto regression term reflects the spatial lag. is

the natural logarithm of the vector of US outward FDI into all countries other than FDI outflows into host market i (j ≠ i) at time t. ρ is the spatial autoregressive coefficient. is equal in form to the weights matrix used in the construction of the SMP variable, but as is

15_{To keep the construction of the SMP variable in line with previous work by Blonigen et al. (2007), we refrain}

here from the normalization procedures used in the construction of the spatial weights for the spatial lag. We will come back to this point below.

16_{Blonigen et al. (2007) explain that it is easier to distinguish between different forms of FDI when host country}

(27)

26 standard in spatial econometrics spatial weights will be normalized (denoted by the * superscript). In line with the bulk of the literature, including Blonigen et al. (2007) and Garretsen and Peeters (2009), we will first perform tests on our baseline models using a row normalized spatial weights matrix, in which each off-diagonal element corresponds to

17 Such a procedure does not only contribute to the weighting matrix' stationarity condition being satisfied, as its largest characteristic root is bounded by +1 (Elhorst, 2010a; 2010b), but also makes that the spatial lag term "has the simple interpretation of row-sums being a proximity-weighted average of FDI into alternative countries" (Blonigen et al., 2007, p1311). However, a disadvantage of row normalizing a spatial weights matrix is that the mutual proportions between the elements are changed. Only relative economic distance matters when each row is divided by a different element, while absolute distances are no longer accurately reflected (Baltagi, 2007b).

In order to avoid this misspecification problem, we will test for an alternative normalization procedure in which W is divided by a single factor. To that extent, we follow a procedure proposed by Keleijan and Prucha (2010) who divide W by its largest characteristic root ( ), defining each off-diagonal element of W as:

. In contrast to a row

normalizing procedure, normalizing W by its largest characteristic root avoids potential misspecification problems by preserving the mutual proportions between the weight elements, while still satisfying stationarity (Elhorst, 2010b). We therefore assume that the later specification more accurately reflects distances and subsequently produces more accurate results on the impact of spatial interactions on FDI.

3.3.1 Sample divisions and alternative specifications

Apart from the results obtained by using aggregate FDI data, we will run regressions including the spatial interactions on samples split by type of FDI (HFDI, VFDI, and FDI involving third country exports), in order to find out if we can give a more precise prediction on the effect of spatial interactions on specific types FDI.

17_{Before row normalizing their spatial weights matrix Blonigen et al (2007), as well as Garretsen and Peeters}

(2009), first multiply inverse distances by the shortest bilateral distance (for example 173km) in their sample. The shortest bilateral distance weight then equals unity, while all other distance weights decrease by .

(28)

27 We will also run two robustness checks.18 First, we will test whether or not our results are robust against the way we measure distance in the off-diagonal elements of the spatial weights matrix. This robustness test is common practice in applied spatial econometric research, as in general the economic literature has not much to say about the specification of

W (Elhorst, 2010a). In line with Baltagi et al. (2007a) we adopt two alternative measures of

distance. In order to reflect a more rapid spatial decay, we will use a W matrix in which the spatial weights are squared. In order to reflect a slower spatial decay, we take the square roots of the spatial weights. Second, we will estimate a more common form of the spatial Durbin specification, in which the exogenous SMP variable will be similar in form to the spatial lag term. That is, we will use a spatial weights matrix standardized by its largest characteristic root and multiply it with the natural logarithm of , instead of using the natural logarithm

of the whole product as in Equation (3). The impactions of this adjustment are unclear and have as such not received attention in the literature on the spatial determinants of FDI. The spatial Durbin model specification is formulated as follows:

(4)

3.4 Diagnostic tests, model comparison and methods of estimation

Equation (1) is estimated with the use of OLS. We will test some of its assumptions by performing basic tests for normality, multicollinearity and hetroscedasticity. To test for the assumption that the residuals are normally distributed, we use a Jarque-Bera (JB) test. The JB test score is calculated out of the skewness and kurtosis coefficients. Large values of the skewness coefficient and kurtosis values that deviate from 3 will increase the value of JB (Hill et al., 2004). The normality assumption is rejected when the JB statistic exceeds the critical value of a chi-square distribution with two degrees of freedom. Tolerance and Variance Inflation Factor (VIF) tests will be used to check for multicollinearity. The rule of thumb is

18

(29)

28 that tolerance levels less than 0.10, and VIF scores higher than 10, are indicative of severe multicollinearity.

After estimating pooled, time-period and two-way fixed effects OLS models, we test for the joint significance of the time-period fixed effects, as well as for the joint significance of the spatial fixed effects, with the use of likelihood ratio (LR) tests. We note that in baseline equations similar to ours, the fixed effects estimator was preferred over a random effects estimator (Garretsen and Peeters, 2009). Although our data violates the random effects assumption of random sampling, we will use a Hausman test to find out whether or not a fixed effects model is better at describing the data than a random effects model.

The following step will be that we use the classic LM test (Anselin et al., 1988) and the robust LM test (Anselin et al., 1996), to test whether a spatial lag model or a spatial error model best describes the data. Anselin et al. (1996) show that the second test is better suited to trace the source of spatial dependence, as it is robust against the possibility that one type of spatial dependence (lag or error) biases the test results of the other type. Both tests are grounded upon the OLS residuals and have a chi-squared distribution with one degree of freedom (Elhorst, 2010a). The test results will give a first indication whether or not the omission of spatial interactions leads to biased results on the determinants of US outward FDI to the EU.

Next, we estimate the various adjusted spatial lag models related to Equation (2). Up to this point we follow a specific to general test procedure proposed by Elhorst (2010a). Elhorst advices further testing on whether or not these specifications should be simplified by dropping spatially weighted variables. 19 We refrain from this procedure. Not only because software limitations do not allow us to perform proposed LR tests that would help us identify which specification best describes the data, but also because we use sign and significance of the two spatially weighted variables to identify the main motivation behind FDI.

19

(30)

29 As is common, maximum likelihood (ML) methods will be used to estimate the spatial models, because of the OLS estimator its inability to estimate a model including a spatial lag. To explain why this is the case we take Equation (2) and rewrite it as:

(5)

whereby all variables are measured in logs and X represents the matrix of control variables ( ) and I the identity matrix of order N. Subscripts on time and country, as well as parameters on time-period and spatial fixed effects, are left behind. The product of the spatial multiplier matrix and the error term shows how one of the main assumptions of the OLS estimator is violated, as each element of the dependent variable FDI is correlated with the error term. In contrast to the OLS estimator, ML methods can take such endogeneity into account. The ML estimator computes the parameters that give the highest likelihood of the sample data. It assumes that the disturbance terms are independent and normally distributed with zero mean and constant variances (Elhorst, 2010b).

Blonigen et al. 2007, as well as Garretsen and Peeters (2009), also use ML estimates to estimate the impact of the spatially weighted variables on FDI. As an alternative, and in line with Baltagi et al. (2007a), instrumental variables or generalized method of moments estimators (IV/GMM) could have been used. An advantage of using IV/GMM estimators is that, in contrast to ML methods, they do not require the disturbances to be normally distributed (Elhorst, 2010a). However, a disadvantage is that the performance of these estimators is sometimes weak due to coefficient estimates outside its parameter space (Elhorst, 2010a). Therefore, when the normality assumption is not violated (for which we will test with the use of JB test statistics), we can safely choose ML methods over IV/GMM techniques.

To estimate the ML spatial lag models we use MATLAB routines provided by Elhorst (2010c).20 These routines take a bias correction into account based on the observation by Lee and Yu (2010) that ML estimates, obtained from two-way fixed effects estimations, do not yield consistent parameter estimates when the number of time-periods and spatial units are large.

An exploration of space in FDI

Master thesis