AN EMPIRICAL RECONCILIATION OF THE KNOWLEDGE – CAPITAL MODEL

AN EMPIRICAL RECONCILIATION OF THE

KNOWLEDGE – CAPITAL MODEL

LE VAN HA

Research Master in International Economics and Business

Faculty of Economics and Business, University of Groningen

An Empirical Reconciliation of

The Knowledge-Capital Model

Le Van Ha

S1713981

Research Master in International Economics and Business

Faculty of Economics and Business, University of Groningen

Supervisors:

Abstract

Contents

1 Introduction 2

2 Theoretical literature review 5

2.1 Firm types in the world economy . . . 5

2.2 Why firms choose to be MNEs . . . 7

2.3 When firms choose to be MNEs . . . 8

2.4 Models on MNEs . . . 10

2.4.1 The Horizontal MNEs model . . . 11

2.4.2 Vertical MNEs models . . . 14

2.4.3 The Knowledge-Capital model . . . 16

3 Empirical Estimation of the Knowledge-Capital model 18 3.1 Empirical literature . . . 18

3.2 New empirical specifications . . . 22

4 Data and results 25 4.1 Data . . . 25

4.2 Results and discussion . . . 26

5 Conclusion 33

1

Introduction

Multinational enterprises (MNEs) are corporations that locate their headquarter offices in one country but have significant fixed investments in other countries, as op-posed to national firms whose activities concentrate in only one location. MNEs are the forces behind increasing flows of foreign direct investment(FDI) between coun-tries. As FDI has become an important sector of almost all national economies, is-sues relating to MNEs have also received more and more attention from researchers. One of the central discussion in this area of research is the motivations behind MNEs’ decision to incur investment abroad. Starting from early 1980s, two theoret-ical models were introduced at quite the same time, they answer the same question, why MNEs can emerge, and they were introduced in the same country, the United States. There seems only one difference between them, they are based on assump-tions that mutually exclude each other, i.e. if one model is accepted, the other cannot hold true.

The first model was introduced by Markusen (1984), named horizontal (HOR) MNEs model. It assumes a flat world in which countries are identical but separated by walls of trade cost. In order to avoid the trade costs, firms choose multiple locations of production for the same goods or services. The other model was introduced in Helpman (1984), named the vertical MNEs (VER) model. Helpman’s model as-sumes that countries are very different in factor endowments but there is no trade costs. Firms therefore opt to break their production process into different stages that can be geographically separated. Each stage then will be located wherever needed factors of production are cheaper. The two models stayed independently for a decade or so until Markusen (1997) brings them into a common framework known as the Knowledge-Capital (KC) model.

The KC model takes advantage of the computer time when simulation methods can be used to shed light on economic agents’ behavior in complicated models. First, a certain set of parameters on trade cost, investment cost, input requirements are chosen. Then relative factor endowments between countries are allowed to vary from one extreme case where countries are identical and trade cost is high (HOR model) to the other extreme case where relative factor endowments are totally different and no trade cost exists (VER model). In this direction 1, we will start with a world where HOR MNEs dominate to a mixed HOR-VER MNEs world before VER MNEs become the dominating production structure. Using different set of parameters, the model can make interesting predictions about the directions of capital flows that mimics the real world patterns.

1_{Of course we can go the other way round, from one extreme where relative factor endowments are totally}

This theoretical beauty has quickly induced substantial efforts from the academic circles in testing its empirical performance. While the theoretical model is clear and attractive, empirical investigation has encountered some serious problems (Blonin-gen et al. (2003), Davies (2008), Mariel et al. (2009)). There have been many papers discussing various issues relating to the estimation of this model since the pioneer work by Brainard in 1997. The general agreement is that the HOR model well fits the data while the VER often fails (Markusen et al. (2003)). It is a very common feeling that VER FDI is present out there. And such a failure really creates a puzzle. Carr et al. (2003) is the first paper to claim that VER FDI is important. However, Markusen and Maskus (2002), Bloningen et al. (2003), Helga (2005) find that the HOR model cannot be rejected in favor of the KC model. Mariel et al. (2009) suggests that the the relationship between VER FDI and relevant explana-tory variables is time variant and therefore normal econometric model cannot find evidence on VER FDI. Davies (2008), Brancoiner (2005) finds VER FDI relevant by including nonlinear terms of factor endowment difference in econometric model. All in all, authors agree that VER model poorly fits the data because the relation-ship between VER FDI and factor endowment differences is highly nonlinear and nonmonotonic (Markusen et al. (2003)).

In this thesis, we contribute our efforts to the current empirical literature on the KC model in solving the missing evidence on VER FDI. The questions needed to be answered are: What is an appropriate econometric specification to test the pre-dictions by the KC model? and how relevant is VER FDI?

By examining the theoretical models, we find that previous authors have focused too much on the nonlinear and nonmonotonic characteristics of the relationship between VER FDI and factor endowment differences. They therefore overlook some serious misspecification of the empirical models. There are also problems with statistical as-sumptions that they made. The contradictory empirical results are the consequences of both econometric misspecification and inappropriate estimation methods.

We come up with a new econometric model that includes the host country fac-tor endowment as an explanafac-tory variable to solve the nonlinear and nonmonotonic problem. We then collect data on FDI of each OECD country to test our new speci-fication. This is a much larger and more comprehensive dataset than any other that has ever been obtained before. We are able to show that, under the new econometric specification and appropriate estimation methods, the nonlinear and nonmonotonic problem is solved. VER FDI is found to exist and is as important as HOR FDI. We are also able to explain the causes of contradictory findings in the past.

2

Theoretical literature review

In this section, we will describe the theoretical models on MNEs, focusing mainly on the KC model. However, as Davies (2008) puts it, the KC Model is extremely complex and a full treatment would require a some 400 page book similar to that of Markusen (2002). Therefore, we resort to a verbal description of the relevant models to give readers an overview on the underlying assumptions, how the models work and what are the main theoretical predictions. Mathematical formulations will be limited to the minimum and interested readers are directed to relevant works by previous authors. The theoretical summary in this section is based on Markusen (2002).

2.1 Firm types in the world economy

Let’s start with a 2x2x2 world that has been very familiar in international economics literature. There are two countries, two goods and two factors of production. The countries are subscribed i and j, and when standing together the indices mean home and host country respectively. Put it more clearly, if someone says f irmij, he or she means this firm is headquartered in country i, the home country, and invest in or export to country j, the host country. The two production factors are skilled labor (S) and a composite factor (L). This composite factor includes all production fac-tors except S. Markusen (2002) notes that empirical evidence suggests that skilled labor is generally a crucial factor in understanding multinationals, thus justifying this dichotomized classification of inputs. Factor of production are assumed to be immobile across borders.

The two goods are named X and Y . Good Y is produced with constant returns to scale technology by a competitive industry, using both L and S. Good X is dis-tinguished from good Y by the assumption that its production is more S-intensive, under increasing returns to scale and being divided into two stages. In the first stage, the MNE must undertake some headquartering activity. This is intended to represent the development of blueprints, technologies, etc. The second stage is the actual physical production of the good. The headquarter and production stages can be geographically separated. Following Markusen (2002), we assume that the skilled labor intensity is ranked as:

headquarter activities > integrated X > X plant activities > Y sector.

T ype − hi: Horizontal multinationals that maintain plants in both countries, with headquarters located in country i.

T ype − hj: Horizontal multinationals that maintain plants in both countries, with headquarters located in country j.

T ype − di: National firms that maintain a single plant, with headquarters in country i. T ype − di firms may or may not export to country j.

T ype − dj: National firms that maintain a single plant, with headquarters in country j. T ype − dj firms may or may not export to country i.

T ype − vi: Vertical multinationals that maintain a single plant in country j, with headquarters in country i. T ype − vi firms may or may not export to country i.

T ype − vj: Vertical multinationals that maintain a single plant in country i, with headquarters in country j. T ype − vj firms may or may not export to country j.

Based on the way their headquarters and production plants are located, firms are generally classified into type − d (domestic), type − h (HOR MNEs) and type − v (VER MNEs) firms. HOR MNEs maintain plants at multiple place to produce the same good or service. They are seen as trade cost jumping MNEs. They have to bare a bigger lay out because they build multiple plants at different places. However, HOR MNEs do not have to pay the trade cost. VER firms are MNEs that locate their headquarter in one country and open production plant in other countries in order to exploit the factor price difference between these locations. This type will have to bare trade cost if they export back to their home country. Finally, domestic firms have both headquarter and plants in one place, producing to serve the local market and, if possible, serve the other market by exporting. Under different situ-ation, one firm type may have advantage over the others. For example, if country i is large relative to country j and the cost to export X from i to j is low, then type − d will definitely have advantage over other two firm types. If two countries are pretty much the same, type − h firms will gain advantage because they can use the service of one headquarter as input to two plants that serves the two markets without baring additional trade costs. If only type − d existed, there would be no FDI investment in the world and there would be no such a term as MNEs. This situation has been studied extensively in new trade theories, giving rich predictions about the direction of international trade flows.

and more important role in the world economy. We are interested in questions such as why and when will firms choose to go multinational? and whether or not the current theoretical models are empirically supported.

2.2 Why firms choose to be MNEs

To answer the question why, we go back to Dunning’s so called ownership, location and internalization (OLI) framework. Dunning (1977, 1981) proposes a unifying framework for determining the extent and pattern of foreign owned activities. It posits that multinational activities are driven by three sets of advantages, namely ownership, location and internalization advantages. It is the configuration of these sets of advantages that either encourages or discourages a firm from undertaking foreign activities and becoming an MNE. The OLI paradigm has proved to be re-markably adaptable. Since the publication of Dunning’s seminal contribution in 1977, it has been developed and extended in many directions (Mudambi, 2004). One important direction is to connect the framework with the firm (technology) and country characteristics in a consistent way (Markusen, 2002). Main contribu-tion to this task are Markusen (1984, 1997), Either (1986), Helpman (1984, 1985), Horstmann and Markusen (1996), and Markusen and Venables (1998, 2000). To better understand following models on MNEs, a further discussion of this frame-work at this moment is of order.

Ownership advantage means that firms must have certain products or production know-hows that other firms do not have. MNEs are observed to be associated with activities that employ high skilled labor such as R&D, marketing, scientific and tech-nical workers, product newness and complexity, and product differentiation. This fact leads to the hypothesis that MNEs are firms that are intensive in the use of knowledge-capital assets (Markusen, 2002). Knowledge-capital asset is a broad term that includes the human capital of the employees, patents, blueprint, procedures, and other propriety knowledge, and finally marketing assets such as trademarks, reputations, and brand names.

abundant countries and/or in large markets. The public-good characteristic means that knowledge-capital can be used jointly or even simultaneously at different plants without depreciation. The production of knowledge-capital may be costly but once they are created, they can be supplied at relatively low cost to foreign plants without reducing their value or productivity in other facilities. This property is crucial in firms’ decision to invest abroad. It is important to both VER and HOR firms be-cause MNEs involve in sectors that production process are geographically separated.

Location advantages are different between VER and HOR MNEs. For HOR firms, lo-cation advantage means that MNEs invest abroad to “domesticalize” their products and thus avoiding transport costs, tariffs, quotas and proximity advantages. High trade cost is the most important determinant of this type of MNEs. HOR firms may also emerge if the destination market is large and plants exhibit increasing returns to scale. On the other hand, VER MNEs export their service or intermediate products to foreign location. They may also import their final product to their home country. Trade cost is, therefore, a barrier to VER MNEs. If the factor prices are different between home and host countries and the production process may be fragmented into different stages, VER MNEs will choose to locate each stage where its inten-sively used factor is abundant and cheap. For VER MNEs, location advantage is the proximity to cheap factor of production.

Internalization is the opposite of the arms-length arrangement with a licensee or contractor. As mentioned before, knowledge-capital has the joint-input, public-goods property that creates ownership advantages. This property also makes it easier to be dissipated to MNEs’ partners. Licensees can easily absorb the knowl-edge capital and then defect from the firm or ruin the firms reputation for short-run profit. As a result, firms transfer knowledge internally, internalizing their knowledge capital assets by opening and owning plants abroad, in order to maintain the value of assets and prevent asset dissipation. With these three advantages, MNEs find their interest to invest abroad.

2.3 When firms choose to be MNEs

has to employ 2Sp to run two plants and Sh to run its head quarter. Denote G as the units of composite factor input that a plant has to employ to produce one unit of output. G and Sp are assumed to be constant across plants , regardless of which firm type they belong to. Finally, denote Ii, Ij as the investment costs, costs that are separated from fixed and variable costs, such as administrative costs, fees, waiting time and so on, that a firm has to pay if it open a plant. We now can calculate the cost function, ck

i, k = d, v, h, of firms with headquarter located in country i: cd_i = zi(Sd+ Sp) + wiG + Ii ch_i = ch_ii+ ch_ij, where ch ii= zi(αSh+ Sp) + wiG + Ii, and ch ij = zi(1 − α)Sh+ zjSp+ wjG + Ij.

Notice how the headquarter cost is divided between two countries in case of hori-zontal MNEs. α is between 0 and 1. It conveys the idea that by choosing to be a horizontal MNE, the headquarter can service two plants at the same time and therefore the headquarter cost is covered by profits from both plants, and

cv

i = ziSv + zjSp+ wjG + Ij.

In Markusen (2002) the investment cost Ii (Ij) is not mentioned. However, it is mentioned in econometric models when the KC model is estimated. We include it here to facilitate later econometric specifications. It can be understood that even when a country has relatively cheap factor prices, firms may not opt to locate their plants there if this investment cost is too high. We see that plant uses both skilled labor and the composite factor but headquarter is assumed to use skilled labor only. This is because firms need a certain amount of skilled labor to run their plants. For plants abroad, MNEs may hire local executives to operate but that is not always necessary, i.e. they can send executives to plants abroad, whichever cheaper. Do-mestic firms set up their headquarter and production plant in one place while VER MNEs have separated headquarter and production plants. We therefore assume that VER MNEs have to bear more headquarter cost to coordinate their operation in comparison with domestic firms, i.e.

ziSd≤ ziSv.

same time so it is also very likely that their headquarter cost should not double that of a domestic firm but more that that of a VER MNE. That is:

2ziSd ≥ ziSh ≥ ziSv

Next consider the profit functions of these firms, denoting Xk

iiand Xijk as the amount of a certain differentiated good X sold in country i and j respectively, k = d, v, h, and pi, pj as the price of X sold in each country. Their profit function can be cal-culated as: πd_i = X_iid(pi− cdi) + Xijd(pj − τjcdi) πh i = Xiihchii+ Xijhchij πv i = Xiiv(pi− τiciv) + Xijv(pj− cvi).

τ here represents transport costs. Firms will have to pay iceberg transport costs by scaling up their production cost by a factor τ ≥ 1 if they want to serve the market other than the location of their production plant.

If we consider firms based their headquarter in country j instead, the same re-sults can be derived. We further assume that within each differentiated product of category X, the technology is the same across plants, input factor markets are perfectly competitive and free entry and exit is allowed. In the long run, the market share of each plant on their home and foreign markets will be equalized.

Under the assumption that firms maximize their profit, an automatic mechanism is set up to decide which firm-type will be chosen based on considerations between firm characteristics and country characteristics. Firms have to pay factors that they employ for production and they will choose the as cheap as possible way to do so. In other words, firms will go multinational when they find doing that more profitable. However, firms’ profitability depends a lot on assumptions on firm and country char-acteristics. It is the difference on these assumptions that lead to the HOR, VER and KC model.

2.4 Models on MNEs

two plants. The advantage of firm level economy of scale may not be very apparent. However, in reality, the number of plants each firm can build may be much larger and therefore, the role of firm level economy of scale is very important. Plant level economy of scale may be an disadvantage to HOR MNEs as it may be more prof-itable for firms to focus the production activities in one plant and export to other markets. However, there have been a growing number of MNEs and therefore we will continue to assume that firm economy of scale is more important than plant economy of scale. Then what are the relevant country characteristics of interest to our analysis?

Examining the profit functions above, we see that the important factors are market size, trade cost, factor price and investment cost. Investment cost is an important determinant on the attractiveness of a country in the eyes of MNEs. However, this factor is very clear and most MNEs model have focused on the other three factors. As an art of model building in economics, in such a multidimensional setting, we of-ten find authors trying to explain the response of economic agents in some directions while keeping others fixed. This is exactly what happens in the literature on MNEs. Prior to the introduction of the KC model, there have been two different models on MNEs, the HOR model, ignoring factor prices difference between countries and VER model ignoring trade cost. A closer look at these model will facilitate our understanding of the KC model that follows.

2.4.1 The Horizontal MNEs model

The first version of the HOR model was introduced in Markusen (1984). In this ver-sion, Markusen assumes a flat world where countries are identical. That is, prices are the same in both countries:

pi = pj, wi = wj, and zi = zj.

Prices of the final goods are the same between countries. If a firm export to another market, it has to pay a trade cost τ ≥ 1. The price of imported product will, there-fore, be higher than the price of the same product produced domestically. If trade cost is prohibitively high, firms will always find opening a production plant abroad to serve the local market more profitable, giving rise to HOR MNEs. HOR MNEs is, as a result, viewed as trade cost jumping MNEs. Now we will compare the profit and cost functions between each firm type, assuming, for simplicity, that each plant will produce one unit:

Type d firms:

cd

cd j = ziSd+ ziSp+ wiG + Ij πd i = Xiid(pi− cdi) = pi− (ziSd+ ziSp+ wiG + Ii) πd j = Xjjd(pj− cdj) = pi− (ziFd+ ziFp+ wiG + Ij) Type h firms: ch = ziSh+ 2ziSp + 2wiG + Ii+ Ij πh _{= 2p} i− ch = 2pi− (ziSh+ 2ziSp+ 2wiG + Ii+ Ij) Type v firms: cv i = ziSv + zjSp+ wjG + Ij

πv_i = pi− (ziSv_{+ ziSp+ wiG + Ij)}

πv_j = pi− (ziSv+ ziSp+ wiG + Ii)

Taking into account the assumption that 2ziSd ≥ ziSh ≥ ziSv, from the profit functions, we have:

πd

i − πiv = ziSv− ziSd+ Ii− Ij > 0,

unless Ij is much higher than Ii, which is not very likely. As a result domestic

firms will out compete vertical MNEs and type − v cannot be chosen in this ”iden-tical” world. We are left with two firm types, d and h, and:

πh_{− π}d

i − πjd= 2ziSd− ziSh > 0.

This inequality means that when the two countries are identical, one HOR MNE can out-compete two domestic firms located in each country. HOR MNEs are more profitable because they have to pay less headquarter cost and no trade cost.

closer to observed patterns of investment (Davies, 2008). Using simulation, the au-thors show that the volume of bilateral FDI is negatively related to the difference in factor endowment between two countries. To see this argument, let’s start with an identical world where home and foreign countries are similar with the same factor price. Suppose now that country i becomes relatively more skilled labor abundant. Country j will lose its attractiveness to the X sector due to two reasons. First, firms will choose to locate their headquarter in country i because i is more skill abundant. The world FDI now, if exist, will flow only one way from country i to country j. Furthermore, as j becomes less skilled abundance, the assumption that X is more skilled labor intensive than Y makes it more profitable to base production plants in country i rather than opening a new plant in country j. As a result, the remaining unilateral FDI flows decreases as country j becomes less skilled abundant. In other words, HOR FDI flows decrease as the skilled labor abundance between home and host countries increases.

In Figure 1 2_{, world MNEs activity level is graphed against the world endowment}

distribution between two countries 3_{. The vertical axis is the world MNEs activity} level. Country i is at the South-West (SW) origin, country j is at the North-East (NE) origin. In the first version of the HOR MNEs model, countries are identical, meaning that they lie on the SW-NE diagonal. Start from the SW origin, world affiliate production is low because if one country (country i) is small, it is less at-tractive to open another horizontal plant there. If we move from the SW origin to the NE origin, the world affiliate production will be largest when countries are very similar, that is countries are in the center of the Edgeworth-box. The same pattern applies if we move from the NE origin. This explains why world affiliate production volume has an inverted U-shape. In the second version, countries are also different in factor endowment, we see that HOR MNEs are still chosen but the world affiliate production falls quickly as the two countries become more different in factor endowment. In this figure, trade cost is set to be very high as it is an important determinant of HOR FDI.

It is also worth to notice the importance of market size to firms’ profit. Firms’ profitability depends positively on the amount of good X that they can sell in each market, the larger the market, the more the volume that they can sell in that mar-ket. In Markusen (2002), the simulation shows that if market size is doubled, the ”hump” in Figure 1 is higher. In conclusion, HOR models predict that HOR FDI is positively correlated with market size and host country’s skilled labor abundance and host country’s trade cost and negatively correlated with market size difference

2_{All figures mentioned in the text are included in the appendix}

3_{This is done in Markusen (2002) where foreign affiliate production is taken as a proxy for MNEs activities.}

and skilled labor abundance difference. For the purpose of later arguments, we want to emphasize on the important result that HOR FDI is positively related to skilled labor abundance of the host country and negatively related to the absolute skilled labor difference.

2.4.2 Vertical MNEs models

The VER MNEs model was introduced the first time in Helpman (1984). The pur-pose of this model is to catch the reality phenomenon that MNEs invest in developing countries to take advantage of factor price differences. As a result, factors that give rise to HOR MNEs such as trade cost are not considered. The important underlying assumption is that firm specific assets associated with marketing, management, and product specific R&D, that is the knowledge capital assets can be used to service production plants in countries other than the country in which these inputs are employed (Helpman, 1984). To make the arguments more intuitive, suppose that country i is more skilled labor abundant than country j, that is zi < zj, wi > wj, and then:

cd_i = ziSd+ ziSp+ wiG + Ii

ch_i = ziSh+ ziSp+ wiG + Ii+ zjSp + wjG + Ij

cv

i = ziSv + zjSp+ wjG + Ij.

Trade costs equal zero allows firm to produce in one place and service both home and foreign markets. The most important determinant now is the factor prices. The zero profit and free entry and exit assumptions ensures that the market share are equal across firms. For simplicity, assume that each firm produce two units and supply each market one unit. Their profit functions are:

πd_i = pi+ pj − (ziSd+ ziSp+ 2wiG + Ii)

πh

i = pi+ pj − (ziSh+ ziSp+ wiG + Ii+ zjSp+ wjG + Ij) πv

i = pi+ pj − (ziSv+ zjSp+ 2wjG + Ij). It is easy to see that:

πd_i − πh

i = zi(Sh− Sd) + (wj− wi)G + zjSp + Ij > 0,

out-compete HOR MNEs if the composite factor prices are not too different. However, in case that wj is much lower than wi, firms will choose to be type − v rather than type − h MNEs because:

πv

i − πih = zi(Sh− Fv) + (wi− wj)G + zjSp+ Ij > 0.

Of course the assumption zi < zj, wi > wj is made for the purpose of exposition only. VER MNEs will have the potential to emerge if zi

wi <

zj

wj. As noted by Davies

(2008), under the zero trade cost and different factor endowment assumptions, the integrated world equilibrium can be achieved through trade in goods alone if factor endowments of the two countries lie within the factor price equalization (FPE) set. There is no need for VER MNEs. However, if factors endowments of the two coun-tries lie outside the FPE set, we need to either allow for trade in factors or trade in headquarter services that give rise to VER MNEs.

The first version VER MNEs model by Helpman (1984) assumes that there are two types of factors, labor and a firm specific asset. The other version by Markusen (2002) further clarifies the firm specific asset as capital assets or headquarter ser-vices. This model then predicts that VER FDI will flow from a skill labor abundant home country to low skilled abundance host country. The volume of FDI flows is positively correlated with skilled labor difference between home and host countries. Before going to discuss the KC model, we want to emphasize again the assumption on skilled labor intensity that:

headquarter activities > integrated X > X plant activities > Y sector.

2.4.3 The Knowledge-Capital model

As noted by Carr et al. (2001), Markusen (2002), Bloningen et al. (2003) and Davies (2008), MNEs models prior to the KC model make assumptions that exclude one of the two MNEs types in equilibrium. The KC model brings both HOR and VER model into one framework and then varies parameters such as trade cost, headquar-ter costs, plant fixed cost and the distribution of world endowment in order to see how firms behave in equilibrium. As a result, if trade cost is set equal to zero and countries are located in the NW - SE diagonal, we arrive at the VER model. If trade cost is really high, and countries are located on the SW-NE diagonal, the KC model becomes the HOR MNEs model. The advantage of the KC model is that, by us-ing simulation method, we can calculate the world’s affiliate production at different world endowment distribution points. At each point of world endowment distribu-tion, the simulation will calculate the world X affiliate production as a measure of the level of MNEs activities. This process produces results that we see in Figure 3, a combination of Figures 1 and 2.

Details about the derivation of this model is presented in Markusen (2002). Here we only discuss the results because the KC model is only a combination of VER and HOR model presented above but the mathematical presentations are rather tedious. Let’s think of the HOR and VER models as the extreme cases where we are at the NW (SE) origin (VER model) or the center (HOR model) of the Edgeworth-box. The question now is how firms behave as we move away from these points, i.e. we relax the assumptions that trade costs are zero and that countries are identical. Consider Figure 3, if we are in the center of the Edgeworth-box, with the presence of trade cost, HOR MNEs will dominate the world. Now moving away from the cen-ter to the NW origin, country i is becoming more and more skilled labor abundant while country j is going into the reverse direction. If there is no trade cost, HOR MNEs will quickly render their advantage to VER MNEs. However, the positive trade cost mitigate this process, forming a region when both VER and HOR FDI can exist before HOR FDI become zero. Figure 4 shows this pattern. We also see from this Figure that when one country is small and skilled deficient, there is no FDI at all.

want to note this fact and assert that if an empirical model is specified to test the prediction of the KC model, it is necessary that host country’s skilled labor level be included as an important explanatory variable.

As we move further to the NW origin, HOR MNEs will lose their advantage to VER MNEs. World affiliate production starts to rise to higher level than that when countries are identical. This is because now firms choose to locate their headquarter in country i and produce in country j to exploit the factor price differences unless trade cost to country i is prohibitively high that firms have to produce in country i to serve the local market. The larger factor price difference is, the more advantage VER MNEs get, that is, the more VER FDI flows. This result comes from the VER MNEs model. Figure 3 shows that world affiliate production is highest when one country is relatively small but skilled abundant. However, we should also notice that when one country is small and skilled deficient, there is no affiliate production at all.

3

Empirical Estimation of the Knowledge-Capital model

3.1 Empirical literature

Empirical investigations on the prediction of models on MNEs starts with Brainard (1997) who finds that overseas production by multinationals increases relative to ex-ports the higher are transport costs and trade barriers and the lower are investment barriers and scale economies at the plant level relative to the corporate level. He also suggests that multinational activity is more likely the more similar are the home and foreign markets, meaning that the HOR MNEs model is empirically supported. As a beautiful model that can bring two different and mutually excluded HOR and VER MNEs models, the KC model has receive much attention with respect to its empirical relevance. Markusen et al. (1999), Carr (2001) and Markusen (2002) were the first authors to suggest an econometric model as follows:

RSALESijt= β0+β1(GDPit+GDPjt)t+β2(GDPit−GDPjt)2_t+β3(SKit−SKjt)t+β4T Cjt+

β5T Cit+ β6Ijt+ β7IN T ER1ijt+ β8IN T ER2ijt+ β9DISTij + uijt(1)

Where:

RSALES: The real sale volumes of all country i0s affiliates in country j. This de-pendent variable is a proxy for the operation of MNEs.

(GDPit+ GDPjt): The total GDP of home and host countries. This variable

cap-tures the idea that affiliate production depends on market demand and therefore the market size. Given a level of trade costs and price difference, MNEs is more profitable the larger the home and host markets. β1 is expected to have a positive sign.

(GDPit− GDPjt)2: The square GDP difference between home and host countries.

The difference is squared to catch the fact that the graph of MNEs activities and market size differences in Figure 1 has a hump shape. Affiliate production is largest

for HOR MNEs when home and host countries are both large and similar. β2 is

expected to have a negative sign.

(SKit − SKjt): The skilled labor abundance difference between home and host

countries. This variable is included to take into account the VER MNEs activities. VER MNEs emerge to take advantage of factor price difference. The larger the difference, the more investment. β3 is expected to have a positive sign.

T Cjt: Trade cost of exporting to country j. This variable is relevant for HOR

firm types. β4 is expected to have a positive sign.

T Cit: Trade cost of exporting to country i. This variable is relevant for VER MNEs

in case that they want to export back to their home countries. β5 is expected to

have a negative sign.

INTER1: IN T ER1 = (GDPit − GDPjt) ∗ (SKit − SKjt), the interaction term

between GDP difference and skilled labor difference between home and host coun-tries. This variable captures the simulation results that when a country is small and skill abundance, VER MNEs will locate there and invest to less skill abundant countries.β7 is expected to have a negative sign.

INTER2: IN T ER2 = (SKit− SKjt)2∗ T Cjt, the interaction term between squared kill abundance difference and skilled labor difference between home and host coun-tries. This variable captures the simulation results that when a country is skill deficient and closed (high trade cost), there will be less FDI inflows. β8 is expected to have a negative sign.

DISTij: The distance between home and host countries. As noted by Carr et

al. (2001), trade cost and investment cost both increase as DISTij increases, there-fore, we do not know which effect is captured by this variable. However, subsequent

authors (Mariel (2009), Davies (2008)) find that DISTij is negatively correlated

with FDI activities. β9 is expected to have a negative sign.

Ijt: The investment cost that firms have to lay out to set up a production plant in country j. It is intended to catch the quality of the investment conditions such as the administrative costs, time needed to start a business or environmental regu-lations. β6 is expected to have a negative sign.

Different slightly from the previous section, in this empirical equation, i is used to denote the home (parent) country and j is used to denote the host country. Carr et al. (2003) then employ data on the US bilateral MNEs operation, that is the US MNEs investing abroad and foreign MNEs investing in the US. They found robust results, especially with respect to the skill difference variable, which has a positive and significant coefficient. The coefficient on skill difference is of special interest because it is this variable that shows the emergence of VER FDI. As noted by Carr et al. (2001), Mariel (2009), among others, the prediction of the HOR model has been strongly confirmed by empirical data. And it is the inclusion of a variable on skilled labor difference that distinguishes the KC model from the HOR model.

lie in a certain corner of the domain (Markusen and Maskus, 2002). Second, the econometric model uses only ordinary least square (OLS) and weighted least square (WLS) methods and did not exploit the advantage of a panel data set. This might lead to overstatement of the precision gains because the t-statistic is greatly inflated as pooled OLS underestimate the standard error of the residuals (Cameron and Trivedi, 2002). Furthermore, it is the fact that some countries are inherently more capable of attracting MNEs than others which suggests the presence of heterogene-ity that may be correlated with other variables on the right hand side of Equation 1, rendering OLS estimation results inconsistent. Third, as being emphasized several times in the theoretical part 4 , the Carr et al. (2001) econometric specification fails to include the host country skill abundance variable to the right hand side of Equation 1. This leads to inconsistent estimation of the coefficient β3. Markusen et al. (2003) attribute this inconsistency to the fact that the relationship between FDI flows and skilled labor abundance difference is highly nonlinear. We argue that it is true that the relationship is nonlinear and nonmonotonic. However, the problem of omitting variable (the host country skilled level) is a more severe. We will focus on addressing these issues in this section.

Bloningen et al. (2003) used the same data set by Carr et al. (2001) together with a new data set on bilateral FDI stock between OECD countries over the 1982 - 1992 as the dependent variable. The reason for taking FDI stock as the dependent variable is the limitation on data on world affiliate production. Only the US collect such data on a regular basis. They also use the econometric specification by Carr et al. (2001). The difference is that they take the absolute value of the skill difference variable and divided the sample into two subsamples, one is for the case that the home country is more skilled labor abundance (the positive subsample) and one is when the home country is less skilled labor abundance (the negative subsample). So their econometric specification should read:

F DIijt = β0+β1(GDPit+GDPjt)t+β2(GDPit−GDPjt)2t+β3(|SKit−SKjt|)t+β4T Cjt+ β5T Cit+ β6Ijt+ β7IN T ER1ijt+ β8IN T ER2ijt+ β9DISTij + uijt, (2)

except that variable INTER1 now should be IN T ER1 = (|GDPit − GDPjt|) ∗

(|SKit− SKjt|). A problem with this specification is the absolute value of GDP dif-ference is taken. The interaction term (GDPit−GDPjt)∗(SKit−SKjt) is included to reflect the simulation result that if the home country is small, (GDPit− GDPjt) < 0 but skill abundant, (SKit − SKjt) > 0, the world affiliate production is

high-est, i.e VER MNEs dominate the world. If |GDPit − GDPjt| is taken to

con-struct variable INTER1, this idea is lost and the reason for taking the interaction is unclear. We therefore suggest that ITER1 should be defined as IN T ER1 = (GDPit− GDPjt) ∗ (|SKit− SKjt|).

One further note here is that the dependent variable now can either be the real sale volume of all foreign affiliate of country i’s MNEs in country j or the FDI stock from country i to country j. This is because only the United State can provide sys-tematic data on foreign affiliate sales by US MNEs abroad and foreign MNEs in the US. Bloningen (2003) therefore collect the data on FDI stock by OECD countries as a proxy for FDI operations.

Bloningen et al. (2003) then argued that, according to the KC model, vertical FDI is larger when the skill difference between host and home countries is larger. Their emphasis is that according to the HOR model, FDI inflows decreases as the host country becomes less skilled labor abundant because X is the skilled labor in-tensive industry. On the other hand, according to the KC model, vertical FDI will rise as the skilled labor abundance difference between home and host countries gets large. They then argued that for the horizontal FDI model to be rejected in favor of the KK model, the coefficients on skill difference should have the same (positive) sign for both subsamples. If they have opposing sign, that is negative sign for the positive subsamples and positive sign for the negative subsample, the horizontal model cannot be rejected.

The interesting point is that using both Markusen et al. dataset and their new OECD dataset, they found that the coefficients are of opposing signs. Bloningen et al. (2003) then concluded that the horizontal model cannot be rejected in favor of the KK model.

Davies (2008) make a further refinement by pointing out that vertical FDI only flows from skilled labor abundant home countries to less skilled labor abundant host countries. For the skill-deficient country, which only has outbound HOR FDI, move-ment away from equal relative endowmove-ments decreases its outbound FDI as before. Davies also proposes a more general empirical specification by introducing a squared term of skilled labor difference , (SKit− SKjt)2, into the model with the hypothesis that “for the negative subset, FDI should fall as (SKit− SKjt) moves towards nega-tive infinity. This implies a posinega-tive coefficient for (SKit− SKjt) and a negative (or small positive) coefficient for (SKit − SKjt)2. For the positive subset, FDI should initially fall and then rise as (SKit− SKjt) rises, yielding a negative coefficient for (SKit − SKjt) but a positive coefficient on (GDPit− GDPjt)2”. His idea can be summarized in a graph as in Figure 5 where the vertical axis is FDI stock and the horizontal axis in skill difference. Davies (2008) equation is:

By using the same dataset by Markusen at el (2001) and Bloningen et al. (2003), the author claims to “find that when the parent is skill-abundant, but only slightly so, FDI is decreasing in the skill difference. This is consistent with HOR FDI domi-nating in this region. As the skill difference rises, this relationship reverses itself and FDI increases in the skill difference. This is consistent with VER FDI dominating investment when the parent is very skill-abundant” and “by using a more general empirical specification, I am able to control for the switching between HOR and VER investment in my estimation, which was a primary goal of the KC model”. What is puzzling is that with the same datasets, different authors find different re-sults and they seem to attribute to the nonlinear and nonmonotonic nature of the relationship between (SKit− SKjt) and FDI.

Branconier et al. (2005) employs data on affiliate production by MNEs from Sweden, the US and Japan to estimate the KC model with a new econometric specification by replacing (SKit− SKjt) with SKit and claims to find significant results. This is

very likely because SKit reflects the potential investment in both HOR and VER

FDI. However, this specification fails to connect the idea that VER MNEs make investment in order to take advantage of factor price differences. Host country skill level is not taken into account and therefore the specification does not really “test” the KC model. Mariel et al. (2009) argues that the previous studies find contra-dictory estimate for coefficient β3 in Equation 1 because the coefficient may not be constant over time and use a time varying coefficient model to estimate this equa-tion. They find that VER FDI is relevant. However, this model does not give a consistent prediction on the relationship between FDI and (SKit− SKjt) over time. Helga (2005) use the bilateral FDI data on Iceland and finds that the KC model of a small country is not supported. Some other authors such as Waldkirch (2003) use FDI data at sector level to test the KC model and also find that VER FDI is indeed relevant. In this study, we will focus only on the country level data.

3.2 New empirical specifications

All previous empirical studies explain the reason that they could not find evidence for vertical FDI due to the non linear and nonmonotonic relationship between MNEs operations and skilled labor abundance difference. For example, Davies (2008) rec-ognized that as (SKit− SKjt) gets large, vertical FDI emerges but horizontal FDI declines. The total result must be ambiguous. For reasons clearly indicated in pre-vious sections5_{, we propose to include the host country skilled labor force to Carr et} al. (2001) specification. The skilled labor force of the host country (SKjt) will cap-ture the horizontal FDI effects. Then (SKit− SKjt) will capture the vertical effect only. With this specification, the coefficient on SKjt should be positive and much

larger in magnitude than the coefficient on SKDIFF because as noted by Markusen et al. (1999), Bloningen et al. (2003), horizontal FDI dominates vertical FDI in reality. One further reason is that the value of FDI projects might also be positively correlated with the quality of the labor force of the host country. This proposition can be seen clearly if we think of the so called East Asian Flying Geese economic model in which there exists a hierarchy of technological transferring process. Most developed countries in the group such as Japan, South Korea will focus on high tech industries and least developed countries will focus on labor intensive activities such as assembly or garment production. As a result, if there is an FDI project in the automobile industry to Thailand and one in the garment and textile industry to Cambodia. It is reasonable to speculate that the former will receive a larger FDI inflow.

Let’s consider again the Bloningen et al. (2003) specification: taking the abso-lute value, named |SKit− SKjt|, and divide the dataset into positive and negative

SKit − SKjt subsamples. They argue that for the HOR model to be rejected in

favor of the KC model, the coefficient on (SKit− SKjt) in both subsamples must be positive. These authors find positive coefficient for the negative subsample and neg-ative for the positive subsample, i.e. negneg-ative sign for |SKit− SKjt| in both cases. The question is, what is the indication of the negative coefficient on |SKit− SKjt| found by Bloningen et al. (2003)?

For the negative subsample, as the value of |SKit− SKjt| increases, the home

coun-tries is losing its advantage as a headquarter location for MNEs. At the same

time, they are losing the potential of a financial source for FDI. It is not surprising that the coefficient on |SKit− SKjt| is negative. For the positive subsample, when

|SKit − SKjt| increases, there are three mechanisms working at the same time to

F DIijt = β0+β1(GDPit+GDPjt)t+β2(GDPit−GDPjt)2t+β3a(SKit−SKjt)t+β3bSKjt + β4T Cjt+ β5T Cit+ β6Ijt+ β7IN T ER1ijt+ β8IN T ER2ijt+ β9DISTij+ uijt, (4).

4

Data and results

4.1 Data

To test our new econometric model, we have collected a dataset on FDI stock be-tween OECD countries and the rest of the world for a 12 year period, from 1995 to 2006. This data is available online at SourceOECD website. Data on each OECD country is complete, i.e. includes all inward to and outward FDI from that country. For countries that are not in the OECD, there is only data on FDI between them and the OECD countries. In other word, data between non OECD countries are not included. As indicated before, MNEs activities can be measured either by affiliate real sales or by FDI stock. Bloningen et al. (2003) shows that these two proxies produce the same result. Davies (2008) even argues that FDI stock is a better mea-sure of MNEs activity due to three reasons: First, factor abundance difference, and thus factor price difference, is a long term consideration by firms. FDI stock better reflect this long term choice. Second, FDI stock is not as volatile to short term economic fluctuation of the host country as affiliate sales. Third, firms may have incentives to misreport their operations in at a certain time.

The independent variables pose some missing value problems, especially the skilled labor abundance. Carr et al. (2001), Markusen (2002), Davies (2008), Mariel (2009), Branconier et al. (2005) use groups 1 and 2 in Table 2C from the ILO Labor Statis-tics Yearbook as a measure of skilled labor abundance of a country. With this

measure, SKit = ((Group 1 + Group 2)/total labor), and therefore, 0 ≤ SKit ≤ 1.

Bloningen et al. (2003) use the linear interpolation data on average years of school-ing obtained from Barro and Lee (2002) dataset as a measure of skilled level. Here we use the first measure because years of schooling might not correctly reflect the fraction of skilled labor force that is relevant to the MNEs. It is a pity that data on skilled labor abundance is not available for the two largest developing markets, China and India. Data on some countries are only available for some years and missing for other years. If that happen, the missing year data is replaced by the average of the whole period.

(2001) specification6.

For the variable Investment cost, we use the Global Competitiveness Index, com-puted by WEF. Investment cost is understood as the cost that an MNE have to pay to set up a new plant. It involves host country characteristics. When an MNEs consider an FDI project to a country, it will take into account the costs related to red tape, the time needed to complete investment procedures, political stability and so on. The Global Competitiveness Index captures these factors. However, the number of countries ranked by WEF is not constant but increases over time and participating countries are ranked from number 1, the best, to the the number equal to the number of participating countries in a given year. This makes it very hard to compare the ranking one countries between two years. As an example, in 2005, there were 117 countries and in 1997, there were 53 countries ranked by the WEF. The Philippines was rank number 73 and 34 respectively in those two years. That does not mean that their competiveness fell sharply from 1997 to 2005 but rather due to the fact that some more competitive economies started to participate. To make this a more reliable measure, we divide the rank of each country by the number of participating countries. The investment cost index of the Philippines is now 0.584 and 0.645 respectively. This is a reasonable result because over the last two decades, foreign direct investment attraction policy have been implemented in most developing countries.

Finally, we use the weighted distance computed by the French Research Center in International Economics, CEPII, as a measure of the distance between two coun-tries. This measure takes into account the distance between the most important economic center between two countries instead of merely measuring the geographi-cal distance between the two capitals. With these variables, we obtain a complete dataset consisting of 26,220 observations. This is three times larger in comparison with Mariel et al. (2009). However, Mariel (2009) data covers only OECD countries for the 22 year period. That means the non OECD countries are not included. This makes our dataset more attractive in the sense that developing countries, which are becoming more and more important to MNEs activities, are also included.

4.2 Results and discussion

Table I presents estimation results of Equations 1 and 4. Column I, II, and III repli-cate the Carr et al. (2001) specification (Equation 1) using pooled Ordinary Least Squares (OLS), Random Effects (RE) and Fixed Effects (FE) methods respectively. The results in column I seem to support the empirical results found by Carr et al. (2001). Except for the case that the coefficient on trade cost of the host country has

6_{Carr et al. (2001) use the Trade cost index and Investment costs computed by the World Economic Forum}

the unexpected positive sign, all other variables are highly significant with expected sign. However, the estimation in column II, using RE method, are dramatically different, especially the coefficient on (SKit− SKjt) is now much smaller in mag-nitude and insignificant. This suggests that we have to pay more attention to the assumptions needed to obtain consistent estimates of the coefficients.

Let’s consider the regression model:

F DIijt = Xijtβ + γ0yt+ cij + uijt (5)

where X is a matrix of explanatory variables, β is the vector of corresponding co-efficients of interests, cij is the unobserved specific effects, yt is a vector of year dummies and uijt is the error terms. Let ijt = cij + uijt. For the results in Col-umn I to be consistent the explanatory variables must be weakly exogenous. That is:

E(ijt|Xijt, yt) = 0 (6)

For the RE estimates in column II to be consistent, the explanatory variables must be strictly exogenous and the unobserved specific effects must be uncorrelated with the explanatory variables or:

E(uijt|Xij, yt, cij) = 0 (7)

Where Xij = [Xij1; ...; XijT], and

E(cij|Xij) = c, (8)

c is a constant. If condition (8) holds, the regression model under consideration is called the constant coefficient model, OLS will be consistent. If all conditions (6), (7) and (8) hold, then pooled OLS and RE yield similar results. The estima-tion results in column I and II are not similar suggests that the unobserved fixed effects, cij, does not satisfy condition (8). Therefore, OLS or RE methods cannot yield consistent estimates. As a result, statistical inference in this case should based on the FE method estimates. Column III reports the estimates of Equation 1

us-ing FE method. The results are quite dismal. Except that (GDPit+ GDPjt) and

(SKit− SKjt)2 ∗ T Cjt have significant coefficients with the expected sign, all other variables have wrong sign or insignificant coefficients.

The most often cited reasons for such findings are that the relationship between (SKit − SKjt) and F DIijt is highly nonlinear and that there is too little within variation in (SKit− SKjt). If (SKit− SKjt) is almost constant overtime, then the FE method will wipe out the variable, yielding insignificant estimates. These two reasons seem to provide a satisfactory explanation for any failure to detect the

argue that this econometric specification is suffering from omitting variable prob-lems. The last subsection points out that it is necessary to include the host country skill abundance as an explanatory variable. This variable will capture the effects of host country skill abundance on HOR FDI. X is assumed to be a skilled labor intensive sector. Its production is therefore, positively correlated with the skill level of the host country. In Equation 4, β3b is expected to be positive.

Column IV gives the results of estimating Equation 4 using FE method, besides the explanatory variable specified in Equation 4, 11 year dummies are also included. The coefficient on the year dummies (not reported) are all significant at 1% level. This suggests that year specific effects are permanent determinant of FDI activities. This result is reasonable because FDI flows to a country are often volitile between years and there was a financial crisis (Asian Financial crisis) during the period. We therefore include the year dummies to all equations.

However, our interest lies in the estimation of the coefficient on (SKit − SKjt). It is this coefficient that decides whether or not the KC model better fits the real

world data than the HOR MNEs model (Bloningen et al. (2003))7. As we can see

from Column IV, the coefficients on (SKit − SKjt) and SKjt is significant at 1% level. What is more interesting is that the coefficient on SKjt is much larger than the coefficient on (SKit− SKjt). As mentioned earlier, when both (SKit− SKjt) and SKjt are included as explanatory variables, the latter will capture the effects of skill labor abundance of the host country on HOR FDI while the former captures the effects of skill difference on VER FDI. This finding once more confirms the empirical finding that HOR FDI dominate FDI flows in the real world. We therefore conclude that the failure to find that VER FDI is relevant is mainly due to econometric model misspecification, not only the nonlinearity of the relationship between (SKit−SKjt) and FDI.

Carr et al. (2003) uses weighted least square estimation to correct for heteroskedas-ticity. In Columns I, II and III, we also apply this correction method for the sake of comparison with their findings. However, the FE estimation method is efficient only if the error terms are identically and independently distributed (iid). Otherwise it is consistent but not efficient. To check this assumption, we run a regression with the the residuals from Equation 4 as dependent variable on its own first lag. The results (not reported here) shows that the residuals are not independent over time. We therefore use the panel-robust covariance matrix to correct for these dependence and heteroskedasticity. From now on, all standard error estimations are corrected for heteroskedasticity and serial correlations in this way.

The results in Column IV deserve some further comments. First, the coefficients on

(GDPit+ GDPjt), (GDPit− GDPjt)2, (SKit− SKjt), SKjt and T Cjt all have the expected signs and are significant at 5% level. However, the coefficients on T Cit and Ijt do not have the expected sign. T Cit is expected to have a positive coefficient because when the home country is more open, it is easier for there MNEs to export back home to serve the home market and therefore they may be more willing to make FDI investment. With a negative sign, the explanation can be that outsourc-ing is not the dominatoutsourc-ing motive for MNEs to invest abroad. MNEs makes FDI investment in order to serve the host market rather than to replace the expensive labor at home by the cheap labor force in the host country. Take the United States as an example, only 8% of US affiliate production abroad are export back to the home country (Markusen (1995)). This small fraction explains why the coefficient

on T Cit is negative. If a home country has certain advantage to export, that is

having a high openness index, its firms may choose to export to other countries rather than to open a new plant there. This is in line with the prediction of the VER MNEs model. The coefficient on Ijt is more puzzling because it is expected to be negative. In column IV, we find a positive and significant coefficient on Ijt. We will have further discussion on this coefficient later.

The coefficient on INTER1, the interaction term between GDP difference and skill difference between home and host countries has a positive sign instead of an ex-pected negative sign. As mentioned before, the reason to include the interaction between skill labor abundance difference and GDP difference is to catch the simu-lation result that the world FDI investment is largest when the home country has a small market but a high skilled labor force, MNEs will base their headquarter in this country and make VER FDI investment to other countries. However, it turns out that when China and India are not included in the sample, most of the OECD countries have a larger GDP that a non OECD country. As a result, in the region

of VER MNEs, (GDPit− GDPjt) and (SKit− SKjt) are always positive, it is not

surprising that the coefficient on INTER1 is positive. It indeed reflects the effects of skilled labor difference on VER FDI investment.

of the independent variables are included to the right hand side of an estimating equation, RE and FE methods yield the same results. The number of observations now increases from 19,191 to 26,220. The estimation results are reported in Col-umn V. The results are very similar to those in ColCol-umn IV. The coefficients on (SKit− SKjt) and SKjt are now larger but the sign pattern and the significance of each coefficients remain the same. The Mundlak terms are jointly significant at 1% level which once more confirms our argument that the individual unobserved effects are correlated with the explanatory variables.

The next step is to solve the puzzle found in Bloningen et al. (2003). In the

previous section, we explain theoretically why Bloningen et at (2003) can find such results. In Table II, we replicate their estimation using observations with positive dependent variable from our dataset. Column I, II, and III are the results from

running Equation 2 using OLS method 8_{. We find exactly the same sign pattern}

like these authors. In column I, when all observation are pooled together in to one estimation, the coefficient on |SKit− SKjt| is negative and significant at 1% level. The dataset is then divided into positive and negative (SKit− SKjt) subsamples. Columnn II shows the esitmates for the positive and Columnn III shows the esti-mates for the negative subsamples. We can see that, similar to claims by Bloningen et al. (2003), the coefficient on (SKit− SKjt) have oppositive signs. Negative for the positive subsample and positive for the negative subsample. These results mean that FDI is decreasing with respect to skill difference between home and host coun-tries, in line with the prediction by the HOR MNEs model. When FE method is applied (nor reported here), the coefficient on (SKit− SKjt) becomes positive for the positive subsample but not significant at 10% level. This means that Bloningen et al. (2003) econometric specification does not allow VER FDI to emerge because (SKit− SKjt) captures various effects at the same time.

In Column IV and V of Table II, we add the host country skill level, SKjt and

year dummies, as one explanatory variables and run Equation 4 to the positive and negative subsample using FE method. For the positive subsample (Column 4), the coefficient on (SKit − SKjt) become positive, significant and much larger in com-parison with that from column I, II, and III. However, the difference between the coefficient on (SKit− SKjt) and SKjt is now smaller. For the negative subsample (Column V), these coefficient are no longer significant at 5% level. Using the Chow test we find that there is a structural break at the point (SKit− SKjt) = 0. These findings suggest some more hints about our econometric specification.

Coming back to the previous discussion on the behavior of FDI flows as the world en-dowment distribution varies, we notice that for the negative subsample, there is only

HOR FDI flows from home country to host country. In this region, (SKit− SKjt)

is smaller, the smaller SKit given a more skilled labor abundant country j, meaning that the two variables are positively correlated. Furthermore, only SKjt is relevant for HOR FDI. The inclusion of (SKit − SKjt) in the estimation in this region is

therefore, redundant. (SKit − SKjt) is only relevant for VER FDI. As a result,

it should only appear in the estimation for the positive subsample. We then drop variable (SKit− SKjt) in Equation 4 when we estimate the negative subsample.

Column I and II in Table III show the results of the modified models. Column I is the estimation of Equation 4 for the negative subsample with out variable

(SKit − SKjt). The coefficient on SKjt is now significant as expected. Column

II shows the estimation results for the positive subsample. They are similar to that of Column IV in Table 1 when Equation 4 is used to estimate the whole dataset. The noticeable difference is that the coefficient on (SKit − SKjt) is much larger now. This observation can be explained that when the sample is divided, the re-verse mechanisms determining the level of FDI are separated and therefore, VER FDI can fully show up.

In Column III we estimate the econometric specification by Davies (2008) (Equa-tion 3). As men(Equa-tioned before, Davies (2008) argues that VER FDI only emerges when (SKit − SKjt) is sufficiently larger than zero. He expects the coefficient on (SKit− SKjt) to be negative and the coefficient on (SKit− SKjt)2 to be positive in this region. If OLS is used to estimate this model, the results (not reported here) are significant and have the sign pattern as he claims. However, as mentioned before, the unobserved fixed effects are correlated with the explanatory variables and OLS estimates are not consistent. We then run this model again using FE methods. As we can see in column III, the sign pattern is still the same but the interested coef-ficients are no longer significant at 5% level. We continue to add the host country skill level variable to Davies (2008) specification. The coefficients on (SKit− SKjt) and (SKit− SKjt)2 become insignificant. This suggests that the inclusion of the squared term of (SKit−SKjt) is not necessary when the model is correctly specified.

One remaining problem is that the coefficient on Ij still does not have the

5

Conclusion

In this thesis we have reviewed both theoretical and empirical literature on the KC model. This process has facilitated our refinement of previous empirical specification to come up with a more complete econometric model that is capable of distinguish-ing different mechanism determindistinguish-ing the flows of FDI. Besides careful discussion on the appropriate estimation methods, the most important modification is to include the host country skill level as an explanatory variable. Previous empirical model fail to do this, leading to the fact that variable (SKit− SKjt) capture reverse impacts of skill labor abundance in home and host countries on FDI. Its coefficient is therefore insignificant and even does not have the expected sign.

The inclusion of the host country skill level as an explanatory variable reveals that VER FDI is relevant. The theoretical model predicts that VER FDI only flows from one country to another if the home country is more skilled labor abundant than the host country, (SKit − SKjt) > 0. When (SKit− SKjt) is negative, there is only

HOR FDI flows. This suggests a structural break at the point (SKit− SKjt) = 0

and (SKit− SKjt) should only be included as an explanatory variable where VER

FDI has the potential to emerge. That is only when (SKit− SKjt) is positive. Our dataset confirms a structural break point at (SKit− SKjt) = 0 and show that the

coefficient on (SKit − SKjt) is much larger when the positive (SKit− SKjt)

sub-sample is estimated separately. The final conclusion is that when the home country is more skill abundant than the host country, VER FDI is relevant.

As MNEs often operate in the skilled labor intensive sectors. This leads to a

some-what consensus that the relationship between VER FDI and (SKit− SKjt) is

non-monotonic and nonlinear. Our empirical results show that when the host country level is included as an explanatory variable, this nonlinear and nonmonotonic rela-tionship can be approximated linearly. Due to the limited variation in (SKit−SKjt), the inclusion of its squared and cubic terms leads to multicollinearity problem and the coefficients on skill difference variables become insignificant. However, we still agree that the relation between VER FDI and (SKit− SKjt) is non linear because our experiments show that the coefficient on (SKit− SKjt) changes dramatically at different region of the domain.

6

Appendix

TABLE I VARIABLES I II III IV V (GDPit+ GDPjt) 5.368*** 3.238*** 7.861*** 5.730*** 5.629*** (0.301) (0.587) (1.992) (1.879) (1.857) (GDPit− GDPjt)2 -0.000240*** 2.24e-05 -0.000188 -9.93e-05 -0.000122 (3.31e-05) (6.03e-05) (0.000128) (0.000115) (0.000115) (SKit− SKjt) 9939*** 980.7 -2643 12250*** 13946*** (1220) (2292) (3550) (4371) (4334) SKjt 31392*** 32351*** (7665) (7601) T Cj 28.45*** 17.80*** -5.920 -36.55*** -28.60*** (2.786) (4.932) (9.783) (10.71) (9.875) T Ci 13.94*** 10.35* -7.507 -43.62*** -26.14** (2.706) (5.616) (10.12) (10.83) (10.21) (GDPit− GDPjt) ∗ (SKit− SKjt) -4.151*** -2.823** 8.323*** 7.145** 6.190** (0.764) (1.375) (2.838) (2.829) (2.699) (SKit− SKjt)2∗ T Cjt -1127*** -816.0*** -764.4*** -713.9*** -656.6*** (83.60) (138.5) (175.9) (193.4) (181.3) Ij -9020*** -3354*** -742.4 3621*** 1849** (367.5) (483.8) (917.0) (817.6) (757.0) DISTij -0.481*** -0.468*** -0.580*** (0.0216) (0.0602) (0.0645) Constant -926.5 -711.4 -10211*** -7056 -36416*** (646.0) (1044) (3205) (4586) (4526) Observations 19191 19191 19191 19191 26220 R2 _{0.237 0.202 0.132 0.147}

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

In this table, we estimate Equation 1 and Equation 4 in the text. Column I, II and III are the estimation results of Equation 1 using pooled OLS, Random Effects and Fixed Effects methods respectively. Column IV gives the estimation results of Equation 4 where (SKjt) is included using Fixed Effects method, the unobserved fixed effect

is (cij). Column V uses the same specification as Column IV but using the Tobit model. In column I, standard

TABLE II

VARIABLES I II III IV V

(GDPit+ GDPjt) 5.610*** 4.336*** 7.145*** 4.456** 9.794***

(0.109) (0.292) (0.570) (1.933) (3.666) (GDPit− GDPjt)2 -0.000314*** -0.000144*** -0.000400*** -1.59e-05 -0.000350*

(1.14e-05) (3.26e-05) (6.13e-05) (0.000125) (0.000208) |SKit− SKjt| -30324*** (3601) (SKit− SKjt) -4099 77320*** 22714** 4576 (3042) (6596) (10451) (15863) SKjt 28806** 26384* (11398) (14351) T Cj 20.79*** 12.02*** 49.05*** -40.31*** -21.25 (3.503) (2.663) (6.996) (14.96) (15.50) T Ci 17.88*** 25.69*** 13.40*** -17.54 -53.02*** (3.147) (3.868) (3.517) (13.07) (16.68) (GDPit− GDPjt) ∗ |SKit− SKjt| 0.298 (0.260) (GDPit− GDPjt) ∗ (SKit− SKjt) -3.149*** -0.974 9.257** 12.55** (0.945) (1.184) (4.586) (5.269) (SKit− SKjt)2∗ T Cjt -120.8 -800.7*** 1330*** -1116** -893.9* (176.8) (148.7) (223.2) (494.0) (513.1) Ij -7411*** -8685*** -10756*** 3556*** 5082*** (518.9) (457.7) (669.1) (1269) (1044) DISTij -0.428*** -0.443*** -0.365*** (0.0271) (0.0285) (0.0309) Constant 43.45 1050 -2656* -6905 -15216 (733.5) (892.0) (1530) (4855) (9353) Observations 19191 10551 8640 10551 8640 R2 _{0.234 0.239 0.259 0.172 0.145}

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

Chow Test: (No structural break at (SKit− SKjt) = 0), p-value: 0.0293

This Table presents results from estimating Equations 1, 2 and 4 in the text. Column I is the results of running Equation 2. Columns II and III are the results of running Equation 1 on positive and negative (SKit− SKjt)