ISSN 1750-4171

DEPARTMENT OF ECONOMICS

DISCUSSION PAPER SERIES

Location of Foreign Direct Investment in the Central and Eastern European Countries: A Mixed and Multilevel Data Approach

Simona Rasciute and Eric J. Pentecost

WP 2008 - 04

Dept Economics Loughborough University Loughborough LE11 3TU United Kingdom Tel: + 44 (0) 1509 222701 Fax: + 44 (0) 1509 223910

http://www.lboro.ac.uk/departments/ec

1

The Location of Foreign Direct Investment in the Central and Eastern European Countries: A Mixed Logit and Multilevel Data Approach

Simona Rasciute and Eric J. Pentecost

Department of Economics Loughborough University Loughborough Leicestershire LE11 3TU UK

First draft May 2008

Abstract

This paper uses the Mixed logit (ML) model and a novel three-level dataset to examine the factors explaining 1,108 foreign direct investment (FDI) location decisions into 13 Central and Eastern European countries (CEECs) over an eleven-year period between 1997 and 2007. The ML model approach is superior to other methods in that it allows for random taste variation, unrestricted substitution patterns and correlation in unobserved factors over time. The highly significant empirical results, based on a general underlying economic model of imperfect competition, show that the responsiveness of the probabilities of choices to invest in a particular country in CEE to country-level variables differs both across sectors and across firms of different sizes and profitability. The results generalise previous studies that used only country-level data or only industry- and firm-level data to give a more accurate explanation of the firm-specific investment location decisions.

JEL classification Nos: F23, P33

Keywords: Mixed logit model, random parameters, foreign direct investment, multi-level data, Halton draws

2

1. Introduction The discrete choice econometric methodology has become an increasingly popular technique for investigating the location decisions of foreign direct investment (FDI), especially in the regional science literature (see for example, Head et al , 1999, Guimaraes et al, 2000 and Kim et al, 2003). This literature, however, is subject to two principal limitations. First, both Multinomial logit (MNL) and Nested logit (NL) models are not sufficiently flexible to satisfactorily model the investment location choices of multinational enterprises (MNEs). The MNL model is subject to restrictive assumptions regarding the substitution patterns across investment location alternatives and the absence of random taste heterogeneity across decision-makers, while the NL model only partially relaxes the independence of irrelevant alternatives (IIA) assumption in order to accommodate the substitution across alternatives to a limited degree. A second limitation is that the existing literature on the determinants of FDI usually uses only country- and industry-specific data and does not incorporate investing firm characteristic, whereas greater estimation efficiency can be achieved by using multi-level data, on the firm, industry and country. This paper therefore generalises the existing literature by making two principal contributions. First, it uses the Mixed logit (ML) model to investigate investment location choices by MNEs for the first time, merely allowing for random taste variation and unrestricted substitution patterns. Second, it makes use of a multi-level data set – allowing firm, industry (or sector) and country effects to simultaneously determine the firm-level FDI location decisions. The data covers 1,108 investment location choices of firms in the EU(15), Norway, Switzerland, Russia, Japan and the USA into 13 Central and Eastern European Counties (CEECs) – the 12 recent EU member states excluding Cyprus and Malta, but including Croatia, Russia and Ukraine - over an eleven year period from 1997 to 2007. The estimation result shows that firms investing in different sectors and firms of different size and profitability benefit from location factors to a different degree. For example, larger firms and firms investing in scale-intensive industries will prefer to invest in larger countries in order to exploit their economies of scale, as compared to smaller firms and firms investing in other sectors.

3 The rest of the paper is set out as follows. Section 2 outlines a new-trade theory model of the firm’s profit function, used to identify the economic variables that determine the FDI decision and Section 3 explains the ML model. Section 4 discusses the dataset and the construction of these variables and Section 5 presents the econometric results. Finally, Section 6 concludes.

2. The Theoretical Model A lot of theoretical literature exist on the extensiveness of firms production, for example, the firms’ decision to serve only domestic market, to serve domestic market and to export or to serve domestic market and to establish a subsidiary abroad (Baldwin, 2005; Helpman, 2006; Helpman et al., 2004; Jean, 2002; Melitz, 2003; Montagna, 2001). However, once the decision to invest abroad has been made, investment location choices have not been theoretically modelled. As a result, a model of imperfect competition has been applied to model where MNEs choose to locate foreign capital. The decision where to invest does not only depend on opportunities offered by foreign markets and industries but also on firms’ individual characteristics. As a result, the model makes three important assumptions: investing firms are heterogeneous across industries in respect to productivity, industries vary in factor intensities and countries differ in relative factor abundance (but identical in terms of preferences and technologies). Each firm produces a different product and the demand function for firm’s product in the foreign market c can be expressed as in (Helpman, 2006): −ε qci = Dc pcis (2) where qci is the quantity of output by firm i in country c and pcis is the price of one unit of output, Dc is a measure of the demand level in country c, and ε is the elasticity of demand defined as ε ≡ 1/(1-λ) . The demand elasticity is assumed to be constant ( ε >1 , as 0<λ<1 ), where λ is a mark-up over marginal costs. Producers are assumed to be small relative to the industry, and therefore, Dc is treated as exogenous by producers.

Firms that locate their capital abroad have productivity θi and variable costs per unit of output of a/θ i. The firm that enters foreign country c also incurs fixed costs af c, where a is the cost of resources and fc is a measure of fixed production costs in terms of

4 resources. Then the profit-maximising price for firm i is pi=a/λθ i. Variable costs vary with firm’s productivity, while fixed foreign market entry costs are the same for all firms, as many fixed costs such as building or equipping a factory with machinery are unlikely to vary substantially with firm productivity ((Bernard et al., 2007)). The presence of fixed production costs implies that, in equilibrium, each firm chooses to produce a unique variety (Bernard et al., 2007). Firms use three factors of production: skilled labour, unskilled labour and capital. It is assumed that skilled labour and capital are allocated to fixed costs (reflecting costs of R&D and costs of factories and equipment), while unskilled labour are allocated to variable costs (capturing standard production). The relative intensity of factor use varies across industries (therefore, α and β will also vary across industries). The minimum total costs can be specified as: 1

u 1−α −β s α β (wcs ) TCics = ()wcs ()rc fc + qci (3) θi s u where wcs is an hourly wage rate of skilled labour in country c and industry s; wcs is an hourly wage rate of unskilled labour in country c and industry s, rc is a return on capital in country c and the α and β parameters are the shares of skilled labour and capital in total cost respectively. The first term on the right hand side of (3) represents the fixed costs of acquiring information about foreign markets, developing appropriate marketing strategies, building distribution networks (Bernard et al., 2007), also acquiring factories and equipment. MNEs enterprises use foreign capital in the form of foreign direct investment and host country skilled and unskilled labour, Ls and Lu respectively. The labour market equilibrium requires equating labour supply, which is s u exogenous ( Lc ), with the total demand for skilled (L c ) and unskilled (L c ) labouring country c:

s u Lc = Lc + Lc (4) The price can be written as a constant mark-up over marginal cost:

1 This is a standard formulation following Krugman (1991), where the total minimum cost is derived from a standard cost minimization problem assuming a Cobb-Douglas production function.

5 u 1−α −β τ cd (wcs ) pi ()θ = (5) λθ i

“Iceberg” type transport costs τcd , are assumed between the source country d and host country c (Samuelson, 1954) 2. The after-tax profit in each location is defined as total revenue ( TR ) less total costs ( TC ) net of tax and less the costs arising from the institutional, legal, political and macroeconomic environment prevalent in the host country. It can be written as:

π cis = 1( − Tc )(TRcis − TCcis )− Gc (6) where Tc is the tax rate in location c and Gc is a term that captures the costs that firms incur due to the macroeconomic investment environment prevalent in the host countries c. TRcis is the revenue received by firm i in country c and industry s from selling output qc , and TCcis are the costs of producing qc . Substituting (4) and (5) into (6) we obtain:

1−ε   u 1−α −β   τ ()w s α β π = ()()1−T  1− λ D  cd cs  − ()w ()r f  − G cis c  c  λθ  cs c  c (7)   i  

This specification of the profit function of investment abroad, in contrast to the majority of the current theoretical literature on FDI, allows for the heterogeneity of firms in different sectors and countries. Country specific equation (7) in respect to T, D, τ, w, r and G will also vary across firms in respect to θ and across industries in respect to α and β. Therefore, it is assumed that particular location advantages do not have the same value for all multinational enterprises, as firms operating in different sectors and firms of different size and profitability benefit from local resources to different degrees. Location advantages vary for MNEs with different characteristics, and the interaction between location and firm together with industry attributes, rather than each of the firm and industry factors independently affects location choices. For example, smaller firms are

2 For example, when goods are shipped from country d to country c , only a fraction 1/τ dc of the original unit is assumed to arrive ( τ cd > 1). Hence, other things being equal, the more remote locations are at a disadvantage.

6 expected to invest in countries with strong historical ties and similar culture and language, while larger firms and firms investing in scale-intensive industries are expected to invest in more remote and larger countries in order to exploit their economies of scale.

3. Econometric Estimation The ML model has been applied in environmental, health and transport studies (Greene and Hensher, 2003; Hole, 2007; Meijer and Rouwendal, 2006), however, its application in economics is still limited. Furthermore, the ML model has never been used to investigate the investment location choices of MNEs. The ML model, which is probably the most flexible discrete choice model, has become applicable only recently with the development of simulation and increased computer speed. The ML 3 model is a very flexible model that approximates any random utility model (McFadden and Train, 2000). In contrast to the MNL model, the ML model allows for random taste variation, unrestricted substitution patterns and correlation in unobserved factors over time. Furthermore, the ML probability can still be derived from utility-maximising behaviour. In contrast to the , it is not restricted to normal distribution. The derivation of ML is straightforward, and simulation of its choice probabilities is computationally simple (Train, 2003). In the most general form, the ML probabilities are the integrals of standard logit probabilities over a density of unobserved random parameters (Train, 2003):

Pic = ∫ Lic (βi ) f (βi )dβi (1) where f(β i) is the random parameter density function, which is specified to be continuous. The mixed logit probability is a weighted average of the logit formula evaluated at different values of βi, with the weights given by the density f(β i) (Train, 2003). The logit probability evaluated at parameters βi is expressed as

' eβ i xic Lic ()βi = ' eβ i xig (2) ∑d

3 In the literature also referred to “random parameter logit”, “mixed multinomial logit”, “kernel logit” and “hybrid logit”.

7 Where xic is a vector of observed choice-specific variables and βi are individual firm- specific parameters. Since the integral cannot be calculated analytically, it has to be approximated through simulation by maximising the simulated log-likelihood function. A value of β is drawn form the distribution f(β│ψ) for a given value of ψ and labelled β r with the r subscript r=1 referring to the first draw. The logit formula L ci (β ) is calculated with this draw. Finally, the two steps are repeated many times, and the results are averaged, which gives the simulated probability:

R ( 1 r Pci = ∑ Lci ()β (3) R r=1 ( where R is a number of draws. Pci is an unbiased estimator of P ci . Its variance decreases ( as R increases. It is strictly positive (ln Pci is defined) and smooth (twice differentiable) in the parameters ψ and the variables x, which facilitates the numerical search for the ( maximum likelihood function and the calculation of elasticities. Furthermore, Pci sums to one over alternatives which is helpful when interpreting the results. The simulated probabilities are substituted into the log-likelihood function to give the simulated log likelihood: C I ( SLL = ∑∑ d ci ln Pci (4) c=1i = 1 where d ci =1 if i chose c and zero otherwise. The maximum simulated likelihood estimator (MSLE) is the value of ψ that maximizes SLL. The ML structure can be derived in two equivalent ways: by allowing flexible substitution patterns across alternatives ( error-components structure ) (Brownstone, 2000) and by accommodating unobserved heterogeneity across individuals ( random-coefficients structure ) (Bhat, 1998, 2002, Revelt and Train,1998, Train, 1998).

4. The Data Set and the Variable Specification Table 1 gives a summary of variable definitions and sources. There are 1,108 firm-level data observations on FDI flows from firms of 20 market economies (EU15 countries,

8 USA, Japan, Russia, Norway and Switzerland) to the firms into 13 transition economies (12 new EU member states (except for Malta and Cyprus) plus Croatia, Russia and Ukraine) from 1997 to 2007. Of all 13 host CEECs in the sample, Poland has been chosen the majority of times by MNEs to locate their investment (about 21 percent) and it was followed by Russia with about 17 percent of foreign investment location choices. Slovenia and Latvia, on the other hand, have received the smallest share foreign capital allocations (2 and 3 percent respectively). The two major source countries for investment in CEE in the sample are Finland and the UK with the shares of approximately 12 and 11 percent respectively. MNEs from Japan and Ireland were at the other end of the scale regarding investment location choices in CEE of about 1 percent each. Almost a third of the firms in the sample are small (with turnover up to €100,000), 30 percent of the firms are medium (with turnover from €100,000 to €1mln), 26 percent of the firms are large (turnover from € 1mln to € 10mln) and 12 percent of the firms are very large firms (turnover above € 10mln). The majority- almost half- of very large firms in the sample have invested in scale-intensive industries. The rest of the firms have selected traditional sectors to locate most of their investment (33 percent of small firms, 41 percent of medium firms and 43 percent of large firms). 14 percent of the firms in the sample have incurred losses, 44 percent of the firms have earned profits up to € 50,000 and 42 percent of the firms have earner profits above € 50,000. Regardless of the investing firms’ profitability traditional sectors received most investment allocations. The investment location characteristics may have a different effect on firms investing in different sectors. Scale-intensive sectors (Scale ) include typical oligopolistic, large firm industries, with high capital intensity, extensive economies of scale and learning, high technical and managerial complexity, for example, automobiles, aircrafts, chemicals, petrol and coal products, shipbuilding, industrial chemicals, drugs and medicines, petrol refineries, non-ferrous metals and railroad equipment (Midelfart- Knarvik et al ., 2000). Science-based sectors (Science ), on the other hand, are characterised by innovative activities directly linked to high R&D expenditures, for example, fine chemicals, electronic components, telecommunications, and aerospace (Midelfart-Knarvik et al ., 2000). Traditional (supplier-dominated) sectors (Tradit ) include such industries as textiles, clothing, furniture, leather and shoes, ceramics, and the

9 simplest metal products. Finally, banking insurance and retail are examples of service sectors (Service ). Firms locating investment in traditional sectors may be more concerned about the availability of unskilled labour and may pay lower wages as compared to the firms in other sectors, for example, science-based industries, where more skilled labour is employed and higher wages are paid for higher skills. Four groups of industries have been chosen to locate foreign investment more or less evenly in the sample. The largest number of foreign capital allocations in CEE took place in traditional sectors (36 percent), followed by scale-intensive industries (24 percent) and service sectors (23 percent). Science-based industries have received the smallest share of FDI (18 percent) in the sample. However, when looking at the distribution of investment location choices among the four groups of industries in separate countries, traditional sectors have not necessarily attracted most foreign capital allocations. For example, in the Czech Republic, Hungary and Slovakia scale-intensive industries have received the largest share of FDI (30 percent, 26 percent and 47 percent respectively), while in Estonia and Lithuania the service sector has attracted most foreign capital allocations (35 percent and 33 percent respectively). The country-specific determinants of FDI into the CEECs can be loosely divided into the traditional determinants and the transition-specific determinants. The transition- specific determinants are proxied by the risk associated with each host country, Gc, in equations (5) and (7). The institutional, legal and political environment, i.e. transparency and effectiveness of legal system, are important for the decision of foreign investors to locate their capital abroad. The Transparency International Corruption Perception Index (TICP) is used as a measure of the extent of corrupt practices in the host country. This index pools information from ten different surveys of business executives, risk analysis and the general public. The TICP index ranks countries in terms of the degree to which corruption is perceived to exist among public officials and politicians and it varies from 1 (high corruption) to 10 (no corruption). In order to make the interpretation of the parameter more intuitive the TICP index is multiplied through by minus one, so that the smaller the number the higher risk. Following the theoretical model described in section 2, the traditional determinants are the market size of the host country, Dc, the cost of capital in the host

10 country, rc , distance τcd , and tax rates Tc in the host country. As Table 1 shows, market size is simply the real GDP of the country and the rate of return is measured as the real discount (interest) rate. Distance can be considered as a measure of the transaction costs of undertaking foreign activities, such as the costs of transport and communications, the costs of dealing with cultural and language differences, the costs of sending personnel abroad, and the informational costs of institutional and legal factors, e.g., local property rights, regulations and tax systems. These kinds of costs are assumed to increase with distance. The corporate income tax rate affects the profitability of foreign direct investment and hence influences the investment location choices of MNEs. Few studies analyse the effect of taxation on the location choices of foreign firms in the CEECs (Bellak and Leibrecht, 2005; Carstensen and Toubal, 2004; Clausing and Dorobantu, 2005; Wei, 2000). The studies that do include tax rates as location choice factors in CEECs usually use statutory corporate income tax rates. Statutory corporate income tax rates are not an appropriate indicator of the tax burden especially in the case of FDI, because they are only one of the determinants of total tax burden, while the tax base is also influenced by depreciation schemes, treatment of losses and valuation of inventories among others. In this paper, in contrast to the majority of other studies, the tax burden in the 13 CEECs is measured as the effective corporate income tax rate which is calculated by dividing revenue taxes paid by corporations and other enterprises by a host country’s GDP. This approach allows comparisons of different tax systems, taking into account such important aspect as untaxed reserves, tax enforcement and the treatment of losses. In addition to the above mentioned factors, three other country-specific factors are included in the empirical model: the national rate of unemployment and two dummy variables, one for European Union (EUD) and another for a common border (CBD) between the investing and the investment receiving country. A dummy variable for common border between the source and the host country is included, as it is expected that the host country is more likely to be chosen to locate investment if it shares the border with the source country. Usually countries sharing the same border have similar culture and language and stronger historical ties.

11 Countries that joined the EU by January 2007 had to satisfy the economic (market economy), political (democracy and human rights) and administrative (well-functioning institutions) criteria set at the Copenhagen European Council in 1993. The accession of a CEE country into EU meant free trade with EU member states and the adoption of Western business and legal environment, which provided foreign investors with confidence in success of each country’s reforms. As a result, the parameter of EU dummy variable is expected to have a positive sign. Although, unemployment is not important for the individual firm’s profit function, it may still be of significance at the country level as an indicator of labour market flexibility and availability of labour force. Countries with high local demand for goods and services and high labour market flexibility are likely to face relatively low rates of unemployment, which may encourage firms to invest in a particular host country. On the other hand, a high unemployment rate may mean that although it is easy to recruit labour, there is low demand locally and labour market rigidities. The impact of unemployment on the investment location decision is therefore strictly ambiguous and it may have a different effect on firms investing in different industries. For example, firms investing in traditional sectors employ less skilled labour and may be more concerned about the availability of workers, while firms investing in science-based industries, which employ more skilled labour, may be discouraged by higher unemployment, as unemployed people loose their skills through time.

Industry-level real wage rates, wcs , are included as a proxy for the average variable costs of firms and they implicitly assume that workers are not fully mobile across sectors, at least in the short run. The profitability of the firm investing abroad is expected to be higher if the labour costs are lower in the chosen country than in the rest of the destination countries. On the other hand, higher wages may reflect higher skill and, therefore, may have a positive effect for firms investing in science-based industries, which employ more skilled labour as compared to other industries. As a result, the sign of the parameter is amgiuous. The characteristics of individual investing firms can also have an influence on the responsiveness of country-level variables. The firm-level variables include the turnover of the investing firm as a proxy for its size ( si ) and earnings before interest and tax as a

12 proxy for its profitability ( ei ). Firms of different sizes and profitability possess different resources and capabilities (Dean et al ., 1998). Small firms are assumed to be characterised by speed, flexibility and niche-filling capabilities due to their structural simplicity and faster decision making, entrepreneurial-orientation and less risk aversion (Woo, 1987). As a result, smaller firms respond quicker to the dynamics of the industry environment. Larger firms, which are usually more profitable, are able to acquire larger market share by exploiting scale economies, bargaining power, patents, reputation and financial resources to deal with shocks and business downturns (Dean et al ., 1998). Larger firms investing in scale-intensive industries are expected to invest in countries with larger markets in order to exploit their economies of scale, while more profitable firms are expected to be less discouraged to invest in remote countries, as more funds are available to cover transaction costs, such as costs of transport and communication, the costs of dealing with cultural and linguistic differences and information costs of institutional and legal factors, etc.

5. Estimation and Results The analysis starts by treating each country level variable and the variable that varies among countries and industries ( wsc ) separately as random by imposing various distributions (Table 2). The random parameters with most appropriate distributions are combined in the final specification, which has the best model fit (the largest log- likelihood value) and which avoids distributions that cause flat log-likelihood at the estimates. Triangular distribution is imposed on the variables for market size in the host country, Dc, the wage variable, wsc , the distance variable, τcd , and the unemployment variable, uc. Restricted uniform distribution is imposed on the two dummy variables: the dummy variable for common border and the dummy variable for EU membership (Table 3). The means of the uniform distribution for the dummy variables are restricted to be equal to their variances; as a result, the sign of the estimate random parameters of the two dummy variables will be the same for all the investing firms. Initially 100 Halton intelligent draws are used to estimate the model, while increasing the number to 1000 for the final model specification.

13 The ML is not only able to determine the existence of heterogeneity around the mean parameter, through the estimation of the standard deviation parameter, but it can also indicate the source of the heterogeneity through the interaction between a random parameters and other attributes (moderator variables). For example, an observed heterogeneity in some country level determinants of investment location choice can be due to differences in industry or/and investing firms’ characteristics. The following interaction terms are statistically significant: the interaction terms between the dummy variable for traditional sectors and unemployment rate in the host country, Tradit×u c; the interaction term between the dummy variable for traditional sectors and the wage rate in the host country, Tradit×w sc ; the interaction term between investing firms profitability and distance between investing and investment receiving countries, ei× τcd, and finally, the interaction term between investing firm’s size and host country’s market size, si×GDP c. Despite the widespread use of interaction terms in Discrete Choice Methodology, the majority of applied researchers misinterpret coefficients of interaction terms (Ai and Norton, 2003). Unlike in linear models, the interaction effect in nonlinear models is a function of not only the coefficient for the interaction, but also the coefficients for each interacted variable and the values of all the variables in the model (Greene, 2008). Therefore, the sign of the interaction coefficient may not indicate the direction of the interaction effect, as the interaction effect may have different signs for different values of covariate. Furthermore, the interpretation of a separately included in the model variable if it is also a part of an interaction term changes (Jaccard, 2001). It does not represent a “main effect” but a conditional effect instead: the effect of the variable when the values of the moderator variable (the other interacted variable) are zero. For example, the variable, wsc is not only included in the model separately but also interacted with the dummy variable,

Tradit . As a result, while the interaction, wsc ×Tradit , represents the effect of wage rate in traditional sectors in a particular host country on the probability of selecting the country to locate foreign capital, the variable, wsc , represents the effect of wage costs in other sectors (Science-based, Service and Scale-intensive sectors ).

14 As neither the sign nor the magnitude of the interactions and separately included in the model variables if they are also included in interaction terms are informative, elasticities and marginal effects have to be estimated for continuous and dummy variables 4 respectively (Table 4) . Negative estimated elasticities for variable, Wage sc , indicate that the higher wage costs are in non-traditional sectors in a host country, the less likely the country will be chosen to locate foreign capital. The effect of the wage rate on the probability of locating investment in a particular country if a firm chooses to invest in traditional sectors is a sum of the estimated elasticities for Wage sc and wsc ×Tradit. The sums of the estimated elasticities are negative and larger in absolute values showing that firms that choose to invest in traditional sectors are more sensitive to higher wages rates in the host country than country, which choose to invest in non-traditional sectors..

The unemployment variable, uc, is not only included in the model separately but also interacted with the dummy variables for traditional sectors, Tradit×u c. Negative elasticities for, uc, indicate that the higher the unemployment rate is in the host country, the less likely the country to be chosen by foreign firms to locate their capital, if they choose to invest in non-traditional sectors. Negative but smaller in absolute values or even positive sums of the elasticities for Tradit×u c indicate that higher unemployment in a host country has a less negative or even positive effect on the probability of selecting the country to locate foreign capital for firms that invest in traditional sectors, as compared to the firms that invest in non-traditional sectors. Typically, traditional sectors employ more unskilled labour, as compared to other sectors, for example, science-based industries, which employ more skilled labour and pay higher wages that reflect a skill premium.

When two continuous variables are interacted, for example, ei× τcd and si×GDP c, the interpretation of the interaction terms is much more complicated. As a result, the changes in market shares in different countries and the changes in a number of firms investing in a particular country are investigated due to the gradual change in the

4 The elasticities for separately included in the model market size and distance variables do not have much explanatory value, as they indicate the effect when moderator variables (investing firm’s size and profitability respectively) are equal to zero. The tax variable and the risk variable do not appear statsitically significant. The standar deviation of the retrun on capital variable is not statistically significant, indicating that all information is captured by the mean. The higher is the return on capital in the ost country, the more likely the country to be chosen by foreign investors.

15 moderator variable with the help of simulation. The results presented in Table 5 show the effect of 1, 10, 50 and 100 percent increase in variables ei and si on the change in the market shares and in the number of firms investing in a particular country. The estimated changes in market shares and in the number of investing firms for both interaction terms show consistently positive and increasing effects. The results for ei× τcd indicate that more profitable investing firm are less likely to be deprived to invest in more remote countries as compared to less profitable firms. More profitable firms usually have more resources to pay for transaction costs associated with investment in more remote countries, for example, cost of transport of communication, the costs of dealing with cultural and linguistic differences, the cost of sending personnel abroad, and information costs of institutional and legal factors. The results for si×GDP indicate that the larger the host country is, the more likely it is to be chosen by an investing firm to locate its capital, and the effects is stronger for larger investing firm, as compared to smaller firms. Larger firms are usually characterized by high economies of scale; therefore, they search for larger foreign markets to exploit these economies. Statistically significant interaction terms between country-level variables and industry-level dummies together with the firm- level variables confirm the existence of heterogeneity revealed by statistically significant standard deviations of the parameters of certain country-level variables. Positive and statistically significant estimated elasticities for EU dummy variable indicate that countries which became members of the EU by January 2007 are more likely to be chosen by foreign investors to locate their capital. An investment receiving country that has a common border with an investing country is more likely to be chosen as an investment location, and this is reflected in the positive elasticities for CBD cd (Table 4). Neighbouring countries usually have strong historical ties and less linguistic and cultural barriers.

6. Conclusions This paper applies the Mixed logit (ML) model, which is probably the most flexible discrete choice model, to investigate investment location choices by MNEs for the first time. It also makes use of a novel multi-level data set – allowing firm, industry (or

16 sector) and country effects to simultaneously determine the firm-level FDI location decisions. The highly significant empirical results support the presence of heterogeneity in the investment location decisions, which is not only revealed by statistically significant interaction terms, but also by statistically significant standard deviations of the random parameters. The results show that firms investing in traditional sectors are less likely to be discouraged to invest in countries with higher unemployment rate but more likely to be discouraged by higher wage rates as compared to MNEs that invest in non-traditional sectors. The larger the host country is the more likely it is to be chosen by foreign investors to locate their capital and the effect is larger for larger investing firms. On the other hand, more profitable firms are less likely to be discouraged to invest in more remote countries, as compared to less profitable firms. This more general approach to the FDI decision shows that to allow for firm heterogeneity is important if robust estimates are to be found for their complex effects. These results cast doubt on the robustness of earlier empirical studies that focused on macroeconomic features of the FDI location decision not incorporating investing firms’ caracteristics, and applied Multinomial logit and/or Nested logit models.

17 References

Ai, C., and Norton, E. C. (2003). Interaction terms in logit and probit models. Economic Letters 80 , 123-129. Baldwin, R. E. (2005). Heterogeneous Firms and Trade: Testable and Untestable Properties of the Melitz model. nbER Working Paper Series, Working Paper 10471 . Bellak, C., and Leibrecht, M. (2005). Do low corporate income tax rates attract FDI? Evidence from Eight Central- and East European Countries. In "Research Paper Series, Globalisation, Productivity and Technology. Research Paper 2005/43", The University of Nottingham. Leverhulme Centre for Research on Globalisation and Economic Policy Bernard, A. B., Redding, S. J., and Schott, P. K. (2007). Comparative advantage and heterogeneous firms. Review of Economic Studies 74 , 31-66. Brownstone, D. (2000). Discrete Choice Modeling for Transportation. In "9th IATBR Travel Behavoir Conference", Australia. Carstensen, K., and Toubal, F. (2004). Foreign direct investment in Central and Eastern European countries: a dynamic panel analysis. Journal of Comparative Economics 32 , 3-22. Clausing, K. A., and Dorobantu, C. L. (2005). Re-entering Europe: Does European Union candidacy boost foreign direct investment? Economics of Transition 13 , 77-103. Dean, T. J., Brown, R. L., and Bamford, C. E. (1998). Differences in large and small firm responses to environmental context: Strategic implications from a comparative analysis of business formations. Strategic Management Journal 19 , 709-728. Greene, W. H. (2008). "Econometric Analysis," 6th/Ed. Pearson-Prentice Hall. Greene, W. H., and Hensher, D. A. (2003). A latent class model for discrete choice analysis: contrasts with mixed logit. Transportation Research Part B- Methodological 37 , 681-698. Helpman, E. (2006). Trade, FDI, and the organization of firms. Journal of Economic Literature 44 , 589-630. Helpman, E., Melitz, M. J., and Yeaple, S. R. (2004). Export versus FDI with heterogeneous firms. American Economic Review 94 , 300-316. Hole, A. R. (2007). Modelling heterogeneity in patients' preferences for the attributes of a general practitioner appointment. National Primary Care Research and Development Centre, Centre for Health Economics, University of York. Jaccard, J. (2001). "Interaction effects in ," Sage Publications, London, New Delhi. Jean, S. (2002). International trade and firms' heterogeneity under monopolistic competition. Open Economies Review 13 , 291-311. McFadden, D., and Train, K. (2000). Mixed MNL models for discrete response. Journal of Applied Econometrics 15 , 447-470. Meijer, E., and Rouwendal, J. (2006). Measuring welfare effects in models with random coefficients. Journal of Applied Econometrics 21 , 227-244. Melitz, M. J. (2003). The impact of trade on intra-industry reallocations and aggregate industry productivity. Econometrica 71 , 1695-1725.

18 Midelfart-Knarvik, K. H., Overman, H. G., Redding, S. J., and Venables, A. J. (2000). The Location of European Industry. European Economy - Economic Papers 142, Commission of the EC, Directorate-General for Economic and Financial Affairs (DG ECFIN) . Montagna, C. (2001). Efficiency gaps, love of variety and international trade. Economica 68 , 27-44. Samuelson, P. (1954). The Transfer Problem and Transport Costs, Economic Journal. Economic Journal 64 , 264-89. Train, K. (2003). "Discrete Choice Methods with Simulation," Cambridge University Press, Cambridge. Wei, S. J. (2000). How taxing is corruption on international investors? Review of Economics and Statistics 82 , 1-11. Woo, C. Y. (1987). Path-Analysis of the Relationship between Market Share, Business- Level Conduct and Risk. Strategic Management Journal 8, 149-168.

19 Table 1: List of variables, definitions and sources Variable Definition Source Choice c a CEEC, in which firm n chooses to locate its Bureau van Dijk Zephyr investment over the period of time from 1997 to database 2007 (it gets the value of 1 if the country received investment and 0 otherwise) τcd distance between the capital cities of the source http://www.indo.com/distance/ country d and the host country c in kilometres Dc Real GDP of the host country c of the year IFS investment took place Gc Corruption perception index of the host country c Transparency International of the year investment took place uc unemployment rate of country c (percentage per IFS annum) of the year investment took place Tc effective corporate income tax rate in country c of Calculated using data from the year investment took place IFS rc the real discount (interest) rate IFS CBD cd a dummy variable that takes a value 1 if both constructed source country d and host country c share a border, and 0 otherwise EU c dummy variable that takes value 1 if country c constructed joined EU before January 2007, and 0 otherwise Scale s dummy variable that takes a value 1 if industry s constructed is a scale-scale industry, and 0 otherwise Science s dummy variable that takes a value 1 if industry s constructed is a science-based industry, and 0 otherwise Tradit s dummy variable that takes a value 1 if industry s constructed is a traditional industry, and 0 otherwise Service s dummy variable that takes a value 1 if industry s constructed is a service sector, and 0 otherwise Wage c hourly real wage rates in the industry s in the International Labour country c of the year investment took place Organisation Size n turnover of the investing firm i in Euros of the Bureau van Dijk Zephyr year investment took place database Earnings n earnings before interest and taxes of the investing Bureau van Dijk Zephyr firm i in Euros of the year investment took place database

20 Table 2 The Imposition of Various Distributions for Country-level Random Variables Distance Risk Coef z-stats Coef z-stats Log-likelihood Coef z-stats Coef z-stats Log-likelihood Normal Log-likelihood is flat 0.3906 {5.354} 0.5161 {6.012} -2507.072 Log-normal Log-likelihood is flat Log-likelihood is flat Triangular -1.1821 {-11.007} 3.0309 {7.569} -2501.440 0.3964 {5.394} 1.2748 {6.218} -2506.846 Restricted Triangular -1.0522 {-11.635} 1.0522 {11.635} -2510.590 0.2894 {4.309} 0.2894 {4.309} -2513.508 Dome -1.1291 {-10.192} 3.1230 {7.490} -2502.545 0.4579 {5.571} 1.6673 {6.261} -2504.697 Erlang Log-likelihood is flat 0.4193 {5.097} 0.5565 {4.884} 2507.592 Weibull Log-likelihood is flat 0.9415 {6.076} 0.3256 {5.190} -2509.200 Exponential Log-likelihood is flat 0.3280 {4.888} 0.3845 {4.454} -2510.319 Unemployment Interest Coef z-stats Coef z-stats Log-likelihood Coef z-stats Coef z-stats Log-likelihood Normal -10.3786 {-5.634} 18.3814 {8.110} -2498.291 0.1900 {1.122} 2.7461 {0.727} -2514.210 Log-normal 1.4548 {1.2147} 1.2147 {4.719} -2508.554 -0.3434 {-0.077} 0.9028 {0.178} -2514.339 Triangular -10.5425 {-5.674} 44.5571 {8.379} -2498.102 1.2036 {1.151} 6.7393 {0.955} -2514.191 Restricted Triangular -5.7529 {-4.164} 5.7529 {4.164} -2513.707 1.0176 {1.029} 1.0176 {1.029} -2514.376 Dome -11.1859 {-5.827} 47.8832 {7.987} -2499.240 1.4166 {1.363} 10.1212 {1.114} -2514.019 Erlang -9.9504 {-5.427} 15.9837 {6.419} -2502.942 1.1047 {1.124} 1.6004 {0.455} -2514.222 Weibull 14.3140 {5.474} 13.5585 {7.337} -2501.263 Log-likelihood is flat Exponential -8.7071 {-4.840} 21.9603 {7.439} -2496.474 1.0517 {1.040} 1.3075 {0.405} -2514.332 Tax Wage Coef z-stats Coef z-stats Log-likelihood Coef z-stats Coef z-stats Log-likelihood Normal -0.0766 {-1.503} 35.1249 {3.971} -2511.830 -0.105 {-1.990} 0.0392 {1.451} -2514.294 Log-normal 0.0975 {0.051} 1.7071 {1.552} -2513.612 -2.9966 {-3.077} 1.9047 {2.696} -2510.452 Triangular -0.8978 {-0.347} 87.1765 {4.013} -2511.847 -0.1071 {-1.990} 0.1191 {1.550} -2514.260 Restricted Triangular 2.3872 {1.181} 2.3872 {1.181} -2514.386 Log-likelihood is flat Dome -0.7759 {-0.282} 99.8594 {4.096} -2511.453 -0.1466 {-2.521} 0.3523 {1.187} -2513.993 Erlang 0.2126 {0.077} 28.6690 {2.416} -2512.262 -0.2053 {-3.152} 0.3396 {3.478} -2511.963 Weibull Log-likelihood is flat 0.0275 {0.228} 0.0894 {0.825} -2513.937

21 Exponential 0.3441 {0.140} 31.1936 {2.704} -2512.805 0.1725 {-2.123} 0.1725 {1.689} -2513.408 GDP Coef z-stats Coef z-stats Log-likelihood Normal 0.4970 {7.635} 0.9543 {6.857} -2485.322 Log-normal Log-likelihood is flat Triangular 0.4956 {7.497} 2.2860 {6.988} -2484.991 Restricted Triangular 0.5579 {11.120} 0.5579 {11.120} -2504.503 Dome 0.4803 {5.752} 2.1914 {5.160} -2490.409 Erlang 0.3918 {4.174} 0.9384 {4.661} -2494.652 Weibull 1.7012 {8.995} 0.6940 {0.1214} -2494.405 Exponential 0.4301 {5.791} 0.8612 {4.652} -2500.163

22

Table 3 The Results of the Mixed Logit Model Estimation

Variables Distribution Mean Stand. Dev. Wage Triangular -0.285 {-4.493} 0.7252 {3.021} GDP Triangular 0.575 {10.760} 1.4868 {5.511} Distance Triangular -1.252 {-10.625} 1.9154 {3.223} Unempl Triangular -3.5644 {-2.684} 20.7788 {2.653} Restricted Border Uniform 0.5651 {4.205} 0.5651 {4.205} Restricted EU Uniform 0.8028 {5.204} 0.8028 {5.204} Interest - 4.5863 {3.945} - - Prof_Dist - 0.1471 {2.922} - - Size_GDP - 0.1334 {4.090} - - Tr_Unemp - 8.4893 {4.308} - - Tr_Wage - -1.4718 {-5.065} - -

Table 4 Elasticities and Marginal Effects Country Wage Tr_Wa Sum Unem Tr_Un Sum GDP Dist Border EU

BG -0.2202 -0.1107 -0.3309 -0.2445 0.3796 0.1351 0.0519 -1.6297 2.7083 7.1613 CZ -0.5055 -0.2631 -0.7686 -0.3289 0.1867 -0.1422 0.3748 -1.7266 1.5842 7.8312 EE -0.3121 -0.1521 -0.4642 -0.1797 0.0607 -0.119 0.0222 -1.7427 0.3026 7.1535 CR -0.4737 -0.3035 -0.7772 -0.193 0.4491 0.2561 0.0798 -2.0011 0 0 HU -0.4693 -0.2257 -0.695 -0.329 0.1733 -0.1557 0.2714 -1.7798 1.2098 6.2476 LT -0.4095 -0.2069 -0.6164 -0.3091 0.198 -0.1111 0.0528 -2.0335 0 4.5774 LV -0.3861 -0.2 -0.5861 -0.3177 0.211 -0.1067 0.0262 -2.0144 0.2689 5.1733 PL -0.3865 -0.2599 -0.6464 -0.076 0.3714 0.2954 1.0206 -1.3697 1.1459 10.8886 RO -0.1836 -0.0732 -0.2568 -0.3222 0.2252 -0.097 0.0956 -1.8359 0 5.0919 RU -0.0886 -0.039 -0.1276 -0.1628 0.1214 -0.0414 1.3678 -1.9985 1.9985 0 SI -0.3268 -0.2823 -0.6091 -0.321 0.1284 -0.1926 0.0686 -1.7367 0.8931 3.96 SK -0.4688 -0.293 -0.7618 -0.1991 0.3965 0.1974 0.093 -1.7468 1.0103 5.8174 UA -0.1212 -0.0617 -0.1829 -0.3373 0.3159 -0.0214 0.1277 -1.9746 0.4912 0

23

Table 5 Simulation Results of Changes in Market shares and the Number of Firms

Size_GDP 1% 10% 50% 100% BG 0.001 0 0.005 0 0.022 0 0.045 1 CZ 0.004 0 0.038 0 0.202 2 0.435 4 EE 0 0 0.002 0 0.011 0 0.023 1 CR 0 0 0.003 0 0.017 0 0.035 1 HU 0.002 0 0.021 0 0.111 1 0.236 3 LT 0.001 0 0.004 0 0.021 1 0.042 0 LV 0.001 0 0.003 0 0.011 0 0.023 1 PL 0.018 0 0.185 2 1.007 11 2.148 24 RO 0.001 0 0.007 0 0.039 1 0.081 1 RU 0.025 0 0.254 2 1.23 3 2.373 26 SI 0 0 0.003 0 0.018 0 0.037 0 SK 0.001 0 0.008 0 0.039 0 0.079 1 UA 0.001 0 0.007 0 0.034 1 0.07 1

Prof_dist 1% 10% 50% 100% BG 0.008 0 0.089 1 0.973 11 1.788 20 CZ 0.007 0 0.09 1 0.899 10 1.64 18 EE 0.007 0 0.071 1 0.732 8 1.723 19 CR 0.003 0 0.034 1 0.538 6 1.21 14 HU 0.007 0 0.085 1 0.914 10 1.667 18 LT 0.006 0 0.063 0 0.747 8 1.639 18 LV 0.006 0 0.062 1 0.738 8 1.633 18 PL 0.022 0 0.261 3 1.33 15 2.164 24 RO 0.008 0 0.102 1 1.051 12 1.905 21 RU 0.032 0 0.3 3 1.267 14 2.103 23 SI 0.003 0 0.037 0 0.508 5 1.134 12 SK 0.005 0 0.059 0 0.727 8 1.476 16 UA 0.005 0 0.06 1 0.833 10 1.626 18

24