Tax Competition in EU - Bénassy-Quéré et al. 2007 PDF

o n Tax Economic ECOP © 0266-4658 Original XXX TAX A. CEPR, BÉNASSY-QUÉRÉ, competition COMPETITION Blackwell Oxford, Article CES, UKPolicyMSH,Ltd Publishing 2007....

o n Tax Economic ECOP © 0266-4658 Original XXX TAX A. CEPR, BÉNASSY-QUÉRÉ, competition COMPETITION Blackwell Oxford, Article CES, UKPolicyMSH,Ltd Publishing 2007. N. GOBALRAJA and A. TRANNOY e t iti m p c o Ta x SUMMARY The debate on tax competition lacks due attention when it comes to the provision of public goods used by firms in their production process. Indeed, firms may accept higher corporate taxation provided they enjoy good infrastructure and public services. We quantify such trade-off, i.e. the extent to which a ‘high tax, high public goods’ strategy is attractive to capital as compared to a ‘low tax, low public goods’ com- bination. We revisit and develop the popular model of tax competition introduced by Zodrow and Mieszkowski (1986) in a way that allows for the testing of its main prediction. The under-provision of public inputs can be tested econometrically by estimating and comparing two simple elasticities: capital with respect to the tax rate, and capital with respect to public inputs. We regress US foreign direct invest- ment in 18 EU countries over 1994–2003 on several variables, including the corporate tax rate and the stock of public capital, used as a proxy for public input. Based on these estimations (−1.1 for the tax elasticity and +0.2 for the public input elasticity), we conclude that raising public input through an increase in the corporate tax rate reduces inward FDI, and that tax competition may indeed lead to an under-provision of public inputs. Furthermore, a ‘high’ equilibrium (high taxation and high level of public input) is not attainable for a country starting from a ‘low’ equilibrium unless households have a strong preference for public inputs. On the whole, the impact of tax competition may be more diverse than a mere ‘race to the bottom’. — Agnès Bénassy-Quéré, Nicolas Gobalraja and Alain Trannoy Economic Policy April 2007 Printed in Great Britain © CEPR, CES, MSH, 2007. TAX COMPETITION 387 Tax and public input competition Agnès Bénassy-Quéré, Nicolas Gobalraja and Alain Trannoy CEPII, Paris, and Université Paris X; Université Paris X; Ecole des Hautes Etudes en Sciences Sociales, IDEP-GREQAM 1. INTRODUCTION The debate on tax competition has been gaining momentum in the EU since the beginning of the new century, particularly since the EU enlargement of 2004. Indeed, within the last ten years (1996– 2006), statutory corporate tax rates declined by an average of 9 percentage points in the EU15 and an average of 11 percentage points for the new member states (NMSs) (Figure 1). Due to the broadening of the tax base, the decrease in the effective tax rates was less significant in the EU151 than in the new member states, where the decline was nearly one percentage point a year on average.2 These figures support the existence of a ‘race to the bottom’ which, when defined in this context, describes any series of competitive and non-cooperative tax cuts made by national governments with the ambition of attracting more foreign capital. We are grateful to Michael Overesch for providing a complete EATR data set, to Brigitte Dormont for a useful suggestion, to Donna Henry-Norker, Michel Le Breton, Philippe Martin, Thierry Mayer, Cecilia Garcia Penelosa, Pierre Pestieau, Stephen Redding, Fred Rychen, Laurent Simula and Tanguy Van Ypersele for useful comments, and to the participants in the Economic Policy panel in Vienna on 21–22 April 2006 for additional comments. The usual disclaimer applies. The Managing Editor in charge of this paper was Philippe Martin. 1 See Devereux et al. (2002). 2 The fact that tax revenues did not converge downwards (see recent evidence by Stewart and Webb, 2006), can at least be partially explained by increased profitability and higher incentive for firms to incorporate (see Griffith and Klemm, 2004; Krogstrup, 2004; Devereux and Sorensen, 2005; Weichenrieder, 2005). Economic Policy April 2007 pp. 385– 430 Printed in Great Britain © CEPR, CES, MSH, 2007. 388 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY Figure 1. Corporate tax rates in the EU25, 1990–2006 (unweighted averages) Source: Devereux, Griffith and Klemm (2002), Eurostat, KPMG for statutory rates; Overesch (2005) for EATR (effective average tax rates). 1.1. A puzzle A contradiction seems to emerge regarding the perceived importance of the corpo- rate tax rate: at odds with the ‘race to the bottom’ scenario are a number of business surveys revealing that corporate taxation is not a primary criterion for location decisions. For instance, corporate taxation is only one of the 18 criteria listed by Ernst & Young (2005), along with domestic market, transportation, infrastructure, telecommunica- tion networks, labour costs, flexibility and skills, R&D, etc. However, European policy debates regarding tax competition between member states have primarily focused on corporate taxation; other issues and policy instruments (e.g. infrastructure, education, public R&D) have been to a large extent overshadowed. The present paper aims at breaking with the truncated debate in favour of a more balanced approach which considers the impact of both sides of government activity on location choices. In this respect, Figure 2 suggests that such bi-dimensional com- petition does exist in the EU. Figure 2 positions 18 EU member states according to their combined level of corporate tax rate and stock of public capital per square kilometre, as measured in 2002. We note that the countries with high corporate tax rates provide a high level of infrastructure, which could mitigate the disadvantages of high taxation. Our data is bounded by the Netherlands and Hungary: the former exhibits both a high tax rate and high public capital, while the latter displays a low combination of both. Interestingly, we observe that Ireland is somewhat of an outlier and positioned close to Hungary, while the majority of EU 15 governments display a relatively high combination of taxation and public capital. The race to the bottom and the bi-dimensional competition suggested by Figures 1 and 2 could be regarded as contradictory. At a first glance, the very nature of TAX COMPETITION 389 Figure 2. Public capital and statutory tax rate in 2002 Note: In some countries, several tax rates are applied. In Estonia, for instance, retained profits are not taxed. In France, small and medium size firms face a reduced rate. Here we use top statutory corporate tax rates, which, in the absence of tax optimization, are those applying to multinationals. For the calculation of public capital, see Section 3, Box 6. Source: Kamps (2004a), World Development Indicators (2005) and same as Figure 1. bi-dimensional competition seems to render any race to the bottom impossible. Indeed, high tax countries could remain attractive to foreign capital through the provi- sion of public goods, such as infrastructure, which are useful to firms. In keeping with this argument, it follows that high tax countries should have little or no incentive to take the ‘race-to-the-bottom’ route. However, many EU countries seem to be tempted by the prospect of lowering their corporate tax rates. For instance, Germany, a high tax and high spending country according to Figure 2, plans on cutting its statutory tax rate by almost 10 percentage points by 2007, and envisages further cuts in the long term. 1.2. The purpose of the paper The aim of the paper is to cast light on the existence of a ‘race to the bottom’ in the EU in spite of the importance of public goods in corporate location choices. It is even more important to solve this puzzle given that two opposite economic policies can be defended, depending on whether the message conveyed by Figure 1 or that conveyed by Figure 2 is favoured. On the one hand, from a normative standpoint tax coordina- tion within the EU can be justified on equity grounds (Figure 1). On the other hand, from a positive standpoint Figure 2 illustrates an impressive diversity of tax and spending profiles, which can be attributed to history, political preferences, or the pressure and lobbying exerted by firms. Judging from this diversity, a normative argument could be made for a laissez-faire policy: unnecessary at best, tax coordination may 390 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY even be harmful since it precludes each nation from expressing its own choice of tax level and public spending. To gain further insight into these contending views, we provide a theoretical frame- work which captures both the bi-dimensional (or ‘dual’) conception of competition as well as ‘race to the bottom’ competition. With this in mind, we revisit and develop the simple and popular model of tax competition introduced by Zodrow and Mieszkowski (1986), with the ambition of testing its predictions. In their model, capital taxation enables the government to provide public goods, such as infrastructure, which are in turn used as factors of production by the firms. In our framework, we call such ‘productive’ public goods public inputs. Their positive impact on marginal productivity mitigates the negative impact of capital taxation on after-tax private return. Consequently, an increase in capital taxation induces less capital outflow. Under reasonable assumptions providing for a ‘well-behaved’ production function, our theoretical model lends credence to the idea that benevolent governments looking to attract foreign capital should either (1) cut taxes while reducing public input, or (2) rely on immobile tax bases in order to fund the public input. Our main methodological contribution is to emphasize that the incentive for ‘a race to the bottom’ is an empirical question that can be econometrically evaluated: it requires comparing the elasticities of capital with respect to the tax rate and with respect to the public input. We then estimate the elasticities of foreign capital with respect to corporate taxation rates and public capital on the basis of US foreign direct investment in 18 EU countries between 1994 and 2003. Standard control variables are used, such as the size of the economies under investigation, the existence of agglomeration effects, and unit labour costs. 1.3. The key findings The dual and ‘race to the bottom’ conceptions of competition are nothing but the two sides of the same coin. On the one hand, inward foreign direct investment is significantly affected in a negative manner by corporate taxation and positively affected by public capital stock. On average, a higher tax rate must be compensated by a higher stock of public capital at the equilibrium. Low-equilibrium and high-equilibrium countries can coexist when the two elasticities are opposite in sign. This supports the evidence of Figure 2. On the other hand, our estimates confirm that funding a marginal increase in public capital through corporate taxation, induces a capital out-flow. Public capital is in fact not productive enough to compensate for the required increase in corporate taxation. This conclusion holds unless labour taxation bears an increasing share in the financing of public inputs. Consequently, even if public infrastructure is appealing to firms, the threat of a race to the bottom being triggered cannot be dismissed. The magnitude of this threat largely TAX COMPETITION 391 depends on the (un)willingness of voters to bear an increase in taxes destined to benefit firms. Figure 1 suggests that voter reluctance has decreased in the recent past. In contrast with the attractiveness of public inputs, household-oriented public goods, such as health or social security spending, seem to be unappealing to foreign investors. Competition could therefore lead to a distortion in the structure of public spending, in favour of firms and at the expense of households. These findings bear important policy implications. National governments are caught in a vice with firms demanding competitive tax cuts on one end, and households making their voices heard by way of the ballot box. From this perspective, tax co-ordination in the EU is justified on both political and social grounds. An underlying argument for coordination is that it may allow citizens to influence the distribution of the tax burden and that of spending, while at the same time mitigating the risk that divergent national preferences create too large a gap between societies and political leadership at the EU level. However, minimum tax rates may not provide the appropriate solution. We suggest that tax coordination should instead consist in agreeing on the minimum share of public inputs to be funded by corporate taxation. Each country would remain free to decide which should be the nature, quantity and quality of its public input. The paper is organized as follows. Section 2 develops the theoretical model, Section 3 provides econometric support for its main implications, Section 4 derives policy implications and Section 5 concludes. 2. A CONCEPTUAL FRAMEWORK FOR TAX AND PUBLIC INPUT COMPETITION It has been acknowledged at least since Tiebout’s celebrated paper (1956) that both local and national jurisdictions compete on taxes and public inputs to attract economic agents. The driving force behind this competition is the mobility of agents (be they households or firms) which leads each jurisdiction to vie for the most attractive combination of tax-level and public spending. Zodrow and Mieszkowski’s (1986) paper is the cornerstone of a wide literature on tax competition,3 which examines the implications of differential mobility between different production factors, with capital being mobile while land or labour is not. In their seminal paper, a public good (which can be either consumed or used as an input) is financed by a distorting source tax on capital. It is shown that the relative cost of providing the public good is greater under perfect capital mobility than in closed economy. Consequently, when capital taxation is the only viable option, capital mobility leads to the under-provision of public goods, as compared to the autarky case. 3 See reviews by Wilson (1999) and Krogstrup (2002). 392 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY This under-provision result has been challenged by Noiset (1995), Sinn (1997, 2003), and Dhillon et al. (2006) when dealing with the case of public inputs. Noiset emphasizes that under-provision comes from an assumption about the impact of the public input on the marginal productivity of capital. Specifically, Zodrow and Mieszkowski (1986) assume that an extra unit of public good raises the marginal productivity of private capital by less than its marginal cost for private investors in terms of additional taxation. When this assumption is relaxed, tax competition can (1) be efficient or (2) can also result in a ‘race to the bottom’ or (3) may even result in a ‘race to the top’, characterized by an over-provision of public goods and excessive tax rates (Dhillon et al., 2006). Naturally, such indeterminacy is a cause for concern because it calls for conflicting policy implications. For instance, if tax competition leads to a ‘race to the top’, then tax co-ordination should consist in capping tax rates rather than introducing a minimum rate. It is shown in this section that the case for under- or over-provision of public goods basically depends on the relative values of two elasticities which can be econometri- cally estimated: (1) the elasticity of capital with respect to the tax rate, with the level of public input held constant; and (2) the elasticity of capital with respect to the public input, with the level of the tax rate held constant. Estimates of these elasticities are presented in the empirical section. It is also argued that the conventional wisdom about under-provision of public goods is confirmed under reasonable assumptions regarding the production process. All proofs are relegated to Appendix A. 2.1. A model with testable implications Let us now consider an extension of Zodrow and Mieszkowski’s (1986) model in which public goods are used by both consumers and firms. In practice, this is true of many public goods, such as transport infrastructure, public education, justice and police, among others. These are to be distinguished from other public goods such as sports, ‘cultural infrastructure’, health expenditures and inter-individual redistribution, which target households rather than firms. For the sake of clarity, a public input is defined herein as a public good used both by firms and households. It is also taken to be a pure public good à la Samuelson within the country’s borders.4 More specifically, for a given output level, the quantity available to each firm is independent of the number of firms. This assumption seems reasonable when considering most public inputs, such as public infrastructure, education or research. In addition, congestion in the use of this public input is ruled out.5 In our simple model, presented in Box 1, all firms are identical. Production for each firm depends on its capital stock and employment, as well as on the public 4 From an international point of view, it is a local public good. 5 The marginal productivity and the marginal utility of public inputs do not depend on the number of users. Clearly, this is a simplifying assumption. For a more nuanced approach to congestion, see Sinn (1997, 2003). TAX COMPETITION 393 Box 1. The model Let F (K, L, G ) be the production function, where K stands for capital, L for (fixed) labour and G for public input. If x denotes the consumption of the private good, the government maximizes the utility of a representative household: Max U(x, G ) (1.1) under the three following constraints: (i) Public budget constraint: a fraction α (0 ≤ α ≤ 1) of the public input G is funded by capital taxation tK: tK = αG (1.2) (ii) Balanced current account: net exports are equal to net capital income receipts (long-run national budget constraint): F (K, L, G ) − x − G = r (K − K ) (1.3) where K denotes households’ wealth and r is the world rate of return. (iii) International arbitrage condition: after-tax return (FK − t) is equal to world return r (profit maximization constraint): FK − t = r (1.4) In a closed economy, constraint (1.3) reduces to F (K, L, G ) = x + G and constraint (1.4) is replaced by: K = K. input. The private good and the public input are produced by the same production process, and the gross opportunity cost of producing one extra unit of public input in terms of the forgone production of the private good is equal to one. Capital is assumed to be internationally mobile, as opposed to labour, which is not. Consequently, under the small open economy assumption, arbitrage implies that the after-tax return of capital is equal to the exogenous world return.6 Public input is partially financed through a source tax t on the capital stock available in the country, K, either owned locally or internationally. Hence, other taxes are levied to finance its provision. Since labour is assumed to be immobile, a tax on this 6 Two alternative assumptions are usually considered in the literature. Under the small country hypothesis, the world capital after-tax return is exogenous whereas it is endogenous in the large country case (cf. Wildasin, 1988; Laussel and Le Breton, 1998; Wooders and Zissismos, 2005). Here, it is assumed that the return of capital in the country has no feedback effect on the world return of capital. This seems a realistic assumption for each European country taken separately. 394 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY factor induces no distortions and is therefore lump-sum. For a basic model in which the government is completely free to set both labour and capital tax rates to reach efficiency, the outcome is quite predictable: the efficient government would indeed rely entirely on labour taxation. A more realistic and nuanced model might help us gain further insight by introducing a binding constraint on the tax rates, which might for instance capture the government’s concern for fairness between labour and capital taxation. Let 0 ≤ α ≤ 1 be the proportion of public input funded by the corporate tax. The two extreme cases correspond to pure capital taxation (α = 1) and to lump-sum taxation (α = 0). The rate of capital taxation is chosen to maximize households’ utility,7 which depends on their private-good consumption and on the public input. In particular, households benefit from public inputs both directly, since they consume it, and indirectly, since it helps firms in producing more private goods. Each household’s wealth endowment can either be invested in local or foreign capital. In each case, it follows from international arbitrage that the after-tax return is the world return. Households also receive labour income equal to the production surplus once the world capital return have been paid by firms. The optimal provision of public good in this open economy is contrasted with the optimal provision under autarky, where capital is immobile. The closed economy case serves as a benchmark, and is therefore presented first. In such an economy, capital and labour resources are fixed. In particular, the domestic capital stock is equal to households’ wealth endowment. Hence, a tax on either factor is equivalent to a lump-sum tax and efficiency of resource allocation is independent of the means by which the public good is financed. An efficient allocation, where household utility is maximized, is characterized by equality between the marginal rate of substitution and the marginal rate of transformation. Here, the marginal rate of substitution is the amount of private good the consumer is willing to give up to consume one extra unit of public input. The marginal rate of transformation is the net opportunity cost of producing one additional unit of public input instead of producing the private good. In our economy, the latter is equal to 1 minus the marginal productivity of the public input. Therefore, in closed economy (Gclosed), the efficient provision of public inputs is reached when the marginal rate of substitution between the public and private goods (MRSG/x) is equal to 1 net of the marginal productivity of the public input (FG). Formally this reads as follows: MRSG/x = 1 − FG (1) In an open economy, capital is assumed to be perfectly mobile between countries, while labour remains immobile. Thus, a tax increase to fund additional public input may lead capital to flow in or out of a given country. This adds a new constraint, or pressure, on government decision-making. At first glance, and all other things being 7 Hence, we do not consider the case of a Leviathan government. See Edwards and Keen (1996). TAX COMPETITION 395 equal, we expect capital to flow out since a tax increase depreciates the domestic return of capital (direct effect). However, the ‘all other things being equal’ clause is not satisfied here. Indeed, more corporate tax receipts translate into more public input, which raises the marginal productivity of capital and thus the domestic return of capital (indirect effect). If the indirect effect prevails, capital may flow in. Consequently, the optimal provision of public goods when opening up the economy depends on the direction in which capital reacts subsequent to a marginal increase in taxation. Specifically, the optimal public-input provision Gopen requires the following condition to hold (see Appendix A): e K /t MRSG/x = 1 − FG − α , (2) 1 + e K /t where eK/t is the elasticity of private capital with respect to the tax rate, accounting for the fact that higher taxation means higher public-input provision. This tax elasticity, formally defined in Box 2, is the percentage change in capital generated by a percentage Box 2. The adjusted tax elasticity of capital The adjusted tax elasticity of capital is defined as: dK /K e K /t = 1 (2.1) dt /t G = tK α As shown in Appendix A, eK/t can be detailed in the following way: ek0/G + ek0/t e K /t =. (2.2) 1 − ek0/G with: dK /K ek0/t = 0 (2.4) dG /G It follows from (2.2) that at the optimum of an open economy, ek/t ≤ 0 if and only if Either: |ek0/t | ≥ ek0/G Or: ek0/G ≥ 1. (This second possibility is ruled out by one of our assumption of decreasing returns.) 396 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY change in the tax rate, with the provision of public input adjusted accordingly. Henceforth, it is referred to as the adjusted tax elasticity of capital. The optimal rule adopted by the government in an open economy (2) differs from that which is adopted in a closed economy (1). The two rules only coincide when the elasticity is zero, that is, when the tax rate has little or no influence on the level of capital, or when there is pure lump-sum funding.8 Comparing these two optimal policies reveals that solving the under-provision issue only requires determining the sign and magnitude of this elasticity. The sign of eK/t, depends on whether capital flows in or out following a tax increase. In the conventional ‘race-to-the-bottom’ scenario, the adjusted tax elasticity is negative, so a tax increase invariably leads to capital outflows. If, in addition, the elasticity’s absolute value is less than 1, the marginal rate of substitution between private and public goods in the country at hand is higher under tax competition than under autarky. Since the MRS is a decreasing function of the public good, a shift against the provision of the public input at the margin is obtained. This may be interpreted as an inverse elasticity rule à la Ramsey: the more elastic the capital, the less taxed it must be, resulting in a lower provision of public inputs. Consequently, in a second- best setting, it is efficient to provide less public inputs because it is funded by a distorting tax. This remains true for whatever fraction of the public input is financed by corporate taxation. If a government intent on attracting foreign capital were to spend one extra euro towards the provision of public inputs, corporate taxation would have to be raised by less than α euros, meaning that corporate taxation is increasingly substituted by labour taxation. Indeed, the model predicts an outflow of capital whenever the relative contribution of the corporate tax towards this extra euro of public goods is greater than or equal to α%. However, if an extra unit of public input has a large positive impact on the marginal productivity of private capital, then the adjusted tax elasticity may be positive. In this case, the positive impact of an additional unit of public input may outweigh the negative impact of extra taxation on the domestic rate of return. Thus, the conclusion is reversed provided eK/t < 1: opening up the economy results in a shift towards more public input at the margin (‘race-to-the top’ scenario). Under any scenario, a heavier reliance on capital taxation (i.e. the higher α ), results in a larger gap between the closed-economy and open-economy variables.9 Conversely, if the public input is funded exclusively through labour taxation (α = 0), then opening up the economy is neutral as regards resource allocation between private and public goods. The issue of under- or over-provision at the margin may be settled by separating eK/t into two elasticities, both of which can be econometrically estimated: (1) the tax 8 This case corresponds to α = 0. Note that it is not the openness of the economy per se which induces inefficiency but the fact that a distorting tax is used. e 9 The term α K /t is increasing in α. eK/t itself depends on α. It has been omitted in the notation for convenience. 1 + e K /t TAX COMPETITION 397 0 elasticity of capital holding public input constant e k/t on the one hand; and (2) the elasticity 0 of capital with respect to public input holding the tax rate constant eK/G on the other hand. These two elasticities provide a clear advantage over the adjusted tax elasticity of capital, since each of them can be estimated because the other element is held constant. The former (1) is the standard tax elasticity of capital which gives the relative variation of the capital stock following a 1% increase in the tax rate. The latter (2) is the percentage change in capital in response to a 1% increase in the public input. Intuitively, it is natural to expect the former to be negative and the latter to be positive. However, their relative magnitude is an empirical issue with major consequences. Indeed, as shown in Box 2, the adjusted tax elasticity is negative if |eK/t 0 | > eK/G 0. As is made explicit in the econometric section (3.3), this condition is satisfied. Ultimately, our findings corroborate the view that tax competition leads to under-provision of public input at the margin. However, it is worth re-emphasizing that Equations (1) and (2) are marginal conditions. Comparing them allows us to determine the direction in which resource allocation is distorted by capital openness at the margin. It does not, however, provide a clear indication on the level of public input provision: this level can increase or decrease under autarky, depending on whether the country is a net capital importer or exporter. Global results are obtained under the assumption of certain decreasing returns, as is made more precise in the next section. 2.2. A case for under-provision It is first argued that, at the optimum of an open economy, plausible assumptions on the production function require that the adjusted tax elasticity of capital is negative, which in turn implies the under-provision of public inputs. A global scenario for the under- provision of public inputs is then derived under the same assumptions and further qualifications. Our assumptions concerning the production function all amount to some form of diminishing returns (Box 3): the marginal productivity of each production factor is positive and decreasing, and when both capital and public input grow at the same rate, the marginal product of capital exhibits diminishing returns to scale. In addition, the marginal productivity of capital increases with public input, which is a com- plement to private capital. However, this effect is lower the greater the level of capital or public input. The plausibility of this assumption – which is crucial for our results – merits illustration. Consider a road network. Without a doubt, adding one extra kilometre of road would increase the productivity of one additional truck. However, intuitively the higher the initial number of trucks or the higher the initial length of the road network, the lower the positive effect. As shown in Box 4, these reasonable assumptions imply that for small, open economies relying solely on the corporate tax, the adjusted tax elasticity of capital must be negative at the optimum. A closer look at the implications of the arbitrage condition on the capital market illuminates this result. 398 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY Box 3. Assumptions on the production function Let F (K, L, G ) be an increasing function of all three of its arguments (FK, FL, FG > 0) and let the marginal productivity of each factor be decreasing (FKK, FLL, FGG < 0). The marginal product of capital exhibits decreasing returns to scale in capital and public input, i.e. FK(L, K (1 + λ), G (1+ λ)) < FK (L, K, G ) for any λ > 0. (This assumption is confirmed by Kamps (2004b) on the basis of estimated augmented Cobb–Douglas functions for OECD countries.) In addi- tion, K and G are complements in the sense that the marginal productivity of capital increases with G (FKG > 0). However, such complementarity declines at higher levels of capital or public input (FKGK, FKGG < 0). FKG tends to infinity when G = 0, and tends to 0 when G approaches infinity. The widely used augmented Cobb–Douglas production function, F (K, L, G ) = AL1−γK γG 1−β with 0 < γ < β < 1, satisfies these conditions. Box 4. The sign of the adjusted tax elasticity of capital Differentiating the arbitrage condition (1.4) and the public budget constraint (1.2) yields the adjusted tax elasticity of capital: t α − KFKG e K /t = (4.1) K αFKK + tFKG The fact that the marginal product of capital exhibits decreasing returns to scale as a function of capital and public input ensures that the denominator is negative (see Appendix A, Claim 6). Hence, eK/t and KFKG − α must be of the same sign. It is negative if the gain in the marginal capital productivity following a marginal increase in public input (FKG) is lower than the corresponding tax increase (α/K ): here, a capital tax increase with a corresponding rise in public input leads to capital outflows. It is shown in Appendix A that this scenario accurately captures an open economy’s equilibrium under the assumptions outlined in Box 3 when α = 1. As depicted in Figure 3, the optimum must thus lie in the downward-sloping section of the graph, which traces the after-tax return of capital as a function of public input. TAX COMPETITION 399 Figure 3. Capital return as a function of public input for a given capital level By definition, the after-tax domestic capital return ρ is the difference between the marginal productivity of capital FK and the budget-balanced tax rate t: G ρ( K , G ) ≡ FK ( L , K , G ) −. (3) K The after-tax domestic capital return is graphed as a function of public input in Figure 3. As emphasized in Appendix A, this graph is hump-shaped for any stock of capital K; an open economy at equilibrium is of no exception and also exhibits an inverted-U shape. Namely, ρ rises with G up to a threshold of public input G *, at which point ρ decreases. Given a fixed level of capital stock, this shape serves to illustrate the diminishing impact of an extra unit of public input on private return, bearing in mind that the cost of an extra unit of public input in terms of additional taxes remains constant for the firm. In general, two levels of public input satisfy the arbitrage condition: a low one and a high one (i.e. points B and A, respectively, where the after-tax domestic return of capital is equal to the world return r). It can be shown that the level of public input chosen by the government turns out to be greater than G *, and corresponds to point A (Appendix A). Seeing as this equilibrium lies in the downward-sloping segment of the graph, it follows that a marginal increase in public spending reduces the after-tax domestic return of capital. The following policy implication arises. In an open economy, it is not possible for a government to attract capital by raising corporate taxes even when spending all of the extra revenues on public input. In sum, although public services are productive, they are not productive enough to raise the after-tax capital return. The different levels of public input provided by a closed and an open economy can now be contrasted. Three possible scenarios are examined in turn. We first investigate a case in which the capital stock is equal to domestic wealth (K = K ). Here, the level of public input in an open economy Gopen is strictly lower than 400 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY that of a closed economy Gclosed. Gopen is therefore on the left of Gclosed, which implies that opening up the economy results in the under-provision of public inputs. The focus is now on a capital exporting country (K < K ). The after-tax return under autarky is thus below the world return, and capital flows out when opening up the economy. If the demand for public inputs increases with income (i.e. if it is a normal good), public input provision increases with capital. Therefore, it follows that a ‘global’ version of the under-provision result is valid as well. We now turn to a capital-importing country (K > K), i.e. a country whose after- tax return under autarky is higher than the world return. The inflow of private capital allows public input provision to increase. However, the inflow of capital must be above a certain threshold to ensure a greater provision of public inputs than under autarky.10 Below this threshold, public input is still under-provided. However, the ‘provision gap’ between the autarky and open economy equilibria is smaller here than under the two other scenarios developed above. The above analysis does not take into account the possibility of agglomeration economies, which would preclude marginal capital productivity from declining when capital accumulates. A number of recent models have indicated that ‘centrally located’ countries may benefit from a geographic rent, which allows them to keep their capital tax rates above those of ‘peripheral countries’.11 These models cogently suggest that geographic diversity is as a source of tax diversity. Additionally, they emphasize that public investment in infrastructures or R&D can favour agglomeration effects. As a result, the agglomeration dimension of the ‘tax & spend’ choice has to be acknowledged and taken into account in the empirical testing of the model. 3. ECONOMETRIC ANALYSIS 3.1. Empirical determinants of US foreign direct investment in Europe As argued in the previous section, the incentive for a ‘race to the bottom’ depends on the relative magnitudes of two simple elasticities: (1) the elasticity of capital with respect to the corporate tax rate, and with public input held constant, and (2) the elasticity of capital with respect to public input, taking the corporate tax rate as constant. These elasticities can be estimated on the basis of direct foreign investment allocation in Europe. We use data issued by the US Bureau of Economic Analysis that provides a measure of US foreign direct investment in EU countries. This data- base, widely used in the literature (see Hines, 1999), is well suited for our exercise because of its large size (US FDI in 18 EU members for ten sectors over the span of 10 Beyond the threshold, the marginal result remains robust, i.e. the provision of public input is reduced at the margin, compared with autarky, but we cannot conclude that the amount of public input provided is lower. 11 Such models are part of the growing ‘new economic geography’ literature. See Krugman, 1991; Ludema and Wooton, 2000; Andersson and Forslid, 2003; Baldwin and Krugman, 2004. TAX COMPETITION 401 Figure 4. Total inward FDI and US position in the European Union, in 2004 (US$ million) Source: Bureau of Economic Analysis, and UNCTAD, World Investment Report. 1994 –2002)12 and its consistency (only one reporting country, one methodology, one taxation scheme of repatriated profits). In addition, the country distribution of US FDI does not contradict that of total inward FDI in EU countries (Figure 4). Our dependent variable is the capital expenditure by US majority-owned affiliates in 18 EU countries in ten industries, during the time span of 1994 to 2003. Capital expendi- ture is defined as an expenditure ‘made to acquire, add to, or improve property, plant, and equipment’ (Bureau of Economic Analysis, 2003, p. 88). Hence, the stock of capital expenditures corresponds to productive capital, i.e. the foreign component of K in our theoretical model.13 For convenience, it is herein called FDI, although FDI actually includes some intra-group capital flows which may be unrelated to physical investment. Both flows and stocks of FDI have been studied in the literature. Working with stocks has several advantages. First, it is coherent with our theoretical framework which models decision-making on the basis of capital levels. Second, stocks are much less volatile than flows, which are sometimes dependent on one or two large take- overs, especially in relatively small countries. Lastly, stocks are never negative or nil in our database. This allows us to work with logarithms and estimate elasticities. To date, the extensive literature on the determinants of FDI14 has not devoted much attention to the impact of public inputs (see Box 5). This silence may partly be 12 The 18 EU countries are the EU15 plus the Czech Republic, Hungary and Poland. The ten sectors are chemical, electric and electronic goods, food, metals, machinery, finance (except banking), wholesale trade, transport equipment, other industries, and services. 13 Wheeler and Mody (1992), among a number of other authors surveyed in Hines (1999) use the same dependent variable. 14 See, for instance, Eaton and Tamura (1996), Wei (2000), or Bloningen and Davies (2002), Head and Mayer (2004), Bloningen (2005). 402 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY Box 5. The empirical literature The combined impact of tax rates and the provision of public inputs on capital location has not yet been extensively explored in the empirical literature. This combined impact can be tested either at the local or international level. At the local level, Gabe and Bell (2004) study the influence of public expenditures and property tax rates on corporate location choices in Maine (US) from 1993 to 1995. They find that raising education spending by 10% leads to a 6% increase in the number of firms settling there. Their study confirms that firms take the provision of public factors into consideration when choosing their location. According to Gabe and Bell, a strategy combining low taxation and a low provision of public factors is less successful than the opposite strategy that couples high taxation with a high provision of public goods. At the international level, there is an extensive literature with the ambition of assessing the impact of corporate taxation on FDI (see the surveys by Hines, 1999 and Devereux and Griffith, 2002). The meta-analyses proposed by de Mooij and Ederveen (2003, 2005) show that, on average, a 1% decline in the corporate statutory tax rate raises inward FDI by 3– 4% (a result gleaned from 371– 427 recorded semi-elasticities of FDI with respect to corporate tax rates). In their overview of the differences arising in the literature, de Mooij and Ederveen account for differences in data, methodology, and control vari- ables, but without including public factors as routine controls. The impact of public infrastructure has been investigated much less often. Wheeler and Mody’s (1992) index of infrastructure quality (drawn from the Country Assessment Service of Business International) reveals a positive and significant impact on US capital expenditures in majority-owned affiliates during 1982– 88. Loree and Guisinger (1995) use two composite infrastructure indicators to assess their relationship with US foreign direct equity investment in 1977 and in 1982. Mody and Srinivasan (1998) obtain a positive and statisti- cally significant relationship between a given country’s production of electricity per dollar of GDP and the level of US FDI in that country. due to the challenging task of identifying an appropriate measure of public infrastruc- ture. In our study we use the stock of public capital per square kilometre as a proxy for public infrastructure. This variable is in agreement with our theoretical model which assumes the absence of congestion in the use of public capital.15 It is denoted LPUBKijt in logarithm for country i, sector j and year t. 15 As a robustness check, we also performed some estimations with public capital per capita as a measure of public input. Due to collinearity between this measure of public capital and external distance, the results were weakly significant but still consistent with the results presented in the paper. TAX COMPETITION 403 The impact of public capital on FDI is to be contrasted with that of household- specific public goods as measured by (1) the logarithm of the ratio of social public expenditures to GDP (LSOCEXPit) and that of (2) the logarithm of the ratio of health public expenditures to GDP (LHEALit). As these two variables are unlikely to have a direct impact on private capital productivity, they are not expected to attract FDI. They could even deter foreign direct investors from investing if they anticipate higher tax pressure resulting from higher social expenditures. Two measures of corporate tax rates are used: statutory rates (LSTATUTit) and effective average tax rates (LEATRit), both expressed in logarithms. While in theory the EATR provides a more accurate measure of corporate taxation, its computation relies on a number of assumptions concerning: the type of investment (share of real estate, machinery, intangibles, inventories and financial assets in the investment project); the source of financing (the share of retained earnings, equity and debt); a fixed inflation rate; a fixed interest rate; and a fixed, pre-tax rate of return on investment.16 In particular, the way investment is financed could be thought of as endogenous to the tax rate (thin capitalization strategy). The EATRs used here were calculated by the ZEW 17 on the basis of a 20% pre-tax return and a 2% inflation rate. The difficulty is that this calibration may not apply to every EU country. In particular, we expect the EU15 countries to have lower pre-tax returns and lower inflation rates than the new member states. However, as argued by Overesch (2005), the main source of difference in EATR across countries pertains to statutory rates. Consequently, we have included these statutory rates in our baseline estimations, and robustness checks are performed with EATRs. In keeping with the literature, two sets of control variables are introduced: economic geography and costs. The economic geography literature (see Baldwin et al., 2003) maintains that market size, distance to markets, and agglomeration economies influence inward FDI because of the combination of returns to scale and trade costs.18 For instance, using these two factors (returns to scale and trade costs) we can formulate an alternate definition for market potential (which is generally defined as the weighted sum of domestic and foreign GDP). In particular, one such definition has been applied in the context of the EU, and involves weighing the two factors according to internal and external distances (see Harris, 1954; Head and Mayer, 2004). Internal distance is taken into account to mirror internal transportation costs. External distance reflects the transportation costs borne when supplying European foreign markets from a given, reference country. However, there is no reason to suspect that FDI should react to GDP, internal and external distances in the same manner. Thus, GDP remains separate from distance in our estimations. 16 The methodology was first proposed by King and Fullerton (1984) and applied to OECD countries by Devereux and Griffith (2002). 17 Zentrum für Europäische Wirtschaftsforschung. We are grateful to Michael Overesch for kindly providing the data. 18 Market access is not considered here because all target countries are part of the EU. 404 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY Both the dependent variable (FDI) and the market size variable (GDP of the host country in logarithm, LGDPit) are expressed at constant domestic prices but in current dollars. So, when currency i appreciates against the US$, the GDP of country i rises in US$. If the volume of FDI stays constant, FDI in current US$ rises by the same proportion as the domestic exchange rate. Nevertheless, US investors may be tempted either to reduce the volume of FDI, because local costs are higher once converted into US$, or to increase it, because of the target market’s higher purchasing. As this is an empirical issue, the nominal exchange rate (LEXCHit in logarithm) is added to the GDP as a separate explanatory variable, which is lagged one year to avoid reverse causality. Conventionally, internal distance is considered as the sum of town to town distances. It is therefore a measure of market density. Indeed, for a given GDP, a larger country is less attractive because of higher internal transportation costs. As for external distance, it corresponds to the log of the sum of distances from country i to each other destination country in the sample. In order to avoid multicol- linearity, one of the two distance variables must be omitted in our estimations. Due to border effects, external distance is expected to matter more than internal distance. Therefore, the former (LDISTEi ) is included rather than the latter. Finally, agglomeration economies are difficult to disentangle from market potential since they also rely on size. Head and Mayer (2004), for instance, use the number of firms in the same sector originating from the same country and which are already located in the region of destination. However, they work on a specific, firm-level dataset of Japanese firms. In our study, the share of employment in the destination country of a given sector (AGGLOit) is used as a proxy for agglomeration economies, an approach that is consistent with Braunerhjelm and Svensson (1996) and Brauner- hjelm et al. (2000). The intuition is that we expect US multinationals to be attracted by countries specialized in their sector. The underlying assumption is that they will benefit from agglomeration economies, irrespective of the nationality of the firms already in place. Turning to cost variables, we include the logarithm of sector-level unit labour costs (LULCijt) in the estimations. As detailed in Box 6, unit labour costs are expressed in current dollars and we ensured that this sector-specific variable was not collinear with the nominal exchange rate (see Appendix B). A measure for labour market flexibility, which captures hiring and firing practices, is also taken into account in the set of controls (FLEXit). The definitions and sources of dependent and explanatory variables are detailed in Box 6. 3.2. Econometric methodology The following equation is estimated: LFDIijt = a0 + a1LGDPit + a2LDISTEi + a3AGGLOijt + a4LEXCHit + a5LULCijt + a6FLEXit + a7LTAXit + a8LPUBKit + uijt (4) TAX COMPETITION 405 Box 6. The variables used in the regressions LFDIijt denotes the logarithm of the stock of capital expenditures by US majority-owned companies located in country i, sector j, year t (with at least 50% ownership), deflated by the local investment price index in current US$ (100 in 2000). Source: Bureau of Economic Analysis (FDI in US$) and OECD (investment price index in local currency). LPUBKit is the logarithm of the net stock of public capital measured in US$ millions at constant prices per square km. Source: areas are from the World Bank, public capital is from Christopher Kamps (2004a), updated for 2002 using Eurostat public investment figures. For the Czech Republic, Hungary and Poland, which are missing in Kamps’ database, we use the index of public infrastructure constructed by Ayalp et al. (2004) for 2002, which we rescale by using the ratio of public capital to public infrastructure in Germany during the same year. We then recover the whole series of public capital in the four countries using Eurostat public fixed capital investment series and the same depreciation factor as in Kamps (2004a). For Luxembourg, which is also omitted in Kamps’ database, the public capital ratio is assumed to be the average between that of the Dutch and the Swiss in 2002. The whole series is then obtained in the same way as for the three NMS. This capital stock is that of ‘the government’ in the traditional sense. It includes public hospitals, education infrastructure, administration offices, public lands and buildings, gross fixed capital of public firms and public equipment, but excludes military capital. LSOCEXPit is the logarithm of social expenditures by public institutions as a share of GDP in country i at time t. Source: Eurostat. LHEALit denotes the logarithm of health expenditures as a share of GDP in country i at time t. Source: OECD. LTAXit is the logarithm of the corporate tax rate in country i and in year t. Two measures are used for this rate: statutory rates (Sources: Michael Devereux’s webpage and Eurostat) and effective average tax rates (EATR; Source: Overesch, 2005). LGDPit is the logarithm of GDP in country i and in year t at year-2000 prices, converted into US$ using current exchange rates. Source: OECD. LDISTEi is the logarithm of the sum of distances from country i to each of the other destination countries in the sample. Source: CEPII’s web page. AGGLOijt is the share of sector j in total employment of country i in year t. Source: STAN-OECD (rev3) database. Due to differences in sector coverage between STAN and BEA databases, some approximations had to be made and the shares do not sum to unity in each country over our sectors. 406 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY LEXCHit denotes the current nominal exchange rate index with respect to US$, base 100 in 2000. Source: OECD. For EMU countries, the exchange rate prior to 1999 is re-calculated with fixed conversion rates between each currency and the euro (conversion rates are available on ECB’s website). The nominal exchange rate is lagged one year to avoid reverse causality from inward FDI to the exchange rate. A rise denotes an appreciation of the domestic currency against the US$. LULCijt is the unit labour cost in country i, sector j, year t, in current US$. Unit labour cost indices in US$ (100 in 2000) are multiplied by each country’s real exchange rate with respect to US$ in 2000 (100 = PPP, > 100 if domestic currency is overvalued compared to PPP, 0, a3 > 0, a2 < 0, a5 < 0, a6 > 0, a7 < 0, a8 > 0, while a4 is a priori ambiguous. More specifically, a7 stands for the elasticity of capital with respect to the corporate tax rate, with public capital held constant. It corresponds to e K0 /t in the theoretical model (Section 2, Box 3). a8 is the elasticity of private capital with respect to public capital density, with the tax rate held constant, i.e. it represents e K0/G.19 Two difficulties arise when estimating Equation (4). The first one stems from potential collinearity among the regressors. Fortunately, the correlation matrix reported in Appendix B provides evidence against multicollinearity: the highest Pearson correlation coefficient is 53% in absolute terms, which is generally considered to lie within acceptable limits (see Cohen et al., 2003). The second difficulty pertains to endogeneity. Some explanatory variables may be affected by inward FDI. This is especially the case regarding the exchange rate because capital inflows are expected to lead to an exchange-rate appreciation. However, other variables such as GDP, unit labour costs, agglomeration, tax rates or public capital, 19 Additional estimations performed on semi-elasticities provide the same order of magnitude for the elasticity of the average tax rate (the results are available from the authors on demand). TAX COMPETITION 407 could also be involved. To resolve this problem, the exchange rate is lagged one year in all regressions and robustness checks are performed by lagging the variables already mentioned. Nevertheless, this is not sufficient since both FDI and explanatory variables could be affected by some other omitted variables. To deal with this unobserved heterogeneity, fixed effects are introduced. Since our panel is three-dimensional, three types of fixed effects have to be considered: time, sector and country. The three types of fixed effects are introduced successively rather than simultaneously, lest we leave too little variance to be explained by economic variables. The variance under scrutiny is different in the three exercises. In the time dimension, a group is defined by the average of the dependent variable for a given year. Then, time-fixed effects capture the between-group variance (variance across the years), which represents a small fraction of the total variance of the dependent variable (11.5%). Therefore, the regression with time-fixed effects is expected to generate approxi- mately the same result as the OLS regression. It should be likewise for the regressions with sector-fixed effects, since they reflect the between-sector variance, which only amounts to 11.8% of the total variance. However, introducing country-fixed effects changes the perspective dramatically since they absorb the variance of the dependent variable across the countries, which represents 64.7% of the total variance. Our FDI data does not depict a large intra-country dispersion. It is all the more so for explanatory variables and obviously for geographic distance. In Austria, Spain and Sweden, the statutory tax rate does not show any variation across time. Moreover, GDP, labour market flexibility and public capital are relatively smooth over time in all countries. As regards public capital, the coefficient of variation over time is often very low (below 5%). Only in new member states does it reach relatively high levels (22% in Poland). By contrast, the cross-country dispersion exceeds 100% in 2002. So, country-fixed effects are likely to be collinear with some explanatory variables. To address this difficulty, external distance, which is both time and sector invariant, is omitted at first. Unsurprisingly, the bulk of our sample’s variance is absorbed by country-fixed effects and the impact of tax and public capital variables cannot be measured, calling for another approach. We adopt a different strategy that involves running two kinds of regressions.20 First, in order to account for the within-sector × country variance (which only represents 8% of the dependent variable’s overall variance), Equation (4) is estimated using first-differentiated variables. The purpose of running this regression is to evaluate how changes in the explanatory variables over time influence FDI flows into and out of a given country-sector. Not much explanatory power is expected due to the low time variance of our explanatory variables.21 Second, Equation (4) is estimated using time averages in hopes of accounting for the remaining variance (the between sector × country variance which represents the bulk of the total variance, 86.5%). From a statistical viewpoint, an observation is now the time average of a country-sector. This 20 We are grateful to Brigitte Dormont for this suggestion. 21 Mody and Srinivasan (1998) face the same problem when performing a within-estimation. 408 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY second type of regression casts light on the impact of cross-sector, cross-country or cross-sector × country differences within the explanatory variables on average inward sector-country FDI. In this regression, country and sector-fixed effects are introduced. When country-fixed effects are taken into account, the regression correctly measures the impact of fiscal variables on inward FDI while controlling for cross-country heterogeneity. In all cases, standard errors are clustered at the country level, (see Moulton, 1990) and the residuals are corrected for heteroscedasticity. 3.3. Econometric results A first set of estimation results is provided by Table 1 below, where the corporate tax variable is the statutory tax rate. Columns (1) to (4) report the results obtained by Table 1. Benchmark results Fixed effects OLS Panel Aggregate Sector- Sector-level Sector- Baseline FDI level balanced level lagsa (1) (2) (3) (4) (5) (6) (7) No No No No Time Sectors Countries LGDP 1.145*** 1.185*** 1.153*** 1.187*** 1.143*** 1.173*** −0.907 0.052 0.062 0.069 0.061 0.058 0.054 0.654 LDISTE −1.148*** −0.609 −1.008* −0.725 −0.718 −0.625 − 0.234 0.453 0.486 0.443 0.425 0.464 AGGLO − 0.033*** 0.037*** 0.034*** 0.032*** 0.031** 0.032*** 0.008 0.010 0.008 0.008 0.012 0.007 LEXCH(−1) 0.067 1.062** 0.955** 1.128** −1.425* 0.741* −0.063 0.378 0.476 0.332 0.498 0.704 0.384 0.564 LULC 1.382*** −0.071 −0.146 −0.185 −0.161 −0.007 −0.409 0.328 0.316 0.324 0.384 0.387 0.373 0.266 FLEX −0.146*** −0.002 −0.024 0.005 −0.053 0.001 0.072 0.039 0.102 0.102 0.104 0.092 0.083 0.074 LTAX −2.044*** −1.479*** −1.321*** −1.463*** −1.446*** −1.505*** −0.223 0.115 0.179 0.196 0.177 0.169 0.179 0.495 LPUBK 0.181*** 0.277*** 0.171** 0.237*** 0.240** 0.293*** 0.624 0.068 0.085 0.080 0.079 0.088 0.080 0.938 CONS −0.583 −2.456 1.965 −0.881 0.128 −3.241 0.542 2.809 4.603 5.078 4.725 4.448 4.783 3.105 No. obs 165 1346 1800 1231 1346 1346 1346 R-squared 0.8511 0.6217 0.6097 0.6211 0.6325 0.7356 0.6694 Root MSE 0.6465 1.206 1.1899 1.2063 1.1925 1.0116 1.1342 Hausman − − − 0.04 − − − testb (0.92) a LGDP, LEXCH, LTAX, LPUBK, LULC and AGGLO are lagged one year. b Statistics and p-value. The null of the Hausman test is non-endogeneity (equal coefficients in columns (4) and (2)). *** significant at 1%; ** significant at 5%; * significant at 10%. Standard errors below coefficients are clustered at country level. TAX COMPETITION 409 OLS, that is, on pooled data. In Column (1), data are aggregated by countries, and consequently the sector dimension of the data is omitted.22 In all other columns, the sector dimension is included, which raises the number of observations from 165 to 1346. In Column (3), the qualitative values of the dependent variable are replaced by their conditional mean23 so that the number of observations rises to 1 800 to obtain a balanced panel. Finally, Column (4) is the same as Column (2), except that a number of explanatory variables are lagged. According to these four columns, inward FDI increases by slightly more than 1% for countries that experience a 1% increase in GDP, whereas a country which is 1% more peripheral suffers from approximately 1% less inward FDI. These orders of magnitude are relatively standard.24 Furthermore, a sectoral employment increase equal to 1% of total employment in a given country leads to a 3.5% increase in inward FDI in this sector/country. We find substantial evidence for the existence of an agglomeration effect, as did Braunerhjelm and Svensson (1996) and Braunerhjelm et al. (2000), who used the same variable for explaining Swedish location choices. In OLS sector-level estimations, the coefficient on the (lagged) nominal exchange rate is positive and significant at least at the 5% level. An exchange-rate appreciation is therefore positively correlated with an increase in inward FDI. One way of inter- preting this result is to say that the demand-side effect (higher purchasing power in the target country) prevails over the supply-side effect (higher costs). Conversely, the coefficient estimated for the sector-level unit labour cost is not significant. Moreover, in the estimation performed on aggregate FDI, the exchange rate coefficient is not significant, but the unit labour cost variable (which accounts for the whole economy in this context) is significant and of positive sign. All in all, there is no evidence suggesting that higher costs serve as a deterrent to inward FDI in the EU. While this may at first run against our intuition, these results are not isolated in the literature (see Loree and Guisinger, 1995; Wei, 2000; Head and Mayer, 2004; Bénassy-Quéré et al., 2005). They can be explained by demand effects or by the mismeasurement of labour productivity (see Devereux and Griffith, 1998 who also find unit labour costs to be non-significant determinants of US FDI in Europe). It should also be noted that the labour flexibility variable is either non-significant or of the wrong sign in this first set of estimations. The coefficient on the statutory tax rate is negative and significant at the 1% level in all four estimations (Columns (1) to (4)). The elasticities obtained (−2.0 on aggre- gate data, −1.3 to −1.5 on sector-level data) are consistent with the meta-analysis 22 The agglomeration variable is omitted since it is the share of each sector in the employment of each country. For unit labour costs, we use country data. 23 There are in fact 454 values of the dependent variable which are expressed as a qualitative variable. We know that FDI has been positive but no aggregate amount is reported. We then replace the unknown amount by the sector-country mean on observable data. 24 Note that the distance variable does not measure distance from the origin country (the United States), but average distance between the host country and the other European countries. 410 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY proposed by de Mooij and Ederveen (2005). As is consistent with theory, more public capital attracts FDI, and the coefficient is significant at least at the 5% level in all four regressions. This strongly suggests that, all things (and in particular the corporate tax rates) being equal, public capital attracts FDI. A 10% increase in public capital will, on average, translate into a 2% increase of FDI inflow. However, the latter elasticity is significantly lower than for the statutory tax rate, and less than 1. This implies that the condition for the under-provision of public capital is met (see Box 2 in Section 2): a 1% drop in corporate taxes with a corresponding decrease in public capital leads to a foreign capital inflow (the elasticity eK/t is negative). These results are not altered by balancing the panel (Column 3) or introducing lags (Column 4). The Hausman test reveals that the null hypothesis of non-endogeneity cannot be rejected at the 1% level (see the bottom of Table 1, Column (4)). The results obtained from panel data estimations are reported in Columns (5) through (7). The regressions with fixed effects on time (Column (5)) or on sectors (Column (6)) lead to similar results as the OLS estimations. The sole difference is that the regression with time-fixed effects yields a negative exchange rate coefficient, that is, a currency appreciation leads to lower inward FDI. Indeed, countries whose currencies appreciate relatively more than others tend to receive less FDI, probably because of a relative loss in cost-competitiveness. When country-fixed effects are introduced (Column (7)), only agglomeration, which is a country × sector × time variable, remains significant. Even GDP fails to be significant in this last regression of Table 1. As already mentioned, one prominent feature and difficulty that charac- terizes our dataset is the limited variance of most explanatory variables in the time dimension – especially when compared to the larger variances found in the sector dimension and, above all, in the country dimension. Hence, panel data analysis with country-fixed effects is unlikely to provide very useful results about the nature or existence of dual competition on tax rates and public inputs. In our estimates from Table 3, the regression using sector-fixed effects is taken as our baseline specification. The results obtained from running regressions on reduced dimensions of the panel are presented in Table 2. In the first column, all variables are in first-difference. As expected, the model is unable to explain the changes in FDI for each sector-country over time, due to the very low variance of explanatory variables. In the other columns, all variables are averaged over time and the OLS regressions are performed on averages. Columns (2) and (3) first present the results without fixed effects under two different specifications. We note that in this case both the tax rate and public capital become significant, of correct sign and close to the benchmark estimation. In Column (4), country-fixed effects are introduced. The software then automatically drops redundant fixed-effects or explanatory variables. Both fiscal variables remain signifi- cant and of correct sign, but the impact of labour flexibility is surprisingly negative and significant. In Column (5), this variable is excluded and the results are unchanged for taxation and public capital. In the last column, both country and sector fixed TAX COMPETITION 411 Table 2. Regressions on reduced dimensions of the panel Fixed First differences Time averages effects (1) (2) (3) (4) (5) (6) No No No Countries Countries Countries and sectors LGDP 0.230 1.161*** 1.130*** 1.036*** 1.124*** 1.032*** 0.537 0.078 0.071 0.007 0.010 0.011 LDISTE − − −0.954*** − −0.201** − 0.243 0.089 AGGLO 0.024 0.036*** 0.034*** 0.036*** 0.036*** 0.013 0.047 0.009 0.009 0.010 0.010 0.012 LEXCH(−1) −0.031 1.189 −0.123 − − − 0.160 1.566 1.811 LULC 0.011 0.023 0.253 −0.992 −0.992 −1.131 0.123 0.694 0.622 0.722 0.687 1.016 FLEX −0.003 0.007 −0.069 −0.241*** − −0.241*** 0.031 0.127 0.110 0.010 0.015 LTAX 0.053 −1.408*** −1.483*** −1.225*** −1.187*** −1.213*** 0.140 0.257 0.198 0.059 0.036 0.087 LPUBK −0.367 0.316*** 0.180** 0.144** 0.179** 0.144* 0.312 0.086 0.077 0.056 0.064 0.080 CONS 0.001 −8.763*** − −2.508 −1.772 −2.922 0.016 2.710 2.983 3.854 4.501 No. obs 1620 180 180 180 180 180 R-squared 0.0006 0.6552 0.7734 0.7183 0.7183 0.8425 Root MSE 0.6176 1.0935 1.07 1.0248 1.0248 0.7889 *** significant at 1%; ** significant at 5%; * significant at 10%. Standard errors below coefficients are clustered at country level. effects are introduced. The results are basically unchanged, except for the agglomer- ation variable which fails to be significant. Columns (4) to (6) correspond to our preferred specifications since they account for unobserved countries heterogeneity. The elasticity of FDI with respect to the tax rate is lower when country-fixed effects are included, which is in line with de Mooij and Ederveen’s (2005) findings.25 It is worth mentioning that the elasticity of FDI with respect to the tax rate is again significantly higher than unity in absolute value. Therefore, raising the corporate tax rates while keeping the stock of public capital constant reduces tax revenue, which would be inconsistent with a government’s optimal decision’ in an open economy.26 It is equally important to note that the coefficient on public capital is lower when country-fixed effects are introduced but remains significant at least at the 10% level 25 De Mooij and Ederveen find a typical semi-elasticity of −1.92 for panel data, compared to −7.81 on a cross-section basis. In a recent paper, Razin and Sadka (2006) obtain a −3.6 semi-elasticity using a Heckman estimation on flow FDI data with country-fixed effects. In Columns (4) and (5), a −1.19 to −1.23 elasticity is found. With an average corporate tax rate of 33%, a semi-elasticity ranging from −3.6 to −3.7 is obtained. 26 See Appendix A. Values for tax elasticity found in the literature are quite high and it is not uncommon to find figures above 1 in absolute value (see the previous footnote). 412 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY Table 3. Panel estimations, robustness checks Baseline 1994 –1999 1998–2003 EATR Lagged Social Health (1) (2) (3) (4) EATR expend. expend. (5) (6) (7) LGDP 1.173*** 1.166*** 1.179*** 1.192*** 1.189*** 1.205*** 1.201*** 0.054 0.041 0.070 0.055 0.054 0.041 0.072 LDISTE −0.625 −0.100 −1.049** −0.494 −0.581 −0.980*** −0.582* 0.464 0.446 0.462 0.469 0.474 0.320 0.455 AGGLO 0.031** 0.033** 0.028** 0.032** 0.031** 0.027** 0.031** 0.012 0.014 0.011 0.013 0.013 0.011 0.012 LEXCH 0.741* −0.454 1.090*** 0.858* 0.978** 0.914*** 0.769* 0.384 0.442 0.346 0.434 0.447 0.313 0.390 LULC −0.007 0.275 −0.181 −0.062 −0.211 −0.151 −0.093 0.373 0.303 0.496 0.368 0.457 0.296 0.366 FLEX 0.001 −0.061 0.017 −0.001 0.013 0.015 −0.005 0.083 0.059 0.100 0.089 0.092 0.061 0.075 LTAX −1.505*** −1.686*** −1.460*** −1.841*** −1.811*** −0.996*** −1.442*** 0.179 0.188 0.180 0.326 0.309 0.150 0.163 LPUBK 0.293*** 0.496*** 0.165* 0.307*** 0.268*** 0.176** 0.297*** 0.080 0.073 0.087 0.079 0.073 0.069 0.089 LSOCEXP − − − − − −1.470*** − 0.427 LHEAL − − − − − − −0.590 0.580 CONS −3.241 −7.542 1.515 −4.942 −3.847 5.042 −2.105 4.783 4.678 4.688 4.966 5.311 4.019 5.066 No. obs 1346 718 889 1346 1231 1346 1346 R-squared 0.7356 0.7567 0.7558 0.7289 0.7302 0.7498 0.7369 Root MSE 1.0116 0.9993 0.960 1.0243 1.0217 0.98448 1.0096 Hausman − − − − 0.78 − − test (1.000) *** significant at 1%; ** significant at 5%; * significant at 10%. Standard errors below coefficients are clustered at country level. in all specifications. Consequently, this step of our econometric methodology corrobor- ates the qualitative predictions offered by our theoretical model, and we can safely conclude that the adjusted tax elasticity (eK/t in the theoretical model) is negative. Table 3 presents several robustness checks for our baseline estimation, and this latter estimation is reported for convenience in the first column. Columns (2) and (3) report the results from the same panel estimation with sector-fixed effects on two sub-samples: 1994 –1999 and 2000– 2003. The corporate tax rate and public capital coefficients are of correct sign and significant at least at the 10% level in the two sub-samples, although we note that both coefficients decline over time. In contrast, the distance coefficient is only significant in the second sub-sample, which is consistent with a trade literature showing that, in spite of decreasing transportation costs, distance is becoming increasingly important (see, for instance, Brun et al., 2005). In Column (4), the statutory tax rate is substituted by the EATR. The elasticity of FDI with respect to the EATR is higher than with respect to the statutory tax rate, which is TAX COMPETITION 413 standard in the literature (see de Mooij and Ederveen, 2005). Similar results are obtained when the EATR is lagged (Column (5)). Finally, the last two columns of Table 3 include social expenditures and healthcare expenditures as additional controls. It is remarkable to note that the coefficients on both variables are negative, which suggests that higher household-specific expenditures have a negative impact on inward FDI. That said, the social expenditure coefficient is the only statistically significant one. In this regression, the public capital coefficient declines to slightly less than 0.2, while the tax rate elasticity decreases to a unitary value. Introducing health expenditures, however, does not affect the regression. The final step of the analysis involves assessing how the impact of taxation depends on such country characteristics as: (1) size, (2) trade openness, (3) the existing stock of public capital and (4) the tax rate. The rationale for examining (1) and (2) can be found in the traditional tax competition literature; the standard argument proposes that small open countries have a greater incentive to cut taxes than their larger, relatively closed counterparts because the perceived elasticity of capital with respect to taxation is higher. In turn, justification for including (3) can be found in our theoretical model, which yields a tax elasticity of capital that is a declining function of public capital stock.27 Hence, capital is less elastic with respect to the tax rate in countries that spend more on public input. The sign for the tax elasticity derivative with respect to the tax rate (4) cannot be a priori determined, but applying a Laffer curve argument suggests that high tax countries are likely to exhibit a higher elasticity than low tax countries. In order to study such heterogeneity, dummy variables are introduced. Each of them divides the sample into two groups of countries, according to their size, openness, public capital and corporate taxation, successively. These dummies are then included with tax rates in the estimations. The list of countries found in each group is provided in Appendix C.28 The results are reported in Table 4. The elasticities obtained when grouping the countries according to their size (Columns (2)) or openness (Column (3)) are all significantly negative, but not significantly different from one another. In contrast, the tax elasticity for ‘high public capital’ countries is not significant, as opposed to ‘low public capital’ countries.29 This result suggests that a tax cut is more likely to attract FDI in low public capital countries.30 The final column of Table 4 shows that the elasticity of FDI with respect to the statutory tax rate is higher in high-tax countries than in low-tax countries, but the difference is not significant. In other words, from an econometric point of view we cannot safely claim anything ∂e K0 /t t FKGK 27 = < 0 under the assumptions detailed in Box 3. ∂G 2 K FKK 28 Trade openness is defined here as the sum of exports and imports over GDP in 2003, source OECD. Size is based on our GDP variable. Public capital is based on our PUBK variable (public capital per square kilometre) and tax on our STATUT variable. 29 See the Wald test on the bottom-most line of Table 4. Only in Column (4) does the test reject the null of equal coefficients on both groups of countries. 30 We have tried to test the robustness of this result by introducing two dummy variables in the multiplicative form in the specification reported in Table 2. The results are inconclusive due to one of these variables being dropped by Stata. 414 A. BÉNASSY-QUÉRÉ, N. GOBALRAJA AND A. TRANNOY Table 4. Panel estimations, country groupings Dependent Baseline Country Trade Public Statutory variable: (1) size openness expenditure tax level LFDI (2) (3) (4) (5) LGDP 1.173*** 1.168*** 1.143*** 1.230*** 1.236*** 0.054 0.092 0.054 0.069 0.067 LDISTE −0.625 −0.619 −0.765 −1.174** −0.595 0.464 0.435 0.489 0.512 0.467 AGGLO 0.031** 0.031** 0.031** 0.031** 0.032** 0.012 0.012 0.011 0.012 0.011 LEXCH 0.741* 0.741* 0.797* 0.438 0.802** 0.384 0.385 0.387 0.374 0.354 LULC −0.007 −0.007 −0.040 −0.090 −0.086 0.373 0.375 0.356 0.356 0.323 FLEX 0.001 0.002 −0.002 0.082 0.056 0.083 0.081 0.079 0.065 0.079 LTAX −1.505*** − − − − 0.179................................................................................................................................................................................................................................................................................................. Small size − −1.503*** − − − 0.179 Large size − −1.522*** − − − 0.350................................................................................................................................................................................................................................................................................................. High openness − − −1.564*** − − 0.231 Low openness − − −1.703*** − −