The Impact of Thin‐Capitalization Rules on the Location of Multinational Firms’ Foreign Affiliates

This paper examines how restrictions on the tax deductibility of interest cost affect location choices of multinational corporations (MNCs). Many countries have introduced so&#8208;called thin&#8208;capitalization rules (TCRs) to prevent MNCs from shifting their tax base to countries with lower tax rates. As of 2012, in our sample of 172 countries, 61 countries have implemented a TCR. Using information on nearly all new foreign investments of German MNCs, we provide a number of new and interesting insights in how TCRs affect the decision of where to locate foreign entities. In particular, stricter TCRs are found to negatively affect location choices of MNCs. Our results include estimates of own&#8208; and cross&#8208;elasticities of location choice and also novel results on the relative importance of tax base vs. tax rate effects. We finally provide estimates for different uncoordinated as well as coordinated policy scenarios.


| 37
MERLO Et aL. the first location choice). Second, we find that policies of one country exert significant externalities on other countries. For example, a 1% more lenient TCR in France would reduce the probability of locating in Argentina by −0.039%. Note that these externalities on other countries are heterogeneous across countries. This implies not only that own optimal policies differ, but also that coordinated action would produce winners and losers. Our estimations suggest that the main losers of a coordinated policy would be Austria, Belgium, Switzerland, and Ireland. The main winners of such a policy would be France, the United Kingdom, and the United States.
Finally, we provide estimates on the relative importance of tax rate vs. tax base effects. We illustrate this using the example of the United States and its policy options. Starting from actual values of tax and TCR policy, we demonstrate that location choices are more sensitive to tax rate changes. For the United States, our estimates imply that a 10 percentage point stricter TCR needs to be matched by a 2.3 percentage point lower corporate tax rate in order to keep the location attractiveness unchanged.
We believe that our paper not only contributes to the discussion about how to prevent profit shifting of MNCs but also to a general literature on the impact of tax and tax-base effects and their relative importance. We provide a number of new and instructive results supporting theoretical work. Given the externalities created by tax policy, our findings suggest that under strategic interaction, tax rates are set too low and TCRs are set too lenient. Coordinated measures against profit shifting by implementing a uniform TCR would therefore be clearly welfare increasing (see Haufler & Runkel, 2012).
The remainder of the paper is structured as follows. Section 2 briefly reviews related literature. Section 3 describes how TCRs work and in Section 4 we discuss how TCRs affect location choices of MNCs. Sections 5 and 6 describe the estimation strategy and our dataset. The results and numerous policy experiments and quantifications are reported in Sections 7 and 8. Additional results are reported in Section 9, while Section 10 discusses policy implications and concludes.

| RELATED LITERATURE
Our paper contributes to several strands of the literature. First, it relates to a growing number of empirical papers providing evidence on profit shifting by MNCs. For example, Swenson (2001), Clausing (2003), Bartelsman and Beetsma (2003), and Cristea and Nguyen (2016) show that firms distort intra-firm transfer prices in a way that is consistent with tax differentials. New evidence provided by Davies, Martin, Parenti, and Toubal (2018) suggests that tax avoidance through transfer pricing, particularly of large firms, is economically sizable. Desai, Foley, and Hines (2004), Huizinga, Laeven and Nicodème (2008), Mintz and Weichenrieder (2010), Wamser (2010, 2014), Møen, Schindler, Schjelderup, and Tropina (2011), Buettner and Wamser (2013), and Egger et al. (2014) present evidence that corporate taxes determine capital structure choices of affiliates of MNCs, which is in line with debt and profit shifting behavior (see also Heckemeyer, Overesch, and Feld, 2013, for a meta-study). Second, beside the contributions on the impact of TCRs, other papers confirm that legislations enacted by European countries to limit the abusive use of transfer pricing are effective (Lohse & Riedel, 2013;Beer & Loeprick, 2015). There is also evidence that controlled foreign company (CFC) legislation has an impact on how MNCs allocate passive assets across countries (Ruf & Weichenrieder, 2012). Our paper contributes to this literature by assessing the impact of TCRs on the location of real corporate activity of multinational firms. To the best of our knowledge, this link has so far largely been ignored.
Our paper is also related to prior work on the impact of corporate taxation on the location decision of MNCs. The large majority of papers on corporate taxation and firm activity analyze corporate tax rate effects on marginal investment decisions (see, e.g., Feld, 2011). The impact of corporate taxes on location choice is, on the contrary, studied by a relatively small number of papers. The seminal paper by Devereux and Griffith (1998) provides evidence that corporate taxation deters the location of subsidiaries of MNCs. Barrios, Huizinga, Laeven, and Nicodème (2012) confirm this finding using rich data on European MNCs. In line with this evidence, our estimates suggest a negative impact of corporate taxes on multinational location decisions and, additionally indicate a negative impact of stricter anti-avoidance rules. Moreover, contrary to most prior work, our analysis accounts for the worldwide location decision of multinational firms and does not restrict the perspective to a limited set of countries in the OECD, Europe or North America. The paper by Gumpert, Hines, and Schnitzer (2016) uses data on German MNCs to analyze the extensive margin of tax haven activity of MNCs.
Finally, a number of papers discuss to what extent the questions raised in the OECD BEPS report require action and what this action should look like. For example, Dharmapala (2014) argues that policy measures to prevent income shifting can not be implemented without having reliable estimates on the magnitude thereof. Hebous and Weichenrieder (2014) reason that measures to prevent profit shifting have been implemented successfully by many countries, but that it is less clear to what extent partial harmonization and coordination of these measures leads to beneficial results, given that tax rates are still set at the national level. Our paper contributes to the policy discussion by quantifying the externalities of uncoordinated anti-avoidance policies in terms of the attractiveness of a location for real investment. We also quantify the trade-off between base-broadening and tax-cutting reforms.

| THIN-CAPITALIZATION RULES
As described in the introductory section, MNCs have an incentive to distort the financial structure of their operations in order to shift income from high-tax to low-tax entities. This is achieved by injecting equity capital in a low-tax affiliate that then lends to related entities located in high-tax countries. As interest payments for intra-firm borrowing are deductible from the corporate tax base, the associated income is stripped out of the high-tax country and taxed at a low or zero rate at the low-tax or tax-haven entity.
The purpose of thin-capitalization rules is to limit the deductibility of interest payments on intrafirm loans from the corporate tax base, thereby reducing the described debt-shifting incentives. Most countries' tax legislation lays down specific safe-haven or safe-harbor debt-equity relations above which interest deduction is not restricted. 3 Once a firm's debt-to-equity ratio is in excess of such a safe-haven ratio, interest is no longer tax-deductible and fully taxed. An example may help to see this. For instance, interest costs of a foreign affiliate located in Canada are fully deductible only if its debt is below 1.5 times its equity. However, suppose a foreign affiliate is financed by a loan of 10 million Canadian dollar (CAD) and by 5 million CAD equity. Then, only 75% of the interest expenses are deductible as the loan exceeds 1.5 times equity by $2.5 million CAD (10-1.5 × 5). Denoting ω as the amount of debt and ϑ as the amount of equity, we can define a safe-haven threshold Θ as Using this definition, the Canadian safe-haven threshold (SHT) amounts to Θ CAN = 1. 5 1.5+1 = 0.6. Equation 1 implies that higher values of Θ are associated with less strict TCRs and lower values of Θ are associated with stricter ones. In the extremes, if interest is non-deductible for all debt, Θ = 0; if interest deduction is not restricted and there is no TCR in place, Θ = 1. 4 (1) Θ ≡ + .

MERLO Et aL.
Our analysis is based on TCR information for a sample of 172 countries (see . In our data, the average SHT conditional on Θ < 1 equals 0.73. Hence, the Canadian SHT is stricter than the average SHT in our data (conditional on Θ < 1). The prevalence of thin-capitalization requirements has increased substantially over our sample period. By 2012, 61 countries had implemented a TCR (111 countries did not have one). From 1996 until 2012, 37 countries have introduced a TCR, six relaxed their rules (an increase in Θ), and 21 countries made their rules stricter (a reduction in Θ). Four countries abolished their TCR between 1996 and 2012. 5

| THE EFFECT OF TCRs ON LOCATION CHOICES
As mentioned in Section 2, corporate taxation is an important determinant of MNCs' location choices. Previous work focused on the effect of profit tax rates on the choice of location. As Devereux and Griffith (1998) show, a firm facing a given number of possible locations will base its location decision on the comparison of after-tax profits arising at each location. The effective average tax rate (total tax payments relative to gross profits) determines the location choice through its effect on average costs. 6 Since TCRs directly determine the effective average tax rate, we expect them to have an effect on location choices. Denoting gross profits by G, the volume of debt financing by D, the statutory tax by τ, and debt interest by ι, we obtain a simple representation of an average effective tax as e measures the proportion of total profit taken in tax and, in line with the discussion above, a higher e reduces ceteris paribus after-tax profits at a given location and thus makes that location less likely to be chosen over other locations. The relevant component for understanding the effect of a TCR on e is the fraction of deductible interest expenses θ, θ ∈ [0,1]. This fraction is always 1 if Θ equals 1 and interest deduction is not restricted. If Θ < 1, the parameter θ may take any value between 0 and 1. A stricter rule (a lower Θ) implies a lower fraction of deductible interest expenses θ. Since e < 0, a stricter TCR implies a higher effective tax rate. This leads us to the following prediction: Hypothesis A laxer TCR (a higher Θ) implemented by a given country reduces the average tax burden faced by MNCs in that country and increases the probability that firms choose that country as a host location.

| ECONOMETRIC APPROACH
We examine the impact of TCRs on MNCs' location decisions using a discrete location choice model, where each choice yields a potential (latent) payoff. Suppose a firm i is concerned with choosing one of J potential locations (countries) to set up its first foreign affiliate. Each of the j = 1, …, J locations is associated with a latent profit * ij and the actual choice of a location C i ∈ {1, 2, … , J} is based on the maximum attainable profit, argmax( * i1 , * i2 , … , * iJ ). We postulate potential profits to depend on observable and unobservable firm and country characteristics as follows: where Θ ij is the safe-haven threshold in country j as defined in Section 3, ij is the statutory corporate tax rate in country j, x ij is a 1 × K vector of country-and country-firm-specific characteristics, and ij is a disturbance term. Note that variables in (2) do not bear a time index t, although we measure all variables in the year of each firm's first location choice. To highlight that firms may face different safe-haven thresholds and statutory tax rates in the same country (because firm's first location choices occur in different years), we index the variables Θ and τ by ij. In other words, a firm's choice is about location in a given year, which makes a country at stake something quasi firm-specific. The parameters γ and those in the vector β are fixed population parameters to be estimated. The parameter on the corporate tax rate i is indexed by i as it is defined as a firm-specific random coefficient and assumed to be normally distributed with parameters a and σ, which are to be estimated. Assuming i ∼ N(a, 2 ) and ij ∼ iid extreme value yields the mixed (or random parameters) logit model. 7 Specifying the coefficient i on the corporate tax rate as random directly relates to the expectation of a large heterogeneity across firms in tax avoidance activities (depending on firm characteristics, products sold, access to finance, etc.), which suggests heterogeneity in tax elasticities.
Alternatively, it is useful to think of i ij as error components, which, together with ij , represent the stochastic part of * ij . This stochastic part ij = i ij + ij is allowed to be correlated across alternatives. Under the assumption of a zero error component, the unobserved proportion of profits for one alternative is not correlated with the unobserved proportion of profits for another alternative. 8 By allowing for correlation in profits over alternatives m and n, we have Train, 2009).
One of the central issues about (2) is specifying the variables that induce correlation among alternatives. One way to proceed is to think about the different determinants of location choice and why they might induce such correlation. It seems natural to consider the tax rate as a variable that causes such correlation as differences in taxes and tax policy across countries induce unobservable tax avoidance activities affecting * ij through different forms of ij-specific tax planning or income shifting. Another interpretation in view of the theoretical tax competition literature is that tax policy is used by one country to attract mobile capital at the expense of other countries. 9

| DATA
To test whether TCRs, affect MNCs' location choices we make use of the German firm-level censustype dataset MiDi (Microdataset Directinvestment) provided by Deutsche Bundesbank. This annual dataset comprises information on direct investment stocks of German enterprises held abroad. Data collection is enforced by German law, which determines reporting mandates for international transactions if investments exceed a balance-sheet threshold of €3 million. 10 MiDi is particularly well suited for exploring the determinants of corporate location choices, as we observe all (directly and indirectly held) new entities established by German firms in foreign countries over a 11-year period between 2002 and 2012.
For the empirical analysis, we restrict our attention to the location choice of the first foreign affiliate. For each firm in the dataset, we observe the country of location of their first foreign affiliate and the year in which it is set up. In the location choice model the firm's choice set consists of all J countries in which we observe first locations. The dependent variable indicating each firm's choice is a binary variable c ij defined for all firm-i and country-j combinations. c ij equals one if firm i locates its first foreign affiliate in country j, that is, c ij = 1, and zero otherwise (i.e., for all other possible J−1 locations). Since firms establish their first foreign affiliate in different years, the choice set of each firm corresponds to | 41 MERLO Et aL. the given set of countries, and the respective characteristics of those countries in the year of the choice. The country-and firm-specific characteristics that determine the choice are correspondingly dated. In our data, 3,574 German MNCs locate their first foreign entity in one of 80 countries in the period between 2002 and 2012. 11 Many of the foreign entities are established in neighboring countries to Germany such as France (283 entities), Austria (263 entities), Poland (248 entities) or Switzerland (196 entities). Other European countries such as the United Kingdom are important as well (216 entities). However, the most important host country in terms of number of new establishments is the United States, where 458 new entities were established between 2002 and 2012. We also count a substantial number of new investments in emerging markets such as China and Russia (177 and 108, respectively).
As outlined above, location choice is determined by all variables that determine * ij . Beside tax determinants, our empirical analysis uses a very rich set of control variables that have been identified in previous studies as determinants of corporate location decisions. 12 Our explanatory variables of interest are a country's safe-haven threshold, SHT (Θ in Equation 2), and statutory corporate tax rate, TAX (τ). Additionally, we include the following variables. The log of a country's GDP, log(GDP), is included to capture local market size and demand conditions. Ceteris paribus, we expect that the location choice probability is positively related to this variable. Moreover, we include the log of GDP per capita, log(GDPPC), as a proxy for a country's labor productivity. As far as log(GDPPC) is positively related to purchasing power and the foreign entity is part of a horizontal FDI strategy, we would expect a positive impact of this variable. If, instead, the foreign entity is part of a vertically integrated firm and the MNC produces intermediate goods in low wage countries, a higher GDP per capita may be associated with higher average wages, which may lead to a lower probability to choose a location. Gross domestic product growth in country j, GDP growth, may be considered as a general measure for the economic attractiveness of a location. We furthermore include the variable DCPS to measure domestic credits provided to the private sector in a country relative to a country's GDP. We expect that DCPS is positively correlated with the quality of a country's financial market. Thus, higher values of DCPS are expected to make host countries more attractive. In addition, we include the log capital-labor ratio of host country j, KLRATIO. This variable should reflect relative factor endowments of countries. To capture fixed investment cost we include COSTBS, which measures costs of business start-up procedures (in percent of GNI per capita) in a potential host country. The cost of starting a business is clearly an entry cost factor for MNCs (irrespective of whether FDI is vertical or horizontal), so its impact is expected to be negative.
Another relevant country characteristic is market j's inflation rate, INFLR. The variables CORRUPTION (freedom from corruption) and PRIGHTS (property rights) measure institutional quality. They can take values between 0 and 100, higher values referring to less corruption and better property rights in a host country. As foreign locations are more attractive for MNCs if they are more integrated in terms of bilateral investment treaties (BITs) and double taxation treaties (DTTs), we condition on the existing treaty network of host countries by including BIT and DTT. BIT refers to the aggregate number of BITs, and DTT refers to the aggregate number of DTTs concluded by host country j with all other countries.
Using information from MiDi, we calculate the variable log(TASSETS) as the sum of total assets of German MNCs in country j in the year before a new investment is established. The idea is to include a variable that measures the general attractiveness of foreign markets for German investors. Note that this variable refers to the aggregate of German FDI in the period before firm i enters a market, but all other explanatory variables are measured in the years a new foreign entity is set up.
Our analysis also accounts for control variables that reflect distance between host locations and the parent country Germany. On the one hand, these measures relate to geographical distance: log(DISTANCE) is the log of the distance (in kilometers) between the most populated cities between Germany and a host country; CONTIG is an indicator variable that equals one if Germany and a potential host country share a common border, and zero otherwise. On the other hand, we include measures that relate to cultural closeness: COLONY is equal to one if the potential host country is a former colony of Germany, and zero otherwise; COMLANG is equal to one if Germany and the foreign country j share a common language. Mean values, standard deviations, definitions and data sources are summarized in Table 1. Table 2 presents our preferred specification of the location choice model. 13 In addition to the variables listed in the previous section, the specification shown in Table 2 additionally includes interactions of the non-tax (fixed) determinants with the sales-to-total-asset ratio (SATA) of the parent. 14 The estimated mean of TAX is significant at 1% and negative. The estimated standard deviation is significant and suggests that there is quite some heterogeneity in how tax rates affect location choices of MNCs.

| Estimation results
Our central result is the finding of a positive and significant coefficient for SHR. Hence, a laxer TCR (an increase in the safe-haven ratio) leads to a higher probability that a country is chosen as first location. We will provide a quantification and interpretation of this result in the next sections.
The estimated coefficients on the other controls are usually in line with what we expect and can be summarized as follows. First, closer countries (in terms of distance, direct neighborhood, but also in terms of historic ties and language) are chosen with a higher probability than ones farther away. Second, higher FDI by German firms in the period before market entry is positively related to location probabilities. Third, the positive coefficient on DCPS and the negative estimate on SATA × DSPS suggests that, while an underdeveloped financial market deters foreign affiliate location, the effect is less severe for larger MNCs that can arguably rely on an internal capital market. Fourth, we cannot find a statistically significant effect for BIT, DTT, INFLR, and COSTBS.
Tables 3 and 4 present alternative specifications of our location choice model. In Table 3 we test whether the omission of the firm-country interactions makes a big difference for the estimated coefficients of TAX and SHT. The results show that the estimates are very similar when compared with the specifications using the additional interactions. In Table 4 we also define the safe-haven ratio as random. However, the estimates suggest that there is no additional heterogeneity in the responses of firms as the standard deviation of SHT is insignificant. Conditional on TAX, this seems very plausible as the differences in taxes across countries, rather than cross-country variation in SHT per se, induce firms to optimize over intra-firm trade or financing. Taken all results together, it appears that the coefficients on SHT are precisely estimated as comparing it across different specifications shows that it hardly differs: 0.437 in Table 2, 0.433 in Table 3, and 0.430 in Table 4.

| Estimated location probabilities
Given the estimated coefficients of our preferred specification (Table 3) we calculate the probability of a firm choosing a given country to locate its first foreign affiliate. The mixed logit model probability of firm i choosing location j is where is the probability conditional on the unobserved firm-specific parameter i . The unconditional probability P ij is obtained integrating L ij ( i ) over all possible values of i . 15 Table 5 reports the estimated base location probabilities for the 80 countries included in our sample. These estimates vary from 0.126 for the United States to values close to zero for Guyana, Jordan, Nicaragua, or Qatar. Note that these base probabilities are important not only when calculating elasticities but also when expressing our findings in terms of number of new affiliates below.

| Own-and cross-SHT-and TAX-elasticities
The mixed logit model allows the calculation of interesting substitution patterns, that is, the own-and cross-country effect of a change in the safe-haven threshold of any given country on the location probabilities. The percentage change in the probability for alternative ℓ given the percentage change in Θ of jurisdiction j is given by where the change in the probability depends on the correlation between L i ( ) and L ij ( ) over different values of α.
Tables 6 and 7 present own-and cross-elasticities for a selected number of countries. In these tables, the entries on the main diagonal refer to the estimated own-elasticities (in boldface). For example, a 1% higher SHT (a 1% laxer safe-haven threshold Θ) in Brazil increases the probability to choose Brazil as a location to set up the first affiliate by 0.4238%. A 1% more lenient SHT in Ireland is associated with a somewhat lower elasticity of 0.2142. The entries off the main diagonal refer to cross-elasticities of a 1% change in the SHT of a country in a column on a country in a row. Table 6 shows that these cross-elasticities are not only estimated to be heterogeneous across countries changing their SHTs (across columns) but also across countries facing externalities exerted by (4)   other countries (in rows). For example, a 1% more lenient SHT in the United States leads to large negative responses in Argentina, Canada, Japan, and Norway. We estimate the smallest (the least negative) elasticity for Russia. The differences in estimated cross-elasticities may reflect differences or similarities in factor endowments or closeness in terms of language, culture, or distance (for Canada). It is also interesting to notice that there is no clear regularity with respect to how countries are recipients of shocks. For example, for a given country (in a given row), whether or not the impact on this country is big or not (compare columns for a given line), is highly dependent on which country is changing its policy. Table 7 presents own-and cross-elasticities for changes in the tax variable. 16 On average, we find larger elasticities compared with changes in the SHT. For example, a 1% lower tax in Canada would lead to a 0.7448% higher probability of locating a new entity there. The cross-tax-elasticities are also larger and highly heterogeneous. It is interesting to interpret these estimates in the light of the SHT  elasticities. For example, we find that a change in the tax in the United States leads to a huge impact on the probability of locating in Ireland (a cross-elasticity of −0.1317), while the estimated SHT-crosselasticity was rather modest. The reason for this finding may be that the tax burden of foreign affiliates in Ireland is not very high, so restrictions on debt financing do not bite. But when other countries benefit from cutting taxes, this comes at the expense of Ireland whose attractiveness as a low-tax country is relatively reduced. This is confirmed when focusing on the row IRL and comparing cross-responses across columns: the negative effect on Ireland is usually one of the largest. We can finally interpret Tables 6 and 7 in light of the theoretical literature. Tax competition models with strategic interaction usually predict that increasing its own tax rate leads to an outflow of capital. A higher safe-haven ratio (a more lenient TCR) would imply an inflow of capital. In this sense, higher taxes exert positive externalities on other countries, while a higher safe-haven ratio exerts a negative externality on other countries. Hence, on average, taxes are too low and TCRs are too lax as countries do not consider these externalities.

| Policy options for the United States
In this section we take a closer look at the policy options of a single country. In particular, we will focus on the United States as it is the most important country in terms of number of new entities in our data. Figure 1 presents estimated probabilities (the vertical axis) and how these depend on the two policy variables we are interested in. Although we know from Tables 6 and 7 that tax elasticities are somewhat larger compared with safe-haven elasticities, it is not clear what this means for a given parameter space and actual policy options. However, it becomes clear in Figure 1. A tax cut would have a massive impact on the location choice probability. The difference in location probabilities between a tax of 40% and a zero tax for a given SHT of 0.5 is more than 0.15. 17 Compared with this, given a tax of 42%, abolishing the TCR would increase the probability of choosing the United States only by −0.024. To see that the impact in terms of real number of foreign affiliates is not that small, suppose the United States abolished its TCR (a discrete jump in Θ from 0.5 to 1). Using the average number of first location decisions per year observed in our data (about 320) and the U.S. specific impact of its TCR, this would imply that the United States attracted about eight additional affiliates of German multinationals, ceteris paribus.
Another interesting experiment examines how the United States would affect other countries by abolishing its TCR completely. For this, we set Θ equal to 1 for the United States. The implications for the 79 other countries included in our dataset are presented in Table 8. Note that countries are sorted in alphabetical order according to their country codes. The estimates suggest that this policy comes mainly at the cost of France, the United Kingdom, and Poland.

| Uncoordinated tax rate and tax-base policy
Over the last 30 years, corporate tax laws in many countries have seen tax-cutting and base-broadening reforms. Devereux, Griffith, and Klemm (2002) show that these reforms had the effect that, on average, effective tax rates remained relatively stable. Concluding from this that the reforms did not change the attractiveness of a location for real investment assumes that the marginal impact of tax and tax-base effects are of similar magnitude. In Table 9 we demonstrate that this is not necessarily the case. The table presents some calculations on the tax rate cut that would be necessary in order to keep the location probability constant if the tax base was broadened by implementing a 10 percentage point  stricter SHT. For the selection of countries from above, the numbers in Table 9 represent percentage point reductions in the tax rate. For example, Singapore would need to cut its tax by 1.44 percentage points if it reduced its SHT by 10 percentage points in order to hold the number of new entities constant. Hence, the table provides information about the relative importance of tax base vs. tax rate effects. It demonstrates that Ireland could easily make its TCR stricter without a large need to cut its tax rate. In contrast, countries like Japan, Spain, or the United States would need to cut taxes by more than 2 percentage points in order to keep the number of new foreign affiliates (additional inward FDI at the extensive margin) constant.

| Coordinated policy action
Our empirical approach also allows us to determine winners and losers of a coordinated policy experiment. Suppose all countries took a coordinated action and set Θ equal to 0.5. This would imply that interest deduction for any amount of debt exceeding equity financing would be denied. A value of Θ = 0.5 refers to the strictest rules we have in our data, but a number of countries use rules that are nearly as strict.
The results of this experiment are summarized in Figure 2. The blue color in this figure denotes losers and orange denotes winners of the coordinated policy. Among the biggest losers are countries such as Austria, Belgium, Switzerland, or Ireland. The loss in probability mass is, however, rather modest. For example, the probability that Austria attracts a new affiliate is reduced from 0.0554 to 0.0503. The impact on the other countries is even smaller. Belgium faces a reduction of −0.0040, Switzerland a reduction of −0.0019, and Ireland a reduction equal to −0.0007 in their estimated probabilities to T A B L E 9 Tax

| Industry-specific growth effects
We may be concerned about industry-specific growth effects, which may lead to biased estimates on SHT. Table 10, where we add such effects to the estimated model, shows that our results remain fully robust as the estimated TCR effect is hardly affected. In particular, to account for industryspecific growth effects, we build the variable GTH as average growth of foreign affiliates' total assets per industry and year. Table 10 includes 16 additional interaction terms between the country-specific variables and the variable GTH. 18 For the latter variable, we first calculate total asset growth at the level of foreign affiliates. We then take the average of this growth variable per industry and year. GTA is finally defined as the one-period lagged value of this industry-year-specific growth variable. 19

| Subsequent investments
So far, our empirical analysis has focused on first investments of MNCs observed in our data. We believe that this produces the most reliable results as we avoid measurement problems related to more complex sequential investment patterns. A concern with this approach might be, however, that the relevance of TCRs could increase in the extent of foreign activity (in the number of foreign investments). TCRs are, of course, relevant for all entities as these rules apply to all subsidiaries of MNCs if internal or total debt exceeds certain threshold levels (so they should be relevant for first investments as well). Table 11 presents results on second investment decisions. It additionally includes the growth variables from above as well as the binary indicator LCHOICE. The latter is an alternative-specific variable equal to one if a country has been chosen as first location by the MNC. If a country has not been the actual choice in the previous decision, LCHOICE equals zero. The results on the second location choice are very convincing as (i) we estimate a positive and significant impact of SHT(Θ) (with a larger coefficient), (ii) LCHOICE = 1 makes it more likely that the same country is chosen, (iii) the effect of the tax rate is negative, but the heterogeneity of this effect seems to have vanished.

| CONCLUSIONS
The purpose of this paper is to assess the impact of TCRs on the location of multinational firms' foreign affiliates. Using unique data on the worldwide activities and particularly on the first new foreign affiliates of German MNCs, we find that TCRs have a significant impact on location decisions of MNCs. 20 Although the impact of TCRs is statistically as well as economically relevant, we can show that location choices are more sensitive to tax rate changes. To the best of our knowledge, our paper is not only the first one to examine the impact of TCRs on the extensive margin of foreign investment activity, it is also the first to provide actual estimates for the relative importance of tax-rate and taxbase effects in this context. We believe that this is a central contribution to the corporate tax literature, as finding out about the quantitative (and relative) effectiveness of policy instruments is crucial for the design of tax policy.
Our results imply that policymakers should be aware of two things. First, imposing restrictions on profit shifting has implications for real investment activity: unilateral measures to "limit base erosion via interest deductions and other financial payments" (OECD, 2013b, Action 4, p. 17) certainly come at the cost of losing real investments. Second, policymakers should focus on organizing coordinated policy action when imposing TCRs. Our analysis suggests that this is welfare improving.

MERLO Et aL.
reported whenever foreign affiliates held more than 50% or more of the shares or voting rights in other foreign enterprises with a balance-sheet total of more than €3 million. The reporting requirements are set by the Foreign Trade and Payments Regulation. For details and a documentation of MiDi, see Lipponer (2009). 11 In the location choice model, each of the 3,574 firms faces 80 potential locations, which gives a total number of observations of 3,574 × 80 = 285,920. As a result of missing values in some country-level explanatory variables for some country-year combinations, our estimation sample has 264,959 observations. 12 Note that most of the following variables are country-j-specific and are allowed to vary in time t. However, as mentioned above, we model location choice as a choice from alternatives at a given t and suppress t and j indices for the sake of simplicity. 13 We have tested a number of different specifications, including ones that define SHR (Θ) as a random variable. Some of the additional robustness estimates are presented below. We have also estimated conditional logit models (under the unfavorable IIA assumption). The results are very robust to this. However, a conditional logit does not allow for calculating meaningful substitution elasticities. 14 Note that the explanatory variables in a mixed logit model need to exhibit variation across alternatives. The way to introduce firm-specific variation is to interact firm-level variables with the alternative-specific (i.e., country-level) variables. 15 The integral in Equation 3 does not have a closed form and has to be approximated through simulation by drawing values of i from a normal distribution with mean and standard deviation as estimated in Table 2 (see Train, 2009). 16 We are only aware of one previous paper that reports cross-tax elasticities. In a recent contribution, Griffith, Miller, and O'Connell (2014) calculate own-and cross-elasticities with respect to variations in corporate tax rates for a sample of 14 countries. Our estimates seem to be on average a little larger, but often relatively similar (for example, for Norway we find an elasticity of 0.7369; the elasticity estimated by Griffith et al., 2014, equals 0.783). 17 Of course, a tax rate of zero is a relatively unrealistic scenario.