Journal of Public Economics 87 (2003) 22252252 www.elsevier.

com / locate / econbase

Why pay more? Corporate tax avoidance through transfer pricing in OECD countries
Eric J. Bartelsman a , Roel M.W.J. Beetsma b , *
a

ESI, Department of Economics, Free University of Amsterdam, De Boelelaan 1005, 1081 HV Amsterdam, The Netherlands b Department of Economics and CEPR, University of Amsterdam, Roetersstraat 11, 1018 WB Amsterdam, The Netherlands

Received 12 July 2000; received in revised form 19 December 2001; accepted 21 December 2001

Abstract This paper presents suggestive evidence of income shifting in response to differences in corporate tax rates for a large selection of OECD countries. We use a new method to disentangle the income shifting effects from the effects of tax rates on real activity. Our baseline estimates suggest that a substantial share of the revenues from a unilateral increase in the corporate tax rate is lost because of a decline in reported income. 2003 Elsevier B.V. All rights reserved.
Keywords: Income shifting; Transfer pricing; Corporate tax rates; STAN database JEL classication: F2; H2

1. Introduction Policymakers are increasingly worried about cross-country differences in tax rates. In particular, they fear losing real economic activity to other countries if tax rates are too high. Many experts, therefore, predict a race to the bottom for taxes on mobile production factors, in particular corporate taxes. Indeed, a recent report by the OECD (2000) tries to identify harmful tax practices and to provide a
* Corresponding author. Tel.: 1 31-20-525-5280; fax: 1 31-20-525-4254. E-mail address: beetsma@fee.uva.nl (R.M.W.J. Beetsma). 0047-2727 / 03 / $ see front matter 2003 Elsevier B.V. All rights reserved. doi:10.1016/S0047-2727(02)00018-X

2226 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

platform for the possible cooperation on tax policies with cross-border effects. However, cross-country differences in corporate tax rates may not only induce real activity shifts, they may also lead to pure accounting income shifts between countries. Such income shifts have received much less prominence in the policy debates. Moreover, if policymakers reveal worries about income shifting, then it is about income shifting from the main industrialized economies to the so-called tax havens, small countries with very low corporate tax rates. Indeed, most the empirical literature has been concerned with income shifting from the US to such tax paradises (e.g. see Hines and Rice, 1994 and Grubert and Slemrod, 1998). The purpose of this paper is three-fold. First, we provide empirical results suggesting that the lack of attention for income shifting among OECD countries and other major economies may be unjustied. That income shifting between these countries could be important should not be surprising given the amount of intra-industry trade among these countries (e.g. Stone and Lee, 1995). Moreover, the relevance of the issue likely will increase because of their ongoing economic integration. Indeed, the ensuing empirical analysis indicates an amount of income shifting that is both statistically and economically signicant across a variety of different empirical specications. Back-of-the-envelope calculations based on our baseline estimates suggest that at the margin more than 65% of the additional revenue from a unilateral tax increase is lost due to a decrease in the reported income tax base. Yet, this gure should only be taken as a rst indication that income shifting among OECD countries may be large.1 We explore a large number of variations and extensions of the baseline empirical framework, generally conrming our baseline ndings. Secondly, we develop a novel method for isolating the pure effects of income shifting, by controlling for the effects of taxes and unobserved productivity on the scale of real economic activity. In essence, we do this by regressing the ratio of total value added to wage payments on tax rate differences. Finally, whereas previous work has either been concerned with individual rm data or country level data (involving the US and a set of tax havens; e.g. Hines and Rice, 1994), this paper takes an intermediate position and considers evidence for disaggregated industrial sectors for a group of OECD countries. The remainder of the paper is structured as follows. Section 2 discusses the channels and incentives for income shifting. Section 3 motivates the empirical model, while Section 4 describes the data. Section 5 presents the baseline empirical results. In Section 6 we investigate the robustness and a number of extensions of the baseline setup. Finally, Section 7 concludes the main body of the paper.
Based on individual import and export transactions, Pak and Zdanowicz (2001) estimate income shifting (through transfer pricing) out of the US for 2000 at $131 billion or about 18% of total reported corporate prots. The top three countries for shifting income to are Canada, Japan and Mexico. The top 25 contains no tax havens.
1

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2227

2. Channels and incentives for income shifting There are various channels through which at least part of the additional revenue from an increase in the corporate tax rate may leak away. First, the capital stock may shrink. For given (sectoral) capitallabor ratios, rms can shift real activity to more labor-intensive sectors and / or to the same sector in other countries.2 Second, for a given activity level within a sector, they may substitute labor for capital. Finally, even when both total real economic activity and the capitallabor ratio in a sector (e.g. the manufacturing sector) are held constant, there are several ways in which corporate income can be suppressed. Firms might want to alter the composition of the compensation for executives and other employees. In particular, an increase in the corporate tax rate relative to the personal income tax rate would lead to an increase in the wage bill, at the expense of the corporate prot. Gordon and Slemrod (1998)s evidence for the US indicates that these shifts are substantial. An increase in the corporate tax rate also reduces (raises) the relative attractiveness of being incorporated for protable (loss-making) rms. The empirical results of Mackie-Mason and Gordon (1997) for the US support this hypothesis. Finally, an increase in the corporate tax rate may lead to corporate income shifting to other countries. In this paper, we study this latter channel, while trying to control as well as possible for other potential channels. There are two major ways in which multinationals can shift income from high-tax countries to low-tax countries. First, the nancing structure of afliates matters for the tax burden of a multinational. In particular, it is relatively more attractive to nance afliates in high-tax countries by debt than by equity, with loans extended from the parent company or afliates in other countries. The second main channel for cross-border income shifting, and the one that will be the focus of our empirical analysis, concerns the prices that are used for intra-rm international trade in goods and services. Although the OECD Model Tax Convention and the OECD Transfer Pricing guidelines recommend the use of the arms length principle,3 strict application of the principle is often problematic in practice. For example, for many intra-rm transactions there exists no comparable market outside the rm. This is in particular the case for intellectual property or knowledge intensive intermediate goods that are developed or produced by one part of the multinational company and used by other parts of the company in other countries. The empirical literature suggests that both channels are important vehicles for income shifting. However, most of the empirical evidence concerns the US and,
2 Hines (1996a), Devereux and Grifth (1998) and Gorter and Parikh (2000), for example, investigate how differences in corporate tax rates affect the allocation of real economic activity between jurisdictions. 3 The arms length principle requires prices for transactions between different parts of the same multinational to be set at same level as prices for similar transactions between unrelated parties.

2228 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

more specically, income shifting from the US to tax havens. In addition, the evidence is based on either country-level data or rm-level data. Excellent surveys of the existing empirical work on income shifting can be found in Hines (1996b, 1999) and Newlon (1997). Hines and Hubbard (1990) provide evidence on the effect of the nancing structure. They nd that the average tax rate for foreign afliates that make interest payments to their US parent is higher than for afliates that do not make interest payments. The US Tax Reform Act of 1986 has introduced limits on the extent to which interest payments can be deducted from the income of foreign afliates. Collins and Shackelford (1992) and Froot and Hines (1995) report individual-rm evidence that is consistent with the changed tax incentives provided by the Tax Reform Act of 1986. Other empirical work focusses on income shifting as a result of transfer pricing alone or in combination with the effect of the choice of the nancing structure. For example, Grubert and Mutti (1991) investigate the sensitivity of total foreign earnings of U.S. afliates to corporate tax rates for a cross-section of other countries. In a similar vein, Hines and Rice (1994) explore the relation between the reported protability of foreign afliates of US-controlled rms and hostcountry tax rates, while controlling for capital and labor inputs. Clausing (1998) nds that intrarm trade balances of US parents with their foreign afliates improve when the foreign effective tax rate increases, which is consistent with the overpricing (underpricing) of goods or services sold to afliates in high-tax (low-tax) countries. Several studies based on rm-level data also provide evidence that is (in principle) consistent with income shifting. Harris (1993) nd that the presence of afliates in low-tax countries is associated with lower tax liabilities in the US. Grubert (1998) detects a negative relationship between reported subsidiary income and the statutory corporate tax rate in the host country. He also explores the substitutability between various income streams (royalties, interest and dividends) to parent companies as a result of changes in their relative tax prices. Finally, Grubert and Slemrod (1998) focus on income shifting of US multinationals to Puerto Rico. Because of the so-called possessions tax credit, income generated in Puerto Rico is in effect completely exempt from US corporate taxation, while taxation by the Puerto Rican authorities is virtually negligible. One of the main features of the analysis is that it treats income shifting and real investment as joint decisions, because more real investment facilitates income shifting. The incentives for income shifting depend in the rst place on the differences in corporate tax rates between countries and the system that residence countries use to avoid double taxation. Multinationals from countries with an exemption system (for example, France and the Netherlands) are exempt from domestic tax payments on income that has already been taxed abroad. These multinationals have an incentive to shift income to the country with the lowest corporate tax rate. For each dollar of income shifting, the gain is the absolute value of $st * 2 td, where t and t * are the domestic and the foreign corporate tax rates, respectively. Under

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2229

the credit and deferral system (for example, Japan, the UK and the US), multinationals receive a credit from their home country for taxes paid abroad. The credit can be subtracted from the tax liability at home. If the credit exceeds the domestic tax liability, the company has an excess foreign tax credit, while it has a decit foreign tax credit when the foreign tax credit is smaller than the domestic tax liability. Usually, domestic taxes on foreign income can be deferred until the income is remitted in the form of dividends. Therefore, when the domestic tax rate exceeds the foreign tax rate, the benet from income shifting arises from the deferral of the domestic tax payment. However, if the foreign tax rate is the higher of the two, shifting an additional dollar of income from the foreign country to the residence country yields a gain equal to $st * 2 td.4,5 In the Supplementary Appendix, Section 1 (available upon request) we present a simple formal model that yields the result that, if (i) the cost of income shifting (for example, in the form of an expected ne) is non-negative and convex in the amount of income shifting and (ii) countries use the exemption system or multinationals have excess foreign tax credits, then an increase in the domestic corporate tax rate leads to a fall in the reported income, while the opposite happens with an increase in the foreign corporate tax rate. Unfortunately, because of the limited time horizon, the model does not allow for the possibility of the deferral of dividend remittances in the case of a decit foreign tax credit, so as to reduce the present value of the tax payments. Other features of a countrys tax system may also play a role. Countries generally impose withholding taxes when dividends are remitted. Some countries (for example, Germany) use a split rate system in which distributed earnings are subject to lower corporate tax rates than retained earnings. There are also countries that integrate corporate and personal taxes. For example, Germany, France, Finland and the United Kingdom employ an imputation system which provides a personal tax credit for tax payments at the corporate level. The credit is usually extended only for domestic corporate tax payments and not for corporate taxes paid by foreign afliates. Nevertheless, the model in the Supplementary Appendix, Section 1 (available upon request) illustrates that in all these cases a relative increase in the corporate tax rate leads to a shift in reported income to other

There are situations where the gain is different. For example, in the case of a royalty which is classied as foreign source by the tax rules of the residence country, the gain is $t * (e.g. Newlon, 1997). The royalty absorbs excess foreign tax credits, without changing domestic tax payments. The specic rules for the treatment of the various items may differ across countries, so that it is impossible to identify all the cases where the gain differs from $st * 2 td. Nevertheless, generally the incentive is to shift income from the higher- to the lower-tax country. 5 Whether a shift in income is the result of an excess foreign tax credit or of a tax-rate difference for a multinational residing in an exemption country does not make a difference empirically. Any shift in income shows up as an increase in the value-added statistics of one country at the expense of value-added in another country, an effect that should be picked up in our regressions (see below).

2230 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 Table 1 Formal enforcement of transfer pricing rules by country Country Australia Austria Belgium Canada Denmark Finland France Germany Italy Japan Netherlands Portugal Spain Sweden UK US Explicit TP rules 07 / 83 07 / 99 01 / 99 01 / 31 09 / 85 02 / 83 12 / 86 04 / 86 01 / 96 07 / 99 01 / 28 Formal TP documentation rules 09 / 95 07 / 99 01 / 99 01 / 99 04 / 96 07 / 99 01 / 94 TP specic penalties 07 / 83 01 / 99 04 / 96 07 / 99 01 / 94

Notes: (1) Source: Ernst and Young (2000); (2) TP stands for transfer pricing; (3) Numbers indicate month and year in which this practice was introduced. Potential changes in formal enforcement mechanisms that have occurred after 1999 or are planned for the near future have not been included in the table.

countries. The basic intuition is that any relative change in corporate tax rates alters the costbenet trade-off associated with income shifting. A nal set of factors possibly inuencing the amount of income shifting is the enforcement of transfer pricing rules by the tax authorities. In principle, all the countries in our sample follow the standard OECD guidelines for transfer pricing and apply the arms length principle. However, countries may differ in the degree of enforcement of these guidelines. Table 1 summarizes the information about enforcement for the countries involved in the ensuing empirical analysis (Ernst and Young, 2000). Most countries have explicit transfer pricing rules (an exception is the Netherlands), while a smaller group of countries uses formal transfer pricing documentation rules and yet an even smaller set of countries imposes transfer pricing specic penalties for the violation of transfer pricing rules.6
6 Ernst and Young (2000) denes the enforcement indicators as follows: Explicit transfer pricing rules: transfer pricing regulatory provisions exist, Formal transfer pricing documentation rules: governing tax authority requires or recommends that taxpayers prepare and maintain written documentation to conrm that the amounts charged in related party transactions are consistent with the arms length standard and Transfer pricing specic penalties: tax authority will impose a transfer pricing specic penalty if the tax payer is found not to be in compliance with the transfer pricing rules imposed by that country.

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2231

3. Empirical motivation Income shifting leads to differences between reported income and the true income generated by real economic activity. Revenue from production is underreported (overreported) in countries with relatively high (low) tax rates, because rms claim lower- (higher-) than-market prices for intra-rm international shipments of their products or the provision of services (possibly by intangibles). Inversely, intermediates purchases may be underreported (overreported) by internationally operating rms in countries with relatively low (high) tax rates. Consequently, reported nominal value-added (NVA 5 revenue less intermediate purchases) in a country is negatively affected by the level of its corporate tax rate relative to that of countries with which intra-rm trade takes place. Income shifting causes statistics on nominal value-added collected by the tax authorities or statistical ofces to be measured with error: NVA 5 P tr Q * where Q * is actual real value-added and P tr is the implicit (and unobserved) price determined by this expression. It differs by some factor from the market price, or arms-length price (P * ): e P tr 5 P * est Dd, where ]] ,0 t D (3.1)

where t D is the difference between the tax rate in the residence country and the country with which intra-rm trade takes place. The statistical ofce uses proper price quotes on market transactions, so that price statistics and actual market, or arms-length, prices coincide: P 5 P * , where P is the product or materials price measured by the statistical ofce. Regardless of the method used to deate value added, prot shifting caused by higher taxes will reduce the implicit price P tr relative to the ofcial deator P. Real value-added statistics ( Q ) are thus contaminated with mismeasurement from income shifting, and, therefore, differ from actual real value-added ( Q * ) by the factor e(t D ): NVA Q 5 ]] 5 Q * est Dd P In order to estimate the magnitude and signicance of income shifting, one would need to determine actual real value-added, Q * , and nd out whether it differs systematically with tax rates from reported real value-added. Unfortunately, two practical problems stand in the way. First, actual real value-added is not observed. Next, actual real value-added may itself be affected by tax rates, because international rms make greeneld and expansion investment decisions based on average and marginal effective tax rates, respectively. The following derivation shows how we control for this in order to empirically identify e(t D ) from observed data.

2232 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

The basic insight for the estimation strategy comes from viewing the transferpricing problem as an errors-in-variables problem in estimating a production function. The identication issue is to disentangle the mismeasurement component and the productivity residual component. Analogous to the approach of Roeger (1995) for disentangling mark-ups and productivity movements, we take a ratio of nominal output and nominal input expenses, in order to cancel out the unobserved productivity shock. We start with a standard production function, and the rst-order condition for prot maximization with respect to the labor inputs. Actual production is given by: Q * 5 AF(K,L ) where K is capital and L is labor input, and A is unobserved (by the econometrician) productivity. We assume that the multinational hires labor until the nominal wage, W, equals the marginal revenue product, valued at market prices: Q * W 5 P * ]] L With a CobbDouglas technology and constant returns to scale, in symbols F(K,L ) 5 K 1 2 a L a , we get the following relationship between the observed (reciprocal of the) labor share and the factor e : PQ 1 ] 5 ] est Dd WL a (3.2)

where a is the output elasticity of labor from the CobbDouglas production function. Note that the unobserved productivity term A cancels out. We refer to the left-hand side of (3.2) as the valuelabor ratio. The advantage of using (3.2) in the empirical analysis is that we control for the effect that tax rates may have on the scale of operations, as measured by the amount of inputs in the production process.7 Under CobbDouglas, the scale of operations and the wagerental ratio will not affect the labor share of income. However, with a constant elasticity of substitution (CES) production technology, in symbols F(K,L ) 5fd K r 1s1 2 dd L rg 1 / r , an increase in the cost of capital as a result of a tax rate change affects the labor share of income: PQ r ] 5 1 1 z 1 /sr 2 1d ] WL w

SD

r / s r 2 1d

G est d
D

(3.3)

denotes the tax-dependent where z ;fd /s1 2 ddg depends on the CES parameters, r JorgensonHall user cost of capital and w is the real wage rate. Because the

Appendix A presents an example of a more rigorous microfoundation for relationship (3.2).

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2233

can functional relation between taxes and the user cost of capital is well known, r be constructed from the available data, and only the CES parameters and the function est Dd need to be estimated.

4. The data We obtain our data from various sources. All the data are available upon request from the authors. Data on labor compensation and value-added are taken from the OECDs Structural Analysis Database (STAN).8 The STAN is a sectoral database constructed at the OECD and is consistent with the national accounts of the individual countries. It contains information for 22 OECD countries, but does not include Ireland and Switzerland. In our empirical analysis, we include the maximum set of countries (16) for which a complete set of disaggregated sectors are available from the STAN. We use 15 sectors, which together make up the entire manufacturing industry. The countries and the sectors are listed in Appendix B. The maximum time series length covered by the STAN is 19701997. However, due to the limited time series length of the corporate tax rate series (see below), we include only the period 19791997. Given that for some countries the corporate tax rate is missing at the start of the period and that a number of observations is missing in the STAN at the beginning or the end of the period for certain sectors in certain countries, the total number of observations in our empirical analysis is 4100. Data on value-added (the sum of labor and capital compensation) equals sales revenues minus intermediates purchases from outside the sector. These intermediates consist of goods and services, with the former including knowledge intensive components and latter including the use of patents, R&D, etc., owned by other parts of the multinational. Compensation for the services of intellectual property often takes place by means of royalty payments. Changing the amount of cross-border royalty payments thus has the same effect on measured value-added data as changing the prices of cross-border delivered intermediate goods. Headline corporate tax rates (on retained earnings) and statutory (highest marginal) tax rates on personal income are obtained from various issues of PriceWaterhouseCoopers International Tax Summaries. We have two sets of countries for which we compute the average corporate tax rate: the STAN tax set, which is used for the baseline regressions, includes only the countries in the STAN, while the large tax set, which is used for some of the robustness checks, adds an additional 25 countries to the STAN tax set. These additional countries include the OECD countries that are not in the STAN (and for which corporate
8 This dataset is widely used in the analysis of international trade or productivity, e.g. Harrigan (1999). Details about the STAN can be found in OECD Statistics (1999) or via the OECD web-site at http: / / www.oecd.org / dsti / sti / stat-ana /.

2234 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

rates are available for a sufciently long time span), the most important economies outside the OECD, and selected tax havens for which data were available. The tax sample period for these countries generally runs from 1982 until 1997. The large tax set is also listed in Appendix B. Variables used to weigh the tax rates from the large tax set to create an average foreign rate come from the World Bank (2000) Development Indicators. We retrieve the long-term interest rate and the price deator for investment from the Annual Economic Indicators database of the OECD.

5. Baseline econometric specication and estimation results

5.1. Specication
The functional form of the transfer-pricing response to taxes est Dd is assumed to take the form of a mark-up which varies linearly in the deviation between country specic and average tax rates, and which may be country or sector specic. For expositional purposes, we rst present the estimating equation based on (3.2), i.e. the one based on the CobbDouglas production specication:
2i jt Vijt 5 c csy *f1 1 g csstit 2 t dg 1 eijt

(5.1)

where Vijt is the observed valuelabor ratio in country i, sector j and time period t, 2i jt g csstit 2 t mark-up, tit is the headline corporate tax rate in STAN country d is the i 2 i at time t and t jt is the weighted average of the period-t headline corporate tax 2i jt rates of foreign countries (excluding country i ) for sector j.9 Both tit and t are measured in percentage points. We use tax rates on retained earnings when dividends and retained earnings are taxed differently at the corporate level. Furthermore, in different specications the coefcients c csy and g cs are interacted with different combinations of country, sector, and / or time dummies. For example, the constant term can be interacted in all three dimensions to sweep out xed effects as follows: c
csy

5c1

d 52

O c sI
c d

Nc

c d , ijt

2I

c 1, ijt

d1

d 52

O c sI
s d

Ns

s d , ijt

2I

s 1, ijt

d1

d 52

O c sI
y d

Ny

y d , ijt

2Iy 1, ijtd (5.2)

where I is an indicator that equals 1 if observation ijt belongs to country d (i.e. if i 5 d ), and equals 0 otherwise (Nc is the number of countries in our sample);
i 2 In our baseline regression, the average foreign tax rate, t jt , varies slightly across sectors, because it is computed using countries sectoral value-added as a share of total value-added for that sector as relative weights. The consideration was that transfer-pricing takes place between afliates within a sector, owing to taxes differing from those in relevant countries. 9

c d , ijt

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2235

Is d , ijt is an indicator that equals 1 if observation ijt belongs to sector d (i.e. if j 5 d ), y and equals 0 otherwise (Ns is the number of sectors); I d , ijt is an indicator that equals 1 if observation ijt belongs to year d (i.e. if t 5 d ), and equals 0 otherwise (Ny is the number of years). In this specication, the standard xed effects specication undergoes a linear transformation to get a direct estimate of c, or the average constant term of the regression. The other xed effect coefcients are estimated as deviations of the constant for that sector, country, or time period from the constant of the relevant omitted category. Similarly, to let the sensitivity of transfer pricing to tax differences vary across countries or sectors, g cs is allowed to interact with the dummy variables as dened above:

g 5g 1

cs

d 52

O g sI
x d

Nx

x d , ijt

2 Ix 1, ijtd

(5.3)

where g is the average response coefcient, x 5 c if the mark-up varies by country, and x 5 s if it is allowed to vary by sector. There are several reasons why the response coefcient g cs might vary across countries. Firstly, in some countries economic activity may be relatively more concentrated than in other countries in sectors where the scope for income shifting is larger. Secondly, countries may differ in the enforcement of the transfer pricing rules. In particular, holding everything else constant, for countries with lax enforcement we would expect the (absolute) value of g cs to be larger than for countries with tight enforcement. Thirdly, the specics of a countrys tax system may affect g cs . The response g cs may depend on the tax treatment of the foreign prots generated by domestic multinationals (credit or exemptionsee also below), the use of a split rate system for dividend taxation and the presence other taxes, such as dividend withholding taxes and royalty withholding taxes. As mentioned earlier (and formally demonstrated in the Supplementary Appendix available upon request), these complications do not overturn the hypothesis that an increase in the corporate tax rate difference leads to more income shifting. However, they do affect the size of the income shifting response. In our base case regressions we try to pick this up by allowing g cs to vary across countries. The response coefcient g cs may also vary across sectors. A priori, one would expect a relatively large amount of income shifting in sectors dominated by large multinationals and a lot of intra-industry trade. In addition, sectors differ in the composition of cross-border trade within multinationals. Income shifting should be easier if more of the intra-rm trade is in products for which it is difcult to establish a proper arms length price (this would, in particular, be the case for the use of intangibles, such as intellectual property rights, or for knowledge intensive intermediates such as semi-conductor components). The estimating equation used in our baseline regressions is based on Eq. (3.3) and, hence, assumes a CES production function:

2236 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

it r csy Vijt 5 1 1sc d 1 /sr 2 1d ] w ijt

S D Gf
r /s r 2 1 d

11g

cs

2i jt stit 2 t dg 1 eijt

(5.4)

with c csy given by (5.2). With (5.4), we try to control both for the scale at which operations take place (thereby disentangling income shifts from real activity shiftscross border and between sectors within a country) and for the possibility that, for a given scale of operations, corporate tax-rate differences may affect the ~ it r it 1 u 2 q it 5 PI , it ]]] factor shares.10 The user cost of capital is calculated as r s1 2 1 2 tit tZ , itd, where r is the long-term interest rate, PI is the investment deator, u is the ~ is the expected appreciation of capital and tZ is the present depreciation rate, q discounted tax value of depreciation deductions. In the empirical work, u 5 0.08 and tZ is calculated assuming static expectations on future tax rates. Furthermore, we use a constant discount rate (0.06) and geometric depreciation (12 year life). The depreciation rate is assumed to be the same for all observations. Of course, in reality, tax depreciation rules and service lives of capital vary by asset type, sector, and over time, as well as across countries. However, since we have no data on capital stocks and, in particular, no data on the composition of capital by type across observations, we assume a uniform depreciation rate. In principle, other renements are possible in the computation of the user cost of capital. In particular, one might include investment tax credits.11 Unfortunately, data limitations prevent us from making this adjustment as well. However, much of the important variance of the wagerental ratio is captured with available data on wage rates, long-term interest, investment deators, and corporate tax rates. Both the substitution between capital and labor, captured via the rst component in square brackets on the right-hand side of (5.4), and the income shifting effect, captured by the second term in square brackets, depend on the corporate tax rate. In the estimations, both effects can be separately identied because the two components are different functions of the corporate tax rate. In addition, most of the variation in the rst component is driven by variations in the interest rate and wages rather than changes in the corporate tax rate.

If we had data on capital stocks, we could directly estimate the degree of substitution between capital and labor as a result of tax rate differences and, possibly, obtain a more accurate estimate of income shifting. 11 If investment tax credits are not systematically related to movements in corporate tax rates, then the results are unaffected when investment tax credits are neglected. If there is a tendency for investment tax credits to fall when corporate tax rates are lowered (as was the case, for example, for the US and the Netherlands from the middle to the end of the 1980s) then, by neglecting investment tax credits, the effect of a fall in the headline corporate tax rate on the increase in the wagerental ratio and thereby on the increase in Vijt , is overestimated. Therefore, the effect of a tax change on income shifting will be under estimated when investment tax credits are neglected.

10

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2237

5.2. Estimation results

i 2 In the baseline setup, the foreign average tax rate, t jt , is computed using only countries that are available in the STAN data sample. Because Vijt responds to the difference between the own tax rate and the average (excluding i ) STAN country tax rate, our regressions can only pick up income shifting among the STAN 2i jt countries. This is easy to see, if one realizes that an increase in tit 2 t should 2k jt for one or more countries k, so that the fall in Vijt is lead to a decrease in tkt 2 t absorbed by a rise in the Vkjt . In an extension, we compute average foreign tax rates with a far larger sample of countries. This is described in Section 6. Table 2 reports our baseline estimates. We include time, sector and country dummies in (5.2). When g cs is restricted to be xed across countries and sectors, its estimate is negative and highly signicant, as one would expect under income 5 1 /(1 2 shifting.12 The (signicant) estimate of r yields a point estimate of s r ) 5 1.026 for the elasticity of substitution between capital and labor. Hence, the substitutability between capital and labor is barely stronger than for a Cobb Douglas production function.13 The next rows of Table 2 (case B) display the sensitivity of the valuelabor ratio with respect to tax changes when the response

Table 2 Baseline estimatesCES production function Case A: g

cs

c xed 2.70 (0.34) 3.57 (0.46) 2.57 (0.32)

g 2 0.0035 (0.0007) 2 0.0042 (0.0010) 2 0.0033 (0.0007)

r 0.025 (0.0092) 0.041 (0.0093) 0.021 (0.0093)

R2 0.57 0.59 0.58

SSE 0.11 0.11 0.11

B: g cs varies over countries C: g cs varies over sectors

5 estimate of c from (5.2) (average constant), Notes: (1) standard errors are in parentheses, c 5 estimate of g from (5.3) (average response coefcient), with variation in stated dimension, g 5 estimate of r, SSE 5 standard error of regression; (2) Numbers of observations is 4100; (3) Sample r period is 19791997.

Here, and in the following, an estimate is referred to as signicant if it is signicant at the 10% condence level for a two-sided test. 13 There is quite a large, albeit somewhat older, literature on the estimation of CES production functions. The estimates of the elasticity of substitution between capital and labor vary with the data and method (time series or cross-section) under consideration, although they are mostly of the same order of magnitude. For example, the estimated elasticity of substitution in Paraskevopoulos (1979) for US manufacturing industries ranges from 0.5 to 1.5 and in Kemfert (1998) for German manufacturing sectors, it ranges from 0.35 to 0.97.

12

2238 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

is allowed to vary across countries. The estimate of g, dened by (5.3), is the average of the estimated responses over all countries. Similarly, the last rows (case C) report the average of the estimated responses over all sectors. In both is signicantly negative. cases, g The suggested size of the shift in reported income is quite large. Table 3 presents a back-of-the envelope calculation of the effect of a unilateral onepercentage point increase in the headline corporate tax rate on corporate tax revenues. The calculation is based on the case in which g cs is xed. It assumes a constant equity / total assets ratio of 0.5 and a corporate tax rate of 37.5%. The equity / total assets ratio is the unweighted average of the corresponding numbers of the G7 countries for 1991 as reported in Rajan and Zingales (1995), Table 3, panel A). The corporate tax rate is the 1991 unweighted average of the G7 headline rates. At the margin, the net returns (possibly adjusted for differences in risk) on debt and equity should be equal. Because it is hard to obtain a reliable gure for the effective tax rate on interest income, we simply assume that interest income at the personal level is taxed at the same rate as corporate prots. We assume that the latter are retained in the company and that the capital gains tax equals zero. With an equity / total assets ratio of 0.5 and equal net returns on debt and equity, in the initial situation, half of what is left over after compensating labor should go to debt providers. The table shows that the increase in the tax rate leads to only a minor increase in the corporate tax revenue, because by far most of the additional revenue vanishes through income shifting. In the absence of income shifting (i.e. the amount of reported income is not affected by the tax increase), corporate tax revenue would have risen to 7.369. Hence, at the margin, the share of the revenue increase that is lost because of the fall in the amount of reported income is estimated at s7.369 2 7.203d /s7.369 2 7.178d 5 0.73.14 Based on the
Table 3 Effect of a unilateral tax rate increase on corporate tax revenues Initial rate 5 37.5% Value-added 1 21 2 c CES labor share 5 1.62 Compensation debt providers (Reported) prots before taxes Corporate tax revenue 100 61.73 s 2d 0.5(100 2 61.73) 5 19.14 s 2d 19.14 0.375*19.14 5 7.178 New rate 5 38.5% d*100 5 99.65 s1 1 g 61.73 s 2d 19.14 s 2d 18.78 0.385*18.78 5 7.230

CES is the average regression t of the rst term in square brackets on the right-hand side of Note: c CES * g *sWLd g * (value added). The ratio of debt to total assets (5.4). The change in value added is c is assumed to be 0.5. Because employment and debt contracts are given, compensation for labor and debt providers remains unchanged with the tax increase.

Some further scenarios are worked out in the Supplementary Appendix (available upon request). With interest income untaxed at the personal level, the share of additional revenue lost falls to 57%.

14

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2239

other cases in Table 2, we obtain a range from 68 to 87% for the additional revenue share that is lost. Although these gures should only be seen as providing a very rough estimate, they suggest that the amount of income shifting as a result of corporate tax rate differences may be substantial (at least for the manufacturing sector). Table 4 reports the individual estimates of the response coefcient for each country, using a linear transformation of the specication in Eq. (5.3). Not all countries have the same degree of sensitivity to tax rate differences, if at all. This is not surprising, given that the costs and benets of income shifting differ across countries. Nevertheless, while only one of the positive coefcient estimates is signicant, we nd signicantly negative estimates for Finland, France, Italy, Japan, the Netherlands and Portugal, suggesting the presence of income shifting for these countries. The coefcient estimates are particularly large (in absolute value) for the latter three countries. A potential explanation for the results of Portugal is that enforcement may be relatively weak.15 Both Japan and the Netherlands have a concentration of large multinationals in the manufacturing sector, while, moreover, the Netherlands uses an exemption system for the

Table 4 Baseline estimates of individual country and sector response coefcients Country Australia Austria Belgium Canada Denmark Finland France Germany Italy Japan Netherlands Portugal Spain Sweden UK US c g 0.0032 2 0.0004 0.0324 2 0.0035 2 0.0025 2 0.0165 2 0.0070 0.0017 2 0.0038 2 0.0289 2 0.0171 2 0.0236 2 0.0000 0.0011 2 0.0034 0.0013 S.E. 0.0035 0.0016 0.0077 0.0078 .0024 0.0028 0.0033 0.0023 0.0014 0.0034 0.0051 0.0040 0.0019 0.0019 0.0026 0.0033 Sector 1. Food, . . . 2. Textiles, . . . 3. Wood . . . 4. Paper . . . 5. Ind. chem. 6. Other chem. 7. Pottery . . . 8. Iron . . . 9. Non-ferrous . . . 10. Metal prod. 11. Machinery . . . 12. Electrical . . . 13. Transport . . . 14. Prof. goods 15. Other man. s g 2 0.0010 2 0.0050 2 0.0033 2 0.0053 2 0.0042 2 0.0010 2 0.0017 2 0.0095 2 0.0050 2 0.0042 2 0.0043 2 0.0045 2 0.0000 2 0.0050 0.0046 S.E. 0.0017 0.0020 0.0020 0.0019 0.0017 0.0017 0.0019 0.0020 0.0018 0.0021 0.0021 0.0021 0.0022 0.0020 0.0019

c, g s 5 estimated individual response coefcient when g cs is restricted across sectors, resp. Note: g countries; S.E. 5 standard error. For a full description of the sectors, see Appendix B.

15 Table 1 is somewhat suggestive to this extent. Moreover, a transfer pricing expert from a large consultant indicated that the Portuguese tax authorities may lack sufcient expertise to deal effectively with transfer pricing issues.

2240 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

treatment of prots generated abroad. Rather surprisingly, the estimate of g cs for Belgium is signicantly positive. Future research based on the forthcoming STAN release with direct data on capital stocks and a longer sample period may help us to shed more light on this issue. Table 4 also reports the individual estimates of the response coefcient for each sector. Except for one of them, the estimated coefcients are all negative and for 10 out of the 15 sectors they are signicantly negative. For these sectors, the coefcient estimates are of roughly the same magnitude, except for Iron and Steel, which comes out with a relatively large coefcient. Other manufacturing comes out signicantly, but with the wrong sign. However, this sector contains a disproportionate number of small rms and is unlikely to have a lot of multinational activity.

6. Robustness and extensions In this section, we investigate a number of variations on our baseline regressions. First, we extend the number of countries used to create average foreign tax rates. Next, we experiment with the variable used to weigh the foreign corporate tax rates. Thirdly, we control for the possibility there is also shifting between corporate and personal income. Fourth, we investigate the effect of specic actions taken by tax authorities to prevent income shifting through transfer pricing. Fifth, we split the sample in two subperiods to see whether the tax response remains stable over time. Finally, we briey discuss the results of some further variations. Table 5 presents results for extending the number of countries (the large tax set) used to compute the average foreign tax rates applied to each country in the analysis. The large tax set is described in Section 4. This results in a larger set of countries for income shifting to and from STAN countries (notice that in (5.4), i is i 2 always a country from the STAN, while now t is computed over all countries in t 16 the large tax set, excluding i ). The data used to weigh the foreign countries rates are retrieved from the World Banks (2000) Development Indicators. The weights are xed over time and computed as the average over the period 19951998. We present the results in steps. The tax rate difference used in the regression equation, (5.4), no longer varies across the sectors of a given country, as the relative weight is the same for each sector. To isolate the effect of extending the set of countries from that of changed weighting, we rst report the results when the average tax rate is computed for the STAN countries only, using
Ideally, as suggested by Eq. (A.2)see Appendix Awe would want to include the entire world into the computation of the average corporate tax rate. However, data limitations make this impractical. But, depending on the specic weighting variable used, coverage by the large tax set is between 87 and 94% of world activity.
16

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2241 Table 5 Alternative weighting schemesCES estimates Weighting Industry VA

g cs varies by
Country Sector

Countries STAN STAN Large Large Large Large Large Large Large Large

c 3.44 (0.46) 2.44 (0.31) 3.62 (0.49) 2.39 (0.31) 3.65 (0.47) 2.51 (0.32) 3.76 (0.48) 2.54 (0.32) 4.20 (0.53) 2.65 (0.34)

g 2 0.0059 (0.0010) 2 0.0038 (0.0006) 2 0.0056 (0.0010) 2 0.0038 (0.0006) 2 0.0049 (0.0010) 2 0.0034 (0.0007) 2 0.0041 (0.0011) 2 0.0030 (0.0007) 2 0.0026 (0.0010) 2 0.0019 (0.0007)

r 0.041 (0.0096) 0.018 (0.0094) 0.046 (0.0096) 0.017 (0.0094) 0.044 (0.0093) 0.020 (0.0093) 0.047 (0.0091) 0.021 (0.0093) 0.055 (0.0090) 0.024 (0.0093)

R2 0.59 0.58 0.59 0.58 0.59 0.58 0.59 0.58 0.59 0.58

SSE 0.11 0.11 0.11 0.11 0.11 0.11 0.11 0.11 0.11 0.11

Industry VA

Country Sector

GDP

Country Sector

Trade

Country Sector

FDI

Country Sector

Note: VA 5 value added, Large 5 large tax set, third column denotes which countries are included in the computation of the average foreign tax rate. For further explanation, see Notes of Table 2.

industrial sector value added from the World Bank for weighting. These estimates are directly comparable to those from the baseline regressions which also used industrial value added. The estimates of g are signicant and for g cs varying over sectors they are close to those reported in Table 2. The estimate when g cs varies over countries is now higher in absolute value, though not signicantly. The second block of Table 5 extends the computation of the average tax rate to the large tax set. The results hardly change. This is not surprising, because the relative weight of the additional countries in the sample is rather small (ranging from 21 to 28% of the new total, depending on the STAN country under consideration). Furthermore, many of the additional countries, especially the tax havens, have a (low) tax rate that hardly varies over time. If there is income shifting between the STAN and non-STAN countries in such a situation, it will primarily show up in the xed effects estimates and not in the estimate of g. Finally, because we do not have non-STAN country tax rates for the period 19791981, during this period the average foreign tax rate is computed over the STAN countries only. Using the large tax set we also try alternative weighting schemes to compute the average foreign tax rate, namely GDP, foreign trade or foreign direct investment. Theoretically, the relative importance of a foreign countrys tax rate

2242 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

for income shifting should reect the magnitude of intermediate goods or service ows to or from foreign afliates of the home countrys rms. Lacking data on all rms and their foreign afliates, the empirical setup calls for proxies. The most general variable, GDP, provides results similar to those for the baseline regressions (see Table 5).17 Because an important channel for income shifting is the mispricing of intermediates, average tax rates based on relative weights in international trade may be more appropriate: if a country with a disproportionally large weight in international trade (compared to value added) lowers its tax rate, this should lead to disproportionally large income shifts. In our empirical model (5.4), the average tax rate will fall by a relatively large amount if it is based on trade weights, thereby inducing a relatively large change in Vijt . The results reported in Table 5 also are close to the baseline results, which should further increase our condence in the latter. Finally, we consider a weighting scheme based on relative FDI shares. The reason is that more FDI, and thus an increased presence in foreign countries, should be conducive to prot shifting. Although the effect of tax rate differences on the valuelabor ratio is a bit weaker (see Table 5), the effect is still highly signicant. Most variations that we discuss from now on raise the number of parameters to be estimated. Given the nonlinearity of the CES estimating equation, we assume the more simple CobbDouglas production function for the estimates presented below. That is, we restrict r to zero and use (variations on the) regression Eq. is signicant in the results reported so far, its value is always (5.1). Although r close to zero. Indeed, as Table 6 (rst block) shows, the CobbDouglas versions of the baseline estimates of Table 2 are hardly affected: g remains signicantly negative and of roughly the same size. As suggested in Section 2, an additional channel through which corporate tax revenues can leak away when the corporate tax rate is raised is through a shift from corporate to personal income. Unfortunately, we cannot control for changes in organizational form, because our data on value added do not distinguish between the production by incorporated and nonincorporated rms within a sector. However, this type of shifting seems unimportant for the manufacturing sector.18 Shifts in the composition of the remuneration of management and other employees could in principle be picked up by our data. To control for this possibility, we include the difference between the own corporate tax rate and the personal tax rate in the standard regression equation:
2i P jt Vijt 5 c csy *f1 1 g csstit 2 t dg 1 eijt d 1 x csstit 2 t it

(6.1)

17 The Supplementary Appendix (available upon request) contains all the estimates when we restrict the set of countries to the STAN. Basically, the results are qualitatively the same as those for the large tax set reported in Table 5. 18 For the US in 1982, (Gravelle and Kotlikoff (1989), Table 1) report that the share of manufacturing sector output produced by incorporated rms was 97.6%.

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2243 Table 6 Variations and extensionsCobbDouglas Case A: g
cs

c xed 1.62 (0.0062) 1.62 (0.0097) 1.62 (0.0062) 1.61 (0.016) 1.62 (0.0062)

g 2 0.0039 (.0008) 2 0.0047 (0.0012) 2 0.0031 (0.0008) 2 0.0028 (0.0011) 2 0.0031 (0.0008) 2 0.0021 (0.0043) 2 0.0032 (0.0009) 2 0.0027 (0.0025) 2 0.0025 (0.0008) 2 0.0050 (0.0014)

R2 0.48 0.49 0.49

SSE 0.13 0.13 0.13 0.12 0.13

B: g cs varies over countries C: g cs varies over sectors Control for corporate personal income shifting Lax enforcement vs. tight enforcement Other countries vs US, France and Australia 19791988 versus 19891997

2 0.00052 (0.00076)

0.51 0.49

1.62 (0.0062)

0.49

0.13

1.62 (0.0062)

0.49

0.13

Note: except when otherwise indicated in the rst block, to achieve consistency we allow g cs and x cs and x for each of the relevant categories. to vary across sectors and report the average estimates g Here, x is dened analogously to g in (5.3). For further explanation, see Notes of Table 2.

where t P it is the highest marginal tax rate on personal income. We choose the highest marginal rate, because the top management, which often holds substantial amounts of shares in their own company, is probably most directly affected and has the power to take (or at least inuence) decisions to shift between corporate and personal income. Unfortunately, the x cs cannot be allowed to vary across the countries. The reason is that the identication of a country-specic response to P stit 2 t it d needs to be completely determined by the time variation in this tax rate, P because stit 2 t itd is constant across the sectors of the country. Generally, for a given country, there would be too little variation in the tax rate over time to obtain reliable results. Table 6 shows the results with g cs and x cs varying by sector. As is signicantly negative, with a magnitude close to that under the before, g is negative, as baseline estimation. Although the average estimated response x would be expected a priori, it is insignicant. A possible explanation is that shifts in the composition of income earned by the top management are relatively small compared to total value added if rms are large on average. Furthermore, such shifts may be rather difcult to hide for a countrys tax authorities if they can link

2244 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

the information about all domestic income sources of individuals. This would reduce the scope for shifting between corporate and personal income.19 Now, we investigate how the sensitivity of the valuelabor ratio with respect to the tax rate difference depends on the degree of formal enforcement. Based on the information in Table 1, we construct an enforcement indicator, E. For each of the enforcement mechanisms Explicit TP rules, Formal TP documentation rules or TP specic penalties that is applicable for country i and year t, Eit increases by one. Hence, Eit 5 0 when none of these enforcement mechanisms is present, while Eit attains its maximum of 3 when they are all present. Then we estimate (5.1), 1 where, now, g cs 5 o d 5 0 gd Id , it , and I0, it 5 1 and I 1, it 5 0 if Eit is 0 or 1 (lax enforcement), while I0, it 5 0 and I1, it 5 1 if Eit is 2 or 3 (tight enforcement). As Table 6 reports, the coefcient of the sensitivity of the valuelabor ratio with respect to the tax rate difference is signicantly negative under lax enforcement, while it is insignicant and much smaller in absolute value under tight enforcement. One should be careful in drawing strong conclusions from this nding, however. As can be seen from Table 1, observations of tight enforcement are concentrated among Australia, France and the US and, for the latter two, only among the past couple of years in the sample. Therefore, we ran another regression in which we classied these three countries as tight enforcement over the entire sample period. As reported in Table 6, the results hardly differ from those in the previous regression. This may be explained by the possibility that the formal enforcement mechanisms adopted by Australia, France and the US were actually an endogenous response to existing enforcement practices or an existing enforcement culture. Institutional or other relevant changes may have affected the incentives for income shifting over the sample period for any given country. Possible examples are the increasing trade linkages among countries and changes in enforcement. A simple way to see whether the incentives for income shifting have changed is to allow the response to the tax rate difference to vary over time. We split the sample in two subperiods, 19791988 and 19891997, and allow g cs to differ between (see Table 6) is the subperiods. The average estimated response coefcient g signicantly negative in both periods and larger in absolute value in the second period, potentially as a result of increased trade linkages. We conclude this section by briey discussing some additional variations for which we report the results in the Supplementary Appendix (available upon request). First, splitting the sample of countries into those with an exemption and a
19

As an alternative to (6.1), we have also estimated an equation with the own and the average foreign tax rate entered separately as regressors. As expected, the coefcient of the former is signicantly negative, while the coefcient of the latter is signicantly positive. However, in contrast to what one would expect (because the coefcient on the own tax rate would pick up both cross-border income shifting and income shifting between corporate and personal income), the coefcient of the average foreign tax rate is the larger in absolute value.

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2245

between the two groups and credit system, we nd hardly any difference in g hardly any difference with the baseline estimates. This may not be surprising given that in the presence of excess foreign tax credits, the incentive to shift income is often the same under the two systems. Moreover, bilateral treaties sometimes ensure the tax-exemption of foreign-generated income, even though a country may apply the credit system as a general rule. Second, because during the sample period Germany used a split rate system in which distributed income was taxed at a lower rate than retained earnings, we reran our baseline regressions either excluding Germany or replacing Germanys tax rate on retained earnings with the tax rate on distributed income. In both cases, the estimates are qualitatively unaffected. Finally, we have investigated how robust our estimates of the sensitivity of income shifting to tax rate differences are for different assumptions about the presence of xed country, sector or time effects. Generally, the strongest average effects of tax rate differences on the valuelabor ratio are obtained if g cs is allowed to vary over countries rather than over sectors. In all of the cases, the had the expected sign and in only one instance g was average estimate g insignicant. This occurred when the constant term (i.e. the relevant technology coefcient of the underlying production function) was not allowed to vary across sectors, which does not seem like an appealing specication.

7. Conclusion In this paper we have investigated the importance of income shifting among the industrialized and other major economies as a result of differences in corporate tax rates. While most of the existing empirical work investigates income shifting to tax havens, income shifting among (relatively) large economies seems to have escaped the attention of both researchers and policymakers. In our empirical analysis we have attempted to demonstrate that this may be unjustied. Our empirical analysis has tried to extend the literature in other directions as well: rstly, we have presented a novel method for isolating the pure effects of income shifting. Secondly, and in contrast to the preceding literature, which uses either macroeconomic data or rm-level data, our empirical analysis was based on sectoral data. Finally, we have explored a large number of variations and extensions of the baseline regression. Examples are the use of alternative weighting schemes for the average foreign tax rate, a control for potential domestic income shifting and the use of an indicator of formal enforcement mechanisms for transfer pricing rules. Our ndings are consistent with the hypothesis that more enforcement, as measured by the indicator, is associated with less income shifting, although the adoption of formal enforcement mechanisms could well be an endogenous response to existing enforcement practices or an existing enforcement culture. The results for our baseline regression, as well as the variations we considered,

2246 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

suggest that income shifting is both statistically and economically signicant: a back-of-the-envelope calculation for our baseline estimates indicates that at the margin more than 65% of the additional revenue resulting from a unilateral tax increase is lost because of income shifting. Moreover, if correct, our estimates would only constitute a lower bound for the effect of tax rate changes on reported income. The reason is that income shifting may also take place through changes in the capital structure, which we have assumed to be constant in our calculation of the amount of income shifting. Nevertheless, because our analysis is one of the rst of income shifting between OECD countries, the estimated magnitude of the effect should only be taken as a rst indication that there might be substantial income shifting between these countries. As described in Section 2, there are many possible channels through which changes in tax rates can affect reported income. Although we have tried to control for these as well as possible, we cannot rule out that the relation between the valuelabor ratio and the tax rates that we found is inuenced by these alternative channels. Even so, as Feldstein (1995) emphasizes (in the context of personal income taxation), what matters for efciency is the total loss of revenues through all channels, when the tax rate is raised. It should be emphasized that our estimates are based solely on the manufacturing sector, of which large parts are indeed dominated by multinationals. Other sectors in the economy may be much less open, in particular the non-traded sector, which constitutes a large share of economic activity in OECD economies. Hence, estimates of income shifting based on manufacturing sector data cannot directly be extrapolated to the whole economy. Nevertheless, cross-border income shifting between OECD countries could be substantial for several reasons. First, the attention of the tax authorities may be directed more towards detecting income shifting to tax havens than to other OECD countries. Second, because of its large share in the world economic activity, the OECD as a whole would offer a substantial scope for income shifting in response to a relative increase in the corporate tax rate. Finally, and related to the previous point, because income shifting through transfer pricing mostly takes place when there is cross-border trade related to real activity, the scope for income shifting among OECD countries would be correspondingly larger than for income shifting to tax havens (which is also conrmed by the results of Pak and Zdanowicz, 2001, for income shifting out of the US). Relatively small distortions of transfer prices on large cross-border transactions are hard to detect by the tax authorities, especially when there exists no outside market for these transactions and when a large component of their value is related to intellectual input. Our ndings suggest that the revenues from tighter enforcement of transfer pricing rules can in principle be quite high. However, the scope for income shifting within a multinational is tightly linked to the scale at which it operates. Tighter enforcement of transfer pricing rules by countries with high taxes may

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2247

reduce the net return on investments, thereby also causing real activity shifts to countries with lower taxes and / or laxer enforcement.20 This suggests a case for the international coordination of transfer pricing policies. Examples are the exchange of information about the activities of multinationals and agreements about minimum enforcement standards or common transfer prices (for example, see Mansori and Weichenrieder, 1999). Our methodology for isolating the pure income shifting effects opens the door for further research. Firstly, when capital stock data become available it is possible to estimate the effect of relative tax changes on real activity shifts (using the methods from this paper to lter out the possible contamination caused by pure-accounting income shifts). Secondly, with information about the composition of intermediates trade in intangibles and tangibles, one could test whether income shifting is larger in the former case, because it is more difcult to establish proper market prices. Finally, our methodology could be applied to enhance the accuracy of international productivity comparisons. The reason is that income shifting between countries for tax purposes introduces distortions in measured productivity: sales in a low-tax country may be over-reported and intermediate inputs underreported. Even without changes in the underlying technology, income shifting will lead to an increase in the measured productivity of a country that lowers its corporate tax rate. In fact, this observation has been made in the comparison of productivity between similar rms in Ireland and the UK (Birnie, 1996).

Acknowledgements We thank two anonymous referees, Ruud de Mooij, Joeri Gorter, seminar participants at the Netherlands Bureau for Economic and Policy Analysis (CPB) and participants at the NAKE Day for helpful comments. We thank Jeroen Hinloopen, Steef Huijbregtse and Alfons Weichenrieder for stimulating discussions. Moreover, we thank Steef Huijbregtse (Ernst and Young) also for providing us with the data on the enforcement of transfer pricing rules and Aydin Hayri (Deloitte) for information on the country-specic tax treatment of foreign-source income. The usual disclaimer applies. Part of this paper was written while Beetsma was visiting CESifo (University of Munich). He thanks his host institution for their hospitality and their excellent research environment.
20 In 1992, Germany imposed a restriction that linked the amount of German-tax exempt passive income generated by Irish subsidiaries of German multinationals to the amount of active income generated by the Irish subsidiaries. Weichenrieder (1996) argues that this could well have induced German multinationals to start using these subsidiaries (which were originally set up for nancial investment purposes) for local production, thereby shifting real activity from Germany to Ireland.

2248 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

Appendix A. A microfoundation for (3.2) Consider a world with m countries. Country i is the country of residence for a mass of pi multinationals. We normalize the total mass of multinationals in the world to unity, i.e. o m i 5 1 pi 5 1. Each multinational has exactly one production facility in each country other than its residence country. In any production facility in country j, production takes place according to the production function:
a 12a Q* j 5 Lj K

where Lj is the labor input and K is a given amount of capital input. Production requires the use of a xed amount of a specialized intermediate good and / or service, H, which is provided by the parent company located in the residence country. Consider a multinational whose residence country is i. Assuming that H can be generated at a xed cost CH , its reported before-tax income generated in country i is:

pi ; zs1 1 vid H 2 CH
where zs1 1 vid is the price that the parent charges its foreign afliates for the use of the intermediates. Here, z is the arms length price (or market price, if there exists a market outside the multinational) for the intermediates, while vi is a mark-up (positive or negative) imposed in order to shift income across borders. zs1 1 vid H includes the revenue from intermediate tangibles, but also the compensation for intermediate services provided by intangibles, such as a patent on the production process that is owned by the parent of the multinational. This type of compensation often takes the form of a royalty payment. For expository purposes, we abstract here from potential royalty withholding taxes. In the Supplementary Appendix, Section 2 (available upon request) we extend the current model and assume that a certain part of zs1 1 vid H consists of royalty payments that are subject to withholding taxes. While the qualitative implications for Eq. (A.2), derived below, are unchanged, the extent of income shifting as a result of corporate tax rate changes becomes smaller if the royalty withholding tax is higher. Purely for the sake of space, we set CH 5 0. The reported before-tax prot generated in country j i is: 1 12a pj ; P f L a 2 WLj 2 ]] zs1 1 vid H j K m21 where P f is the price of the multinationals nal product (we assume that all multinationals produce the same nal product in each country and, hence, charge the same price for their product) and W is the (nominal) wage rate (assumed to be equal across countries, for example because of perfect international labor mobility or because of identical labor market circumstances in each country). The parent charges each afliate the same price for the use of the intermediate. Charging

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2249

different prices for the same intermediate provided to the various afliates would be a too blatant violation of the accounting rules to escape the attention of the tax authorities. The multinational maximizes the sum of after-tax worldwide prots:

p m,i ;

j 51

Os1 2 t d p 2 csv d
m j j i

where tj is the corporate tax rate in country j. Each country uses an exemption system for the taxation of foreign prots. The function csvid is twice continuously differentiable, strictly convex and attains a minimum of 0 for vi 5 0. It can be thought of as the (expected) ne for prot shifting. For example, the more vi deviates from 0, the higher the chance that the multinational will be caught and / or the larger the ne for not using the arms length price. If vi , 0, prots are shifted to foreign countries and potential nes are incurred in the home country, while, if vi . 0, foreign countries lose tax revenues and impose a potential ne. However, the function csvid can also be given a broader interpretation. In particular, it may not only capture expected nes, but also efciency losses that may arise because charging the wrong price for the use of an intermediate could distort the incentives of local managers and workers. 2 1 Specifying csvid 5 ] 2 v i , the rst-order conditions for Lj and vi can be written as, respectively:
21 12a a P fL a K 5 W, j j i

j 5 1, . . . ,m
i i

p Os1 2 t d ] 1 1 2 t d zH 5 v v s
j j i

The last rst-order condition can be rewritten as: 2 i 2 tid 5 vi zHst (A.1)

1 2 i ; ]] where t m 2 1 o j i tj is the average tax rate over all countries excluding i. 2 i or a decrease in ti leads to prot shifting into country i, Hence, an increase in t because it implies an increase in vi . Using these results, we shall now derive the ratio of total (of all rms) value-added (TVA) and total labor compensation (TLC ) for country i. Total value added is given by:

TVA i 5 pi zs1 1 vid H 1

O p FP L K
f j

a i

12a

j i

1 2 ]] zs1 1 vjd H m21

where the rst term on the right-hand side is the value-added generated by the rms residing in country i, while the second term is the value-added generated by the foreign rms producing in country i, a number pj of which have their parent company in country j. Total labor compensation in country i is o j i pj WLi 5s1 2

2250 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252

12a pid a P f L a . Under the assumption that all countries are equally large, i.e. i K pj 5 p,; j, and using (A.1), we can derive (see the Supplementary Appendix available upon requestfor the details):

TVA i 1 ]] 2 idg, 5 ]f1 1 gsti 2 t TLCi a

m ps ]] m 2 1d where g 5 2sHzd 2 ]]]]] f a 12a s1 2 pd P L i K

(A.2)

2 i leads to a fall in (TVA i / TLCi ) because of Hence, an increase in ti relative to t income shifting to other countries.

Appendix B. The data Countries included in the STAN: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, Portugal, Spain, Sweden, the United Kingdom and the United States. Countries included in the large tax set: All countries in the STAN plus Argentina, Barbados, Bahamas, Bermuda, Brazil, Cayman Islands, China, Cyprus, Greece, Hong Kong, Iceland, India, Indonesia, Ireland, South Korea, Luxembourg, Malaysia, Malta, Mexico, New Zealand, Norway, Puerto Rico, Singapore, Switzerland and Turkey. Included industrial sectors: 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. Food, beverages, tobacco Textiles, wearing apparel, leather and leather products, footwear Wood products, furniture and xtures Paper products, printing and publishing Industrial chemicals Other chemicals Pottery and china, glass products, non-metallic products nec Iron and steel Non-ferrous metals Metal products Machinery nec Electrical machinery Transport equipment Professional goods Other manufacturing

E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 2251

References
Birnie, J.E., 1996. Comparative productivity in Ireland: The impact of transfer pricing and foreign ownership. In: Wagner, K., van Ark, B. (Eds.), International Productivity Differences; Measurement and Explanations. Elsevier, Amsterdam. Clausing, K., 1998. The Impact of Transfer Pricing on Intrarm Trade. NBER Working Paper No. 6688. Collins, J.H., Shackelford, D.A., 1992. Foreign tax credit limitations and preferred stock issuances. Journal of Accounting Research 30 (Supplement), 103124. Ernst and Young, 2000. Transfer Pricing at-a-Glance Guide. Ernst and Young, Rotterdam. Feldstein, M.S., 1995. Tax Avoidance and the Deadweight Loss of the Income Tax. NBER Working Paper No. 5055. Froot, K.A., Hines, J.R., 1995. Interest allocation rules, nancing patterns, and the operations of US multinationals. In: Feldstein, M.S., Hines, J.R., Hubbard, R.G. (Eds.), The Effects of Taxation on Multinational Corporations. University of Chicago Press, Chicago. Gordon, R.H., Slemrod, J., 1998. Are Real Responses to Taxes Simply Income Shifting between Corporate and Personal Tax Bases?. NBER Working Paper No. 6576. Gorter, J., Parikh, A., 2000. How Mobile is Capital within the European Union. Research Memorandum No. 172. CPB Netherlands Bureau for Economic Policy Analysis, The Hague. Gravelle, J., Kotlikoff, L.J., 1989. The incidence and efciency costs of corporate taxation when corporate and noncorporate rms produce the same good. Journal of Political Economy 97 (4), 749780. Grubert, H., 1998. Taxes and the division of foreign operating income among royalties, interest, dividends and retained earnings. Journal of Public Economics 68, 269290. Grubert, H., Mutti, J., 1991. Taxes, tariffs and transfer pricing in multinational corporate decision making. Review of Economics and Statistics 68, 285293. Grubert, H., Slemrod, J., 1998. The effect of taxes on investment and income shifting into Puerto Rico. Review of Economics and Statistics 75, 365373. Harrigan, J., 1999. Estimation of cross-country differences in industry production functions. Journal of International Economics 47, 267293. Harris, D.G., 1993. The impact of US tax law revision on multinational corporations capital location and income-shifting decisions. Journal of Accounting Research 31 (Supplement), 111140. Hines, J.R., 1996a. Altered states: Taxes and the location of foreign direct investment in America. American Economic Review 86, 10761094. Hines, J.R., 1996b. Tax Policy and the Activities of Multinational Corporations. NBER Working Paper No. 5589. Hines, J.R., 1999. Lessons from behavioral responses to international taxation. National Tax Journal 52 (2), 305322. Hines, J.R., Hubbard, R.G., 1990. Coming home to America: Dividend repatriations by US multinationals. In: Razin, A., Hubbard, R.G. (Eds.), Taxation in the Global Economy. University of Chicago Press, Chicago. Hines, J.R., Rice, E.M., 1994. Fiscal paradise: Foreign tax havens and American business. Quarterly Journal of Economics 109, 149182. Kemfert, C., 1998. Estimated substitution elasticities of a nested CES production function approach for Germany. Energy Economics 20 (3), 249264. Mackie-Mason, J., Gordon, R.H., 1997. How much do taxes discourage incorporation? Journal of Finance 52, 477505. Mansori, K.S., Weichenrieder, A.J., 1999. Tax Competition and Transfer Pricing Disputes. Mimeo. Colby College and CES / University of Munich. Newlon, T.S., 1997. Transfer Pricing and Income Shifting in Integrating Economies. Mimeo. US Department of the Treasury.

2252 E. J. Bartelsman, R.M.W. J. Beetsma / Journal of Public Economics 87 (2003) 22252252 OECD, 1999. The OECD STAN Database for Industrial Analysis. OECD, Paris. OECD, 2000. Towards Global Tax Cooperation. Report to the 2000 Ministerial Council Meeting and Recommendations by the Committee on Fiscal Affairs. OECD, Paris. Pak, S.J., Zdanowicz, J.S., 2001. Executive Summary of US Trade with the World: An Estimate of 2000 Lost US Federal Income Tax Revenues due to Over-Invoiced Imports and Under-Invoiced Exports. Mimeo. Florida International University. Paraskevopoulos, C.C., 1979. Alternative estimates of the elasticity of substitution: An inter-metropolitan CES production function analysis of US manufacturing industries, 19581972. Review of Economics and Statistics 61 (3), 439442. Rajan, R.G., Zingales, L., 1995. What do we know about capital structure? Some evidence from international data. Journal of Finance 50 (5), 14211460. Roeger, W., 1995. Can imperfect competition explain the difference between primal and dual productivity measures? Estimates for US manufacturing. Journal of Political Economy 103 (2), 316330. Stone, J.A., Lee, H.-H., 1995. Determinants of intra-industry trade: A longitudinal, cross-country analysis. Weltwirtschaftliches Archiv 131, 6785. Weichenrieder, A.J., 1996. Fighting international tax avoidance: The case of Germany. Fiscal Studies 17 (1), 3758. World Bank, 2000. Development Indicators, April 2000. World Bank, Washington, DC.