Multinational Firms and Tax Havens

Transcription

1 Unversty of Mchgan Law School Unversty of Mchgan Law School Scholarshp Repostory Artcles Faculty Scholarshp 016 Multnatonal Frms and Tax Havens James R. Hnes Jr. Unversty of Mchgan Law School, Anna Gumpert Ludwg-Maxmlans-Unverstät München, Monka Schntzer Unversty of Munch, Department of Economcs, Follow ths and addtonal works at: Part of the Busness Organzatons Law Commons, Comparatve and Foregn Law Commons, Economcs Commons, and the Taxaton-Transnatonal Commons Recommended Ctaton Hnes, James R., Jr., co-author. "Multnatonal Frms and Tax Havens." A. Gumpert and M. Schntzer, co-authors. The Rev. of Econ. and Stat. 98, no. 4 (016): Ths Artcle s brought to you for free and open access by the Faculty Scholarshp at Unversty of Mchgan Law School Scholarshp Repostory. It has been accepted for ncluson n Artcles by an authorzed admnstrator of Unversty of Mchgan Law School Scholarshp Repostory. For more nformaton, please contact mlaw.repostory@umch.edu.

2 MULTINATIONAL FIRMS AND TAX HAVENS Anna Gumpert, James R. Hnes Jr., and Monka Schntzer* Abstract Multnatonal frms wth operatons n hgh-tax countres can beneft the most from reallocatng taxable ncome to tax havens, though ths s suffcently dffcult and costly that only 0.4% of German multnatonal frms have any tax haven afflates. Among German manufacturng frms, a 1 percentage pont hgher foregn tax rate s assocated wth a.3% greater lkelhood of ownng a tax haven afflate. Ths s consstent wth tax avodance ncentves and contrasts wth earler evdence for U.S. frms. The relatonshp s less strong for frms n servce ndustres, possbly reflectng the dffculty of reallocatng taxable servce ncome. I. Introducton TAX havens are typcally small, well-governed states that mpose low or zero tax rates on foregn nvestors (Dharmapala & Hnes, 009). Sophstcated tax avodance strateges nvolvng tax havens have receved consderable attenton n the meda (Drucker, 010; Wake Up and Smell the Coffee, 01; Lucas, Jopson, & Houlder, 01), and tax havens have repeatedly been the focus of natonal and nternatonal polcy measures. To name a few examples, the OECD launched ts Harmful Tax Competton project n 1996 to pressure tax havens to abolsh harmful tax provsons and practces. France announced plans to ntroduce a 50% tax on ncome earned by French afflates n tax havens n February 010. The U.S. House Commttee on Ways and Means held a background hearng on the transfer prcng practces of U.S. taxpayers, wth an emphass on ncome reallocaton to offshore tax havens. The Unted Kngdom Parlament Commttee of Publc Accounts held wdely publczed hearngs n 01 concernng the use of tax havens by foregn multnatonal frms operatng n Brtan. Partly as a consequence, the G-0 fnance mnsters n ther February 013 meetng pledged to take collectve acton to develop measures to address tax base eroson and proft shftng, taskng the OECD wth recommendng polces that could be Receved for publcaton August 8, 013. Revson accepted for publcaton November 3, 015. Edtor: Asm I. Khwaja. *Gumpert: LMU Munch and CESfo; Hnes: Unversty of Mchgan and NBER; Schntzer: LMU Munch, CESfo, and CEPR. Ths paper was partly wrtten durng vsts of the frst and thrd authors to the research center of the Deutsche Bundesbank. The hosptalty of the Bundesbank, the constructve support of the staff, as well as access to ts Mcrodatabase Drect nvestment (MD) are gratefully acknowledged. The project has benefted from fnancal support through the Deutsche Forschungsgemenschaft (German Scence Foundaton) under SFB-Transrego 15 and GRK 801. We thank Johannes Becker, Thess Buettner, Ronald Daves, Dhammka Dharmapala, Rachel Grffth, Andreas Haufler, Shafk Hebous, Nels Johannesen, Irs Kesternch, Nadne Redel, Erk Röder, Rchard Schmdtke, Emmanuelle Taugourdeau, Tll von Wachter, and Martn Watznger, as well as conference partcpants at the the CESfo Conference on Corporate Taxes and Corporate Governance, the Tax Havens Conference at the Max-Planck Insttute Munch, the 10th Journées Lous-André Gérard-Varet, the 67th IIPF Congress, the meetng of the SFB- Transrego 15, the frst CBT Doctoral Meetng, the 9th Bundesbank MD Workshop, and the Royal Economc Socety 01 Annual Conference for helpful comments. Asm Khwaja and three anonymous referees offered many constructve suggestons that are ncorporated n ths paper. A supplemental appendx s avalable onlne at journals.org/do/suppl/10.116/rest_a_ enacted by natonal governments, whch resulted n OECD recommendatons to the G0 fnance mnsters n October 015. What drves the polcy nterest s the concern that tax havens are wdely used to avod tax oblgatons, partcularly those due to hgh-tax countres. Ths concern s consstent wth at least some of the evdence: the U.S. Congress Jont Commttee on Taxaton (010) report dentfes specfc methods by whch U.S. frms use transactons wth tax haven afflates to avod tax oblgatons, and Desa, Foley, and Hnes (006) offer evdence that ownershp of a regonal tax haven afflate s assocated wth reduced foregn tax payments. It s clear, however, that corporatons cannot use tax haven operatons to avod all taxes easly and comprehensvely, as reflected n persstent sgnfcant corporate tax collectons by hgh-tax countres. Furthermore, Desa et al. (006) offers evdence that n 1999, just 59% of U.S. multnatonals wth sgnfcant foregn operatons had any tax haven afflates. Ths suggests that substantal tax avodance opportuntes through the use of tax haven operatons are selectvely avalable only to certan frms, ndustres, or actvtes, though the avalable evdence descrbes only U.S. frms. The purpose of ths paper s to dentfy factors assocated wth demand by German multnatonal frms for tax haven operatons. German frms are major foregn nvestors, so ther use of tax havens to avod taxes s potentally qute mportant; furthermore, evdence of ther tax haven use s a valuable complement to exstng studes of the behavor of U.S. frms. The data ndcate that among manufacturng frms, those that are larger and more productve are the most lkely to have tax haven afflates. Notably, manufacturng frms whose nonhaven foregn operatons are located n hghtax countres are more lkely than others to have tax haven afflates. These patterns are consstent wth a smple model of proft reallocaton, n whch some frms fnd that the benefts of beng able to reallocate taxable ncome from hgh-tax jursdctons exceeds the cost of establshng tax haven afflates, whereas other multnatonal frms, wth fewer profts to reallocate or fewer taxes to save, do not. Among frms n servce ndustres, tax haven use s less closely assocated wth hgh foregn tax rates, qute possbly reflectng the more lmted scope for taxable ncome reallocaton among frms n servce ndustres. It s noteworthy that relatvely few German frms use tax havens. In the sample of all German multnatonal frms from 00 to 008, only 0.4% have tax haven afflates. Ths s consderably smaller than the fracton of U.S. frms usng tax havens and s smaller stll when the sample s lmted to German manufacturng frms. Ths pattern mples that cost-effectve tax avodance opportuntes through the use of tax havens may be qute lmted for German frms and The Revew of Economcs and Statstcs, October 016, 98(4): by the Presdent and Fellows of Harvard College and the Massachusetts Insttute of Technology do:10.116/rest_a_00591

3 714 THE REVIEW OF ECONOMICS AND STATISTICS that the U.S. evdence of more wdespread, though nonetheless stll lmted, tax haven use may be atypcal on the hgh sde. It s partcularly strkng that German frms would be less lkely to use tax havens than are U.S. frms, gven that the U.S. worldwde tax system does more than does the German terrtoral tax system to lmt the benefts of reallocatng taxable ncome from hgh-tax countres to low-tax countres. Many frms could beneft from reallocatng taxable ncome to afflates located n nonhaven countres that have moderately low tax rates, though the most aggressve and successful tax avodance s commonly thought to nvolve tax havens. As a result, ths study focuses on tax haven operatons. Much of the avalable evdence on the determnants of tax haven use by multnatonal frms comes from studes of U.S. frms, whose tax treatment dffers from those of frms based n almost any other major captal-exportng country. Income earned by foregn subsdares of U.S. frms s subject to U.S. taxaton when repatrated, at whch tme U.S. taxpayers can clam credts for ncome taxes pad to foregn governments. Ths system reduces the ncentve to use tax haven operatons to avod foregn taxes, snce foregn tax reductons ental fewer foregn tax credts, and therefore greater U.S. tax lablty, when ncome s ultmately repatrated. The system does, however, encourage the use of tax havens to the extent that they facltate deferral of home country taxes, such as by servng as coordnaton centers to drect foregn profts to new foregn nvestments. Germany taxes only 5% of the actve foregn busness profts of ts resdent corporatons. In ths respect, the German tax system s smlar to those n the Unted Kngdom, Japan, France, Canada, Italy, and most OECD countres, partcularly the major captal exporters. German frms consequently have strong ncentves to avod foregn taxes, snce foregn tax savngs do not ental greater home country taxes when ncome s repatrated. Furthermore, German frms do not have ncentves to structure ther foregn operatons n ways that avod repatratng ncome. Therefore, the tax ncentves for German frms to establsh tax haven afflates are lkely to dffer from those of U.S. frms and bear strong smlartes to those of other G-7 and OECD frms. To gauge the mpact of foregn taxaton on tax haven nvestment by German frms, the paper estmates a lnear probablty model of tax haven nvestment usng the szes of a frm s foregn and domestc actvtes, and the frm s R&D ntensty, as addtonal control varables. The emprcal strategy uses the panel nature of the data to account for the fact that the tax rates a frm faces at ts foregn locatons may be endogenous to ts decson to nvest n a tax haven: specfcally, the estmaton reles on statutory foregn tax changes subsequent to a base year. The estmates ndcate that a 1 percentage pont hgher foregn tax rate s assocated wth a.3% greater lkelhood of havng a tax haven afflate, presumably reflectng the greater beneft of reallocatng taxable ncome from hgh-tax countres. Ths paper s related to two strands of the lterature: one on the use of tax havens by multnatonal frms and the other on proft reallocaton. Harrs et al. (1993) analyze a fveyear panel of 00 large U.S. manufacturng frms and fnd that U.S. tax labltes of U.S. frms holdng afflates n Ireland or one of the four low-tax dragon Asan countres are systematcally lower than those of U.S. frms wthout such actvtes. Hnes and Rce (1994) analyze a cross-secton of country-level data on the actvtes of U.S. multnatonal frms, fndng that U.S. multnatonals report dsproportonate shares of profts n tax havens, whch suggests that ncome may be reallocated for tax purposes. Grubert and Slemrod (1998) use a cross-secton of data and estmate a jont model of the nvestment and proft-shftng decson of U.S. multnatonals n Puerto Rco, whch, due to ts specal status, can serve as a tax haven for U.S. frms. They fnd that frms wth ntangble assets are more lkely than others to nvest n Puerto Rco. Desa et al. (006), who are closest to our analyss, use an afflate-level data set on U.S. multnatonals foregn actvtes n four years between 198 and They estmate a logt model of tax haven nvestment gven parent characterstcs and take nto account the endogenety of the foregn nonhaven tax rate due to smultanety of a parent s locaton decsons. They fnd a negatve effect of the average foregn nonhaven tax rate on the probablty of nvestng n a tax haven, nterpretng ther fndng as evdence of the mpact of ncentves nduced by the ablty to defer home country taxaton of unrepatrated foregn profts. Thus, t s partcularly nterestng to compare the U.S. evdence wth the tax haven nvestment behavor of frms that are subject to a tax exempton regme, as German frms are, whch have clear ncentves to use tax haven operatons to reallocate taxable ncome. There s a vast lterature on nternatonal proft shftng, so for brevty, t s helpful to revew just recent examples of dfferent strands of the lterature. That taxes nfluence reported profts has been documented by Huznga and Laeven (008), among others. They use a cross-secton of European multnatonal enterprses (MNEs) and fnd evdence for substantal proft shftng between dfferent countres n Europe, whch fts nternatonal proft-shftng ncentves that arse from tax dfferences between the parent and host country and among dfferent afflate locatons. Wechenreder (009) analyzes a panel data set of German nbound and outbound FDI and dentfes emprcal patterns that are consstent wth proft shftng n both cases. Wth respect to dfferent proft-shftng strateges, Clausng (001, 003, 006) provdes emprcal evdence that taxes exert a substantal mpact on transfer prces and ntrafrm trade flows of U.S. frms, whle Vcard (014) provdes complementary evdence for French frms. Dschnger and Redel (011) offer evdence from a panel data set of European frms that MNEs prefer locatng ntangble assets n low-tax locatons, arguably dong so because they are able to choose favorable transfer prces for ntangble

4 MULTINATIONAL FIRMS AND TAX HAVENS 715 assets. Karknsky and Redel (01) report smlar fndngs wth respect to patent locaton wthn MNEs. Egger et al. (010) compare the debt-to-asset ratos of domestcally and foregn-owned European frms and dentfy a gap n the ratos systematcally related to corporate tax rates. Buettner et al. (009) provde further evdence on tax-motvated choce of captal structure usng a panel data set of German MNEs. Usng the same data, Buettner and Wamser (013) analyze the use of ntrafrm loans for proft shftng, but fnd that they have rather small tax revenue effects. Wechenreder and Mntz (010), as well as Wamser (011), show, usng data on German MNEs, that frms tactcally locate ther drect and ndrect afflates and strategcally use ownershp chans n a way that facltates tax avodance. Many of these studes, as well as others (e.g., Desa et al., 009) analyzng the behavor of multnatonal frms, confne ther analyss to manufacturng frms. Ths s done both due to the mportance of manufacturng and n an effort to control for frm heterogenety. As a result, dfferences n proft-shftng opportuntes between frms n dfferent ndustres have been much less extensvely studed. II. Incentves to Establsh Tax Haven Operatons In ths secton we lay out a stylzed theoretcal framework to descrbe the ncentves of a multnatonal frm to nvest n a tax haven and derve the man emprcal predcton we test. Consder a multnatonal frm that can nvest n a range of countres = 0,..., n, ncludng a tax haven, whch s denoted as country 0. Startng a foregn afflate nvolves fxed setup cost c. Let ρ denote before-tax profts earned n country by the afflate once t s nstalled. Reported profts are taxed at rate τ n country. Wthout loss of generalty, we assume that τ 0 = 0; that s, there s no taxaton n the tax haven. Frms can reallocate an amount ψ of ther actual profts n country to a country that taxes reported profts at a lower tax rate, most notably to the tax haven country for example, by adjustng ther transfer prces. Ths s possble only at some cost. Frms may need to set up addtonal facltes to make transfer prces seem plausble, neffcent relocaton of producton and ntrafrm trade may be needed to arrange ncome reallocaton, and transacton costs, lke legal expenses, are ncurred. We assume that ncome reallocaton gets ncreasngly expensve as the amount reallocated ncreases relatve to ncome earned n country. Followng Hnes and Rce (1994), these ncome reallocaton costs are assumed to be (a/)(ψ /ρ ). 1 Parameter a captures how much the cost of ncome reallocaton ncreases wth the amount reallocated. Note that a s a frm-specfc parameter because ncome reallocaton costs may vary wth frm-specfc characterstcs such as ownershp of ntangble assets. In the followng, we 1 For smplcty, we assume that the costs of reallocatng ncome to a tax haven and another nonhaven country are equal. Ths assumpton does not affect the man ntuton of the model but renders notaton far more tractable. assume that the ncome reallocaton cost are ncurred n the country from whch ncome s shfted, though the model s mplcatons are qute smlar f these costs are ncurred n the country to whch the ncome s transferred. The reported proft n country, π, after fxed cost c s sunk, s thus π = ρ ψ a ψ. (1) ρ Consder now the opton of settng up an afflate n a tax haven at cost c 0. To save on notaton, we set ρ 0 = 0 and let c 0 capture the net cost of nvestng n a tax haven, after deductng any profts that arse genunely n ths country. Note that due to antdeferral regulatons, frms may have to ncur consderable fxed costs to set up a tax haven afflate that may be used for proft reallocaton purposes. 3 For c 0 < 0, the multnatonal has an nterest n nvestng n a tax haven country and does so, ndependent of nvestments n other countres. Ths nterest could arse from plans to reallocate ncome from the home country. However, snce our data set contans nformaton on parent frms from only one home country, t s not possble to gauge the mpact of ths tax ncentve emprcally. Thus, we focus on multnatonals that nvest n nonhaven countres as well. To evaluate the ncentve to nvest n a tax haven, consder frst the stuaton of a multnatonal wth a tax haven afflate. The frm chooses n whch other countres to locate afflates and how much of ther profts to reallocate to the tax haven. Thus, the nvestor s maxmzaton problem, gven that t has a tax haven afflate, s max d,ψ n ( d [ψ + (1 τ ) ρ ψ a =1 wth d {0, 1}, s.t. ψ ρ ) c ], ρ ψ a ψ 0 = 1,..., n. (3) ρ The followng lemma descrbes the soluton to ths maxmzaton problem. For ease of presentaton, we restrct consderaton to parameter cases such that an nteror soluton s obtaned. 4 Lemma 1. Suppose the nvestor has a tax haven afflate. Then the optmal amount of proft shftng s () ψ th = τ ρ a(1 τ ). (4) Ths and other extensons to the model, along wth tables reportng supplementary regressons, are presented n supplementary materal avalable at 3 For detals on the taxaton of foregn profts of German multnatonals and the antdeferral regulatons, refer to the onlne supplementary materal. 4 Ths s the case f τ 1 suffcently large. 1 a+1, whch s trvally fulflled for a

5 716 THE REVIEW OF ECONOMICS AND STATISTICS Let countres be numbered such that country = 1 yelds the hghest after-tax proft, ncludng the fxed cost of settng up the afflate, and country = n yelds the lowest proft. Then the multnatonal chooses d = 1 for all countres = 1,...,ñ, where ñ s determned by the condton ( ψñ + (1 τñ) ρñ ψñ a ψ ) ñ cñ 0 > ρñ ( ) ψñ+1 + (1 τñ+1 ) ρñ+1 ψñ+1 a ψ ñ+1 cñ+1. ρñ+1 (5) Proof. See the onlne supplementary materal. Consder now the multnatonal s stuaton f t has no tax haven afflate. In ths case, proft shftng has to be drected to the country chargng the lowest tax rate among those n whch the multnatonal holds an afflate. 5 Let τ denote the mnmum of all tax rates charged n countres n whch the multnatonal nvests. In the supplementary materal, we derve the optmal amount of proft shftng ψ nth and descrbe the optmal number of countres ˆn n whch to set up a foregn afflate. It s straghtforward to show that ñ ˆn, snce the profts realzed from each country are potentally larger f t s possble to reduce taxes by reallocatng ncome to a tax haven. To determne the ncentve to set up a tax haven nvestment, let Π th = ψ th + (1 τ ) [ ρ ψ th a ψ th ρ ] c (6) denote the maxmal after-tax proft generated n nonhaven country when the company nvests n a tax haven, and smlarly, let Π nth = (1 τ)ψ nth [ + (1 τ ) ρ ψ nth a ψ nth ρ ] c (7) denote the maxmal after-tax proft generated n a nonhaven country when the company does not nvest n a tax haven but can shft ncome to a nonhaven country wth tax rate τ. Let Inc th denote the net beneft ( Incentve ) from nvestng n a tax haven at a setup cost of c 0. Then nvestng n a tax haven s optmal f and only f Inc th ˆn =1 [Π th Π nth ]+ ñ =ˆn+1 Π th c 0 > 0. (8) The man focus of our paper s to determne the mpact of tax rates on the ncentve to set up a tax haven. Proposton 1 descrbes how a change n tax rates n countres n 5 We assume for smplcty that the multnatonal shfts profts to one country only. Gvng up ths assumpton would yeld computaton far more complcated but would not affect our results qualtatvely. whch the multnatonal s actve ndependent of a tax haven nvestment, that s, = 1,...,ˆn, affects the ncentve to set up a tax haven. Proposton 1. The larger the tax rate n countres n whch the multnatonal s actve ndependent of a tax haven nvestment, the larger s the ncentve to establsh a tax haven afflate: dinc th dτ = d(πth [ + a Π nth ) dτ ψ th ρ a = (ψ th ψ nth ρ ψ nth ) ] > 0 = 1,...,ˆn. Proof. See the onlne supplementary materal. Intutvely, when a company has a tax haven afflate, t avods profts n nonhaven countres more than when t does not. Thus, the larger are the tax rates n the nonhaven countres, the larger s the multnatonal s ncentve to avod taxes. Ths s the man predcton of the model tested n secton VI. For the nterpretaton of our emprcal results, t s nstructve to evaluate also the effect of tax changes n countres n whch the multnatonal s actve only n case of a tax haven nvestment (see the supplementary materal for a dervaton): [ ] dinc th = dπth = ρ ψ th a ψ th < 0 dτ dτ ρ = ˆn + 1,...,ñ. (10) Ths result has the notable mplcaton that a multnatonal may n fact be tempted to nvest n a tax haven followng a tax reducton n a country n whch t has not been present so far. Ths prma face counterntutve stuaton can arse f ths tax reducton makes an nvestment n ths country attractve and hence adds to the potental base for proft shftng. It remans the case that, condtonal on foregn pretax ncome, hgher foregn tax rates ncrease a frm s demand for tax haven operatons, but foregn tax changes nfluence the producton of pretax ncome n a way that can produce an uncondtonal assocaton of hgher foregn tax rates wth reduced tax haven demand. III. Takng Theory to the Data In ths secton we dscuss some conceptual ssues to be consdered n order to test the predcton lad out n proposton 1 and nterpret the emprcal results n sectons VI and VII. Frst, the multnatonal can be engaged n several countres, and hence the tax rates of all these countres matter, but potentally to a dfferent extent. Thus, ndvdual tax rates have to be weghted. Second, we are not able to observe actual profts, only reported after-tax profts. Thrd, the foregn tax rates observed are potentally affected by tax (9)

6 MULTINATIONAL FIRMS AND TAX HAVENS 717 haven nvestments, so the endogenety needs to be addressed. Fourth, frms may exhbt heterogeneous costs of proft shftng and hence may dffer n ther senstvty to tax rate changes. A. Weghtng Indvdual Tax Rate Changes When attemptng to dentfy the effect of foregn tax rates on the tax haven decson, we need to take nto account that the multnatonal s potentally engaged n several countres and therefore the tax rates of all these countres matter. As equaton (11) shows, they do so to a dfferent extent, however, dependng on the proftablty of the ndvdual afflates (see the supplementary materal): d Inc th dτ dρ = τ( τ) > 0 = 1,...,ˆn; (11) a(1 τ ) that s, hgher profts ncrease the effect of the foregn tax rate. We capture ths by nvestgatng the mpact of the average nonhaven tax rate, where all the foregn tax rates are weghted by the proftablty of the ndvdual afflate. If the multnatonal has not nvested n a tax haven, ths average foregn tax rate s determned by the tax rates n countres = 1,...,ˆn and s gven by ˆn =1 τ ρ. (1) ˆn =1 ρ B. Unobservablty of Actual Profts In our emprcal analyss, we encounter the dffculty that we are not able to observe the actual profts ρ n country, only reported after-tax profts (1 τ )π. These reported profts are dstorted due to taxaton and ncome reallocaton. In case of a tax haven nvestment, they are gven by ( (1 τ )π = (1 τ ) ρ ψ a ) ψ ρ [ = (1 τ ) 1 τ ] ( τ ) ρ a(1 τ ). (13) Inspecton shows that ths dstorton rses wth the country s tax rate τ. Thus, we requre approprate proxes to capture the effect of an afflate s proftablty on the decson to nvest n a tax haven. We wll proxy the afflate s proftablty by the number of employees, as dscussed n more detal below. C. Endogenety of Observed Tax Rates to Tax Haven Investments Furthermore, we need to account for the fact that the average foregn tax rate we observe s potentally affected by the very fact of whether the multnatonal holds a tax haven afflate. The tax haven nvestment may make t proftable to nvest n foregn countres = ˆn + 1,...,ñ, that would not have been attractve destnatons for nvestments wthout the ncome reallocaton opportuntes created by the tax haven nvestment. Consder a change n tax rates Δ 0 n countres = 1,...,ñ such that the nvestor chooses to nvest n a tax haven after ths change n tax rates but would not do so before. Accordng to equatons (9) and (10), both an ncrease n the tax rates at locatons = 1,...,ˆn, where the multnatonal already holds an afflate, and a decrease n the tax rates at locatons = ˆn + 1,...,ñ, whch become attractve only after tax haven nvestment, could render tax haven nvestment optmal. The average nonhaven tax rate for the nvestor changes from the status quo descrbed n equaton (1) to the new average nonhaven tax rate: ñ =1 (τ + Δ )ρ ñ =1 ρ. (14) Takng the dfference between equatons (14) and (1) yelds the observed change n the nonhaven average tax rate, whch can be rewrtten as ˆn =1 Δ ρ + ñ =1 ρ ñ =ˆn+1 ρ ( ) ñ=ˆn+1 (τ ρ +Δ ρ ) ˆn=1 τ ρ ñ=ˆn+1 ρ ˆn=1 ρ ñ =1 ρ. (15) In our emprcal analyss, we are nterested n dentfyng the effect of exogenous changes n tax rates n countres = 1,...,ˆn, that s, countres n whch the multnatonal would nvest even wthout a tax haven afflate. Ths exogenous change n tax rates s captured by the frst term. As shown n equaton (9), nvestng n a tax haven s postvely nfluenced by an ncrease n the tax rates of the countres n whch the multnatonal already holds afflates. Thus, when estmatng the mpact of foregn tax rates, hgher tax rates n countres n whch a multnatonal frm would nvest under any crcumstances should stmulate greater demand for tax haven afflates. The second term captures the change n the observed nonhaven tax rate that s due to the endogenety of the multnatonal s nvestment decson. Evaluatng the numerator of the second term, we fnd that the observed change n the average nonhaven tax rate exceeds the change of nterest f the new afflates the multnatonal opens due to the tax haven nvestment are located n countres that exhbt on average hgher tax rates than the prevous average tax rate, and the converse. Ths has mportant mplcatons for the nterpretaton of the causal effects of tax changes. In partcular, OLS results overestmate the true effects, as captured by the IV estmates, f the tax rates at the frm s new locatons ncrease the frm s average foregn nonhaven tax rate, and underestmate the true effects f the tax rates faced at the new locatons are lower than the prevous average foregn nonhaven tax rate. Consequently, there are several reasons that smple correlatons of foregn nonhaven tax rates wth tax haven afflate

7 718 THE REVIEW OF ECONOMICS AND STATISTICS ownershp mght not reflect the drect effect of hgher foregn tax rates on tax haven demand. The frst, as noted n equaton (10), s that lower foregn tax rates may encourage frms to earn greater taxable foregn ncome, whch drectly ncreases ther demand for tax haven operatons. The second, as reflected n the second term of equaton (15), s that ownershp of a tax haven afflate changes the pattern of nonhaven foregn nvestment, resultng n hgher or lower measured average foregn nonhaven tax rates. A thrd reason, not captured by ths model, s that frms wth nvestments n tax havens may generally nvest n other low-tax countres n order to reallocate some of ther ncome to countres wth tax rates not qute as low as those offered by tax havens. These consderatons suggest that OLS estmates of the effect of nonhaven foregn tax rates on demand for tax haven afflates may understate the true effect. The emprcal work presented n secton VI accounts for the potental endogenety of the observed tax rate usng an nstrumentaton strategy based on the locatons of foregn nonhaven afflates at the start of the sample perod and restrcts attenton to changes n observed tax rates for these locatons only. D. Heterogeneous Cost of Proft Shftng Our emprcal strategy also needs to take nto account that frms may dffer wth respect to ther cost of reallocatng profts between countres. As equaton (16) shows, the more dffcult proft shftng s for the multnatonal, the less senstve wll be ts reacton to foregn tax rate changes (see the supplementary materal): d Inc th dτ da = ρ τ( τ) < 0 = 1,...,ˆn. (16) a (1 τ ) Average foregn tax rates and values of the shftng cost parameter are lkely to dffer among frms and may vary systematcally among ndustres. Industres may dffer n average values of the shftng cost parameter a, reflectng dfferences n the mportance of ntangble assets and other busness features that facltate proft reallocaton; ndustres may also dffer n the extent to whch a vares among frms n the ndustry. Dfferentatng equaton (16) wth respect to a ndcates that d 3 Inc th dτ d a = ρ τ( τ) > 0. (17) a 3 (1 τ ) Snce the expresson n equaton (17) s postve, t follows that the effect of a on dinc dτ s nonlnear and, more specfcally, that a mean-preservng spread n the dstrbuton of a produces a greater average value of dinc dτ. Consequently, ndustres n whch frms have very dfferent costs of proft reallocaton should be expected to dsplay greater average senstvty of tax haven demand to nonhaven tax rates than do other ndustres, although average costs of proft reallocaton do not dffer. We use frm-fxed effects n the baselne econometrc analyss to control for dfferences n margnal costs of ncome reallocaton and dstngush frms by ndustral sectors to proxy for cost dfferences that vary wth ndustry. IV. Data and Descrptve Statstcs Our analyss s based on the Mcrodatabase Drect nvestment (MD) provded by the Bundesbank, the German central bank. We use the nformaton on outward foregn drect nvestment by German companes. The database conssts of a panel of yearly nformaton on the foregn afflates of German frms for the perod from 1996 untl 008. By the German Foregn Trade and Payment Regulaton (Aussenwrtschaftsverordnung), any resdent who holds shares or votng rghts of at least 10% n a company wth a balance sheet total of more than 3 mllon euros s oblged to report nformaton on the fnancal characterstcs of these afflates to the Bundesbank (Lpponer, 009). 6 The same nformaton has to be provded on branches or permanent establshments abroad f ther operatng assets exceed 3 mllon euro. The comprehensveness of these data suggests that they can be used to draw a very relable pcture of the foregn nvestment of German companes. The MD nformaton on parent companes starts n 00, so the analyss s restrcted to 00 to 008 n order to avod the effects of the crash of 008 and ts aftermath. 7 Durng the perod, the MD contans 173,31 afflateyear observatons. Some afflates are reported several tmes, because multple nvestors hold partcpatng nterests n them. We focus our analyss on drectly held foregn afflates and thus abstract from more complex ncentve structures that may exst n multlevel holdng chans. 8 Ths lmts the analyss to 117,585 afflate-year observatons. For consstency across parents, we delete 18 observatons for whch the degree of partcpaton of the parent s smaller than the reportng requrement of 10%. In addton, we drop observatons on parents n a number of sectors, ncludng government nsttutons and prvate households. We drop observatons on parents n the fnancal sector, because they are subject to specal balancng requrements. We delete the sectors housng enterprses and other real estate actvtes, as they report nether sales nor employees, whch we wll use as the sze measure n our analyss. Smlarly, we drop the sector holdng companes as reported sales and employees are very often zero, even though these companes are 6 The reportng thresholds have changed several tmes n the past. We refer only to the reportng threshold as of 00 that s relevant to us. 7 Pror to 00 Germany taxed corporatons usng a splt-rate system that mposed a lower rate of tax on profts dstrbuted to shareholders as dvdends and taxed dvdends usng an mputaton system that granted credts to shareholders for taxes pad by corporatons on ther domestc (but not foregn) profts. Ths dvdend mputaton system may have reduced ncentves to reallocate taxable ncome from Germany to foregn tax havens but was abolshed n For an n-depth dscusson of the complex determnants of ownershp chans, see Wechenreder and Mntz (010).

8 MULTINATIONAL FIRMS AND TAX HAVENS 719 Table 1. Choce of Tax Havens, by Sectoral Group Parent Sector Manufacturng Servce Fnancal Total number of parent years 11,603 6,733,506 Of whch wth tax haven afflate 1,976 1, Of whch nternatonally actve parents 77.3% 59.01% 57.19% wth more than one tax haven afflate.87% 17.50% Tax Tax Tax Nonhaven Haven Nonhaven Haven Nonhaven Haven Number of afflate years 33,03,89 14,47 1,768 7,897,94 Of whch In manufacturng sector 51.19% 3.63% 1.08% 4.81% 3.89% 0.74% In servce sector 46.69% 63.56% 8.0% 90.16% 15.35% 18.09% In fnancal sector 1.38% 3.39% 4.76% 4.58% 79.84% 81.17% Other 0.75% 0.4% 0.96% 0.45% 0.9% Mean number of afflates per parent Choce of Haven Manufacturng Servce Fnancal Afflate Parent Afflate Parent Afflate Parent Years Years Years Years Years Years Bg havens (more than 1 mllon nhabtants) Hong Kong Ireland Lebanon Lbera Panama Sngapore Swtzerland 1,368 1, Small havens (fewer than 1 mllon nhabtants) Bermuda Brtsh Vrgn Islands Cayman Islands Cyprus Channel Islands Luxembourg Malta Other Total,89,579 1,768 1,588,94 1,69 Empty cells denote tax havens where fewer than three afflate-years or parent-years are observed, so the exact number of nvestments must not be reported for confdentalty reasons. Manufacturng frms: frms classfed NACE ; servce frms: frms classfed NACE , wth the sample restrctons already noted; fnancal frms: frms classfed NACE If a parent nvests n several tax havens, t s counted multple tmes (once per tax haven). not small. 9 We later remove ths restrcton as a robustness check and fnd that our results are unaffected. We fnally obtan a sample of 54,367 afflate-year observatons that correspond to 19,165 parent-year observatons. The observatons are dstrbuted evenly across years, wth a mnmum of,639 observatons and a maxmum of,875 observatons. We augment the MD wth nformaton on statutory tax rates manly from the Internatonal Bureau of Fscal Documentaton (00 006, ) and nformaton on GDP from the Internatonal Monetary Fund (IMF). We use the defnton of tax havens derved by Hnes and Rce (1994), whch s wdely accepted n the lterature and was only recently used by Dharmapala and Hnes (009). 10 Alternatvely, we could have used the defnton propagated by 9 We also delete 331 afflate-year observatons for parents that are not classfed holdngs, but are de facto holdngs after consultatons wth the staff of the Bundesbank. 10 Swtzerland s ncluded among the tax havens on ths lst and s commonly consdered one of the world s most mportant tax havens from the standpont of corporate tax plannng. Of the sx examples of sophstcated nternatonal tax avodance by U.S. frms analyzed n detal by the Jont Commttee on Taxaton (010), three rely on the use of Swss afflates. the OECD (000). We chose Hnes and Rce s (1994) tax haven defnton to derve results that are comparable to the lterature, n partcular the study by Desa, Foley, and Hnes (006). Further, no OECD member countres appear on the OECD s tax haven lst, whch thereby omts a number of tax havens popular wth German frms, such as Swtzerland. Table 1 presents descrptve evdence on the use of tax havens by sectoral group. For comparatve purposes, nformaton on fnancal frms s provded n addton to nformaton on frms n the manufacturng and servce sector, whch we analyze n later sectons. On average, a tax haven afflate s held n 0.4% of parentyears (17.9% excludng fnancal companes). Ths fgure seems low by nternatonal standards: Desa et al. (006) report that tax haven nvestment s observed for 59% of U.S. multnatonal parent companes n Ths dfference We are unable to dstngush nvestment n Monaco and Sant Martn from nvestment n France, so these tax havens are neglected n our analyss. 11 Desa, Foley, and Hnes (004) use the same data as Desa et al. (006) and report that U.S. parents own between 7.5 and 7.8 afflates on average n the years 198, 1989, and Parents n our sample average only.8 foregn afflates (4.0 f ndrectly held afflates are ncluded).

9 70 THE REVIEW OF ECONOMICS AND STATISTICS reflects, n part, the nclusveness of the MD data, n that the sze thresholds for reportng are much lower than n the U.S. data analyzed by Desa et al. (006), resultng n a hgher proporton of small frms and those wth relatvely small foregn operatons. The proporton of frms ownng tax haven afflates s hgher for servce frms (19.9%) than for manufacturng frms (17.0%), and a larger proporton of servce frms own a tax haven afflate but are not nternatonally actve n nonhaven countres. About a ffth of both manufacturng and servce frms that are present n tax havens own more than one tax haven afflate, and the mean number of tax haven afflates s also approxmately equal. In contrast, 37.% of fnancal frms hold afflates n tax havens, and they own on average twce as many tax haven afflates as do manufacturng and servce frms. The share of afflates n tax havens that are n the servce sector s dsproportonately hgh. For manufacturng frms, the share of servce afflates n tax havens s about 17 percentage ponts hgher than ther overall share of afflates n the servce sector, and for servce frms, t s 8 percentage ponts hgher. Also for fnancal companes, nvestment n servce afflates s more common n tax havens than n nonhaven countres. The lower panel of table 1 reports the number of afflateyear and parent-year observatons by tax haven and sectoral group of the parent frm. It shows that the preferred tax haven destnaton vares by sectoral group. Manufacturng frms clearly prefer the bg tax havens. More than 90% of observatons are accumulated there about 48% n Swtzerland alone. The sland tax havens, n partcular Bermuda, the Cayman Islands, and the Channel Islands, are very rare nvestment destnatons. Swtzerland s smlarly popular among servce frms; about half of ther tax haven afflates are located there. Servce frms more extensvely use the small havens, where almost a ffth of tax haven afflates are located, most promnently 9% n Luxembourg. For fnancal companes, Luxembourg s dstnctly the most popular destnaton, wth 38% of afflate-year observatons n tax havens. The Cayman Islands are ther fourth most mportant tax haven destnaton: 10% of afflate-year observatons n tax havens are located there. Evdently the attractveness of tax havens strongly depends on sector characterstcs. Fgures 1 and provde local polynomal plots of the relatonshp of tax haven nvestment and foregn nonhaven taxaton for manufacturng and servce frms wth at least two foregn afflates n the pooled sample. The x-axs shows the average of the statutory tax rates faced by a parent frm s foregn afflates n nonhaven countres weghted by number of employees, whch s the measure of foregn nonhaven taxaton used n the emprcal analyss. 1 Foregn tax rates are weghted by number of employees to reflect szes of foregn operatons, and therefore potental producton of taxable 1 The number of employees s adjusted by the partcpaton of the parent n the afflate. Tax haven use Fgure 1. Manufacturng Frms Average foregn non-haven tax rate Ths fgure dsplays the graph resultng from a kernel-weghted local polynomal regresson of the relatonshp of tax haven use and the average foregn nonhaven tax rate by frms wth at least two foregn afflates for manufacturng frms. The average foregn nonhaven tax rate s measured n percent, and tax haven use s dsplayed as a fracton of frms; the shaded regon s the 95% confdence nterval. Observatons n the lowest and hghest percentle of the average foregn nonhaven tax rate are excluded. Tax haven use Fgure. Servce Frms Average foregn non-haven tax rate Ths fgure dsplays the graph resultng from a kernel-weghted local polynomal regresson of the relatonshp of tax haven use and the average foregn nonhaven tax rate by frms wth at least two foregn afflates for servce frms. The average foregn nonhaven tax rate s measured n percent; tax haven use s dsplayed as fracton of frms; the shaded regon s the 95% confdence nterval. Observatons n the lowest and hghest percentle of the average foregn nonhaven tax rate are excluded. ncome subject to dfferent tax rates; weghtng foregn tax rates by assets or sales produces smlar results, though wth smaller samples due to data lmtatons. The varable on the y-axs s tax haven use, a bnary varable that s equal to 1 whenever a parent nvested n at least one tax haven n a gven year. For both servce frms and manufacturng frms, the cross-sectonal relatonshp of taxaton and tax haven use ncreases over a wde range of tax rates. The ncrease for manufacturng frms s steeper than for servce frms, though both largely level off at average foregn tax rates above 5% to 8% We explored an emprcal specfcaton ncludng a quadratc term of the foregn nonhaven tax rate, but a quadratc relatonshp s not robustly supported by the data.

10 MULTINATIONAL FIRMS AND TAX HAVENS 71 Table. Summary Statstcs, Regresson Sample Full Sample Manufacturng Frms Servce Frms Mean Mean Mean Mean SD Dfference Mean SD Dfference Mean SD Dfference Have haven [wthn varaton] Number of parent employees (n 1,000) Ln (number of parent employees) Number of nonhaven employees (n 1,000) Ln (number of nonhaven employees) Average foregn nonhaven tax rate Average tax rate at 001 nonhaven locatons Average market sze Average dstance Average regulatory qualty Average rule of law Average control of corrupton Observatons 1,755 8,533 3,751 Full sample ncludes parent frms n the sectors agrculture, mnng, electrcty and water supply, and constructon, n addton to manufacturng and servce frms. Mean dfference by haven status; base category: does not hold an afflate n tax haven. *p < 0.10, **p < 0.05, ***p < Manufacturng frms: frms classfed NACE ; servce frms: frms classfed NACE , wth the already noted sample restrctons. Have haven: ndcator varable; 1 f parent frm owns at least one afflate n at least one tax haven n a gven year, 0 otherwse. Number of parent employees: number of employees, parent frm. Number of nonhaven employees: sum of number of employees n afflates not located n tax havens, reduced accordng to share of partcpatng nterests. Average foregn nonhaven tax rate: Average of the statutory tax rates faced by foregn afflates n nonhaven locatons weghted by afflate number of employees, adjusted by share of partcpaton where approprate. Average tax rate at 001 nonhaven locatons: Average of the statutory tax rates faced by foregn afflates n nonhaven locatons of 001 weghted by GDP. Average market sze: mean GDP of foregn nonhaven countres where frms nvest. Average dstance: mean dstance of these foregn nonhaven countres (source: CEPII). Average regulatory qualty, rule of law, control of corrupton: mean scores of these foregn nonhaven countres from the World Governance Indcators (WGI) for the ndcators regulatory qualty, rule of law, and control of corrupton. The numbers of observatons are 1,731, 8,518, and 3,74 for the WGI ndcators because data are not avalable for Tawan. The top part of table provdes an overvew of the man varables used n our regresson analyss for the full sample and the two subgroups we consder. We explan the varables n detal n the next secton. The proporton of frms nvestng n a tax haven s lower (around 15%) than for the full data and equal across sectoral groups, because frms nvestng only n a tax haven drop from the regresson sample. As frms wth 0 sales or employees drop, the average sze of the frms used n our regressons s slghtly hgher than the average sze of all frms n the sample. The statstcs of the average foregn nonhaven tax rate and the nstruments for the regresson sample are smlar. 14 The thrd columns for every group report mean dfference tests of the man regressors by the dependent varable. Frms that nvest n a tax haven are on average sgnfcantly larger, both domestcally and nternatonally. In addton, they face sgnfcantly hgher average foregn tax rates, whch s consstent wth the ncentves dscussed earler. The bottom part of table compares the characterstcs of foregn nonhaven nvestment destnatons of frms wth and wthout tax haven afflates. Frms wth tax haven afflates appear to nvest n countres that are smaller on average, though there s no sgnfcant dfference n market sze for servce frms. They nvest n countres that are more dstant from Germany. Concernng the nsttutonal qualty as measured by the World Governance Indcators (WGI), ther nvestment destnatons exhbt a sgnfcantly lower regulatory qualty and, for servce frms, a sgnfcantly lower 14 We drop 4.1% of manufacturng frms and 8.6% of servce frms because only nvestment n tax havens s observed. We drop 4.0% of the remanng manufacturng frms and 18.0% of servce frms due to ther 0 number of employees. Table C.1 n the supplementary materal presents the summary statstcs for the full data. percepton of the rule of law. The control of corrupton seems to be slghtly better n destnaton countres of frms that nvest n tax havens, at least for the full sample and manufacturng frms. V. Emprcal Approach As outlned n secton II, a multnatonal frm s decson to nvest n a tax haven depends on the taxaton t faces at ts foregn nonhaven locatons, ts margnal cost of reallocatng taxable ncome, and the fxed cost of tax haven nvestment. Condtonal on pretax foregn ncome, the probablty of tax haven nvestment should ncrease as foregn nonhaven tax rates rse, wth ths effect beng strongest for frms wth low costs of reallocatng profts. We specfy the followng lnear probablty model: y jt = β 0 + β 1 τ jt + β p jt + β 3 p jt + β 4nh jt + β 5 nh jt + α j + γ t + u jt. (18) The dependent varable y jt s a dummy that s equal to 1 f a frm j holds at least one afflate n at least one tax haven n a year t. Our ndependent varables are τ jt, the average of the statutory tax rates faced by j s nonhaven afflates n t weghted by the number of afflate employees; p jt, the natural log of the sze of company j n perod t and ts square, p jt ; nh jt, the natural log of the sze of j s foregn nonhaven actvtes n t and ts square, nh jt ; α, the frm-specfc costs of reallocatng profts across countres; and γ t, a year fxed effect. The coeffcent of man nterest s β 1. As ndcated above, we expect β 1 > 0. It captures the effect of the taxes leved on the profts of a multnatonal s foregn nonhaven afflates on the probablty that t nvests n a tax haven. Equaton (11)

11 7 THE REVIEW OF ECONOMICS AND STATISTICS mples that greater frm proftablty ncreases the mpact of nonhaven tax rates on the lkelhood of nvestng n a tax haven afflate. Thus, we use a weghted, not a smple, average of the foregn nonhaven tax rates. For ths purpose, we use the number of employees of the afflate, adjusted by the degree of partcpaton of the parent, to approxmate the relatve mportance of an afflate for a multnatonal group. We also explore alternatve weghts and fnd that our results are largely unaffected (see supplementary table C.5). There are two potentally mportant challenges n estmatng equaton (18): the endogenety of the observed foregn nonhaven tax rates and unobserved frm heterogenety n the cost of nvestng n a tax haven. The varable cost of advantageously usng a tax haven afflate should vary wth frm-specfc characterstcs such as the extent to whch a frm holds and uses ntangble assets such as ntellectual property. The locaton of ntangble assets, lcense arrangements, and royalty payments have been shown to be used as ncome reallocaton tools (e.g., Dschnger & Redel, 011; Karknsky & Redel, 01). 15 A frm wth larger ntangble assets should have greater dscreton n choosng transfer prces due to the lack of comparable market transactons. Thus, and as suggested by equaton (16), the response to changes n foregn taxaton should vary across frms dependng on ther margnal cost of ncome reallocaton. These frm-specfc characterstcs are, however, unobservable. We take two measures to address ths ssue. Frst, we conduct our analyss separately for the group of manufacturng frms (NACE ) and for the group of servce frms (NACE , wth the already noted restrctons), 16 out of concern that manufacturng frms may rely to a sgnfcantly greater degree on ntangble property. Furthermore, manufacturng and servce frms may have dfferng tax and nontax determnants of foregn locaton and may dffer n the extent to whch foregn locatons can substtute for each other. One observable ndcator of dfferences between manufacturng and servce frms s that manufacturng frms are more R&D ntensve. Usng sector-level data from the Innovaton Survey of the Center for European Economc Research (Zentrum für Europäsche Wrtschaftsforschung, ZEW) for the years 1996 to 008, we fnd that the average R&D ntensty for the manufacturng sector s twce as hgh as the R&D ntensty for servce sectors. At the same tme, the descrptve evdence provded n secton IV shows that the proporton of servce frms ownng tax haven afflates and the share of servce frms afflate years observed n tax havens are hgher than the correspondng statstcs for manufacturng frms. Frms n servce ndustres have operatons n tax havens n part to provde busness servces to other frms located there; consequently, some of the attracton may stem not from ther own tax avodance but from 15 In addton, a varant of ths type of strategy s part of all sx case studes of the report by the Jont Commttee on Taxaton (010) prepared for the publc hearng before the House Commttee on Ways and Means. 16 Ths mples that we do not consder parent frms n agrculture, mnng, electrcty and water supply, and constructon n our analyss. that of others. Furthermore, there may be lower fxed costs of tax haven nvestment for servce frms, so not only taxaton but also parent sze could have a dfferental mpact on tax haven nvestment by manufacturng and servce frms. In addton, the share of servce afflates of manufacturng parents located n tax havens s dsproportonately hgher than the share located n nonhaven countres. Overall, there s thus reason to beleve that the processes governng tax haven nvestment by servce frms and by manufacturng frms may dffer sgnfcantly, whch suggests that they should be analyzed separately. We report the results of a full-sample analyss n the supplementary materal n table C.. The second way n whch the analyss addresses unobserved frm heterogenety s through the use of frm fxed effects to capture the nfluence of frm-specfc dfferences n the margnal cost of ncome reallocaton, at least to the extent that they are approxmately constant over the sample perod. Fxed effects also account for unobserved frm-specfc characterstcs such as the degree of tax senstvty, that s, the mportance that a frm assgns to the amount of ts tax payments, whch may render frms ex ante more or less lkely to nvest n tax havens. 17 Lkewse, the data provde nformaton on the sector of the afflates mostly at the two-dgt NACE Rev. 1 level, so partcular ncentve schemes for frms n subsectors cannot be taken nto account, and we do not have nformaton on the subnatonal locaton of frms, so we cannot account for local taxaton. The use of frm-fxed effects controls for tme-nvarant aspects of these frm attrbutes. In estmatng equaton (18), t s necessary to take nto account that the average foregn nonhaven tax rate s endogenous because entry n a tax haven may have a feedback effect on the optmal locaton decsons of a frm, as dscussed n secton III. To address ths ssue, we use an nstrumentaton strategy based on the ntal locaton decsons of the frm. We fx the locaton decsons of the frm n the year 001, the year pror to our analyss perod, and nstrument the observed average foregn nonhaven tax rate wth a hypothetcal average foregn nonhaven tax rate had the frm not changed the locatons of ts foregn afflates. The nstrument thus captures only changes n frms foregn nonhaven taxaton that result from changes n tax rates, not from changes n locaton decsons. 18 To counter any endogenety that may be left on the ntensve margn, we use GDP nstead of the number of employees as weghts when calculatng our nstrument. 19 Our nstrumentaton strategy thus elmnates the endogenety n the observed average foregn nonhaven tax rate by leveragng the nsghts drawn from our theoretcal analyss: wth the endogenety stemmng from frms adjustng ther nonhaven nvestment due to tax haven nvestment, 17 Ths ssue has been rased but not addressed n Desa et al. (006). 18 Sxty-nne of the 101 countres n the 00 sample changed ther tax rates over the perod. In most cases, these changes were reductons, though there s suffcent cross-sectonal varaton (ncludng some tax rate ncreases) to dentfy tax rate effects. 19 Earler studes ndcate that GDP correlates very closely wth foregn nvestment and foregn proftablty n both an aggregate cross secton (e.g., Hnes & Rce, 1994) and a frm-level panel (e.g., Desa et al., 006).