research paper series

research paper seres Research Paper 2004/10 Intra-Industry Trade wth Multnatonal Frms: Theory, Measurement and Determnants by Hartmut Egger, Peter Egger and Davd Greenaway The entre acknowledges fnancal support from The Leverhulme Trust under Programme Grant F114/BF

The Authors Hartmut Egger s Senor Assstant at the Unversty of Zurch. Peter Egger s Professor of Economcs at the Unversty of Innsbruck. Davd Greenaway s Professor of Economcs at the Unversty of Nottngham and Drector of the Leverhulme entre for Research on Globalsaton and Economc Polcy (GEP). Acknowledgements Davd Greenaway acknowledges support from the Leverhulme Trust under programme grant F114/BF.

Intra-Industry Trade wth Multnatonal Frms: Theory, Measurement and Determnants by Hartmut Egger, Peter Egger and Davd Greenaway Abstract A number of recent developments, ncludng the analyss of frm level adustment to fallng trade costs, have contrbuted to a revval of nterest n ntra-ndustry trade. Most emprcal work stll reles on the standard Grubel-Lloyd measure. Ths however refers only to nternatonal trade, dsregardng ncome flows stmulated by repatrated profts. Gven the overwhelmng mportance of the latter, ths s a maor shortcomng. We provde a gude to measurement and estmaton of the determnants of blateral ntra-ndustry trade shares from the perspectve of new trade theory wth multnatonal frms. We develop an analytcally solvable general equlbrum model to nvestgate nvestment costs, multnatonal actvtes and ncome flows from repatrated profts. The robustness of our fndngs are nvestgated n fve smulaton analyses. We also dscuss and quantfy bases of dfferent Grubel-Lloyd ndces n an emprcal assessment of ntra-ndustry trade shares and dentfy repatrated proft flows of multnatonals as a key determnant of based measurement. To overcome ths, we provde several alternatve, bas-corrected versons of the Grubel-Lloyd ndex. Fnally, we demonstrate that the determnants motvated by our theoretcal analyss offer mportant nsghts nto varatons n the Grubel-Lloyd ndex. Our new specfcaton outperforms any other prevously estmated model as llustrated n regressons on numercally generated data. JEL classfcaton: F12, F23 Keywords: ntra-ndustry trade, multnatonals Outlne 1. Introducton 2. Theoretcal background 3. Emprcal analyss 4. onclusons

Non-Techncal Summary The publcaton of Grubel and Lloyd (1975) stmulated enormous nterest n ntra-ndustry trade (IIT), for two reasons. Frst, the emprcal phenomenon of hgh levels of trade n products from smlar ndustres between countres wth smlar factor endowments seemed to be at odds wth the standard Heckscher- Ohln-Samuelson (HOS) workhorse model of nternatonal trade. Second, the observed ncrease n ntrandustry trade concded wth what appeared to be relatvely panless adustment to economc ntegraton n western Europe. The dslocaton antcpated as nter-ndustry specalsaton occurred dd not materalse, gvng rse to the so-called smooth adustment hypothess. In the decade that followed Grubel and Lloyd (1975) the lterature exploded. Emprcal analyss focused prmarly on three thngs. Frst whether the phenomenon survved data dsaggregaton. Second, was IIT a pecularty of trade n western Europe? Thrd, what were the drvers of the phenomenon? Recent years have seen a revval of nterest n ntra-ndustry trade, stmulated by fronter work on trade costs, economc geography and a range of aspects of frm level adustment to globalsaton. One focus of ths, from both a theoretcal and measurement standpont s ntra-ndustry trade n a settng wth multnatonal frms. Ths s a very mportant development from a theoretcal standpont because we have known for a long tme that both phenomena co-exst, ndeed are often co-termnous and we need good models for explanng ths. But t s also mportant from a measurement perspectve because of the mportance of nternatonal producton and ntra-frm trade relatve to armslength trade. FDI has grown about twce as fast as trade over the last decade. The prncpal sources and hosts are ndustralsed countres and two-way trade s closely assocated wth two-way FDI. Ths paper contrbutes to ths new lterature n several ways. Frst, t generates a proof that the standard and stll wdely used Grubel-Lloyd ndex has to be adusted to reflect more than the ntra-ndustry trade share n a narrow sense. We buld a general equlbrum model whch shows that wth multnatonal frms, both unbalanced proft repatraton and trade costs dstort the ndex. We expose the bases resultng from these emprcally relevant phenomena and construct several new versons of bas-corrected Grubel-Lloyd ndces. Second, we develop a three-factor general equlbrum model of trade and multnatonals to provde a detaled analyss of the role of nvestment cost dfferences between countres as a determnant of FDI and, hence, ntra-ndustry trade. By ntroducng three factors, we emphasse the dstncton between two mportant characterstcs of headquarters: ther provson of physcal captal to set up plants, and the human-captal ntensve generaton of frm-specfc assets through brand prolferaton. Besdes ths more complete descrpton of headquarter servces, there s an advantage of analytcal tractablty snce there are as many actvtes (homogeneous goods producton, exporter and multnatonal producton of manufactures) as there are factors (physcal captal, sklled labour, unsklled labour). In ths settng, we are able to evaluate not only the role of nvestment cost levels and dfferences n general but also ther nteracton wth labour and captal endowments, dependng on whether horzontal or vertcal multnatonals are actve. Thrd, a large number of numercal smulatons of our model allow us to evaluate the robustness of our analytcal fndngs wth respect to smplfyng assumptons as well as tradtonal determnants such as country sze, captal-labour ratos and sklled-unsklled ratos.

Fnally, we mplement and report on an extensve emprcal analyss, where uncorrected and bascorrected versons of the Grubel-Lloyd ndex are used as regressors. Ths yelds several conclusons. We fnd that bases not only affect the overall magntude of the Grubel-Lloyd ndex but also systematcally affect parameter estmates; cross-secton estmates tend to be nconsstent f country-specfc effects are excluded; the determnants generated by our theoretcal model account for more than 50% of the varaton n ntra-ndustry trade-share data, mplyng that less than half of ther varaton s explaned by tradtonally used varables. Gven the crucal mportance of estmatng accurately ntra-relatve to nter-ndustry trade, ths s very sgnfcant.

1 Introducton The publcaton of Grubel and Lloyd (1975) stmulated enormous nterest n ntra-ndustry trade (IIT), for two reasons. Frst, the emprcal phenomenon of hgh levels of trade n products from smlar ndustres between countres wth relatvely smlar factor endowments seemed to be at odds wth the standard Heckscher-Ohln-Samuelson (HOS) workhorse model of nternatonal trade. Second, the observed ncrease n ntra-ndustry trade concded wth what appeared to be relatvely panless adustment to economc ntegraton n western Europe. The dslocaton antcpated as nter-ndustry specalsaton occurred dd not materalse, gvng rse to the so-called smooth adustment hypothess. In the decade that followed Grubel and Lloyd (1975) the lterature exploded. Emprcal analyss focused prmarly on three thngs. Frst whether the phenomenon survved data dsaggregaton. Fnger (1975) famously descrbed IIT as a statstcal artefact, a mrage created by the vagares of statstcal classfcaton. Greenaway and Mlner (1983) among others showed that although shares of IIT n total trade declned as trade data became more fnely dsaggregated, t dd not dsappear. In fact t remaned prevalent. Second, was IIT a pecularty of trade n western Europe. Studes n Tharakan (1983) demonstrated that t was not. Although average levels were lower n developng, and what are now referred to as transton economes, they were non-trval. Thrd, what were the drvers of the phenomenon? Early cross-secton work such as Loertscher and Wolter (1980) and Greenaway and Mlner (1984) ponted to varous aspects of ndustral organsaton but fndngs were not robust. Indeed, an applcaton by Torstensson (1996) of extreme bounds analyss confrmed that the cross-ndustry determnants were very fragle. Ths, and other work, progressed thnkng on measurement and to a lesser extent explanaton. Innovatons on the theoretcal front were much more dramatc, wth the development and refnement of models of monopolstc competton and nternatonal trade (most notably Lancaster 1980, Krugman 1979 and 1980 and Helpman and Krugman 1985) as well as strategc nteracton and ntra-ndustry trade (eg Brander 1981 and Brander and Krugman 1982). These offered convncng explanatons of the market structures under whch we would expect IIT to be generated and have proved to be of lastng value. Many, and n partcular Krugman (1981), focused on dstrbutonal consequences, emphassng the lkelhood of 1

greater symmetry between expandng and declnng actvtes than n an HOS world and offerng a theoretcal underpnnng to the potental for lower adustment costs n an IIT settng as compared to HOS. Recent years have seen a revval of nterest n ntra-ndustry trade, stmulated by fronter work on trade costs, economc geography and a range of aspects of frm level adustment to globalzaton. One mportant focus of ths, from both a theoretcal and measurement standpont s ntra-ndustry trade n a settng wth multnatonal frms. Ths s a very mportant development from a theoretcal standpont because we have known for a long tme that both phenomena co-exst, ndeed are often co-termnous and we need good models for explanng ths. But t s also mportant from a measurement perspectve because of the mportance of nternatonal producton and ntra-frm trade relatve to armslength trade. FDI has grown about twce as fast as trade over the last decade. The prncpal sources and hosts are ndustralsed countres and two-way trade s closely assocated wth two-way FDI. An mportant development n understandng the relatonshp between IIT and ntra-ndustry afflate producton s Markusen and Maskus (2001). From a specfcaton based on numercal smulatons of a two-factor knowledge captal model (assocated wth arr et al., 2001 and Markusen, 2002), they fnd that ntra-ndustry trade between the US and partner economes tends to decrease wth greater smlarty n sze, whch s at odds wth the fndngs of Helpman (1987), Bergstrand (1990) or Hummels and Levnsohn (1995). They also found t decreased wth the blateral trade cost level, but ncreased wth the blateral level of nvestment costs. However, apart from these papers, ths ssue remans largely unexplored. Ths paper contrbutes to ths new lterature n several ways. Frst, t generates a proof that the standard and stll wdely used Grubel-Lloyd ndex has to be adusted to reflect more than the ntra-ndustry trade share n a narrow sense. We buld a general equlbrum model whch shows that wth multnatonal frms, both unbalanced proft repatraton and trade costs dstort the ndex. We expose the bases resultng from these emprcally relevant phenomena and construct several new versons of bas-corrected Grubel-Lloyd ndces. Second, we develop a three-factor general equlbrum model of trade and multnatonals to provde a detaled analyss of the role of nvestment cost dfferences between countres as a determnant of FDI and, hence, ntra-ndustry trade. By ntroducng three factors, we emphasse the dstncton between two mportant characterstcs of headquarters: ther provson of physcal captal to 2

set up plants, and the human-captal ntensve generaton of frm-specfc assets through brand prolferaton. Besdes ths more complete descrpton of headquarter servces, there s an advantage of analytcal tractablty snce there are as many actvtes (homogeneous goods producton, exporter and multnatonal producton of manufactures) as there are factors (physcal captal, sklled labour, unsklled labour). In ths settng, we are able to evaluate not only the role of nvestment cost levels and dfferences n general, but also ther nteracton wth labour and captal endowments, dependng on whether horzontal or vertcal multnatonals are actve. Thrd, a large number of numercal smulatons of our model allow us to evaluate the robustness of our analytcal fndngs wth respect to smplfyng assumptons as well as tradtonal determnants such as country sze, captal-labour ratos and sklled-unsklled ratos. Fnally, we mplement and report on an extensve emprcal analyss, where uncorrected and bas-corrected versons of the Grubel-Lloyd ndex are used as regressors. Ths yelds several conclusons. We fnd that bases not only affect the overall magntude of the Grubel-Lloyd ndex but also systematcally affect parameter estmates; cross-secton estmates tend to be nconsstent f country-specfc effects are excluded; the determnants generated by our theoretcal model account for more than 50% of the varaton n ntra-ndustry trade-share data, mplyng that less than half of ther varaton s explaned by tradtonally used varables. Gven the crucal mportance of estmatng accurately ntra-relatve to nter-ndustry trade, ths s very sgnfcant. The remander of the paper s organzed as follows: Secton 2 sets out our theoretcal model of ntra-ndustry trade wth nvestment costs and ntroduces a corrected Grubel-Lloyd ndex. Ths s subected to smulaton analyss and a number of theoretcal propostons are derved. Secton 3 sets up our econometrc analyss, reports our results and subects them to senstvty analyss. Secton 4 concludes. 3

2 Theoretcal background 2.1 The Grubel-Lloyd ndex The Grubel and Lloyd (1971) ndex has become the standard measure for the ntensty of ntra-ndustry trade. In the two-country case, ths s defned as 1 GLI = ( EX IM ) 2 mn, where EXk s the value of country s exports of good k. k k, (1) k EX k k + IM k k IM k represents expendtures for country s mports of good k. Although ths has been the ndex of choce for most researchers n ths area for over 30 years, t s an napproprate measure f there are multnatonal actvtes because GLI does not account for (unbalanced) repatrated profts of multnatonal frms and, therefore, underestmates the ntra-ndustry trade share. For convenence, we use the term trade mbalance bas to refer to ths measurement error. 2 To see ths bas, consder the case of two economes wth one sector of producton and multnatonal actvtes of country frms n country. From payments balance t follows that 2 m n EX, ( IM ) < EX + IM, f there are flows of repatrated profts due to multnatonal actvtes of country frms n. Thus, GLI < 1, accordng to (1). However, n a one-sector model there s by defnton only ntra-ndustry trade, so that the correct GLI must equal one. To obtan an approprate measure of the IIT share, we have to adust the Grubel-Lloyd ndex for all ncome flows not due to goods trade, lke repatrated profts. 3 More precsely, we correct the denomnator of GLI for all output flows that are balanced by ncome flows not drectly related to exports and mports. Ths gves a hypothetcal measure of balanced trade n the denomnator of GLI. 4 The corrected Grubel-Lloyd ndex for the two-country, mult-sector case s then: 1 We do not dstngush between c..f and f.o.b data for the moment. For a rgorous dscusson on dfferent emprcal specfcatons of the Grubel-Lloyd ndex see Subsecton 3.1. 2 Note that ths has an entrely dfferent motvaton than the case made by Aquno (1978) for a correcton for aggregate payments mbalance. As Greenaway and Mlner (1981) showed ths s nether defensble on theoretcal grounds nor practcable. 3 (See Subsecton 3.1 and Appendx for the quantfcaton of ths and other bases). 4 Ths adustment method was n fact frst suggested by Grubel and Lloyd (1975). However, they dd not develop t on the grounds that t lacked a clear theoretcal motvaton. 4

GLI = ( EX IM ) 2 mn, k k, (2) k EX k k + IM k k EX k k IM k k In our thought experment wth two one-sector economes and multnatonal actvtes of country frms n country, GLI gves a correct measure of the ntra-ndustry trade share,.e. GLI = 1. 5 Accordng to (1) and (2), we obtan GLI EXk IM k k k SHI : = = 1+ GLI EX + IM EX IM k k k k k k k k as a measure of the trade mbalance bas n relatve terms. >1 (3) In what follows we are nterested n the role of multnatonal actvtes and repatrated profts. In partcular we nvestgate how changes n the fxed for ncome flows EXk IMk k k costs of multnatonal actvtes as one key determnant of FDI-flows (see Amt and Wakeln, 2003) affect the corrected Grubel-Lloyd ndex gven n (2) and the rato of the corrected and uncorrected ndces as n (3). To dentfy the basc economc mechansms, we start wth two analytcally solvable general equlbrum models, whch account for horzontal and vertcal multnatonal actvtes, then provde smulaton analyses of fve varants of new trade theory models wth multnatonal frms. 2.2 Two analytcally solvable models onsder two countres wth two sectors, whch dffer only wth respect to factor endowments. In the ndustral X-sector dfferentated goods are produced, whle output n agrcultural Y- sector s homogeneous. Preferences of consumers are dentcal and represented by a obb- Douglas utlty functon: U α 1 α = X Y, 0< α < 1 (4) ε /( ε 1) ( ε 1/ ) ε where X : = x k, ε > 1, s a ES-ndex, that accounts for home-produced and k mported varetes of the ndustral good. 6 Producton technologes n the two sectors are gven by x = L and Y = L, respectvely, where L s unsklled labour. In addton, producton n the X-sector requres fxed set-up costs through the use of captal K and sklled non- 5 Noteworthy, we can substtute EX k = IM k n (2) f f.o.b. measures are used n the calculatons of GLI. Ths wll be mportant n our analytcal nvestgaton below. 6 ountry ndces are neglected for the moment. 5

producton labour S. We choose unsklled labour of country as the numérare and thus, set w = 1. Exportng dfferentated ndustral output gves rse to ceberg transport costs of 1- L 1/t>0 (n real terms). Trade n the homogeneous good does not nduce any trade frctons. Horzontal multnatonal enterprses In a symmetrc equlbrum wth dentcal unsklled wages n the two economes, demand n country for a sngle varant of the dfferentated good s gven by where x x ε α Ep = and x = xτ, (5) P s a varety produced and consumed n country, whle x s produced n and exported to. 7 and 1 E : = L + w K + w S K S ( ) 1 s total factor ncome (total expendtures) of country P = p ε h + h + n + n p ε s a prce ndex. n, n and h, h are exporters and horzontal multnatonals of countres and, respectvely. τ = t s a measure of ceberg transport costs. It s well-known from the lterature that proft maxmzaton leads to a constant prce-markup and, therefore to prces p = ε /( ε 1) and p tε /( ε 1) 1 ε =. 8 To set up an exportng frm (n) requres one unt of captal and one of sklled labour, whlst one unt of sklled labour and g > unts of captal are requred to set up a horzontal multnatonal (h) n wth one plant n and one n. Thus, n equlbrum, zero-proft condtons of country frms are gven by due to w L = w = 1 L condtons n country are gven by π π 2 9 1 = x + τx w w ε 1 n K S 1 = x + x g w w ε 1 h K S = 0, (6) = 0, (7) n the case of dversfcaton. Fnally, the three factor market clearng x 1/ 7 If unts of the ndustral good are produced n n country, only ( ) t x unts arrve n country, due to the exstence of ceberg transport costs. 1 ε Hence, the prce ndex s gven by P : = p h + h + n + nτ f wl = wl = 1. 9 Eqs. (6) and (7) buld upon two smplfyng assumptons, namely that () fxed costs of exporters and horzontal multnatonals only dffer wth respect to the requrement of captal and that () only factors of country are used to set up country frms (and ther plants). 6

( ) L = h + h + n x + τ nx + Y, (8) S = n + h, (9) From (6)-(10), we obtan K = n + gh. (10) w K 1 1 τ = x, ε 1g 1 for equlbrum wage rates n country and K S h =, g 1 1 1 gτ ws = x x (11) ε 1 g 1 n = gs for the equlbrum numbers of horzontal multnatonals and exporters n country. Equvalent expressons are obtaned for wages and frm numbers n country, f both sectors X and Y are actve n both economes. g K 1 (12) For the uncorrected and corrected Grubel-Lloyd ndces we obtan, from (1) and (2), and GLI 2ετ mn nx, nx = ετ nx nx ετ n h x ετ n h x ( + ) + ( + ) ( + ) (1a) GLI 2ετ mn nx, nx = ετ nx + nx + ετ n+ h x ετ n + h x hx hx where ( ετ ) ( ετ ) ( ) ( ) ( ), (2a) n + h x n + h x s Y-trade 10, accordng to the balance of payment condton. 11 Moreover, hx h x s the balance of repatrated profts for whch the denomnator of GL I s adusted. The respectve share SHI s gven by SHI = 1+ hx h x ( ) ( ) ( ) ετ nx + nx + ετ n+ h x ετ n + h x hx hx, (3a) 10 By assumpton, consumers prefer the home-suppled homogenous good n the case of dentcal prces. Ths mples a unque value of Y-trade n the absence of any trade frcton for homogenous goods. 11 Note that we consder f.o.b. trade flows (net of any ceberg transport costs) n eqs. (1a)-(3a) and throughout the rest of the theoretcal analyss. Ths mples that EX k = IM k (see Footnote 5). For a rgorous dscusson on dfferent concepts of the Grubel-Lloyd ndex, see Subsecton 3.1. 7

For smplcty, we assume symmetry wth respect to endowments 12 of K and S but allow for dfferences n endowments of unsklled labour L. Moreover, we assume the two economes are ex-ante equvalent wth respect to cost parameter g, capturng physcal captal related FDI-costs. Startng from ths equlbrum we nvestgate how a margnal change n (for gven ) affects the IIT share GLI and assess the trade mbalance bas n relatve terms by g g calculatng the mpact of g on SHI. Two scenaros can be dstngushed: Scenaro I - L < L : 13 Defne x% : = nx and x% : = nx. Then, usng (11), (12) and E, P n (5) gves and equvalently x% x% K S + τ x% + L gs K α gs K = ε α 1 1 + + ε ( ε 1) ( ) ( g ) S ( K S) τ ( g S K) K S + τ x% + L gs K α gs K = ε α 1 1 + + ε ( ε 1) ( ) ( g ) S ( K S) τ ( g S K), (13). (14) From (13) and (14) t s obvous that L < L mples x% > x%. Hence, we fnd 14 g= g g = g GLI accordng to (2a), and 2ετ x% x% SI = = h n x% h n x x% h n x% h n x% % ( 2 ετ + / ) ( / ) ( / ) ( / ), (15) accordng to (3a). SHI SI h / n h / n x% = 1+ 1 2 ετ h / n x%, (16) 12 These symmetry assumptons wll be relaxed n the smulaton analyses of Subsecton 2.3. Remember our assumpton that both sectors are actve n the two countres. Ths requres that L and L are not too dfferent. 14 Index SI refers to Scenaro I. 8

Result 1. onsder L < L and (ex ante) g rases the ntra-ndustry trade share,.e. bas n relatve terms,.e. dshisi / dg > 0. Proof. See Appendx. =. Then, a margnal ncrease of g (over g dgli SI / dg > 0, and rases the trade mbalance g ) For L < L, an ncrease n (for gven ) makes the two economes more smlar, or n other words reduces country s home-market advantage due to ts better endowment of L. It s well-known that the ntra-ndustry trade share ncreases n the smlarty of countres (see Helpman, 1987, Bergstrand, 1990, Hummels and Levnsohn, 1995), so that g ncreases n. The aforementoned effect tends to reduce SHI, snce the balance of repatrated profts, ( ) ( ) g.e. h / n x% h / n x% > 0 becomes more equal, accordng to (15) and (16). 15 However, there s a second, counteractng effect. An ncrease n g GLI reduces the number of country s horzontal multnatonals (and ncreases ts exporters). Ths lowers the flows of repatrated profts from to and, therefore, rases g ( / ) % ( / ) h n x h n x% and stmulates the trade mbalance bas SHI. In sum, the frm number effect domnates and explans a negatve mpact of g on SHI. Or, put dfferently, f n terms of ther goods trade and therefore, rases GL I L < L an ncrease of, makes countres more smlar dssmlar n terms of ther repatrated profts, whch mples a hgher SHI. Scenaro II - L > L : g, but countres become more From (13) and (14) t s clear that L > L mples x% < x%. Hence, we fnd 16 g= g g = g GLI SII accordng to (2a), and 2ετ x% x% = = ( 2 ετ + h / n) x% ( h / n) x% ( h / n) x% ( h / n) x% x%, (17) accordng to (3a). SHI SII h / n h / n x% = 1+ 2 ετ h / n x%, (18) 15 (One should keep n mnd that repatrated profts are balanced f two economes are dentcal, mplyng k EX k = IMk.) k 16 Index SII refers to Scenaro II. 9

Result 2. onsder L > L and (ex ante) g = g. Then, a margnal ncrease of g (over g ) reduces the ntra-ndustry trade share,.e. SI dgli I / dg < 0, and lowers the trade mbalance bas n relatve terms,.e. dshisii / dg < 0. Proof. See Appendx. Under Scenaro II, an ncrease n g renforces s home-market advantage due to ts better endowment of L. As a consequence, the dssmlarty between countres ncreases wth, whch reduces the ntra-ndustry trade share GL I repatrated profts,.e. ( h / n) x ( h n). Ths stmulates SHI, snce the balance of % / x% > 0 becomes less equal, accordng to (15) and (16). However, the nduced declne n the number of country s horzontal multnatonal frms counteracts and domnates, so that ( / ) x% ( h / n) h n x% smlar n terms of repatrated profts. Ths reduces SHI. g declnes, makng countres more Vertcal multnatonal enterprses It s well-known from the lterature that vertcal multnatonals (v) are more lkely where countres dffer suffcently n ther factor endowments or producton technologes. In a two country model, vertcal multnatonals can only be actve n one economy. We take the smplest possble framework that allows for vertcal multnatonals n country, by assumng the followng parameter constellaton: K > K = S = S. Agan, settng up an exportng frm requres one unt of captal and one of sklled labour; whle one unt of sklled labour and γ > 1 unts of captal are requred for settng up a vertcal multnatonal enterprse n country wth a sngle producton plant n. 17 In equlbrum, the zero proft condtons of exporters and vertcal multnatonals n are gven by 18 π 1 = x + τx w w ε 1 n K S = 0, (19) 17 We use γ nstead of g to refer to the sze of FDI-costs n the case of vertcal multnatonal frms. The reason s that set-up costs of vertcal multnatonals fundamentally dffer from set-up costs of horzontal multnatonals, snce n the former case only one producton plant s requred, whle n the latter case two plants are operated. 18 By assumpton the endowments wth unsklled labour are such that both the X-sector and the Y-sector are actve n the two economes and that vertcal multnatonals as well as exportng frms survve n country. Then, w L = wl = 1, so that n ths model vertcal multnatonal actvtes are drven by a home-market effect (.e. absolute sze dfferences) and not by dfferences n unsklled wages. 10

1 π = x + τx γw w ε 1 v K S = 0, (20) respectvely. (Note the smlarty between (6) and (19).) In country only exportng frms are actve wth profts π 1 = x + τx w w ε 1 n K S The three factor market clearng condtons n country are And those n country are From (19), (20) and (22)-(24) we obtan 19 and w K ( τ ) = 0. (21) L = n x + x + Y, (22) S n v = +, (23) = + γ. (24) K n v ( )( τ ) L = n + v x + x + Y ( 1 τ )( x x ) 1 =, ε 1 γ 1 n γ K K = γ 1 (25) K = S = n. (26), w S ( τγ 1) x + ( γ τ ) 1 x = (27) ε 1 γ 1 v K K = γ 1 for equlbrum wage rates and frm numbers n country. Snce only one frm type s actve n, we cannot dstngush between w and w. Hence, equlbrum wages n are gven by K S 1 w + w = x + τ x ε 1 K S accordng to (21). The equlbrum frm number n s determned by (26). (28), (29) = + + 1 ε P = p n + ( v + n) Usng E w K w K L, K S ( (5) as well as E = wk + ws ) K + L, τ and p /( = ε ε 1) n demand 1 ε P = p n + v + ( ) nτ and p = ε / ε 1 n the respectve expresson for country gves after straghtforward calculatons explct solutons 19 S = K s used n (27) and (28). 11

x 2 ε ( K K ) + τ ( γk K ) L + ML α = ( 1)( 1) α γ ε, (30) 2 ε α NM ( γ 1) τk ( K K ) + τ ( γk K) ε x = x ε + α ( γ ) 1 τk L NL ( ) τ ( γ ) ε α K K + K K L + ML wth N : = ( 1 α / ε + τ)( γ 1) K ( 1 α / ε)( 1 τ)( ) ( 1 τ )( K K ). + Fact 1. Eqs. (30) and (31) are only consstent wth postve wages ε α accordng to (27), f () ( ) ( hold. N K K τ γk K ), (31) K K, M = ( α ε + τ)( γ ) : 1 / 1 K w K > 0,.e. wth x > x, > + and () L > L smultaneously In the remander of our analyss, we focus on postve wage equlbra wth suffcently large 20 τ and L, accordng to Fact 1 and the defnton of N. In addton τγ > 1 s suffcent for w > 0. S w K > 0,.e. For the case of vertcal multnatonals n country we can rewrte the Grubel-Lloyd ndces n (1) and (2) as: 2ετ mn ( n + v) x, nx ( n + v ) x + n x + ( n + v ) x n x v ( x + x ) GLI = ετ ετ τ and (1b) 20 Usng the defnton of N allows us to rewrte condton () of Fact 1 as { } ( ε / α) ( τ α / ε) τ ( 1 α / ε) ( γ 1) K ( 1 τ)( γk K) + + > 0, whch mples that τ > α s suffcent for condton (). Moreover, f condton () s fulflled, then ( ε / α) M τ ( K K ) + τ ( γk K) ( ε / α) ( ) + τ ( γ ) L > > 1 L N K K K K guarantees x > x and thus, w K > 0 n equlbrum. 12

GLI 2ετ mn ( n + v) x, nx ( n + v ) x + n x + ( n + v ) x n x v ( x + x ) v ( x + x ) = ετ ετ τ τ where ετ ( n + v ) x n x v ( τ x + x ) ( payments condton. Moreover, v τ x + x are ncome flows from country to country,,(2b) s Y-trade 21, accordng to the balance of due to vertcal multnatonal actvtes. Accordng to (1b) and (2b), SHI smplfes to ) v ( τ x + x) ( n + v ) x + n x + ( n + v ) x n x v ( x + x ) v ( x + x ) SHI = 1+ ετ ετ τ τ Three scenaros can be dstngushed: Scenaro I - ( n v ) x n x, country s a net exporter of X-goods + < 22.(3b) In ths case, we obtan GLI SI ( K K ) v nx 2 x = 1+ = 1+, (32) n nx γ K K x accordng to (2b). Snce there are ncome flows from country to country, the balance of payments condton requres that exports homogenous good Y f ( ) Moreover, accordng to (3b), we obtan SHI Ths mples Result 3. SI ( τ ) n + v x < n x holds. 1 v x + x 1 K K x = 1+ = 1+ τ + 2 ετ nx 2ετ γ K K 1. (33) x Result 3. onsder ( ) n + v x < n x. Then, an ncrease of nvestment cost parameter γ has a negatve mpact on ntra-ndustry trade,.e. dglisi / dγ < 0. Moreover, a hgher γ leads to a lower trade mbalance bas,.e. dshi SI / dγ < 0. 21 onsder Footnote 10 on our assumptons regardng Y-trade. 22 ( ) One can defne Ω= : n + v / n x / x to fnd Ω K > 0 and ( L) Ω / L = Ω / ( ) L / L > 0, accordng to (26)-(31) and Fact 1. Roughly spoken, ths mples that Scenaro I s more lkely f K and L are not too hgh and L s not too low, motvatng nterestng / nteracton effects that are accounted for n the econometrc analyss below. 13

Proof. See Appendx. An ncrease n γ tends to make vertcal multnatonal actvtes less attractve and therefore reduces X-mports of country. Ths mples a reducton of the IIT share snce was already a net exporter of dfferentated goods. The ntuton for the SHI-effect s as follows. Remember that the dfference between GL I and GL I arses due to the exstence of vertcal multnatonals n. However, multnatonal actvtes become less attractve f γ ncreases. As a consequence, an ncrease of γ reduces flows of repatrated profts from to and reduces the downward bas of ntra-ndustry trade flows f GLI nstead of GL I s used. Ths gves rse to dshisi / dγ < 0. Scenaro II - ountry s a net mporter of both goods 23 In ths case, we obtan accordng to (2b). Indeed, GLI SII ( ) nx < n + v x 2ετ nx = = 1, (34) 2ετ nx mples that country s a net mporter of the dfferentated X-good. If country also mports the homogenous good, there s no nterndustry trade snce net mports of are equal to repatrated profts due to multnatonal actvtes of country frms n country. Moreover, SHI SII = SHI, gven by (33). SI Result 4. onsder ( ) n + v x > nx andετ ( n + v ) x n x v ( τ x + x ) < 0. Then, GLI SII = 1, so that a margnal change of γ has no mpact on the ntra-ndustry trade share. The mpact of γ on SHI SII s negatve. Proof. Frst, use (34) to see that γ has no mpact on GL I. Second, dshi SII / dγ < 0 follows from Result 3. The ntuton for the dscussed below Result 3. SII SHISII -effect of γ s analogous to the ntuton of the SHI SI -effect 23 ετ ountry s a net mporter of both types of goods f both ( n + v ) x n x v ( τ x + x ) < 0 there are proft flows from country to country. ( ) nx < n + v x and smultaneously hold. Such an outcome s only possble f 14

Scenaro III - ountry s a net mporter of X-goods and exports the Y-good 24 In ths case, both ( ) smultaneously hold. Thus, we obtan GLI n + v x > nx and ετ ( n + v ) x n x v ( τ x + x ) > 0 must ετ nx ετ n / v n x + vx vx n v x x + x x 1, (35) SIII = = ετ ( ε 1 ) τ ετ ( / ) / ( ε 1 ) τ / accordng to (2b), and SHI SIII 1 vτ x + vx = 1+ 2 ετ nx + ε 1 τ vx vx ( ) 1 vτ x / x + v = 1+ 2 ετ nx / x + ε 1 τ vx / x v ( ), (36) accordng to (3b). The requrement of balanced payments guarantees SHI SIII > 0. GLI SIII > 0 and + + > Result 5. onsder ετ ( n v ) x n x v ( τ x x ) 0. Then, an ncrease of nvestment cost parameter γ rases the ntra-ndustry trade share,.e. dglisiii / dγ > 0. Moreover, the mpact of γ on the trade mbalance bas n relatve terms s negatve,.e. dshi SIII / dγ << 0, f τ s suffcently large.25 Proof. See Appendx. Agan, an ncrease n γ makes vertcal multnatonal actvtes less attractve. Ths mples a reducton of low-sklled labour n the producton of dfferentated goods n. The resultng expanson of the Y-sector n reduces mports of homogenous goods, leadng to less nterndustry trade. Moreover, there s an ncrease (a declne) of country s dfferentated goods exports to (mports from) country. Both effects rase GL I. The ntuton for the SHI - effect s not straghtforward. An ncrease n γ makes multnatonal actvtes less attractve, thereby reducng the flows of repatrated profts. Ths tends to reduce the trade mbalance bas and thus, SHI (see the ntuton of Result 3). However, a declne n overall trade flows nduces a hgher weght of repatrated profts and ncreases SHI, accordng to (36). It s dffcult to SII SIII 24 Scenaro III s more lkely f K and L are not too low and L s not too hgh, see Footnote 22. 25 Remember our dscusson below Fact 1. A suffcent condton for dshi SIII /dγ << 0 s derved n the Appendx. 15

determne whch effect domnates. However, we can show that a negatve mpact of nvestment costs γ on SHI s guaranteed f τ s suffcently large (see Appendx). 2.3 Smulaton analyss As a complement to our analytcal results, we assess the mpact of nvestment costs and determnants of IIT based on numercally solved versons of the models of vertcal and horzontal multnatonals. Gven the nherent non-lneartes of Dxt - Stgltz type models n general and possble nonmonotonctes due to complementary slackness of general equlbrum models of trade and MNEs n partcular (Markusen, 2002), we mplement numercal solutons to yeld nsghts nto approprate specfcaton choce and robustness. We smulate varous versons of our model. In so dong, we stck to the noton that both the model of vertcal MNEs (Helpman, 1984, Helpman and Krugman, 1985) and that of horzontal MNEs (Markusen, 1984, Markusen and Venables, 1998, 2000) are restrcted varants of the knowledge-captal model, where both types of frms may endogenously arse (arr, Markusen and Maskus, 2001, Markusen, 2002). However, a pure horzontal model and a pure vertcal one are also calbrated. Altogether, we set up fve dfferent models: a KKmodel based on a Leontef technology n the X-sector; a KK-model based on a obb-douglas technology n the X-sector; a KK-model based on a ES-technology n the X-sector assumng a more realstc techncal rate of substtuton of between 0 and 1 (see Sharma, 2002; we choose a relatvely low value of 0.1); a horzontal Leontef-based model; and a vertcal Leontef-based model. 26 In sum, we compute the equlbrum Grubel-Lloyd ndex for 21 21 21=9261 cells of the factor cube and 5 dfferent levels of country s fxed FDI-related nvestment costs (country s nvestment costs are always set at a fxed value). Ths gves 46305 equlbrum values for each model wthout trade cost dfferences. Addtonally, we smulate a set of equlbra, where trade costs for exports from country to amount to 5%-25%, leavng those of exports from to always at 15%. Where countres dffer n trade costs, there are a further 4 9261=37044 equlbrum values. Poolng the two sets of equlbra allows us to search for the preferred specfcaton n the emprcal analyss, accountng for the same varables. Altogether, there are 83349 observatons for each model. Specfcally, we estmate the followng models: 26 Table A.1 n the Appendx provdes detals on the calbraton of the model. 16

= α0 + α1 + LGLI ln( GDP GDP ) ( GDP ( GDP GDP) ) GDP ( GDP GDP) 3 4 ( ) + α2 ln 1 / + / + + α ln K / L - K / L + α ln ( S / L - S / L ) 2 2 + α ln (1 + INV ) - (1 + INV ) + α ln T T +ζ 5 6 2 2 = β0 + β1ln( + ) + β2( - ) + β3( / - / ) LGLI GDP GDP GDP GDP K L K L 2 2 2 4( S / L - S / L) + β5( GDP - GDP) ( S / L - S / L) + β ( INV ) ( INV ) ( T T ) + β ln 0.5 1+ + 0.5 1+ + β ln 0.5 + 0.5 + ζ 6 7 = χ0 + χ1 + χ2 LGLI max{ln GDP, ln GDP } mn{ln GDP, ln GDP } + χ ln ( K / L - K / L ) + χ ln ( S / L - S / L ) 3 4 ( ) + χ ln (1 + INV ) - (1 + INV ) + χ ln 0.5T + 0.5T + ζ 5 6 (M1) (M2) (M3) = δ0 + δ1 + δ2 LGLI max{ln GDP, ln GDP } mn{ln GDP, ln GDP } + δ max{ln( K / L ),ln( K / L )} + δ mn{ln( K / L ),ln( K / L )} 3 4 + δ max{ln( S / L ),ln( S / L )} + δ mn{ln( S / L ),ln( S / L )} 5 6 + δ max{ln(1 + INV ), ln(1 + INV )} + δ mn{ln(1 + INV ), ln(1 + INV )} 7 8 { ( T ) ( T )} ( T ) ( T ) { } + δ max ln, ln + δ mn ln, ln + ζ 9 10 (M4) LGLI denotes the logstcally transformed, corrected Grubel-Lloyd ndex, INV refers to nvestment costs g and γ of our theoretcal analyss, respectvely, and T (T ) s a measure of transport costs for shppng dfferentated goods from country ( ). 27 ( ) to country M1 s closest to Helpman (1987) but wth the addton of dfferences n nvestment costs; M2 s closest to Markusen and Maskus (2001), extended by the squared dfference n captallabour ratos; M3 s n the sprt of Hummels and Levnsohn (1995), wth the addton of absolute dfference n sklled-to-unsklled labour ratos; and M4 extends ther dea of allowng for asymmetrc nfluences between maxmum and mnmum levels of all varables. 27 In terms of our analytcal model, = ( 2 1/ t ) T gves the volume of producton that s necessary f one unt of the dfferentated good s consumed abroad. 17

Runnng those four specfcatons results n the followng adusted R 2 fgures: Horzontal Vertcal KK-Leontef KK-D KK-ES M1 0.0115 0.2345 0.1469 0.0082 0.0568 M2 0.0685 0.3507 0.1419 0.0569 0.0624 M3 0.0113 0.3157 0.1147 0.0113 0.0557 M4 0.1056 0.5085 0.1563 0.0933 0.0880 Wth the excepton of the vertcal model, the reported R 2 fgures are relatvely low, reflectng the hgh degree of non-lnearty n these type of models. Of course, omttng the sklls and frcton varables n M1 or M3 would lead to specfcatons whch are closer to Helpman (1987) and Hummels and Levsohn (1995), but nferor n terms of explanatory power. Smlarly, omttng the captal terms n M2 would render the model closer to Markusen and Maskus (2001) but also nferor. On the other hand, usng the maxmum and mnmum values of both trade and nvestment frctons n every model reduces the dfference n adusted R 2 fgures, but wthout changng ther rankng. Emprcally, the repeated observaton of each country n a blateral settng and the use of country-specfc effects mprove the ft. As can be seen, M4 consstently outperforms M1-M3. Wth regard to the estmated coeffcents of M4, two weak hypotheses can be formulated. Frst, δ 1 <0 and δ 2 <0 are more lkely f horzontal MNEs domnate 28 Note that horzontal MNE actvty s market-seekng,.e. growng wth market sze, and crowds out two-way trade n dfferentated goods, explanng the expected sgn of δ 1. The result δ 2 <0 s due to the non-lneartes, caused by complementary slackness. Suppose that there s ntally a very small country, so that t does not pay to set up horzontal MNEs. In such an economy, ntra-ndustry trade accounts for a large share of trade. As we reallocate absolute factor endowments to ths economy, at some pont t s proftable to establsh horzontal MNEs and ntra-ndustry trade falls. As countres become more smlar, the IIT share rses agan. In our case, the effect nduced by the complementary slackness domnates and δ 2 s negatve. However, δ 1 <0 ndcates that smlarty n sze s mportant and tends to ncrease the IIT share f δ 1 domnates δ 2. Vertcal MNEs tend to foster ntra-ndustry trade but they are stmulated by dssmlartes n country sze, n lne wth our analytcal nvestgaton (see Fact 1). 28 See Markusen and Maskus (2000) and arr et al. (2001) for strong emprcal support of horzontal MNEs. 18

Second, for smlar reasons the maxmum nvestment cost coeffcent δ 7 s negatve wth twoway horzontal FDI (see Result 2). Hence, the share of IIT tends to rse wth the smlarty n nvestment costs. Fnally, the dfference between the maxmum value and the correspondng mnmum value of the sklled to unsklled labour coeffcents ( δ 5 δ 6 ) tends by and large to be negatve, whch supports the common fndng that IIT s hgher between economes wth more smlar factor endowments (Helpman, 1987, Bergstrand, 1990, Hummels and Levnsohn, 1995). Wth regard to the mpact of captal-labour ratos, remember that settng up horzontal MNEs s the most captal ntensve actvty. Horzontal MNEs seem emprcally mportant and note that the share of ntra-ndustry trade n total mbalance-corrected trade tends to rse f horzontal FDI ncreases. However, wth co-exstng horzontal and vertcal or only vertcal FDI the mpact of captal-labour ratos gets less clear-cut. 2.4 Summary of the theoretcal hypotheses for the Grubel-Lloyd ndex From our analytcal nvestgaton we obtan the followng hypotheses. Wth horzontal MNEs an ncrease n nvestment costs g tends to reduce GL I country. In contrast, f g rses n the country wth scarce L supply,, f g ncreases n the L-abundant ncreases. Wth vertcal MNEs, an ncrease n FDI-costs tends to reduce the ntra-ndustry trade share f n the country that hosts the vertcal multnatonal frms factor L s relatvely scarce (so that ths economy s a net exporter of the dfferentated good. In contrast, f the country that hosts the vertcal multnatonals s relatvely L-abundant and, therefore, s a net mporter of the dfferentated good, GL I tends to be postvely (non-negatvely) affected by an ncrease n FDI-costs. The smulaton exercse generates two addtonal hypotheses. Frst, the country sze (δ1,δ 2 ) coeffcents are more lkely negatve, f horzontal MNEs domnate at a reasonable GLI value of the elastcty of substtuton (see Feenstra, 1994, for detaled emprcal evdence). Second, ntra-ndustry trade s by and large hgher between economes wth more smlar sklled-to-unsklled labour endowments and nvestment costs wth two-way horzontal FDI, but less lkely the more mportant vertcal FDI s. 3 Emprcal analyss 3.1 The Grubel-Lloyd ndex n the emprcal trade lterature Grubel and Lloyd (1971) had n mnd a model wth zero transport costs and no multnatonal frms. Both transport costs and MNE actvty are now understood as essental characterstcs 19

of nternatonal exchange. However, ther consequences for the measurement (and determnants) of ntra-ndustry trade shares has to the best of our knowledge not yet been rgorously studed. Below, we provde several alternatve versons of the Grubel-Lloyd ndex, whch can cope wth both transport costs and MNE actvty. (We also explctly dscuss ssues such as the nterpretaton of mssng values n the dsaggregated trade data, the ndex s based on.) Table 1 summarzes. 29 > Table 1 < It seems sensble to start wth the orgnal formulaton of the ndex as also appled n Helpman (1987), Hummels and Levnsohn (1995), or Markusen and Maskus (2001). In the case of a two-country, new trade theory model wth zero transport costs and no MNE actvty, 1 2 3 4 GLI GLI GLI GLI GLI GLI5. Wth multnatonal frms, trade s not necessarly (or even lkely to be) balanced. To see the relevance of ths, consder the smple thought experment of two one-sector economes wth MNEs. Not accountng for ncome flows due to repatrated profts leads to a downward bas of the Grubel-Lloyd ndex,.e. GLI GLI1 < 0, GLI2 GLI3 < 0 and GL I GLI <, whch we refer to as the trade mbalance bas n absolute terms (n contrast to the relatve measure of ths bas, SHI, calculated above). 30 Hence, there reman three canddates for measurng the ntra-ndustry trade share: GL I1, GLI and GL I5 whch dffer f transport costs are postve. 3 4 5 0 Now consder the mpact of transport costs, but stck for the moment to the usual assumpton that t = t. In ths case, GLI1 GLI 3 GLI 5. Note that GLI1 GLI 3, because the denomnator of GLI 1 s hgher than the denomnator of GLI 3 due to transport costs ncluded 29 The Grubel-Lloyd ndces n Table 1 measure blateral ntra-ndustry trade n a mult-country world. Hence, EX ndustres. are country s exports to and IM are country s mports from country. Index k ndcates dfferent 30 The arguments n Greenaway et al. (2001) are related to our arguments. Bergstrand (1983) correctly ponts out that blateral trade tends to be unbalanced also n a multlateral settng wthout MNEs. Our approach also covers ths phenomenon. 20

n c..f. mports IM but not n f.o.b. exports. But also the numerator of GLI s hgher EX 1 unless EX k < IM k k. Ths transport cost level bas n absolute terms 1 GLI GLI 3 I3 5 appears to be a non-lnear functon of t. Moreover, GL and GL I share an advantage over GLI 1, snce they lead to the same ndex values for the two economes (.e., GL I GLI ), whle GLI 1, GLI, 1 f transport costs are larger than zero. 31 In addton, f transport costs dffer,.e. f t t, whch s the emprcally relevant case, also GLI dffers from GLI. 3 5 An approxmaton of ths transport cost dfference bas n absolute terms s GLI 3 GLI5. We focus on two versons of the corrected Grubel-Lloyd ndex n the emprcal analyss: we use GLI 3 as our preferred measure, snce t s derved from our theoretcal model and gves dentcal ndces for the two economes. We also use GL I, snce t s closest to the dea of the tradtonal Grubel-Lloyd ndex GLI, but avodng the trade mbalance bas. 1 Furthermore, there s a mssng value nterpretaton bas at the most dsaggregated level (k). There are two opportuntes to handle ths. One could nterpret them as mssng n a narrow sense and skp all l mssng observatons, before determnng the mnmum export and reexport (from partner statstcs; or mport values). Suppose that the data are sorted so that all l mssng observatons come frst and that EX 1,..., EX l s mssng wth the true 0 EX < EX k k k l. If we gnore any trade mbalance for the moment and use AGLI 3 to assess the share of ntra-ndustry trade, then 32 l EX EX k k k k > K ( EX + EX ) l ( EX + EX ) l k= 1 k= 1 k= + 1 k= l+ 1 K EX EX mples a mssng value nterpretaton downward bas. Otherwse, AGLI 3 s upward based. I1, 1, 31 In our sample, the dfference between GL and GL I amounts to 43 percentage ponts for 5-dgt-based data, whch s about 312% of the correspondng mean. 32 The same problems arse f AGLI1 nstead of AGLI3 were used. 21

The alternatve s to replace mssng values by zeros. However, f EX k = 0 s used, although the true value were < EX < EX k l, there s a downward bas n GLI. In 0 k k addton, the magntude of trade mbalance may be based under both correcton methods. Snce no nformaton on the true values s avalable, we have to rely on assumptons to favor one approach over the other. Here, we stck to the workng hypothess that, on average, the true mssng values are very small, f not zero.summng up, we can label ndces GLI and, 3 33 especally, GLI 3 and not AGLI1 and AGLI3 as the preferred measures. 1 3.2 Econometrc analyss We estmate our M4 model on varous concepts of the Grubel-Lloyd ndex, focusng on GL I 1 as the tradtonal measure and GLI, GLI as our preferred measures. Moreover, we also account for AGLI, AGLI1 and AGLI3 as a robustness check on mssng values. 3 > Table 2 < Our data base comprses 422 observatons of 1990-2000 blateral average IIT share data of OED countres for GLI3 and AGLI3 after excludng mssng values, whle there are about twce as many observatons for GL I, AGLI and GLI1, AGLI1, due to ther asymmetry between the -to- and the -to- trade flow defnton. (A detaled data descrpton s n Appendx D.) Table 2 summarzes our fndngs. Frst, those varables not usually consdered n emprcal analyss but motvated by our theoretcal analyss ( δ5 -δ 10 ) account for 41%-69% of the regresson models explanatory power. Ths agan emphasses the relevance of the MNE-related new trade theory lterature for core emprcal ssues of nternatonal trade. Second, n lne wth the regressons on the data obtaned from the numercally solved models wth two-way horzontal FDI, coeffcents for a varable s blateral maxmum and mnmum value wth the excepton of country sze (GDP) and captal-labour ratos tend to be dfferent n terms of both ther sgn and absolute value. By and large, the evdence suggests that smlarty 33 For completeness, we should menton the so-called Fnger bas whch refers to the problem of potentally upward based ntra-ndustry trade fgures due to the use of hgher aggregated data than avalable and possble statstcal measurement bas due to false reportng by the natonal statstcal offces. A detaled emprcal assessment of the sze of these s presented n Appendx. 22

(though n a non-lnear way) n sklled-to-unsklled labour endowments and n both trade and nvestment mpedments s n favor of more ntra-ndustry trade n total trade. 34 Also the captal-labour rato coeffcents lends support to the horzontal model. Only the strong postve sgn of the mnmum GDP coeffcent contradcts the smulated, purely horzontal model of FDI. Thrd, nvestment cost effects are manly represented by the negatve, sgnfcant coeffcent of maxmum blateral nvestment costs, whch s well n lne wth our theoretcal fndngs. In the analytcal model we fnd a negatve mpact of FDI-costs on the corrected Grubel-Lloyd ndex for some factor endowments (see Results 2 and 3). It turns out that n the regressons based on the trade-mbalance uncorrected measures of the ntra-ndustry trade share (GLI, AGLI) the role of transport costs seems to be over-estmated at the expense of relatve factor endowments, as compared to ther corrected counterparts. If mssng observatons at the dsaggregated level really reflect very low values of trade rather than confdental nformaton from a few, large frms perspectve, the results suggest that skppng mssng values results n a downward bas (n absolute terms) of the effects of captal-labour ratos, maxmum nvestment costs, and transport costs (to see ths compare the coeffcents for AGLI, AGLI1, AGLI3 wth the correspondng ones to ther left). Hence, our fndngs llustrate that measurement bases n IIT share ndces do not only affect the mean (pcked up by the constant), but there s some systematc bas, whch s correlated wth the most mportant explanatory varables. In sum, the results are well n lne wth the model of horzontal two-way FDI, but they lend less support to the exstence of vertcal FDI, rrespectve of whether we use the preferred or the based ndces. Ths s not surprsng gven the composton of our country sample. 3.3 Senstvty analyss We check the senstvty of our results wth respect to the excluson of extreme outlers and ncluson of exporter and mporter fxed effects. Wth regard to outlers, we follow Belsley et al. (1980) and exclude all observatons wth absolute resduals exceedng two standard errors of the regresson. On average, only 2% of observatons have to be elmnated. Fxed countryspecfc effects are able to control for all other unobserved varables, especally, multlateral trade resstance terms n a mult-country settng (see Anderson and van Wncoop, 2003). Note that the parameters of the varables can stll be estmated, snce by defnton there s enough 34 Ths smlarty aspect s also n lne wth our fndngs for the analytcally solvable horzontal multnatonal 23

varaton n maxmum and mnmum values. More precsely, n a sample such as ours t s mpossble that each country exhbts the maxmum or mnmum value n all varables wth respect to all ncluded tradng partners. > Tables 3 and 4 < omparng the results across columns n the upper part of Table 3 confrms that the Fngerbas n our country sample tends not to systematcally bas the parameter estmates. Ths concluson s based on results that exclude extreme outlers but reles on the assumpton that there are no omtted country-specfc nfluences, whch are correlated wth the explanatory varables. The ncluson of country effects tends to reduce the collnearty among regressors and controls for all omtted country-specfc nfluences, whch may otherwse be pcked up by the parameters of nterest (see Baltag, 2001). Ths has two mportant consequences. Frst, both sze coeffcents are now more supportve of the two-way horzontal MNE model than of ts vertcal counterpart (see the frst two columns n the lower part of Tables 3 and 4, respectvely). Second, the fxed effects model parameters tend to be much more senstve to the mpact of the Fnger bas or the mssng varable nterpretaton bas than ther nconsstent counterparts. To see ths, compare the columns n the lower part of Table 3 and note that the frst comprses the preferred specfcaton f our assumpton about mssng values holds. The Fnger bas even changes the sgn of δ 5 f we compute the GL or GLI on the bass of 4- dgt (3-dgt) rather than of 5-dgt data. 35 I1 3 3.4 Extensons Here we provde nsghts nto two addtonal ssues: the role of dfferences n endowments wth labour and physcal captal, respectvely, and the mpact of nvestment costs on the rato of the trade-mbalance corrected-to-uncorrected Grubel-Lloyd ndces SHI. > Table 5 < frms model n Subsecton 2.2. See our dscusson below Results 1 and 2. 35 Note that the reported F-tests on the parameters ndcate that, by and large, usng a smple measure of smlarty or also the average of blateral sze, factor endowments, and trade and nvestment mpedments s nferor to the chosen strategy of ncludng each varable s blateral maxmum and mnmum value separately. 24

Our analytcal results suggest that the mpact of an ncrease n nvestment costs on more lkely to be postve the larger L s compared to L. To assess ths, we construct an nteracton term between the dfference of maxmum and mnmum log nvestment costs and the correspondng log-dfference n L (see Table 5). Accordng to our theoretcal results for two-way horzontal multnatonals, we expect a negatve sgn of the maxmum nvestment cost effect 36 (δ 7 ) but a postve one of the nteracton term (δ 11 ). GLI s As wth the case of labour endowment dfferences, we formulate an nteracton term between the mpact of nvestment costs and dfferences n the endowment wth physcal captal. 37 Gven all other endowments, our model suggests that maxmum mnus mnmum nvestment costs are lkely to have a postve mpact on the ntra-ndustry trade share, the larger the correspondng dfference n captal endowments. Agan, we expect a postve sgn for the nteracton term δ 12. As the pont estmates n Table 5 ndcate, our emprcal fndngs strongly support our theoretcal hypotheses, rrespectve of whch of the preferred GLI concepts s used. However, one caveat apples. It s mpossble to nclude smultaneously both nteracton terms n the specfcatons. The reason s that captal stock levels are large n countres wth large labour forces. Hence, sze dfferences strongly domnate relatve factor endowment dfferences n our sample, renderng the log dfference n respectve captal stocks and that n absolute labour endowments hghly collnear. > Table 6 < Regardng SHI, we know that ths rato should fall wth the dfference between maxmum and mnmum foregn nvestment costs, n partcular, f the country wth the maxmum nvestment 36 ompare the fndngs of the smulaton analyses n Subsecton 2.3 and the summary of our theoretcal hypotheses n Subsecton 2.4. 37 Ths nteracton term s motvated by our analytcal nvestgaton n Subsecton 0 for the case of vertcal multnatonal frms, see Footnotes 22 and 24. Unfortunately, there s no comparable predcton for such an nteracton term f horzontal multnatonals are consdered. Ths s due to our symmetry assumptons n Subsecton 0. 25

costs s less well endowed wth labour than ts counterpart. 38 Ths s nvestgated n Table 6 for the two preferred concepts of the Grubel-Lloyd ndex. The results offer two nsghts. Frst, the pont estmates of both effects exhbt the expected sgns. Second, country-specfc effects are mportant, ndcatng that blateral trade-mbalances are a common phenomenon. Thrd, we have to concede that nvestment costs explan a relatvely small though sgnfcant share of the devaton between the two ndces as ndcated by the R 2 fgures. The other explanatory varables used n the prevous tables only contrbute nsgnfcantly. Hence, other macroeconomc varables, not accounted for n the above theoretcal model and the emprcal specfcatons are probably relevant n ths regard. However, to study ther mpact s beyond the scope of ths analyss. 4 onclusons In a revew of the emprcal analyss of nternatonal trade flows spannng the last 50 years, Leamer (1994) dentfes the extensve amount of ntra-ndustry trade catalogued by Grubel and Lloyd (1975).. as.. one of the only two maor emprcal fndngs (whch) seem to have had a maor mpact on the way (trade) economsts thnk (p.68). That concluson would no doubt be revsed n lght of the growng nfluence of the frm level adustment lterature. Be that as t may, Leamer s concluson artculates a wdely accepted vew that the apparent pervasveness of ntra-ndustry trade stmulated a revoluton n the theoretcal and emprcal modellng of nternatonal trade. From the standpont of emprcal nvestgaton, t s obvously vtal that the ntra-ndustry trade share s measured as accurately as possble. Thrty years after the publcaton of Grubel and Lloyd (1975), ther famous ndex remans the measure of choce for most nvestgators. Yet we know that t s grounded n the assumpton of arms length trade but multnatonal actvty s a feature of the landscape whch should not be gnored. In ths paper we have brought ther presence to centre-stage. We have constructed a three factor general equlbrum model of trade wth both horzontal and vertcal multnatonals, to dentfy precsely the mpact of nvestment costs and multnatonal actvty on ntra-ndustry trade. 38 For the case of vertcal multnatonals, Results 3, 4 and 5 predct a negatve mpact of nvestment costs on SHI. Moreover, as far as horzontal multnatonals are consdered, Result 2 shows that the SHI-effect s negatve f L > L so that the Grubel-Llyod ndex GLI declnes n the nvestment cost parameterm. 26

The model and the measures of ntra-ndustry trade derved from t have been subected to extensve smulaton analyss and rgorous econometrc analyss. The latter focuses on the trade flows of 31 OED countres. Our analyss demonstrates clearly the role of nvestment costs and the bases nherent n the Grubel-Lloyd ndex when we fal to account for the presence of multnatonals. Our econometrc analyss confrms the superorty of our new corrected measures. It also shows that t s mportant to account for varous new determnants of IIT alongsde more tradtonal explanatory varables. Fnally, our analyss lends further support to the relatve mportance of horzontal multnatonals. We hope that the theoretcal underpnnng provded for our new measures and ther robust emprcal performance wll commend ther wder use. 27

Appendx A. Analytcal appendx Proof of Result 1 We defne K S + τ x% + ( ε 1) L ( g S K) gs h α K Γ : = x% = 0, (A1) ε α 1 ( g 1) S + ( K S) + τ ( gs K) ε accordng to (13), and K S + τ x% + ( ε 1) L ( gs K) gs h α K Γ : = x% = 0, (A2) ε α 1 ( g 1) S + ( K S) + τ ( gs K) ε accordng to (14). Eqs. (A1) and (A2) mply system h h h Γ dx% Γ dx% Γ + + = 0 x% dg x% dg g h dx h h Γ % Γ dx% Γ + + = 0. x% dg x% dg g (A3) Γ h Γ Straghtforward calculatons allow us to wrte = = 1, x% x% h g= g g= g ( K S) τ ( gs K) Γ h Γ α + = = x% x% ε B h h g= g g = g Γ K S Sx < 1 and 0 g = α % ε gs K B <, g= g Γ g h 2 α / ε Γ = > 0 α / ε g g= g g= g h α B : = 1 g 1 S + K S + gs K) ε, where ( ) ( ) τ ( and g g = g have been used. Applyng ramer s rule to system (A3), we therefore obtan h h h h h h h Γ Γ Γ Γ Γ Γ Γ + dx% g x% g x% g g x% = = < 0, (A4) dg h h h h g= g Γ Γ Γ Γ 1+ 1 1+ 1 x% x% x% x% 28

accordng to (A3). h h h h h h Γ h Γ Γ Γ Γ Γ Γ + dx% g x% x% g g g x% = = dg h h h h g= g Γ Γ Γ Γ 1+ 1 1+ 1 x% x% x% x% > 0, (A5) SI / Next, we dfferentate GLI = x% x%, accordng to (15), wth respect to g and obtan dgli dx dx% x % x % > 0, (A6) dg dg dg SI 1 % = 2 x% g = g whch s postve, accordng to (A4) and (A5). To determne the mpact of g on SHI, we dfferentate (16) wth respect to g. Ths gves dshi h / SI n x% S 1 dx% dx% x% = dg 2ετ x% g g gs K x dg dg x = % %. (A7) Substtutng h g= g ( K S) τ ( gs K) Γ α + = % x ε B, Γ g h g= g α K S Sx% = ε gs K B and Γ g h 2 α / ε Γ = α / ε g h g= g g= g n (A4) and (A5) t follows that the bracket expresson on the rght-hand sde of (A7) s strctly decreasng n x% / x%, accordng to (A4) and (A5). Thus, dshisi / dg g= g s postve for all x% > x%, f t s postve for x g= g g = g = x % %. We therefore, calculate 1 dx% dx% 1 / / Γ g Γ g α K S S = = 2 1 x dg dg x h, (A8) % % 1 + Γ / x% ε B gs K g = g h accordng to (A4), (A5) and our consderatons above. Snce ( α ε)( ) h 21 / K S / B < 1, t follows that dshisi / dg g= g > 0 for all possble x% > x%, snce g= g g = g dshisi / dg 0 > for g = g x% = x%. Ths completes the proof of Result 1. g= g g = g 29

Proof of Result 2 Frst, note that dshisii / dg g= g < 0 drectly follows from (17), (A4) and (A5). Second, regardng the mpact of g on SHI we calculate SII dshi h / n SII S 1 dx% x% dx% = dg 2ετ gs K x g g % dg x% dg =, (A9) accordng to (18). The rght hand sde of (A9) s strctly ncreasng n x% / x%, accordng to (A4) and (A5). (For detals see the proof of Result 1.) Hence, dshisii / dg g= g s negatve for all x% < x%, f t s negatve for x% = x%. Ths follows mmedately from the g= g g = g proof of Result 1 and completes the proof of Result 2. Proof of Result 3 We use the defntons of M and N and dfferentate ( ) + τ ( γ ) + ( ε / α) x K K K K L ML = x K L NL ( γ 1 ) τ + ( ε / α), (A10) accordng to (31), wth respect to γ. Ths gves after straghtforward calculatons ( / ) x τ + ( ε / α)( 1 α / ε + τ ) = K γ ( ) + τ ( γ ) + ( ε / α) d x x L L d x K K K K L ML L + ( / )( 1 / + ) ( γ 1 ) τk L ( ε / α) τ ε α α ε τ L NL + (A11) and thus, where ( / ) x ( )( 1 ) d x x K K τ 2 ε 2 ε α = K L τ + L ( 1+ τ) 1 + τ dγ x φ ψ α α ε (A12) 2 ε α α + LL 1 + τ 2 + τ < 0, α ε ε ε α ε α φ : = τl + 1 + τ L ( γk K) + L + 2 L ( K α ε α ε K ), (A13) ε α ε α ψ : = τl + 1 + τ L ( γk K) + τl + 2 τl ( K α ε α ε K ) (A14) 30

have been consdered. Fnally, usng gves Result 3. Proof of Result 5 Dfferentatng GL I Usng ( ) ( ) d x / x / dγ < 0 n the frst dervatves of (32) and (33) wth respect to γ SIII, accordng to (35), wth respect to γ gves ( ) SIII ( ) ( ) ( / ) / + ( 1 ) / 1 ετ ( ) + ( ε ) τ GLISIII ετ ( ) ( ε ) τ dgli ετ d n / v / dγ GLI x / x d n / v / dγ SIII = dγ ετ n v x x ε τ x x ( ) ( ( ) n / v 1 d x / x n / v x / x + 1 x / x 1 dγ. (A15) n / v = γ K K / K K ) and n / v = n / v + 1 (whch mples d n / v / dγ = d n / v / dγ > 0), accordng to (28), and notng that GL I < and x / x < 1 ( ) must hold n a postve wage equlbrum, t s straghtforward that the frst term on the rght hand sde of (A15) must be postve. Together wth accordng to (A12), ths mples dgli / dγ > 0. SIII SIII 1 ( ) d x / x / dγ < 0, Next, we calculate the frst dervatve of SHI wth respect to v and obtan ( SHI 1) SIII SIII SIII ετ nx / x SHI = > 0, (A16) v v ετ n x / x + ε 1 τ v x / x v ( ) accordng to (36). Moreover, dfferentatng SHI wth respect to x / x gves SIII ( SHISIII 1) ετ ( n + v ) { ( ) v } SHISIII = ( x / x ) 1 + τx / x ετn x / x + ε 1 τv x / x. (A17) In vew of (26), (28), (A16) and (A17), we therefore obtan ( 1) ( x / x ) ( / x ) dshisiii SHI d x SIII dv SHISIII = + dγ v dγ dγ ( / x ) SHISIII ετ x x K d x = 1+ τ + ( n + v) 1 + τ x / x x 1 d x Q γ γ (A18) 31

wth Q: = ετ nx / x + ( ε 1 ) τ vx / x v. We substtute d( x / x ) / dγ, accordng to (A12), n (A18). Thereby, we consder φ and ψ, accordng to (A13) and (A14), and note that x / x φ / ψ =, accordng to (31). Moreover, we use n K ( ) ( )/( 1) v = K K γ = γ 1/ ( γ 1) and, accordng to (26) and (28). After tedous calculatons we then obtan ( SHI ) dshisiii SIII 1 ετ K x = D, dγ φψ 1 τ x / 1 x x Q γ + 2 2 2 { γ γ τ } wth D: T1( K K ) T2( K K )( K K) T3( K K) = + + + φ and accordng to (36). Thereby, ( SHI ) SIII 1 ετ K φψ 1 + τ x / 1 x x Q γ ( ) x > 0, ε α ε α α T1 : = τ 2τ L 2 1 τ 1 τ 2τ L α ε α ε ε 2 2 2 2 ε α α α 2 2( 1 τ) 1 τ 2τ τ ( 2τ 1) LL α ε ε ε (A19), (A20) 2 2 2 2 : 2 ε 1 α ε T L 2 α ε L 2 α 1 α = τ + τ + τ + + τ + τ LL, (A21) α ε α ε α ε ε T3 : ε L 1 α ε L L 1 α τ τ = + + τ + + L α ε τ α ε (A22) have been consdered. (T > and T 3 > 0 hold for all τ > 0.) Functon D has the followng propertes: and D τ = > 0, 1 2 0 2 2 ( ) ( ) ( ) ε α D 1 2L 2 ε α α = K K L L 2 2 K K 1 K K τ = 0 + γ α ε α ε ε D > 0 dd / d D τ = 0 0 τ D τ = 0 > for all τ > 0. Thus, we can dstngush two cases: If and thus, 0 2 D τ = > 0, then dshi SIII /dγ < 0, accordng to (A19), for all τ (0,1). However, f <, then there exsts a unque ( 0,1) and ( / dγ ) τ= τ τ such that dshi SIII / dγ < 0 (D > 0 ) τ > τ dshi = 0 ( D = 0 ). Ths follows mmedately from (A19)-(A22) and the fact SIII 32

that D τ = > 0. Hence, n the case of 1 completes the proof of Result 5. D τ = < 0, τ > τ guarantees dshi SIII / dγ < 0.39 Ths 0 B. Smulaton appendx Table A.2 provdes detals on the assumptons about the chosen parameter values n the numercal smulaton exercse. > Table A.2 < Our choce of the parameter related to the techncal rate of substtuton ponts to a complementary relatonshp between factors of producton, whch s n lne wth recent evdence (see Sharma, 2002). The choce of the elastcty of substtuton parameter between varetes s well n lne wth the fndngs n Feenstra (1994), and that one of the factor shares broadly reflects the fndngs n Mankw et al. (1992). The assumpton that ceberg trade costs vary around 15% s well n lne wth the stylzed facts (see Baer and Bergstrand, 2001).. Descrptve statstcs on dfferent measurement bases To provde a complete pcture of the sze of both ntra-ndustry trade shares and the varous bases dscussed, we report descrptve statstcs of blateral Grubel-Lloyd ndces accordng to each concept, computed on the bass of three dfferent levels of aggregaton (5-dgt, 4- dgt, and 3-dgt) as publshed by the OED usng the Standard Internatonal Trade lassfcaton. > Table A.2-A.3 < The fgures n Table A.2 llustrate that the average uncorrected ntra-ndustry trade share amounts to about 14-21% for the average blateral OED relatonshp between 1990 and 2000, dependng of whch level of aggregaton s used. Trade mbalance corrected fgures, of course, tend to be consderably hgher. In almost all cases, rrespectve of whch concept or aggregaton level s chosen, the standard error n the share s about as large as the mean. As 39 A more detaled proof s relegated to a supplement, made avalable n the GEP workng paper verson of the paper. 33

the last column n the table ndcates, the maor part of ths varaton s due to the crosssecton rather than the tme dmenson. However, all concepts where mssng values at the dsaggregated level are nterpreted as reflectng zero trade, tend to exhbt much more tme varaton than the others. For ths reason, cross-sectonal rather than tme seres (or panel data) analyss seems better suted for ntra-ndustry trade share measurement, snce measurement errors n the tme dmenson are lkely to cancel out. Table A.3 dsplays the correlaton matrx between all measurement concepts. Obvously, the varous correctons are strong enough showng up n correlaton coeffcents as small as 0.14 between GLI and the (not preferred) 3 AGLI, but also that one between the preferred GLI3 (GL I1 ) and the usually used GLI amounts only to 0.36 (0.59). Although Tables A.2 and A.3 provde frst nsghts nto the relatve sze of the varous bases dscussed, Table A.4 focuses more drectly on ths ssue and summarzes average bas fgures. > Table A.4 < The reported bases are computed n the followng way. To quantfy the trade mbalance 1 2 3 4 5 bas, we calculate GLI GLI, GLI GLI and GLI GLI, as ndcated n Subsecton 3.1. It s obvous that ths contrbutes more than any other bas. At the average aggregaton level, the uncorrected ntra-ndustry trade share s downward based by about 14 percentage ponts, whch s about 51%-81% of the mean. Of course, the trade mbalance bas s related to the level of MNE actvty. For nstance, when regressng the (logt-transformed) absolute value of the bas on the log absolute dfference between two partner countres world outward FDI stocks, we obtan a coeffcent of about 0.10, whch s sgnfcant at 10%. For the transport cost level bas, we subtract the respectve export based ntra-ndustry trade share ndces from ther uncorrected counterparts. (In Table A.4, we treat GL I as the preferred measure of the ntra-ndustry trade share.) In partcular, GL I GLI, AGL AGL 1 3 1 3 and GL I GLI, AGLI AGLI have been computed. The bas s always dsplayed wth the ntra-ndustry trade share concept t s affectng. From Table A.4, we see that the transport cost bas s relatvely small, amountng to 0.6 percentage ponts on average. Transport costs tend to upward bas (by about 7%-10%) the uncorrected ntra-ndustry trade share, whereas ther mpact on trade-mbalance corrected ntra-ndustry trade s on average almost neglgble (between -0.9% and -1.7% of the correspondng ntra-ndustry trade share). 34 3 2 2

However, we would expect that ths bas to be much larger n a sample of non-oed economes. As mentoned above, the transport cost dfference bas drves a wedge between the mportbased concepts and the export-based concepts of ntra-ndustry trade share measurement. 4 2 4 2 5 3 AGLI5 AGLI3 Accordngly, only GLI GLI, AGLI AGLI and GL I GLI, are computed to estmate ths. In our sample, ths bas s even smaller than the transport cost level bas, ndcatng that the asymmetry between two tradng partners transport costs s relatvely small. (Agan, a much larger bas of ths type mght be present f we consdered non-oed countres.) The mssng value nterpretaton bas has to be nterpreted wth care, snce no nformaton s avalable on whether GLI s closer to the true value or AGLI (and smlarly for the corrected fgures). As our workng hypothess, we take the extreme poston and assume that all mssng values ndcate zero trade flows. In ths case, the mssng value nterpretaton bas amounts to 8.2 percentage ponts (about 47%) of ntra-ndustry trade shares on average. For the Fnger bas, we subtract for each concept the ntra-ndustry trade share data of the respectve hgher level of aggregaton from ts next lower counterpart (.e., SIT 4-dgt based shares mnus 5-dgt based ones and 3-dgt based ones mnus ther 4-dgt based counterparts). Then, we average the resultng dfferences over the two aggregaton levels and all country pars and years. Accordng to our results, usng 4-dgt nstead of 5-dgt data exerts an upward bas of about 3.4 percentage ponts on the average ntra-ndustry trade share (.e., about 25%). Of course, usng 3-dgt data nstead causes an upward bas by about twce as much. In a fnal step, we aggregate the aforementoned bases, takng GL I at the SIT 5-dgt level as the preferred measure of the ntra-ndustry trade share. Of course, the dscussed bases do not smply add up, snce they exhbt a non-zero covarance. The overall bases are reported n the last two columns of Table A.4, Importantly, the last column of Table A.4 s ndependent of the Fnger bas, snce all bas fgures are wth respect to 5-dgt based ntra-ndustry trade share measures. We see that the tradtonally used Grubel-Lloyd ndex s downward based by about 10 percentage ponts, whch s about 43% of the corrected value. 3 35

D. Data appendx Data sources and defnton We use blateral export and mport flow data at the Standard Internatonal Trade lassfcaton 5-dgt, 4-dgt and 3-dgt level as publshed by the OED (Internatonal Trade by ommodty Statstcs, 1990-2000). Blateral transport costs are based on trade-weghted averages of c..f./f.o.b. fgures from ths source. Real GDP fgures are from the World Bank s World Development Indcators and measured n constant US dollars of 1995. aptal stock data had to be computed by the perpetual nventory method as dscussed n Leamer (1984, pp.232-234). Snce no data on deprecaton rates are avalable for our countres, the same value as n Leamer (.e., 13.3%) s assumed. Data on human captal measure the average years of schoolng of partcpants n the actve labour force (see Baer, Dwyer and Tamura, 2002, for more detals). Endowment data were kndly provded by Scott Baer. Investment cost data are based on score varables publshed n the World Economc Forum s Global ompettveness Report. Amt and Wakeln (2003) provde a detaled descrpton. The data were kndly provded by Keth Maskus. Table A.5 provdes the correlaton matrx and summary statstcs for the explanatory varables. > Table A.5 < ountry sample The country sample the regresson results are based on conssts of blateral trade flows between the followng 31 countres: Australa, Austra, Belgum, anada, hna, zech Republc, Denmark, Fnland, France, Germany, Greece, Hong Kong, Hungary, Iceland, Ireland, Italy, Japan, Republc of Korea, Mexco, Netherlands, New Zealand, Norway, Poland, Portugal, Slovak Republc, Span, Sweden, Swtzerland, Turkey, Unted Kngdom, USA. 36

References Amt, Mary and Katharne Wakeln, Investment Lberalzaton and Internatonal Trade, Journal of Internatonal Economcs 61 (1), 101-126. Anderson. James E. and Erc van Wncoop (2003), Gravty wth Gravtas: A Soluton to the Border Puzzle, Amercan Economc Revew 93 (1), 170-192. Anderson, James E. and Erc van Wncoop (2004), Trade osts, prelmnary draft, n preparaton for the Journal of Economc Lterature. Aquno, Antono (1978), Inta-Industry Trade and Inter-Industry Specalzaton as oncurrent Sources n Internatonal Trade n Manufactures, Weltwrtschaftlches Archv 114 (2), 275-296. Baer, Scott L. and Jeffrey H. Bergstrand (2001), The Growth of World Trade: Tarffs, Transport osts, and Income Smlarty, Journal of Internatonal Economcs 53 (1), 1-27. Baer, Scott L., Gerald P. Dwyer and Robert Tamura (2002), How Important Are aptal and Total Factor Productvty for Economc Growth?, unpublshed manuscrpt. Baltag, Bad H. (2001), Econometrc Analyss of Panel Data, second edton, Wley, hchester. Belsley, D. A., Edwn Kuh, and Roy E. Welsch (1980), Regresson Dagnostcs, Wley, New York. Bergstrand, Jeffrey H. (1983), Measurement and Determnants of Intra-Industry Internatonal Trade, n P. K. Matthew Tharakan (ed.), Intra-ndustry trade: Emprcal and Methodologcal Aspects, Elsever Scence & Technology Books. Bergstrand, Jeffrey H. (1990), The Heckscher-Ohln-Samuelson Model, the Lnder Hypothess, and the Determnants of Blateral Intra-ndustry trade, Economc Journal, 100 (403), 1216-1229. Bhagwat, Jagdsh and Donald R. Davs (1999), Intrandustry Trade: Issues and Theory, n: Melvn, James R., James. Moore, and Raymond Rezman, eds., Trade Theory and Econometrcs: Essays n Honor of John S. hpman, Studes n the Modern World Economy, vol. 15., Routledge, London and New York, 1-16. Brander, James A. (1981), Intra-Industry Trade n Identcal ommodtes, Journal of Internatonal Economcs 11 (1), 1-14. Brander, James A. and Paul R. Krugman (1983), A Recprocal dumpng Model of Internatonal Trade, Journal of Internatonal Economcs 15 (3-4), 313-321. arr, Davd, Keth Maskus and James R. Markusen (2001), Estmatng the Knowledge- aptal Model of the Multnatonal Enterprse, Amercan Economc Revew 91 (2001), 693-708. Davs, Donald R. (1995), Intra-ndustry Trade: A Heckscher-Ohln-Rcardo Approach, Journal of Internatonal Economcs 39 (3-4), 201-226. 37

Davs, Donald R. (1997), rtcal Evdence on omparatve Advantage: North-North Trade n a Multlateral World, Journal of Poltcal Economy 105 (4), 1051-1060. Dxt, Avnash and Joseph Stgltz (1977), Monopolstc ompetton and Optmum Product Dversty; Amercan Economc Revew 67 (3), 297-308. Durkn, John T. and Markus Kryger (1998), omparatve Advantage and the Pattern of Trade wthn Industres, Revew of Internatonal Economcs, 6 (2), 292-306. Durkn, John T. and Markus Kryger (2000), Dfferences n GDP Per apta and the Share of Intrandustry Trade: The Role of Vertcally Dfferentated Trade, Revew of Internatonal Economcs, 8 (4), 760-774. Feenstra Robert. (1994), New Product Varetes and the Measurement of Internatonal Prces, Amercan Economc Revew 84 (1), 157-177. Fnger, J. Mchael (1975), Trade Overlap and Intra-ndustry trade, Economc Inqury 13 (4), 581-589. Greenaway, Davd and hrs Mlner (1981), Trade Imbalance Effects n the Measurement of Intra-Industry Trade, Weltwrtschaftlches Archv 117 (4), 756-762. Greenaway, Davd and hrs Mlner (1983), On the Measurement of Intra-ndustry trade, Economc Journal 93 (372), 900-908. Greenaway, Davd and hrs Mlner (1984), A ross-secton Analyss of Intra-Industry Trade n the UK, European Economc Revew, 25, 319-344. Greenaway, Davd and hrs Mlner (1986), The Economcs of Intra-Industry Trade, Blackwell, Oxford. Greenaway, Davd, Peter Lloyd and hrs Mlner (2001), New oncepts and Measures of the Globalsaton of Producton, Economcs Letters 73 (1), 57-63. Grubel, Herbert G. and Peter J. Lloyd, (1971), The Emprcal Measurement of Intra- Industry Trade, Economc Record 47 (120), 494-517. Helpman, Elhanan (1984), A Smple Theory of Internatonal Trade wth Multnatonal orporatons, Journal of Poltcal Economy 32 (3), 451-71. Helpman, Elhanan (1987), Imperfect ompetton and Internatonal Trade: Evdence from Fourteen Industral ountres, Journal of the Japanese and the Internatonal Economes 1 (1), 62-81. Helpman, Elhanan and Paul R. Krugman (1985), Market Structure and Foregn Trade, MIT Press, ambrdge, Mass. Hummels, Davd and James Levnsohn (1995), Monopolstc ompetton and Internatonal Trade: Reconsderng the Evdence, Quarterly Journal of Economcs 110 (3), 799-836. Krugman, Paul R. (1979), Increasng Returns, Monopolstc ompetton, and Internatonal Trade, Journal of Internatonal Economcs 9 (4), 469-479. Krugman, Paul R. (1980), Scale Economes, Product dfferentaton, and the Pattern of Trade, Amercan Economc Revew 70 (5), 950-959. 38

Krugman, Paul R. (1981), Intra-ndustry Specalsaton and the Gans from Trade, Journal of Poltcal Economy, 89, 959-973. Lancaster, Kelvn (1980), Intra-Industry Trade under Perfect Monopolstc ompetton, Journal of Internatonal Economcs 10 (2), 151-175. Leamer, Edward E. (1984), Sources of Internatonal omparatve Advantage, MIT Press, ambrdge, MA. Leamer, Edward E. (1994), Testng Trade Theory, n Davd Greenaway and L. Alan Wnters Surveys n Internatonal Trade, Blackwell, Oxford. Loertscher, Rudolf and Frank Wolter (1980), Determnants of Intra-Industry Trade: Among ountres and across Industres, Weltwrtschaftlches Archv 116 (2), 280-293. Mankw, Gregory N., Davd Romer, and Davd N. Wel (1992), A ontrbuton to the Emprcs of Economc Growth, Quarterly Journal of Economcs 107 (2), 407-437. Markusen, James R. (1995), The Boundares of Multnatonal Enterprses and the Theory of Internatonal Trade, Journal of Economc Perspectves 9 (2), 169-189. Markusen, James R. (2002), Multnatonal Frms and the Theory of Internatonal Trade, MIT Press, ambrdge. Markusen, James R. and Keth Maskus (2001), A Unfed Approach to Intra-ndustry trade and Drect Foregn Investment, NBER Workng Paper No. 8335. Markusen, James R. and Anthony Venabales (1998), Multnatonal Frms and the New Trade Theory, Journal of Internatonal Economcs 46 (2), 183-204. Markusen, James R. and Anthony Venables (2000), The Theory of Endowment, Intra- Industry, and Multnatonal Trade, Journal of Internatonal Economcs 52 (2), 209-234. Torstensson, Johan (1996), How Robust are Models of Intra-Industry Trade?, Oxford Bulletn of Economcs and Statstcs, 58, 507-524. Venables, Anthony J., Patrca G. Rce and Martn Stewart (2003), The Geography of Intra- Industry Trade: Emprcs, Topcs n Economc Analyss & Polcy 3 (11), http://www.bepress.com/beeap/topcs/vol3/ss1/art11. 39

Table 1 Alternatve Defntons of the Grubel-Lloyd Index Label Defnton Interpretaton GLI 2 mn( EXk, Mk ) ( EX + M ) AGLI GLI 1 AGLI 1 GLI 2 AGLI 2 GLI 3 AGLI 3 GLI 4 AGLI 4 GLI 5 AGLI 5 Where EX EX k k = are aggregate f.o.b exports of and M = Mk are k the correspondng c..f mports of country. Mssng values at the dsaggregated level are treated as 0. As GLI, but mssng values at the dsaggregated level are skpped and not nterpreted as 0. mn( EXk, Mk ) mn( EX, M ) As GLI, but takng nto account that part of the trade volume serves to balance k mbalanced trade n nvsbles as nduced by the presence of MNEs. As GLI 1, but mssng values at the dsaggregated level are skpped and not nterpreted as 0. 2 mn( EXk, EX k ) ( EX + EX ) As GLI, but only consderng trade flows at f.o.b. Wth postve transport costs, k M k EX k and M EX. As GLI2, but mssng values at the dsaggregated level are skpped and not nterpreted as 0. mn( EXk, EX k ) mn( EX, EX ) As GLI 2, but takng nto account that part of the trade volume serves to k balance mbalanced trade n nvsbles as nduced by the presence of MNEs. As GLI3, but mssng values at the dsaggregated level are skpped and not nterpreted as 0. 2 mn( Mk, M k ) ( M + M ) As GLI2, but based on trade flows at c..f. nstead of f.o.b. GLI4 dffers from k GLI f t. k mn( M, M ) mn( M, M ) k k 2 t As GLI4, but mssng values at the dsaggregated level are skpped and not nterpreted as 0. As GLI4, but takng nto account that part of the trade volume serves to balance mbalanced trade n nvsbles as nduced by the presence of MNEs. GLI dffers from GLI f t t. 5 3 As GLI5, but mssng values at the dsaggregated level are skpped and not nterpreted as 0. k

Table 2 - The Determnants of Intra-Industry Trade Shares (Between Regresson Results; 1990-2000 Data; Left-Hand-Sde Varable s Logt-Transformed) (All Left-Hand-Sde Varables are Based on 5-dgt SIT Fgures) GLI AGLI GLI 1 AGLI 1 GLI 3 AGLI 3 Maxmum GDP: max{ln(gdp ),ln(gdp )} δ 1 0.066 * -0.021 0.035-0.047 * -0.181 *** -0.242 *** 1.89 0.87 1.00 1.79 2.78 5.20 Mnmum GDP: mn{ln(gdp ),ln(gdp )} δ 2 0.498 *** 0.136 *** 0.470 *** 0.144 *** 0.483 *** 0.147 *** 13.18 5.16 12.35 5.02 6.74 2.88 Maxmum aptal-labor Rato: max{ln(k /L ),ln(k/l )} δ 3 0.255 *** -0.019 0.304 *** -0.001 0.898 *** 0.366 *** 2.64 0.29 3.12 0.01 4.89 2.78 Mnmum aptal-labor Rato: mn{ln(k /L ),ln(k/l )} δ 4 0.283 *** 0.071 0.346 *** 0.165 *** 0.199-0.038 4.35 1.57 5.23 3.35 1.60 0.43 Maxmum Endowment wth Sklled Labor: max{s /L,S /L } δ 5-1.047 * -1.397 *** -2.070 *** -2.141 *** -3.918 *** -2.824 *** 1.72 3.28 3.34 4.60 3.43 3.37 Mnmum Endowment wth Sklled Labor: mn{s /L,S /L } δ 6 0.112 1.210 *** -0.278 0.590 * 0.428 1.781 *** 0.24 3.72 0.59 1.67 0.48 2.83 Maxmum Investment osts: max{ln(inv ),ln(inv )} δ 7-1.039 *** -0.510 *** -0.878 *** -0.396 *** -1.155 *** -0.892 *** 5.74 4.03 4.79 2.90 3.36 3.65 Mnmum Investment osts: mn{ln(inv ),ln(inv )} δ 8 0.043 0.171 0.166 0.345 ** 0.474 0.619 ** 0.24 1.33 0.90 2.49 1.38 2.52 Maxmum Transport osts: max{ln(t ),ln(t )} δ 9-0.735 *** -0.338 *** -0.544 *** -0.147 *** -0.654 *** -0.162 * 11.11 7.31 8.15 3.05 5.46 1.93 Mnmum Transport osts: mn{ln(t ),ln(t )} δ 10 0.187 ** 0.178 *** 0.142 * 0.064 0.246 * 0.179 * 2.47 3.38 1.87 1.14 1.82 1.87 onstant δ 0-20.650 *** -4.840 *** -17.838 *** -2.369-14.549 *** 1.583 10.64 3.56 9.10 1.64 3.94 0.62 Observatons 866 866 866 866 422 422 R 2 0.52 0.24 0.43 0.14 0.35 0.21 Share of R 2, accounted for by varables 5-10 n % 48.96 68.70 41.39 50.64 47.88 49.17 F-tests (p-values): δ 1 =-δ 2 0.000 *** 0.000 *** 0.000 *** 0.003 *** 0.000 *** 0.096 * δ 3 =-δ 4 0.000 *** 0.487 0.000 *** 0.044 ** 0.000 *** 0.025 ** δ 5 =-δ 6 0.153 0.683 0.000 *** 0.002 *** 0.006 *** 0.255 δ 7 =-δ 8 0.000 *** 0.004 *** 0.000 *** 0.687 0.036 ** 0.240 δ 9 =-δ 10 0.000 *** 0.000 *** 0.000 *** 0.007 *** 0.000 *** 0.756 Absolute t-statstcs below coeffcents. *** sgnfcant at 1%; ** sgnfcant at 5%; * sgnfcant at 10%.

Table 3 - Senstvty Analyss of Preferred Models (Tradtonal, Trade-Imbalance-Adusted Indces; Assumng that Trade osts Generate Exporter Income) Estmates are Based on Between Models and Exclude Extreme Outlers Based on 5-dgt data Based on 4-dgt data Based on 3-dgt data GLI 1 AGLI 1 GLI 1 AGLI 1 GLI 1 AGLI 1 Maxmum GDP: max{ln(gdp ),ln(gdp )} δ 1 0.064 *** -0.004-0.075 *** -0.119 *** -0.174 *** -0.109 *** 3.70 0.25 2.76 8.08 6.75 6.78 Mnmum GDP: mn{ln(gdp ),ln(gdp )} δ 2 0.441 *** 0.160 *** 0.476 *** 0.233 *** 0.436 *** 0.277 *** 23.91 9.24 16.95 14.64 16.40 16.34 Maxmum aptal-labor Rato: max{ln(k /L ),ln(k/l )} δ 3 0.218 *** -0.064 0.678 *** 0.207 *** 0.674 *** 0.197 *** 4.57 1.43 9.18 5.09 9.74 4.51 Mnmum aptal-labor Rato: mn{ln(k /L ),ln(k/l )} δ 4 0.357 *** 0.170 *** 0.138 *** 0.058 ** 0.072 0.095 *** 11.48 5.69 3.03 2.23 1.68 3.41 Maxmum Endowment wth Sklled Labor: max{s /L,S /L } δ 5-1.337 *** -1.900 *** -1.777 *** -1.925 *** -1.790 *** -1.894 *** 4.38 6.60 4.00 7.70 4.31 7.09 Mnmum Endowment wth Sklled Labor: mn{s /L,S /L } δ 6-0.274 0.740 *** 0.403 1.329 *** 0.736 ** 0.870 *** 1.20 3.43 1.17 6.96 2.28 4.24 Maxmum Investment osts: max{ln(inv ),ln(inv )} δ 7-0.809 *** -0.600 *** -1.023 *** -0.771 *** -0.893 *** -1.020 *** 9.22 7.28 7.78 10.37 7.20 12.68 Mnmum Investment osts: mn{ln(inv ),ln(inv )} δ 8 0.152 * 0.432 *** 0.175 0.297 *** 0.191 0.397 *** 1.69 5.15 1.34 3.96 1.54 4.89 Maxmum Transport osts: max{ln(t ),ln(t )} δ 9-0.479 *** -0.149 *** -0.649 *** -0.365 *** -0.642 *** -0.437 *** 15.23 5.25 14.45 14.08 15.20 15.60 Mnmum Transport osts: mn{ln(t ),ln(t )} δ 10 0.116 *** 0.053 * 0.091 * 0.100 *** 0.156 *** 0.165 *** 3.22 1.65 1.81 3.45 3.29 5.19 onstant δ 0-20.102 *** -3.923 *** -19.250 *** -7.010 *** -14.958 *** -6.399 *** 22.91 4.67 14.80 9.40 12.09 8.16 Observatons 859 849 834 840 834 840 R 2 0.82 0.47 0.76 0.73 0.77 0.73 Share of R 2, accounted for by varables 5-10 n % 46.10 54.34 53.52 61.09 54.04 59.52 F-tests (p-values): δ 1 =-δ 2 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** δ 3 =-δ 4 0.000 *** 0.030 ** 0.000 *** 0.000 *** 0.000 *** 0.000 *** δ 5 =-δ 6 0.000 *** 0.000 *** 0.004 *** 0.026 ** 0.019 ** 0.000 *** δ 7 =-δ 8 0.000 *** 0.029 ** 0.000 *** 0.000 *** 0.000 *** 0.000 *** δ 9 =-δ 10 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** Estmates Include Fxed Exporter and Importer Effects and Exclude Extreme Outlers Based on 5-dgt data Based on 4-dgt data Based on 3-dgt data GLI 1 AGLI 1 GLI 1 AGLI 1 GLI 1 AGLI 1 Maxmum GDP: max{gdp,gdp } δ 1-0.254 * -0.656 ** -3.470 *** -3.250 *** -4.254 *** -3.346 *** 1.75 1.98 8.19 11.44 10.99 12.07 Mnmum GDP: mn{gdp,gdp } δ 2-0.204 * -0.598 * -3.665 *** -3.096 *** -4.332 *** -3.055 *** 1.81 1.82 8.67 10.91 11.30 11.07 Maxmum aptal-labor Rato: max{k /L,K/L } δ 3 1.786 *** -0.384 3.636 *** 3.969 *** 1.525 ** 3.702 *** 2.81 0.68 4.67 7.42 2.08 7.42 Mnmum aptal-labor Rato: mn{k /L,K/L } δ 4 1.796 *** -0.027 3.245 *** 4.047 *** 1.339 * 3.697 *** 2.83 0.05 4.20 7.61 1.84 7.45 Maxmum Endowment wth Sklled Labor: max{s /L,S /L } δ 5-0.559-1.356 ** 3.216 *** -3.820 *** 1.870 *** -3.735 *** 1.00 2.35 4.54 8.15 2.75 8.48 Mnmum Endowment wth Sklled Labor: mn{s /L,S /L } δ 6 0.636 0.173 6.063 *** -2.189 ** 4.423 *** -2.246 *** 1.03 0.28 7.70 4.18 5.80 4.62 Maxmum Investment osts: max{ln(inv ),ln(inv )} δ 7-0.482 *** -0.163 * -0.639 *** -0.077 * -0.771 *** -0.195 ** 4.13 1.76 4.43 1.81 5.62 2.09 Mnmum Investment osts: mn{ln(inv ),ln(inv )} δ 8 0.585 *** 0.565 *** 0.520 *** 0.533 *** 0.526 *** 0.770 *** 4.71 4.91 3.44 5.21 3.63 7.80 Maxmum Transport osts: max{ln(t ),ln(t )} δ 9-0.219 *** 0.004-0.177 *** -0.127-0.281 *** -0.148 *** 7.37 0.13 4.78 5.19 8.04 6.28 Mnmum Transport osts: mn{ln(t ),ln(t )} δ 10 0.036 0.018-0.157 *** -0.034-0.092 *** -0.032 1.16 0.66 4.25 1.37 2.65 1.34 onstant δ 0-32.393 ** 37.157 *** 91.088 *** 85.284 *** 180.820 *** 94.208 *** 2.09 2.69 4.41 6.27 9.55 7.15 Observatons 857 848 838 834 838 836 R 2 0.90 0.70 0.91 0.85 0.92 0.89 Share of R 2, accounted for by varables 5-10 n % 51.02 50.40 51.42 50.63 51.34 50.79 F-tests (p-values): δ 1 =-δ 2 0.496 0.058 * 0.000 *** 0.000 *** 0.000 *** 0.000 *** δ 3 =-δ 4 0.005 *** 0.714 0.000 *** 0.000 *** 0.050 ** 0.000 *** δ 5 =-δ 6 0.943 0.288 0.000 *** 0.000 *** 0.000 *** 0.000 *** δ 7 =-δ 8 0.556 0.015 ** 0.574 0.001 *** 0.226 0.000 *** δ 9 =-δ 10 0.000 *** 0.268 0.000 *** 0.000 *** 0.000 *** 0.000 *** Fxed exporter effects 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** Fxed mporter effects 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** Absolute t-statstcs below coeffcents. *** sgnfcant at 1%; ** sgnfcant at 5%; * sgnfcant at 10%.

Table 4 - Senstvty Analyss of Preferred Models (Export-Based, Trade-Imbalance-Adusted Indces; Assumng that Trade osts Do Not Generate Income) Estmates are Based on Between Models and Exclude Extreme Outlers Based on 5-dgt data Based on 4-dgt data Based on 3-dgt data GLI 3 AGLI 3 GLI 3 AGLI 3 GLI 3 AGLI 3 Maxmum GDP: max{gdp,gdp } δ 1-0.106 *** -0.147 *** -0.173 *** -0.191 *** -0.253 *** -0.194 *** 3.09 5.49 4.25 7.27 6.69 6.96 Mnmum GDP: mn{gdp,gdp } δ 2 0.503 *** 0.144 *** 0.528 *** 0.242 *** 0.447 *** 0.340 *** 14.01 5.15 12.82 8.85 11.75 11.93 Maxmum aptal-labor Rato: max{k /L,K/L } δ 3 0.760 *** 0.269 *** 0.690 *** 0.292 *** 0.668 *** 0.205 *** 8.08 3.70 6.47 4.13 6.80 2.76 Mnmum aptal-labor Rato: mn{k /L,K/L } δ 4 0.231 *** -0.030 0.256 *** 0.154 *** 0.201 *** 0.222 *** 3.84 0.60 3.69 3.28 3.19 4.47 Maxmum Endowment wth Sklled Labor: max{s /L,S /L } δ 5-2.968 *** -2.713 *** -2.300 *** -3.010 *** -2.589 *** -2.759 *** 4.85 5.74 3.44 6.64 4.25 5.84 Mnmum Endowment wth Sklled Labor: mn{s /L,S /L } δ 6 0.104 1.717 *** -0.106 1.014 *** 0.280 0.459 0.24 4.88 0.21 2.97 0.60 1.28 Maxmum Investment osts: max{ln(inv ),ln(inv )} δ 7-0.951 *** -0.704 *** -0.861 *** -0.593 *** -0.912 *** -0.924 *** 5.62 5.31 4.42 4.43 5.12 6.62 Mnmum Investment osts: mn{ln(inv ),ln(inv )} δ 8 0.465 *** 0.478 *** 0.600 *** 0.637 *** 0.634 *** 0.792 *** 2.78 3.53 3.11 4.79 3.58 5.68 Maxmum Transport osts: max{ln(t ),ln(t )} δ 9-0.603 *** -0.192 *** -0.480 *** -0.173 *** -0.396 *** -0.232 *** 10.16 4.21 7.01 3.86 6.55 4.78 Mnmum Transport osts: mn{ln(t ),ln(t )} δ 10 0.266 *** 0.183 *** 0.069 0.024 0.013 0.104 * 4.03 3.54 0.90 0.47 0.20 1.91 onstant δ 0-15.706 *** -0.147-16.643 *** -5.373 *** -11.550 *** -3.805 *** 9.29 0.11 8.83 4.08 6.59 2.86 Observatons 413 413 415 415 415 415 R 2 0.74 0.47 0.69 0.57 0.75 0.28 Share of R 2, accounted for by varables 5-10 n % 51.45 52.06 50.66 49.55 50.56 50.29 F-tests (p-values): δ 1 =-δ 2 0.000 *** 0.938 0.000 *** 0.113 0.000 *** 0.000 *** δ 3 =-δ 4 0.000 *** 0.003 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** δ 5 =-δ 6 0.000 *** 0.055 0.001 *** 0.000 *** 0.000 *** 0.000 *** δ 7 =-δ 8 0.003 *** 0.080 * 0.160 0.733 0.100 * 0.319 δ 9 =-δ 10 0.000 *** 0.768 0.000 *** 0.000 *** 0.000 *** 0.000 *** Estmates Include Fxed Exporter and Importer Effects and Exclude Extreme Outlers Based on 5-dgt data Based on 4-dgt data Based on 3-dgt data GLI 3 AGLI 3 GLI 3 AGLI 3 GLI 3 AGLI 3 Maxmum GDP: max{ln(gdp ),ln(gdp )} δ 1-2.181 *** -1.579 ** -4.195 *** -3.442 *** -4.455 *** -2.987 *** 3.68 2.43 5.59 6.07 5.38 5.28 Mnmum GDP: mn{ln(gdp ),ln(gdp )} δ 2-2.265 *** -1.387 ** -4.253 *** -3.276 *** -4.411 *** -2.685 *** 3.85 2.14 5.70 5.80 5.34 4.77 Maxmum aptal-labor Rato: max{ln(k /L ),ln(k/l )} δ 3 3.386 *** 1.512 5.855 *** 5.413 *** 5.062 *** 3.248 *** 3.08 1.40 4.57 5.33 3.53 3.23 Mnmum aptal-labor Rato: mn{ln(k /L ),ln(k/l )} δ 4 2.698 ** 1.509 5.371 *** 5.400 *** 4.896 *** 3.290 *** 2.47 1.40 4.20 5.33 3.42 3.27 Maxmum Endowment wth Sklled Labor: max{s /L,S /L } δ 5-0.996-0.690 2.700 ** -0.441 1.500-1.399 0.93 0.49 2.43 0.47 1.30 1.58 Mnmum Endowment wth Sklled Labor: mn{s /L,S /L } δ 6 3.759 *** 2.406 * 5.466 *** 0.566 3.277 ** 0.123 3.29 1.73 4.51 0.56 2.56 0.13 Maxmum Investment osts: max{ln(inv ),ln(inv )} δ 7-1.080 *** -0.481 ** -0.806 *** -0.549 *** -0.822 *** -0.866 *** 4.78 2.23 3.42 2.86 3.25 4.40 Mnmum Investment osts: mn{ln(inv ),ln(inv )} δ 8 0.906 *** 0.774 *** 1.192 *** 1.017 *** 0.992 *** 0.999 *** 3.94 3.40 4.66 4.89 3.58 4.61 Maxmum Transport osts: max{ln(t ),ln(t )} δ 9-0.147 *** -0.043-0.100 0.004-0.051-0.010 2.64 0.81 1.67 0.08 0.80 0.20 Mnmum Transport osts: mn{ln(t ),ln(t )} δ 10-0.073 0.074-0.081 0.001-0.156 ** -0.060 1.32 1.39 1.37 0.02 2.54 1.17 onstant δ 0 55.571 ** 47.727 103.984 *** 73.757 *** 132.360 *** 90.169 *** 2.02 1.66 3.05 3.00 3.59 3.49 Observatons 413 411 413 413 411 413 R 2 0.89 0.62 0.89 0.78 0.88 0.81 Share of R 2, accounted for by varables 5-10 n % 51.45 52.06 50.66 49.55 50.56 50.29 F-tests (p-values): δ 1 =-δ 2 0.000 *** 0.023 ** 0.000 *** 0.000 *** 0.000 *** 0.000 *** δ 3 =-δ 4 0.006 *** 0.161 0.000 *** 0.000 *** 0.001 *** 0.001 *** δ 5 =-δ 6 0.174 0.519 0.000 *** 0.945 0.032 ** 0.450 δ 7 =-δ 8 0.601 0.372 0.276 0.105 0.658 0.657 δ 9 =-δ 10 0.000 *** 0.371 0.000 *** 0.883 0.000 *** 0.032 ** Fxed exporter effects 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** Fxed mporter effects 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** 0.000 *** Absolute t-statstcs below coeffcents. *** sgnfcant at 1%; ** sgnfcant at 5%; * sgnfcant at 10%.

Table 5 - The Role of Labor and aptal Endowments for the Impact of Investment osts (5-Dgt Data Based; All Regressons Include ountry Effects and Exclude Outlers) The Role of Labor Endowments The Role of Physcal aptal Endowments GLI 1 AGLI 1 GLI 3 AGLI 3 GLI 1 AGLI 1 GLI 3 AGLI 3 Maxmum GDP: max{ln(gdp ),ln(gdp )} δ 1-0.249 *** -0.691 ** -2.042 *** -1.343 ** -0.312 *** -0.679 ** -2.151 *** -1.191 * 10.76 2.13 3.43 2.18 10.95 2.09 3.56 1.92 Mnmum GDP: mn{ln(gdp ),ln(gdp )} δ 2-0.185 *** -0.633 * -2.125 *** -1.129 * -0.250 *** -0.623 * -2.224 *** -0.976 10.57 1.96 3.59 1.83 10.77 1.92 3.70 1.58 Maxmum aptal-labor Rato: max{ln(k /L ),ln(k/l )} δ 3 2.024 *** -0.326 3.185 *** 1.683 2.158 *** -0.335 3.300 *** 1.552 3.27 0.59 2.89 1.58 3.49 0.60 2.97 1.45 Mnmum aptal-labor Rato: mn{ln(k /L ),ln(k/l )} δ 4 2.047 *** 0.003 2.540 ** 1.619 2.181 *** -0.004 2.611 ** 1.479 3.31 0.01 2.31 1.52 3.53 0.01 2.36 1.38 Maxmum Endowment wth Sklled Labor: max{s /L,S /L } δ 5-0.575-1.466 ** -1.281-2.029-0.558-1.471 ** -1.290-2.130 1.04 2.56 1.18 1.46 1.01 2.56 1.18 1.52 Mnmum Endowment wth Sklled Labor: mn{s /L,S /L } δ 6 0.556 0.184 3.213 *** 1.784 0.543 0.175 3.091 *** 1.708 0.91 0.30 2.79 1.30 0.89 0.29 2.66 1.24 Maxmum Investment osts: max{ln(inv ),ln(inv )} δ 7-0.682 *** -0.317 ** -1.401 *** -1.613 *** -0.683 *** -0.311 ** -1.432 *** -1.620 *** 4.80 2.34 4.52 5.55 4.80 2.29 4.57 5.54 Mnmum Investment osts: mn{ln(inv ),ln(inv )} δ 8 0.930 *** 0.700 *** 1.292 *** 1.690 *** 0.937 *** 0.687 *** 1.351 *** 1.688 *** 6.11 4.91 4.36 5.91 6.17 4.82 4.52 5.88 Maxmum Transport osts: max{ln(t ),ln(t )} δ 9-0.22 *** 0.01-0.17 *** -0.10 * -0.22 *** 0.01-0.18 *** -0.10 * 7.36 0.27 3.07 1.95 7.46 0.35 3.13 1.84 Mnmum Transport osts: mn{ln(t ),ln(t )} δ 10 0.025 0.010-0.038 0.122 ** 0.026 0.010-0.026 0.118 ** 0.83 0.37 0.68 2.27 0.87 0.35 0.47 2.19 Interacton: {ln(inv ),ln(inv )} ln(l ) f INV >INV, else {ln(inv ),ln(inv )} ln(l ) δ 11 0.033 *** 0.019 ** 0.037 * 0.103 *** - - - - 3.24 1.99 1.78 5.23 - - - - Interacton: {ln(inv ),ln(inv )} ln(k ) f INV >INV, else {ln(inv ),ln(inv )} ln(k ) δ 12 - - - - 0.02 *** 0.01 * 0.03 ** 0.07 *** - - - - 3.25 1.91 2.03 5.28 onstant δ 0-38.86 ** 38.31 *** 52.93 * 35.78-38.38 ** 37.94 *** 56.89 ** 30.37 2.55 2.80 1.90 1.29 2.52 2.77 2.03 1.09 Observatons 857 848 413 411 857 848 413 411 R 2 0.90 0.71 0.89 0.65 0.90 0.70 0.88 0.64 {ln(inv ),ln(inv )} s defned as max{ln(inv ),ln(inv )} - mn{ln(inv ),ln(inv )}. Absolute t-statstcs below coeffcents. *** sgnfcant at 1%; ** sgnfcant at 5%; * sgnfcant at 10%.

Table 6 - Explanng GLI /GLI GLI 1/GLI 1 GLI 3/GLI 3 {ln(inv ),ln(inv )} -2.191-89.663 ** 1.52 2.12 Interacton: {ln(inv ),ln(inv )} ln(l ) f INV >INV, else {ln(inv ),ln(inv )} ln(l ) 1.551 * 7.719 ** 1.80 2.06 onstant -2.452-9.184 0.20 0.18 Observatons 857 413 R 2 0.11 0.03 F-tests (p-values): Jont sgnfcance of all other explanatory varables (see Footnote) 0.199 0.433 Fxed exporter effects 0.000 *** 0.000 *** Fxed mporter effects 0.011 ** 0.000 *** {ln(inv ),ln(inv )} s defned as max{ln(inv ),ln(inv )} - mn{ln(inv ),ln(inv )}. oeffcents of max{ln(gdp ),ln(gdp )}, mn{ln(gdp ),ln(gdp )}, max{ln(k /L ),ln(k /L )}, mn{ln(k /L ),ln(k /L )}, max{ln(s /L ),ln(s /L )}, mn{ln(s /L ),ln(s /L )}, max{ln(t ), ln(t )}, mn{ln(t ),ln(t )} not reported due to ther nsgnfcance (see the F-statstcs). Absolute t-statstcs below coeffcents. *** sgnfcant at 1%; ** sgnfcant at 5%; * sgnfcant at 10%.

Table A.1 - Smulaton Set-up Vertcal Model Horzontal Model Knowledge-aptal Model Leontef obb-douglas ES V H KK1 KK2 KK3 Endowments of : Share of K [0.62,0.77] [0.45,0.55] [0.40,0.60] [0.40,0.60] [0.40,0.60] Share of S [0.48,0.52] [0.45,0.55] [0.40,0.60] [0.40,0.60] [0.40,0.60] Share of L [0.15,0.25] [0.45,0.55] [0.40,0.60] [0.40,0.60] [0.40,0.60] Input coeffcents: a KX 0 0 0.3 a SX 0 0 0.2 a LX 0.5 0.5 0.5 see footnote see footnote see footnote Investment costs: Addtonal foregn nvestment costs of [1.10,1.30] [0.10,0.30] [0.10,0.30] [0.10,0.30] [0.10,0.30] Addtonal foregn nvestment costs of 1.2 0.2 0.2 0.2 0.2 Trade costs of dfferentated goods: Iceberg parameter T [1.05,1.25] [1.05,1.25] [1.05,1.25] [1.05,1.25] [1.05,1.25] Iceberg parameter of T 1.15 1.15 1.15 1.15 1.15 In all experments, we set ε=6 (see Feenstra, 1994) and α=0.8 (accordng to UN omtrade data for 1990-2000). The stepwdth between mnmum and maxmum addtonal foregn nvestment costs of country s always 0.05. We assume the followng values for world endowments: K=60; S=40; L=100. In Model KK2 S=80, and n models KK1 and KK3 S=200 and K=300 are asuumed to ensure that exporters and horzontal multnatonals co-exst n the center of the factor cube. The factor box s always splt nto 21 segments of equal sze n any of the two dmensons, so that there are 21µ21µ21 equlbra to be solved for each level of nvestment costs. For the obb- Douglas case, we assume the producton technology z =K 0.3 S 0.2 L 0.5 wth =1,2. For the more general case of a onstant Elastcty of Substtuton Technology, we assume z =[0.3K ρ +0.2S ρ +0.5L ρ ] 1/ρ wth ρ=-10 and =1,2.

Table A.2 - Summary Statstcs for Dfferent oncenpts of the Grubel-Lloyd Index Observatons Mean Std. Dev. Mnmum Maxmum Tme Invar. a) 5-dgt SIT data GLI (usual defnton; mssngs=0) 8429 0.14 0.13 0.00 0.64 0.92 AGLI (mssngs 0) 8429 0.21 0.14 0.00 1.00 0.69 GLI 1 (GLI balance-adusted) 8429 0.23 0.23 0.00 1.00 0.91 AGLI 1 (AGLI balance-adusted) 8429 0.35 0.24 0.00 1.00 0.72 GLI 2 (GLI export-based) 7259 0.13 0.14 0.00 1.00 0.88 AGLI 2 (AGLI export-based) 7259 0.20 0.14 0.00 1.00 0.69 GLI 3 (GLI 2 balance adusted) 7259 0.24 0.27 0.00 1.00 0.94 AGLI 3 (AGLI 2 balance adusted) 7259 0.36 0.28 0.00 1.00 0.77 GLI 4 (GLI mport-based) 7429 0.14 0.15 0.00 1.00 0.95 AGLI 4 (GLI mport-based) 7429 0.21 0.15 0.00 1.00 0.81 GLI 5 (GLI 4 balance adusted) 7429 0.25 0.27 0.00 1.00 0.96 AGLI 5 (AGLI 4 balance adusted) 7429 0.36 0.27 0.00 1.00 0.83 4-dgt SIT data GLI (usual defnton; mssngs=0) 8495 0.17 0.15 0.00 0.71 0.91 AGLI (mssngs 0) 8495 0.23 0.15 0.00 1.00 0.75 GLI 1 (GLI balance-adusted) 8495 0.27 0.24 0.00 1.00 0.90 AGLI 1 (AGLI balance-adusted) 8495 0.38 0.24 0.00 1.00 0.73 GLI 2 (GLI export-based) 7345 0.15 0.15 0.00 1.00 0.87 AGLI 2 (AGLI export-based) 7345 0.21 0.15 0.00 1.00 0.73 GLI 3 (GLI 2 balance adusted) 6878 0.13 0.19 0.00 1.00 0.92 AGLI 3 (AGLI 2 balance adusted) 6878 0.24 0.17 0.00 1.00 0.78 GLI 4 (GLI mport-based) 7345 0.28 0.28 0.00 1.00 0.95 AGLI 4 (GLI mport-based) 7345 0.38 0.28 0.00 1.00 0.86 GLI 5 (GLI 4 balance adusted) 6878 0.25 0.31 0.00 1.00 0.96 AGLI 5 (AGLI 4 balance adusted) 6878 0.41 0.27 0.00 1.00 0.84 3-dgt SIT data GLI (usual defnton; mssngs=0) 8491 0.21 0.17 0.00 0.78 0.91 AGLI (mssngs 0) 8491 0.26 0.17 0.00 1.00 0.80 GLI 1 (GLI balance-adusted) 8491 0.33 0.25 0.00 1.00 0.88 AGLI 1 (AGLI balance-adusted) 8491 0.41 0.25 0.00 1.00 0.75 GLI 2 (GLI export-based) 7337 0.19 0.17 0.00 1.00 0.86 AGLI 2 (AGLI export-based) 7337 0.24 0.17 0.00 1.00 0.77 GLI 3 (GLI 2 balance adusted) 7472 0.21 0.18 0.00 1.00 0.89 AGLI 3 (AGLI 2 balance adusted) 7472 0.25 0.17 0.00 1.00 0.78 GLI 4 (GLI mport-based) 7337 0.34 0.29 0.00 1.00 0.92 AGLI 4 (GLI mport-based) 7337 0.41 0.28 0.00 1.00 0.86 GLI 5 (GLI 4 balance adusted) 7472 0.35 0.27 0.00 1.00 0.92 AGLI 5 (AGLI 4 balance adusted) 7472 0.42 0.27 0.00 1.00 0.82 a) Ths s the share of tme-nvarant nformaton n the data.

Table A.3 - orrelaton Matrx for Dfferent oncenpts of the Grubel-Lloyd Index (OED Annual Statstcs of Foregn Trade Data, 1990-2000) Indces based on 5-dgt SIT data GL AGL GL 1 AGL 1 GL 2 AGL 2 GL 3 AGL 3 GL 4 AGL 4 GL 5 AGL 5 5-dgt SIT data GLI (usual defnton; mssngs=0) 1.00 AGLI (mssngs 0) 0.71 1.00 GLI 1 (GLI balance-adusted) 0.59 0.39 1.00 AGLI 1 (AGLI balance-adusted) 0.28 0.56 0.70 1.00 GLI 2 (GLI export-based) 0.87 0.60 0.53 0.23 1.00 AGLI 2 (AGLI export-based) 0.63 0.61 0.35 0.30 0.77 1.00 GLI 3 (GLI 2 balance adusted) 0.36 0.18 0.74 0.51 0.46 0.32 1.00 AGLI 3 (AGLI 2 balance adusted) 0.14 0.17 0.54 0.61 0.26 0.45 0.79 1.00 GLI 4 (GLI mport-based) 0.87 0.56 0.50 0.18 0.88 0.63 0.36 0.14 1.00 AGLI 4 (GLI mport-based) 0.69 0.62 0.36 0.26 0.73 0.65 0.24 0.17 0.85 1.00 GLI 5 (GLI 4 balance adusted) 0.37 0.18 0.73 0.47 0.40 0.23 0.92 0.69 0.50 0.40 1.00 AGLI 5 (AGLI 4 balance adusted) 0.16 0.18 0.56 0.59 0.21 0.21 0.75 0.73 0.30 0.46 0.81 1.00 4-dgt SIT data GLI (usual defnton; mssngs=0) 0.96 0.64 0.56 0.23 0.84 0.58 0.36 0.12 0.84 0.63 0.37 0.14 AGLI (mssngs 0) 0.80 0.82 0.44 0.38 0.68 0.56 0.24 0.11 0.65 0.58 0.23 0.11 GLI 1 (GLI balance-adusted) 0.59 0.36 0.95 0.64 0.54 0.33 0.73 0.51 0.51 0.34 0.71 0.51 AGLI 1 (AGLI balance-adusted) 0.36 0.46 0.75 0.85 0.31 0.29 0.54 0.55 0.26 0.25 0.50 0.52 GLI 2 (GLI export-based) 0.83 0.53 0.50 0.19 0.96 0.72 0.46 0.23 0.85 0.67 0.39 0.19 AGLI 2 (AGLI export-based) 0.70 0.56 0.40 0.24 0.85 0.85 0.37 0.32 0.71 0.66 0.30 0.22 GLI 3 (GLI 2 balance adusted) 0.55 0.34 0.27 0.06 0.61 0.44 0.20 0.06 0.71 0.61 0.30 0.17 AGLI 3 (AGLI 2 balance adusted) 0.69 0.53 0.32 0.14 0.74 0.62 0.21 0.09 0.84 0.87 0.34 0.30 GLI 4 (GLI mport-based) 0.38 0.17 0.72 0.46 0.49 0.31 0.97 0.75 0.38 0.25 0.89 0.70 AGLI 4 (GLI mport-based) 0.21 0.16 0.58 0.55 0.33 0.38 0.83 0.89 0.21 0.21 0.73 0.71 GLI 5 (GLI 4 balance adusted) 0.22 0.07 0.57 0.36 0.28 0.16 0.78 0.60 0.35 0.26 0.83 0.67 AGLI 5 (AGLI 4 balance adusted) 0.24 0.17 0.58 0.51 0.30 0.25 0.76 0.69 0.37 0.43 0.82 0.90 3-dgt SIT data GLI (usual defnton; mssngs=0) 0.93 0.63 0.54 0.20 0.83 0.57 0.35 0.10 0.82 0.61 0.36 0.12 AGLI (mssngs 0) 0.85 0.76 0.47 0.31 0.73 0.57 0.27 0.10 0.72 0.62 0.27 0.12 GLI 1 (GLI balance-adusted) 0.61 0.37 0.91 0.61 0.56 0.34 0.70 0.48 0.53 0.35 0.68 0.48 AGLI 1 (AGLI balance-adusted) 0.43 0.45 0.77 0.79 0.39 0.32 0.56 0.53 0.34 0.31 0.53 0.53 GLI 2 (GLI export-based) 0.81 0.51 0.48 0.16 0.93 0.69 0.44 0.21 0.83 0.65 0.38 0.17 AGLI 2 (AGLI export-based) 0.74 0.55 0.42 0.20 0.87 0.77 0.38 0.26 0.77 0.70 0.33 0.22 GLI 3 (GLI 2 balance adusted) 0.83 0.49 0.46 0.11 0.84 0.59 0.35 0.11 0.95 0.78 0.47 0.24 AGLI 3 (AGLI 2 balance adusted) 0.77 0.54 0.40 0.14 0.80 0.63 0.29 0.12 0.90 0.84 0.42 0.28 GLI 4 (GLI mport-based) 0.41 0.17 0.71 0.44 0.52 0.33 0.93 0.72 0.41 0.27 0.85 0.67 AGLI 4 (GLI mport-based) 0.28 0.18 0.60 0.52 0.39 0.37 0.83 0.83 0.29 0.27 0.74 0.70 GLI 5 (GLI 4 balance adusted) 0.43 0.18 0.71 0.41 0.47 0.27 0.87 0.63 0.54 0.40 0.94 0.74 AGLI 5 (AGLI 4 balance adusted) 0.31 0.20 0.62 0.51 0.36 0.27 0.79 0.68 0.43 0.41 0.85 0.85 Indces based on 4-dgt SIT data GL AGL GL 1 AGL 1 GL 2 AGL 2 GL 3 AGL 3 GL 4 AGL 4 GL 4 AGL 4 4-dgt SIT data GLI (usual defnton; mssngs=0) 1.00 AGLI (mssngs 0) 0.81 1.00 GLI 1 (GLI balance-adusted) 0.64 0.49 1.00 AGLI 1 (AGLI balance-adusted) 0.37 0.57 0.76 1.00 GLI 2 (GLI export-based) 0.86 0.67 0.57 0.31 1.00 AGLI 2 (AGLI export-based) 0.71 0.65 0.44 0.35 0.86 1.00 GLI 3 (GLI 2 balance adusted) 0.56 0.43 0.31 0.14 0.61 0.52 1.00 AGLI 3 (AGLI 2 balance adusted) 0.67 0.58 0.34 0.20 0.71 0.67 0.84 1.00 GLI 4 (GLI mport-based) 0.41 0.27 0.75 0.54 0.53 0.43 0.23 0.23 1.00 AGLI 4 (GLI mport-based) 0.23 0.20 0.59 0.61 0.35 0.48 0.12 0.15 0.83 1.00 GLI 5 (GLI 4 balance adusted) 0.25 0.14 0.57 0.40 0.30 0.23 0.58 0.41 0.77 0.64 1.00 AGLI 5 (AGLI 4 balance adusted) 0.24 0.18 0.57 0.52 0.30 0.29 0.41 0.49 0.74 0.71 0.82 1.00 3-dgt SIT data GLI (usual defnton; mssngs=0) 0.98 0.80 0.62 0.35 0.85 0.70 0.56 0.67 0.41 0.21 0.25 0.23 AGLI (mssngs 0) 0.87 0.93 0.53 0.48 0.73 0.68 0.48 0.64 0.30 0.21 0.17 0.21 GLI 1 (GLI balance-adusted) 0.66 0.51 0.98 0.74 0.59 0.46 0.34 0.37 0.74 0.57 0.55 0.55 AGLI 1 (AGLI balance-adusted) 0.45 0.56 0.79 0.93 0.39 0.39 0.21 0.27 0.57 0.61 0.42 0.54 GLI 2 (GLI export-based) 0.85 0.65 0.55 0.28 0.98 0.84 0.61 0.71 0.52 0.34 0.30 0.29 AGLI 2 (AGLI export-based) 0.76 0.66 0.47 0.32 0.89 0.93 0.56 0.73 0.45 0.42 0.26 0.31 GLI 3 (GLI 2 balance adusted) 0.85 0.63 0.52 0.22 0.86 0.72 0.71 0.83 0.40 0.22 0.36 0.35 AGLI 3 (AGLI 2 balance adusted) 0.78 0.66 0.44 0.26 0.80 0.73 0.68 0.86 0.33 0.23 0.31 0.38 GLI 4 (GLI mport-based) 0.45 0.28 0.75 0.52 0.57 0.46 0.27 0.28 0.98 0.82 0.74 0.73 AGLI 4 (GLI mport-based) 0.30 0.24 0.62 0.60 0.42 0.48 0.18 0.23 0.85 0.94 0.64 0.72 GLI 5 (GLI 4 balance adusted) 0.47 0.28 0.74 0.49 0.50 0.38 0.36 0.40 0.89 0.71 0.80 0.80 AGLI 5 (AGLI 4 balance adusted) 0.33 0.25 0.63 0.57 0.37 0.36 0.28 0.38 0.78 0.74 0.72 0.86 Indces based on 3-dgt SIT data GL AGL GL 1 AGL 1 GL 2 AGL 2 GL 3 AGL 3 GL 4 AGL 4 GL 4 AGL 4 3-dgt SIT data GLI (usual defnton; mssngs=0) 1.00 AGLI (mssngs 0) 0.87 1.00 GLI 1 (GLI balance-adusted) 0.67 0.56 1.00 AGLI 1 (AGLI balance-adusted) 0.44 0.59 0.80 1.00 GLI 2 (GLI export-based) 0.86 0.73 0.60 0.37 1.00 AGLI 2 (AGLI export-based) 0.77 0.74 0.50 0.42 0.89 1.00 GLI 3 (GLI 2 balance adusted) 0.86 0.73 0.56 0.33 0.87 0.80 1.00 AGLI 3 (AGLI 2 balance adusted) 0.79 0.75 0.48 0.37 0.80 0.80 0.94 1.00 GLI 4 (GLI mport-based) 0.46 0.34 0.76 0.58 0.58 0.50 0.46 0.38 1.00 AGLI 4 (GLI mport-based) 0.30 0.29 0.62 0.65 0.42 0.53 0.32 0.32 0.86 1.00 GLI 5 (GLI 4 balance adusted) 0.48 0.35 0.75 0.56 0.50 0.44 0.59 0.52 0.89 0.75 1.00 AGLI 5 (AGLI 4 balance adusted) 0.33 0.31 0.63 0.63 0.37 0.39 0.45 0.51 0.77 0.77 0.88 1.00

Table A.4 - Quantfyng the Varous Sources of Bas n Intra-Industry Trade Shares (Bas Fgures are Averaged over Tme, Blateral Relatonshps and the Three Aggregaton Levels) Transport cost Transport cost Trade mbalance Mssng value Fnger (1975) Total f 5-dgt level bas dfference bas bas nterpret. bas bas per dgt data used a) GLI (usual defnton; mssngs=0) 0.017-0.104 0.039-0.104 AGLI (mssngs 0) 0.016-0.146 0.059 0.023-0.031 GLI 1 (GLI balance-adusted) -0.005 0.055-0.013 AGLI 1 (AGLI balance-adusted) -0.003 0.101 0.030 0.110 GLI 2 (GLI export-based) -0.128 0.035-0.117 AGLI 2 (AGLI export-based) -0.167 0.057 0.020-0.044 GLI 3 (GLI 2 balance adusted) 0.047 0.000 AGLI 3 (AGLI 2 balance adusted) 0.096 0.023 0.120 GLI 4 (GLI mport-based) 0.001-0.127 0.037-0.108 AGLI 4 (GLI mport-based) 0.011-0.168 0.070 0.022-0.045 GLI 5 (GLI 4 balance adusted) -0.004 0.049 0.007 AGLI 5 (AGLI 4 balance adusted) 0.007 0.110 0.026 0.112 Weghted average (# of obs. weght) 0.006 0.004-0.139 0.082 0.034-0.009 All fgures assume n accordance wth our theoretcal model that LI 3 s the correct ndex. - a) Excludng the Fnger bas.

Table A.5 - orrelaton Matrx and Descrptve Statstcs of Explanatory Varables (Varables n Logs) Max GDP Mn GDP Max K/L Mn K/L Max S/L Mn S/L Max INV Mn INV Max T Mn T Maxmum GDP: max{ln(gdp ),ln(gdp )} 1.00 Mnmum GDP: mn{ln(gdp ),ln(gdp )} 0.45 1.00 Maxmum aptal-labor Rato: max{ln(k /L ),ln(k/l )} 0.28 0.16 1.00 Mnmum aptal-labor Rato: mn{ln(k /L ),ln(k/l )} 0.13 0.11 0.45 1.00 Maxmum Endowment wth Sklled Labor: max{ln(s /L ),ln(s /L )} 0.41 0.27 0.44 0.36 1.00 Mnmum Endowment wth Sklled Labor: mn{ln(s /L ),ln(s /L )} 0.17-0.05 0.37 0.70 0.50 1.00 Maxmum Investment osts: max{ln(inv ),ln(inv )} 0.05 0.20-0.23-0.02-0.26-0.20 1.00 Mnmum Investment osts: mn{ln(inv ),ln(inv )} -0.02-0.06-0.39-0.29-0.37-0.23 0.56 1.00 Maxmum Transport osts: max{ln(t ),ln(t )} -0.09-0.15-0.12-0.24-0.15-0.26-0.12-0.01 1.00 Mnmum Transport osts: mn{ln(t ),ln(t )} -0.14-0.17-0.12-0.17-0.13-0.17-0.07-0.01 0.71 1.00 Descrptve Statstcs Mean 27.04 25.46 11.13 10.23 1.90 1.77 3.58 3.34-1.46-2.09 Standard Devaton 1.28 1.21 0.57 1.04 0.08 0.14 0.29 0.30 1.21 1.29