Trade Costs and Intra-Industry Trade

Transcription

1 Trade Costs and Intra-Industry Trade Jeffrey H. Bergstrand Department of Fnance Mendoza School of Busness Unversty of Notre Dame Notre Dame, IN USA E-mal: Peter Egger Department of Economcs Unversty of Innsbruck Unverstätsstrasse 15 A-6020 Innsbruck Austra E-mal: Abstract Formal economc modelng of ntra-ndustry trade essentally began wth Krugman (1979, 1980, 1981), Lancaster (1980), and Helpman (1981). Helpman (1987) s generally credted wth the frst econometrc test of these theores of ntra-ndustry trade, wth subsequent notable contrbutons from Hummels and Levnsohn (1995) and Debaere (2004). However, all these theores and emprcal models gnore transportaton or, more broadly, trade costs. Yet, as Anderson and van Wncoop (2004) suggest, trade costs are qute large. Ths paper extends work by Bergstrand (1990) that addressed ntrandustry trade n the explct presence of trade costs. In the context of a Helpman-Krugman cum trade costs model, we derve four emprcally testable hypotheses regardng ntra-ndustry trade and trade costs. These hypotheses are nvestgated emprcally usng a cross-secton of blateral OECD Grubel-Lloyd ndexes. The results are strongly n accordance wth the hypotheses, ndcatng the mportance of a more rgorous and systematc treatment of trade costs n the ntra-ndustry trade lterature. 1

2 Trade Costs and Intra-Industry Trade By Jeffrey H. Bergstrand and Peter Egger CONTENTS: I. Introducton II. Theoretcal Issues III. Data IV. Econometrc Analyss V. Senstvty Analyss VI. Conclusons. I. Introducton Trade costs have economcally sensble magntudes and patterns across countres and regons and across goods, suggestng useful hypotheses for deeper understandng (Anderson and van Wncoop, 2004, p. 1). Grubel and Lloyd (1975) created an ndustry n the nternatonal trade lterature. Ther systematc emprcal nvestgaton of trade flows yelded the semnal observaton that the bulk of nternatonal trade certanly among ndustralzed natons was ntra-ndustry, not nter-ndustry. Ths was a startlng observaton for nternatonal trade economsts whose prevalng theores of nternatonal trade at that tme the Rcardan and Heckscher-Ohln theores could only explan nter-ndustry trade. These facts motvated several nsghtful trade theorsts to combne the ndustral organzaton and nternatonal trade lteratures to offer formal theores of ntra-ndustry trade. Notably, Krugman (1979, 1980, 1981), Lancaster (1980), and Helpman (1981) are generally cted as the most nfluental papers n ths regard. Helpman and Krugman (1985) s a semnal book syntheszng and enhancng ths theory. Of course, the absence of a formal theoretcal foundaton for ntra-ndustry trade (IIT) certanly dd not prevent emprcal trade economsts from estmatng econometrc models of the determnants of ntra-ndustry trade pror to However, Helpman (1987) s generally cted as provdng the frst testable hypotheses of ntra-ndustry trade based upon an explct general equlbrum model. Among other papers, several semnal artcles have re-evaluated Helpman s emprcal propostons n the context of formal theores, ncludng Hummels and Levnsohn (1995), Evenett and Keller (2002), and Debaere (2004). 2

3 However, each of the papers just noted have evaluated ntra-ndustry trade n the context of a model wth zero trade costs. As Anderson and van Wncoop (2004) remnd us convncngly, trade costs are large and matter. Even the large emprcal lterature on determnants of ntra-ndustry trade lackng formal theoretcal foundatons found farly systematcally that dstance sgnfcantly reduces ntra-ndustry trade, economcally and statstcally. However, whle the nternatonal trade lterature (especally, work usng the gravty equaton) has provded convncng ratonales for the negatve relatonshp between dstance as a proxy for trade costs and the volume of trade, there s not yet a well accepted ratonale for why dstance should have a strong negatve emprcal correlaton wth the share of ntra-ndustry trade, especally after accountng for countres wth common land borders (.e., cross-haulng ). One paper that dd try to address theoretcally and emprcally the mportance of transport costs n the context of a twosector model of Heckscher-Ohln nter-ndustry and Helpman-Krugman ntra-ndustry trade s Bergstrand (1990). The purpose of the present paper s to advance some new theoretcal and emprcal nsghts nto the relatonshp between ntra-ndustry trade and trade costs. Anderson and van Wncoop (2004) s an excellent survey of nternatonal trade costs, and among other goals dscusses n partcular the relatonshp between trade costs and the volume of trade. Our paper s amed at enhancng our knowledge of the relatonshp between trade costs and the share of ntra-ndustry trade. We also address ndrectly an mportant ssue rased n Davs (1998) on the relatonshp between absolute trade costs vs. relatve trade costs (between two ndustres products) for nternatonal trade and the home-market effect. In ths paper, we enhance the standard two-country, two-good, two-factor Helpman- Krugman model to ncorporate explct transport costs for both the dfferentated and homogeneous products. In the presence of postve transport costs, analytcal solutons can only be obtaned by focusng the analyss on a lmted (and often unrealstc) set of parameter domans. Consequently, we provde numercal solutons to the nonlnear relatonshps between trade costs and Grubel-Lloyd ntra-ndustry trade ndexes (GLI). Specfcally, we motvate four testable hypotheses. Frst, an ncrease n trade costs assocated wth only dfferentated goods should reduce both the volume of ntra-ndustry trade n dfferentated goods and the share of such ntra-ndustry trade n overall trade. Second, a proportonal ncrease n trade costs (across both sectors) wll tend to reduce the overall G-L ndex as well. Thrd, the presence of explct trade costs ntroduces nonlneartes nto the model that can nfluence potentally the senstvtes of relatonshps 3

4 among trade costs and the share of ntra-ndustry trade to economc sze and relatve factor proportons. We rely upon numercal solutons from a computable general equlbrum (CGE) verson of our theoretcal model to show, for nstance, that the effect of a proportonal ncrease n trade costs s senstve to the level of dfferences n relatve factor endowments. Fourth, we show also that the margnal effect of an ncrease n only dfferentated goods trade costs s also senstve to relatve factor endowment dfferences. Fnally, we nvestgate these four hypotheses emprcally usng a large cross-secton of blateral G-L ndexes. The results confrm our theoretcal hypotheses. The remander of the paper s as follows. Secton 2 outlnes the theoretcal model and the four emprcally testable hypotheses. Secton 3 dscusses our database. Secton 4 presents the man emprcal results. Secton 5 presents the results of a senstvty analyss. The last secton concludes. II. Theoretcal Issues A. The Model To llustrate the role of trade costs for ntra-ndustry trade, consder a two-country, two-sector, two-factor model à la Helpman and Krugman (1985). One of the two sectors produces a Dxt and Stgltz (1977) constant-elastcty-of-substtuton (CES) type dfferentated good (X), and the other sector produces a homogeneous good (Y). We assume fxed endowments of two factors, captal (K) and labor (L), n each of two countres (,j). In country s dfferentated sector, n frms engage n (large-numbers-case) monopolstc competton. Each frm faces a demand x n the domestc market and x j n the foregn market. These demands are gven by: x = p P αe ; x = p t P αe, (1) ε 1 ε 1 ε 1 X X j X X Xj where p X s the prce of each dfferentated varety n country, a s the expendture share on dfferentated goods (hence, consumers spend a share of 1-a on the homogeneous good), E =w L +r K s total ncome of labor and captal (w and r denote the respectve factor rewards), P X s the CES prce ndex gven by: 1 ε X + 1 ( t p ) ε P = n p n, (2) and ε denotes the elastcty of substtuton between varetes. We assume ceberg-type transport costs n both sectors (see Samuelson, 1952). We consder non-zero and dfferent transport costs n both sectors. Assume t X -1 (t Y -1) unts of each dfferentated varety (of the homogeneous good) melt durng transportaton of goods to foregn consumers. In ths regard, our approach dffers from several prevous analyses of the Helpman-Krugman 4 j X Xj j

5 model, where t has been typcally assumed that t X = t Y = 1 (Helpman, 1987, Hummels and Levnsohn, 1995; Evenett and Keller, 2002; Debaere, 2004). However, Davs (1998) studes the role of transport costs for the home-bas and focuses on two specfc confguratons of t X and t Y, namely t X = t Y 1 and t X 1 but t Y = 1. An assumpton of factor market clearng guarantees: L = a K LX = a KX n ( x n ( x + x ) + a j j LY + x ) + a KY ( Y We are nterested n the comparatve statc results of the G-L ndex wth respect to t X and t Y n partcular. A comparatve statc analyss n models of monopolstc competton s generally messy, and analytcal results can only be obtaned by focusng on certan (often 5 ( Y + t Y ) + a Y Y j j Ln + t Y ) + a We wll make a few plausble assumptons regardng technology to smplfy the analyss. Frst, we assume frm setup s captal ntensve relatve to producton of goods. Second, we assume that producton of the dfferentated good s captal ntensve relatve to that of the homogeneous good. Ths s guaranteed formally by a a Ln a a LX a a LY n Kn n. (3) Kn KX KY > >. (4) We ensure the latter nequalty by assumng a KY = 0. Further, we assume a world-wde dentcal Leontef technology, whch rules out the possblty of factor ntensty reversals. For convenence, we express X n terms of produced unts and Y n consumed unts. An assumpton of free entry and ext guarantees zero profts n the dfferentated goods sector: ( p c )( x + x ) = a w + a r (5) X X j where c X denotes margnal costs (average varable costs) n the X-sector. Large-number monopolstc competton leads to a constant mark-up over margnal costs, so that we can wrte the prcng condtons applyng to both sectors as Ln ε p X = c X ; py = cy = w ε 1 Kn. (6) Choosng the prce of Y n market as the numérare then mples w = 1. The volume of trade n ths model s gven by VT = p n x + p n x + t p Y Y, (7) X j Xj j j and the trade overlap (Fnger, 1975) expressed as a share of the trade volume hence, the G-L ndex s X j { p n x, p n x } 2mn X j Xj j j GLI = = 1 p n x + p n x + t p Y Y Xj j j Y Yj j j Y Yj p X j n x j j p X + p n x Xj n j j p x j Xj n + t Y j x p j Yj Y j Y j. (8)

6 unrealstc) parameter domans. In our case, for nstance, choosng ε = 0 would allow such an analyss. To avod ths dlemma, we wll provde some numercal solutons later. B. Changes n Relatve Trade Costs and Intra-Industry Trade In ths secton, we consder the relatonshp between a change n trade costs n the dfferentated goods sector (holdng constant trade costs for the homogeneous good) and the theoretcal mpact on our overall ndex of ntra-ndustry trade, GLI. We consder an ncrease n the (gross) trade-cost factor n ndustry X, t X > 0. A rse n good X s trade costs wll make mports of X by each country more expensve, lowerng mport demand and the value of both countres trade flows n X. Usng the frst equalty n eq. (8), ths tends to lower both the numerator and denomnator n (8). However, n general equlbrum, and wth asymmetrc economc szes and relatve factor endowments, the full mpact of a rse n t X s unclear. 1. Economc Intuton To analyze the mpact, we make a few assumptons. Assume that the two countres are equal n economc sze (real GDP), but country (j) s relatvely abundant n captal (labor), the factor used relatvely ntensvely to produce X (Y). Consequently, country (j) s the net exporter of X (Y); both countres export X, but only j exports Y. 1 Gven that country s the net exporter of X and produces a larger share of X n the world, a rse n t X causes the relatve prce of X to consumers n country j (the net mporter of X) to rse dramatcally, reducng real ncome n country j. Due to the love of varety for X, the bulk of X consumed n j s mported. Ths reducton n j s real ncome s equvalent economcally to a loss of factor endowments, whch should rase factor prces n country j. However, the prce of labor n j (w j ) cannot rse. Frst, the prce of labor n (w ) s the numerare; consequently, gven the model s structure, p Y = 1. Consequently, n country j, the wage rate s unchanged; snce proft maxmzaton ensures the wage rate equals the producer prce of homogeneous good Y n country j, and the latter s lnked (adjusted for Y s trade cost factor) to the prce of good Y n (the numerare, p Y = w ), w j s unchanged. So the prces of captal n both countres (r, r j ) bear the brunt of adjustment. The mpled scarcty of captal n j drves up ts prce (r j ). However, snce s the net exporter of X, the fall n demand for X leads to an excess supply of captal, and the prce of captal n (r ) actually falls. On net, the relatve wage-rental rato n country rses relatve to the 1 We assume a suffcently large relatve factor endowment dfference to yeld ths outcome. 6

7 relatve wage-rental rato n j, causng the relatve prce of X to Y n to fall relatve to that n j. The wdenng of relatve prces n the two countres ncreases ndustry specalzaton, dmnshng the overall share of ntra-ndustry trade (GLI). 2. Edgeworth Box Approach In lght of the 2x2x2 dmensons of our model, we can llustrate relatonshps between trade costs, real GDPs, relatve factor endowments, and G-L ndexes usng a tradtonal Edgeworth box. Fgure 1 provdes an llustraton. > Fgure 1 HERE < In ths fgure, we depct an so-gdp-share lne, whch reflects the same share of world (real) GDP correspondng to a gven set of values for transport costs. Assume that the sold so-gdp-share lne s assocated wth t X = t Y and wth equally szed countres. Wth non-zero trade costs (t X = t Y >1), the sold so-gdp-share lne s knked (n contrast to a world wthn the factor-prce-equalzaton set wth zero trade costs). The reason s that, at the southeast end of the lne, country s relatvely labor abundant and wll be the sole producer and exporter of homogeneous good Y, whle producng and exportng a small amount of dfferentated good X. As country s K-L rato ncreases for a gven real GDP, at some pont near the dagonal (dependng upon the values of t X and t Y ), ncomplete specalzaton n producton of Y results; ths creates the ntal knk n the lne movng northwest. At an even hgher K-L rato for country (just above the dagonal), country wll produce none of good Y and wll specalze n the producton and export of ts dfferentated varetes of X; ths s the second knk n the lne movng northwest. Fgure 1 also llustrates two sold so-grubel-lloyd-index (so-gli) lnes assocated wth two alternatve relatve factor endowments for the two countres. At pont A, for example, countres have dentcal GDPs but dfferent relatve factor endowments. The G- L ndex s less than unty; analogously, for pont B. We now consder the effects of changes n transport costs on these loc. Consder an ncrease n trade costs n good X. Ths ncrease causes country s so-gdp-share lne to tlt as ndcated. If good Y uses labor relatvely ntensvely n producton, wth an ncrease n t X the orgnal (sold) so-gdp-share lne s now assocated wth a lower relatve real ncome n j compared wth. More mportantly, Fgure 1 llustrates the effect on the so-gli lne. The sold lne reflects a constant G-L ndex level at varous K-L ratos for country assumng that t X = t Y > 1. As dscussed above, the fall n the relatve prce of good X to good Y n country 7

8 relatve to country j mples that ndustry specalzaton wll ncrease for the two countres, lowerng the overall share of ntra-ndustry trade. Fgure 1 llustrates that, n the northwest quadrant, the so-gli lne shfts to the rght. That s, the orgnal (sold) so-gli lne s now assocated wth a lower level of GLI. 3. Numercal Smulatons Because of extensve nonlneartes n the model, we wll fnd t useful to create a computable general equlbrum (CGE) verson of our model. Ths wll enable us to generate expected theoretcal relatonshps more closely related to the econometrc model. For nstance, t s well known n ths class of (Helpman-Krugman-type) models that the share of ntra-ndustry trade wll ncrease the larger (and more smlar) n economc sze are two countres. As just shown, n general equlbrum trade cost changes affect both GLI and economc szes. In the emprcal model to follow, the ncluson n regressons of GDPs as well as trade costs mples that the estmated relatonshp between trade costs and GLI s holdng constant varaton n GDPs. We would lke to know theoretcally the effect of trade costs on GLI holdng GDPs constant. However, we can use the CGE verson of our model to generate a theoretcal relatonshp between trade costs and GLI holdng constant relatve economc szes. We now descrbe the methodology for the CGE model. Frst, we reallocate captal so that country holds between 50 percent and percent of the world captal endowment, and we reallocate labor to ensure that ths country holds between 0.05 percent and 50 percent of the world labor endowment. Hence, we focus n the smulatons on the northwest quadrant of the factor box n Fgure 1, where country s captal abundant. Second, we choose an extremely fne grd and compute =10,000 equlbra. Thrd, we choose a partcular value of country s share of world GDP (n our case, 54 percent) and select all factor endowment confguratons out of the 10,000 whch (approxmately) produce ths (endogenous) share of world GDP; hence, our smulated relatonshp between trade costs and GLI wll hold constant relatve GDPs. However, the effects of trade cost changes on GLI wll be senstve to the values of parameters. In the context of our model, the parameters are the Leontef nput requrements (a s), the share of expendtures devoted to the dfferentated good, captal requrements for frm setups, and countres factor endowments. In the remander of the paper, we demonstrate theoretcally (usng the CGE model) and later emprcally (usng regresson analyss) the effect of changes n each sector s transport cost factor on the 8

9 aggregate GLI ndex for the country par and the senstvty of ths effect to relatve factor endowments. It wll be useful to defne the absolute dfference n the logs of relatve factor endowments by DRLFAC K j = ln ln. L K L j We now demonstrate theoretcally (usng the CGE model) the relatonshps between trade costs and ntra-ndustry trade. We wll dsplay a GLI-DRLFAC locus for four dfferent values of transport costs, always holdng the chosen share of world GDP constant, as we do n the emprcal analyss of GLI later, where GDP sze and smlarty enter as determnants. In Fgure 2, we focus n partcular on a range of relatve factor endowment dfferences (DRLFAC) that s emprcally plausble and where countres are mperfectly specalzed (so that GLI > 0). In partcular, the relatonshps shown wll hold for a range of the rato of relatve factor endowments for two countres from unty (dentcal relatve factor endowments) to one country havng fve tmes the K/L rato of the other, whch represents the OECD countres. To produce the fgure, we bascally leave the brd s eye vew of Fgure 1 but rather look at the GLI assocated wth a specfc level of relatve GDP at dfferent confguratons of t X and t Y. Specfcally, we assume a conventonal value for the elastcty of substtuton among manufactures (ε = 6; see Anderson and van Wncoop, 2004) and base our nsghts on numercal solutons of the model. 2 > Fgure 2 HERE < Wth ths background, we consder the effect of the rse n trade costs n sector X. Fgure 2 llustrates four lnes. We are concerned here wth only two: the lne representng the relatonshp between GLI and DRLFAC for trade costs of (t X = 1.1, t Y = 1.1) and the lne for trade costs of (t X = 1.3, t Y = 1.1). As dscussed n secton 2B above, our frst hypothess s that a rse n t X wll lower the overall share of ntra-ndustry trade. The lne for (t X = 1.3, t Y = 1.1) s lower relatve to that for (t X = 1.1, t Y = 1.1), as expected. 2 Concernng the nput coeffcents, we choose a LX = 0.6, a LY = 1, a KX = 0.8, a KY = 0, a Ln = 0, a Kn = 1. The expendture share on dfferentated goods s set at a = 0.8. Further, we assume K = 60 and L = 100 for world endowments. In the ntal equlbra, transport costs are set at t X = t Y = 1.1. To assess the mpact of alternatve transport costs, we choose a value of t X = 1.3 and t Y = 1.3 when ndcated. 9

10 C. Proportonal Changes n Trade Costs and Intra-Industry Trade We now consder the theoretcal effect of a proportonal change n trade costs n both sectors on the GLI ndex for the country par. Fgure 2 llustrates that a rse n trade costs from (t X = 1.1, t Y = 1.1) to (t X = 1.3, t Y = 1.3) reduces the GLI ndex of ntra-ndustry trade. Note that the shft downward of the locus s of less magntude than the downward shft n the case of a rse n only X s trade costs. The reason s the followng. In ths case, w j falls dramatcally because the cost of tradng Y has rsen sharply, lowerng Y s prce on the world market (excludng trade costs). Even though the prce of captal falls n both countres, the relatve factor prces n the two countres do not wden as much as n the prevous case. Consequently, nterndustry specalzaton does not ncrease as much, and ntra-ndustry trade does not decrease as much. D. Changes n Relatve Trade Costs and Relatve Factor Endowments The nonlneartes generated n the model by the ntroducton of trade costs lkely make the effect of trade cost changes on the Grubel-Lloyd ndex of ntra-ndustry trade senstve to parameters, ncludng ntal levels of endowments. A pror, t s dffcult to predct analytcally how these nonlneartes wll affect the mpact of trade costs on GLI at dfferent parameter levels; ths s why we construct CGE models. We now use Fgure 2 to gude us n understandng the varyng senstvty of the fall n GLI to the level of dfferences n relatve factor endowments. Careful examnaton of Fgure 2 reveals that the effect of a rse n relatve trade costs n X on the reducton of the GLI ndex s greater the larger s the absolute dfference of (log) relatve factor endowments for the two countres. Snce we know a rse n the relatve trade cost of the dfferentated good mpacts the X sector dsproportonately and leads to greater nter-ndustry specalzaton, ths effect s exacerbated the wder s the ntal level of nter-ndustry specalzaton due to a large dfference n relatve factor endowments. Ths s confrmed by the fact that at the LHS of Fgure 2 when relatve factor endowments are nearly dentcal GLI falls by about 0.15 (approxmately, 20 percent) wth a rse n t X from 1.1 to 1.3, but GLI falls by about 0.20 (approxmately, 30 percent) when one country s relatve factor endowment s about fve tmes that of the other. E. Changes n Proportonal Trade Costs and Relatve Factor Endowments 10

11 Analogously, the effect of proportonal trade cost changes across sectors on the Grubel-Lloyd ndex wll be senstve to the ntal dfference n relatve factor endowments. We know from secton C above that the GLI ndex falls less wth a proportonal ncrease n trade costs compared wth a rse n only X s relatve trade cost. However, n ths case as well, relatve factor prces change and wden, ncreasng nterndustry specalzaton. The wder the ntal level of relatve factor endowments (and, consequently, nter-ndustry specalzaton), the greater ths ncrease n nter-ndustry specalzaton wll have upon the overall level of ntra-ndustry trade. Consequently, we would expect that a proportonal change n trade costs would have a larger negatve mpact on GLI the larger s DRLFAC the dfference n relatve factor endowments. Fgure 2 confrms ths clearly; the effect on GLI s much larger at hgh values of DRLFAC than at low values. We now evaluate emprcally these four hypotheses. III. Data Our data base conssts of ntra-ndustry trade share fgures based on 3-dgt blateral trade data n Standard Internatonal Trade Classfcaton Revson 2 as avalable from the OECD (Internatonal Trade by Commodty Statstcs, ). For each country par n the sample 3, we weght the 3-dgt based GLI fgures to construct a sngle, aggregate GLI value. To elmnate the nfluence of outlers n the tme dmenson, we average the blateral data across years. Although our theoretcal model addresses trade cost factors, n the sprt of Anderson and van Wncoop (2004), the actual measurement of such costs s extraordnarly dffcult, as ther paper emphaszes. For emprcal purposes, we adopt a narrower defnton of trade costs, n partcular, transport costs. We defne total blateral transport costs accordng to the c..f./f.o.b. rato. Usng our data, we construct gross c..f./f.o.b. factors for all the country pars n our sample for both homogeneous goods and dfferentated goods. We employ a narrow defnton of homogeneous goods trade; we classfy the 1-dgt categores 0, 2, and 3 as homogeneous goods. We exclude beverages and tobacco 3 Australa, Austra, Belgum, Canada, Chna, Czech Republc, Denmark, Fnland, France, Germany, Greece, Hong Kong, Hungary, Iceland, Ireland, Italy, Japan, Korea (Republc of), Mexco, Netherlands, New Zealand, Norway, Poland, Portugal, Slovak Republc, Span, Sweden, Swtzerland, Turkey, Unted Kngdom, USA. 11

12 (category 1 ) from ths defnton, snce earler research not to menton our own colleagues suggest consderable product dfferentaton wthn subcategores coverng wne 11241, Whsky 11241, and beer Applyng ths defnton, we end up wth a share of homogeneous goods trade n total trade of 12 percent on average; see Table 1. > Table 1 HERE < All other data come from the World Bank s World Development Indcators. Specfcally, we use real GDP (base year s 1995), labor force, and gross fxed captal formaton. The latter are used to compute captal stocks accordng to the perpetual nventory method, assumng a deprecaton rate of 13.3% as suggested n Leamer (1984). Table 1 provdes detals on the medan, mean, and standard devaton of all varables n use. We would lke to hghlght that trade costs of homogeneous goods are hgher than those of dfferentated goods by fve percentage ponts on average, and ths dfference s sgnfcant at 1 percent accordng to a pared t-test. IV. Econometrc Analyss In the econometrc analyss, we estmate ntally the followng fve specfcatons of cross-secton regressons: GLI j = g 0 + g 1 GDT j + g 2 SIMI j + g 3 DRLFAC j + g 4 [ln(t Xj )-ln(t Yj )] + g 5 ln(t Yj ) + u j (9) GLI j = g 0 + g 1 GDT j + g 2 SIMI j + g 3 DRLFAC j + g 4 [ln(t Xj )-ln(t Yj )] + g 5 ln(t Yj ) + g 6 DRLFAC j µ[ln(t Xj )-ln(t Yj )] + g 7 DRLFAC j µln(t Yj ) + u j (10) GLI j = g 0 + g 1 GDT j + g 2 SIMI j + g 3 DRLFAC j + g 5 ln(t Yj ) + g 6 DRLFAC j µ[ln(t Xj )- ln(t Yj )] + u j (11) GLI j = g 0 + g 1 GDT j + g 2 SIMI j + g 3 DRLFAC j + g 4 [ln(t Xj )-ln(t Yj )] + g 7 DRLFAC j µln(t Yj ) + u j (12) GLI j = g 0 + g 1 GDT j + g 2 SIMI j + g 3 DRLFAC j + g 6 DRLFAC j µ[ln(t Xj )-ln(t Yj )] + g 7 DRLFAC j µln(t Yj ) + u j (13) 12

13 where GDT j =ln(gdp +GDP j ), SIMI j =(GDP GDP j )/(GDP +GDP j ) 2 s the chosen formulaton of smlarty n country sze (see Helpman, 1987, and Bergstrand, 1990, for two alternatve specfcatons), DRLFAC j s defned n Secton 2, and ln(t Xj ) and ln(t Yj ) are the logs of the c..f./f.o.b. blateral transport costs of dfferentated and homogeneous goods, respectvely. Let u j be a classcal error term. Note that any varaton n ln(t Yj ) s representng varaton n total trade costs, as the ncluson of [ln(t Xj )-ln(t Yj )] s holdng constant dfferences n trade costs between sectors. We take nto account that GLI s a lmted dependent varable. Accordngly, we use the logstcally transformed ndex, defned as ln(gli/[1-gli]), n the regressons to ensure that the model predcton of GLI les n the [0,1] nterval (see also Bergstrand, 1983, 1990; Hummels and Levnsohn, 1995). > Table 2 HERE < Table 2 summarzes the results. Frst, we consder the coeffcent estmates for the varables representng economc sze, smlarty, and relatve factor endowment dfferences. The sum of the two countres GDPs has the expected postve effect on GLI and the coeffcent estmates are statstcally sgnfcant. GDP smlarty also has the expected postve relatonshp wth GLI, although coeffcent estmates generally lack statstcal sgnfcance at conventonal levels. Dfferences n relatve factor endowments have the expected negatve relatonshp wth the share of ntra-ndustry trade; coeffcent estmates are statstcally sgnfcant. Thus, n all fve specfcatons, economc sze, economc smlarty, and relatve factor endowment dfferences have the expected correlatons wth GLI. In examnng the emprcal relatonshps between GLI, the transport cost varables, and the nteracton terms, we consder each of the fve specfcatons n turn. Model 1 consders frst the effects of absolute and relatve trade cost changes on GLI n the absence of nteractons wth relatve factor endowment dfferences. In the presence of ln(t Yj ), varaton n [ln(t Xj )-ln(t Yj )] represents changes n X s transport cost only. As expected based upon our theory, ncreases n the relatve transport cost factor n X have a negatve relatonshp wth GLI. Also, as expected, ncreases n absolute transport costs ln(t Yj ), holdng varaton n [ln(t Xj )-ln(t Yj )] constant, decrease GLI. Thus, the coeffcent estmates n Model 1 are remarkably consstent wth the theoretcal model. 13

14 In Model 2, we nclude as well two nteracton terms between the transport cost varables and DRLFAC. As shown n Table 2, the coeffcent estmate for DRLFAC j µln(t Yj ) has the expected negatve sgn, but s not statstcally sgnfcant. The coeffcent estmate for DRLFAC j µ[ln(t Xj )-ln(t Yj )] does not have the expected sgn, but also s not statstcally sgnfcant. The explanaton for these nteracton term coeffcent results s collnearty among subsets of the regressors. In partcular, ln(t Yj ) and DRLFAC j µln(t Yj ) are hghly collnear (correlaton coeffcent of 0.71) and [ln(t Xj )- ln(t Yj )] and DRLFAC j µ[ln(t Xj )-ln(t Yj )] are hghly collnear (correlaton coeffcent of 0.82). To account for ths collnearty, we also ran Models 3, 4 and 5, as shown above. In Model 3, we nclude the three core varables GDT, SIMI, and DRLFAC wth only ln(t Yj ) and DRLFAC j µ[ln(t Xj )-ln(t Yj )], as the correlaton coeffcent between these two varables s only As shown n Table 2, both varables absolute transport costs and relatve transport costs have the expected negatve coeffcent estmates; these results are spared multcollnearty. Moreover, the nteracton of [ln(t Xj )-ln(t Yj )] wth DRLFAC wll stll allow estmatng the margnal mpact of relatve trade costs on the transformed GLI at varous levels of relatve factor endowment dfferences. These estmates wll be summarzed later. In Model 4, we nclude the three core varables wth only [ln(t Xj )-ln(t Yj )] and DRLFAC j µln(t Yj ). As shown n Table 2, both varables have the expected negatve relatonshp wth GLI. The nteracton of ln(t Yj ) wth DRLFAC wll allow estmatng the margnal mpact of absolute trade costs on the transformed GLI at varous levels of relatve factor endowment dfferences, holdng constant relatve trade costs. For completeness, Model 5 ncludes the three core varables wth only DRLFAC j µ ln(t Yj ) and DRLFAC j µ[ln(t Xj )-ln(t Yj )]. Once agan, both nteracton terms have coeffcent estmates wth the expected negatve sgns. We wll also be able to retreve estmates of the margnal mpacts at varous levels of DRLFAC. > Tables 3a and 3b HERE < Tables 3a and 3b provde estmates of the margnal mpacts of the two transport cost varables on the transformed GLI at varous levels of relatve factor endowment dfferences. Table 3a provdes estmates for Models 3 and 4. As our theory llustrated n Fgure 2 suggests, the negatve margnal effects become larger (n absolute terms) wth 14

15 larger dfferences n relatve factor endowments. The results n Table 3b for Model 5 confrm these results. The only dfference of the estmated margnal effects from the theory s that the margnal effects for the trade cost dfference varable are larger (n absolute terms) than those for the total trade cost varable. Careful examnaton of Fgure 2 reveals that the lne for (t X =1.3, t Y =1.1) les systematcally below that for (t X =1.3, t Y =1.3). Whle the estmated margnal effects are not statstcally sgnfcant, we wll fnd later n the senstvty analyss that a slghtly dfferent specfcaton reverses ths outcome. V. Senstvty Analyss We nvestgate the robustness of our fndngs n Models 1-5 n several respects; n several areas, we omt Model 2 smply for brevty and ease of presentaton. 5.1 Measurng Economc Sze and Smlarty The basc specfcaton descrbed above n equatons (9)-(13) has one frequently used alternatve. Both Helpman (1987) and Hummels and Levnsohn (1995) used max(lngdp, lngdp j ) and mn(lngdp, lngdp j ) to represent economc sze and smlarty, rather than GDT and SIMI used earler. Ths alternatve specfcaton s presented n Table 4. > Table 4 HERE < Table 4 shows that the basc results are largely nsenstve to the change n specfcaton. However, we notce one mprovement. As Model 1 reveals, for example, the relatve coeffcent szes (n absolute terms) for the transport cost dfference varable and the absolute transport cost varable change. The change n relatve coeffcent szes s revealed also n estmates of the margnal effects for the transport cost varables usng the alternatve specfcaton shown n Tables 5a and 5b. > Tables 5a and 5b HERE < Tables 5a and 5b ndcate that the estmated margnal effects for changes n relatve transport costs are now larger n absolute terms than those for changes n absolute transport costs. Ths s consstent wth the theoretcal mplcatons of the model, as shown n Fgure 2. For the remanng results n ths paper, we analyze Models 1, 3, 4, and 5 only; we exclude Model 2 for brevty and convenence. 15

16 5.2 Treatment for Influental Observatons We conducted two senstvty analyses to detect the possblty of our results beng drven ether by outlers or leverage ponts. For outlers, we follow Belsley, Kuh and Welsch (1980) and run OLS on all four models excludng observatons wth an absolute error term larger than two standard errors of the regresson. For leverage observatons, we run medan regressons, cf., Greene (2000), to determne how senstve the results are to nfluental observatons. As shown n Table 6, the coeffcent estmates for the relevant transport cost (or nteracton) varables are largely the same as n Table 2. (All the results shown n Table 6 use the orgnal specfcaton wth GDT and SIMI.) We also consdered the possble nfluence of the results beng drven by the smallest country par s observaton or the largest country par s observaton. As Table 6 ndcates, the coeffcent estmates for the relevant varables were nsenstve to the excluson of ether country par. 5.3 Influental Observatons for Partcular Parameters We follow Efron and Tbshran (1993) and conduct a jackknfe analyss to assess the maxmum mpact of cross-sectonal observatons on each transport cost (or nteracton) varable s coeffcent estmate. Specfcally, we nvestgate the maxmum postve and negatve devaton from our orgnal coeffcent estmates n Table 2 as a result of excludng a sngle country-par. In general, the results are robust. In every model, the maxmum and mnmum coeffcent estmates are economcally very close to the respectve coeffcent estmates reported n Table 2. For example, n Model 1, the coeffcent estmate for [ln(t Xj )-ln(t Yj )] s n Table 2 whereas the mnmum (maxmum) coeffcent estmate for ths varable n Table 6 s (-0.175). The coeffcent estmate for ln(t Yj ) s n Table 2 whereas the mnmum (maxmum) coeffcent estmate for ths varable n Table 6 s (-0.305). Thus, the results are robust to a jackknfe analyss. 5.4 Trade-Imbalance-Adjusted GLI As ponted out n earler and more recent research, the use of blateral tradembalance-adjusted GLI ndces often s preferable over unadjusted ones; see Bergstrand, 1983; Greenaway and Mlner, 1986; Egger, Egger, and Greenaway, Accordngly, we also estmated our fve specfcatons employng trade-mbalance-adjusted GLI 16

17 measures. We used blateral aggregate OECD trade fgures to compute adjusted GLI ndces. There s no one wdely-adopted method for adjustng G-L ndexes. For convenence and n the nterest of a senstvty analyss, we adjusted the blateral trade flows to reflect blateral aggregate trade balance. As shown n the last lne of Table 6, the results reported n Table 2 are generally robust to the alternatve use of adjusted GLI ndces. VI. Conclusons Anderson and van Wncoop (2004) have recently challenged nternatonal trade economsts to lend much more consderaton to the mportance of trade costs n nfluencng the pattern of nternatonal trade as well as nternatonal prce dspartes. Ther work suggests that the average mpled markup attrbutable to the costs of nternatonal transacton may be approxmately as hgh as 170 percent! Despte ths, nternatonal trade economsts have devoted lttle attenton to ths mportant noton. Researchers n the determnants of ntra-ndustry trade have shared n underemphaszng the mportance and role of trade costs n nfluencng Grubel-Lloyd measures of such trade. Ths paper departs from earler models of ntra-ndustry trade such as the work of Helpman, Krugman, Hummels and Levnsohn, and Evenett and Keller by focusng theoretcally and emprcally on the nonlnear relatonshp between trade costs and the determnants of ntra-ndustry trade. Because of nonlnear relatonshps between economc sze, relatve factor endowments, and trade costs, we developed a smple computable general equlbrum model to llustrate under plausble parameter values the nfluence of trade costs on Grubel-Lloyd measures of ntra-ndustry trade (GLI). Our theoretcal results suggest that the level of trade costs should negatvely mpact the share of ntra-ndustry trade, that dfferences n trade costs between dfferentated goods and homogeneous goods should affect GLI, and that the margnal effects of ether of these varables on GLI are hghly senstve to the level of relatve factor endowment dfferences. In a large cross-secton of blateral ntra-ndustry trade shares based on OECD data, we nvestgate these hypotheses emprcally. The fndngs are strongly n support of our vew. Ths llustrates how as Anderson and van Wncoop (2004) suggest a more realstc treatment of transport costs n our standard models of trade could help to put forward new and nterestng hypotheses and could become a cornerstone for subsequent emprcal research n nternatonal economcs. 17

18 References Anderson, J.E., and E. van Wncoop (2004). Trade Costs, unpublshed manuscrpt, Boston College. Belsley, D.A., A.E. Kuh, and R.E. Welsch (1980). Regresson Dagnostcs: Identfyng Influental Data and Sources of Collnearty, John Wley and Sons, New York. Bergstrand, J.H. (1983). Measurement and Determnants of Intra-Industry Internatonal Trade. In P.K.M. Tharakan (ed.), Intra-Industry Trade: Emprcal and Methodologcal Aspects, Elsever Scence & Technology Books. Bergstrand, J.H. (1990). The Heckscher-Ohln-Samuelson Model, the Lnder Hypothess, and the Determnants of Blateral Intra-ndustry Trade. Economc Journal 100(403): Davs, D. (1998). The Home Market, Trade, and Industral Structure. Amercan Economc Revew 88(5): Debaere, P. (2004), Monopolstc Competton and Trade, Revsted: Testng the Model wthout Testng for Gravty. Journal of Internatonal Economcs, forthcomng. Dxt, A., and J. Stgltz, (1977). Monopolstc Competton and Optmum Product Dversty. Amercan Economc Revew 67(3): Egger, H., P. Egger, and D. Greenaway (2004). Intra-Industry Trade wth Multnatonal Frms: Theory, Measurement and Determnants, GEP Dscusson Paper No 2004/10, Unversty of Nottngham. Efron, B., and R.J. Tbshran (1993). An Introducton to the Boostrap, New York: Chapman & Hall. Evenett, S., and W. Keller (2002), On Theores Explanng the Success of the Gravty Equaton. Journal of Poltcal Economy 110(2): Fnger, J.M. (1975). Trade Overlap and Intra-ndustry Trade. Economc Inqury 13(4): Greenaway, D., and C. Mlner (1986), The Economcs of Intra-Industry Trade. Oxford: Basl Blackwell. Greene, W.H. (2000). Econometrc Analyss, Fourth Edton. New Jersey: Prentce-Hall. Grubel, H.G., and P.J. Lloyd (1975). Intra-Industry Trade, New York: John Wley. Helpman, E. (1981). Internatonal Trade n the Presence of Product Dfferentaton, Economes of Scale and Monopolstc Competton: A Chamberln-Heckscher-Ohln Approach. Journal of Internatonal Economcs 11(3):

19 Helpman, E. (1987). Imperfect Competton and Internatonal Trade: Evdence from Fourteen Industral Countres. Journal of the Japanese and the Internatonal Economes 1(1): Helpman, E., and P.R. Krugman (1985). Market Structure and Foregn Trade, Cambrdge, MA: MIT Press. Hummels, D., and J. Levnsohn (1995). Monopolstc Competton and Internatonal Trade: Reconsderng the Evdence. Quarterly Journal of Economcs 110(3): Krugman, P.R. (1979). A Model of Innovaton, Technology Transfer, and the World Dstrbuton of Income. Journal of Poltcal Economy 87(2): Krugman, P.R. (1980). Scale Economes, Product Dfferentaton, and the Pattern of Trade. Amercan Economc Revew 70(5): Krugman, P.R. (1981). Intra-Industry Specalzaton and the Gans from Trade. Journal of Poltcal Economy 89(5): Lancaster, K. (1980). Intra-Industry Trade under Perfect Monopolstc Competton. Journal of Internatonal Economcs 10(2): Leamer, E.E. (1984). Sources of Internatonal Comparatve Advantage. Cambrdge, MA: MIT Press. Samuelson, P.A. (1952). The Transfer Problem and Transport Costs: The Terms of Trade When Impedments are Absent. Economc Journal 62(246):

20 Fgure 1: Iso-GLI and Iso-GDP-Share Lnes K =0.8K; L =0.2L K =0.8K; L =0.8L (t X ) A (t X ) Iso-GLI (t X ) B Iso-GDP-share K =0.2K; L =0.2L K =0.2K; L =0.8L

21 Fgure 2: Factor endowment dfferences, transport costs and the Grubel-Lloyd ndex t X =1.1; t Y =1.3 GLI t X =t Y =1.3 t X =t Y = t X =1.3; t Y = DRLFAC

22 Table 1: Descrptve statstcs ( averages) Varable Medan Mean Std. dev. Grubel-Lloyd Index (GLI) Log blateral maxmum real GDP (MAXG) Log blateral mnmum real GDP (MING) Log blateral sum of real GDP (GDT) Blateral smlarty ndex (SIMI; real GDP-based) Absolute log dfference n captal-labor ratos (DRLFAC) blateral dfferentated goods trade costs (t X ) blateral homogeneous goods trade costs (t Y ) Blateral share of homogeneous goods

23 Table 2: Regresson results Model 1 Model 2 Model 3 Model 4 Model 5 Log blateral sum of GDP: GDT *** *** *** *** *** Log smlarty n GDP: SIMI * * Absolute log dfference n blateral labor ratos: DRLFAC *** *** *** *** *** Log dfference n dfferentated and homogeneous goods transport costs: ln(t X )-ln(t Y ) *** ** *** Log 1+blateral homogeneous transport costs: ln(t Y ) *** * *** Interacton term: DRLFACµ[ln(t X )-ln(t Y )] * * Interacton term: DRLFACµln(t Y ) ** *** Constant *** *** *** *** *** Observatons Between R The dependent varable s the logstc transformaton of GLI. Fgures below coeffcents are standard errors. Two-taled t-tests: * sgnfcant at 10 percent, ** sgnfcant at 5 percent, *** sgnfcant at 1 percent.

24 Table 3a: Margnal effect of trade costs Trade cost dfference Total trade costs Model 3 Model 4 Lowest decle of DRLFAC (0.124) * ** Mean of DRLFAC (0.946) * ** Hghest decle of DRLFAC (2.157) * ** The dependent varable s the logstc transformaton of GLI. Margnal effects refer to the transformed GLI. Fgures below margnal effects are standard errors. Two-taled t-tests: * sgnfcant at 10 percent, ** sgnfcant at 5 percent, *** sgnfcant at 1 percent. Table 3b: Margnal effect of trade costs Trade cost dfference Total trade costs Model 5 Lowest decle of DRLFAC (0.124) * *** Mean of DRLFAC (0.946) * *** Hghest decle of DRLFAC (2.157) * *** The dependent varable s the logstc transformaton of GLI. Margnal effects refer to the transformed GLI. Fgures below margnal effects are standard errors. Two-taled t-tests: * sgnfcant at 10 percent, ** sgnfcant at 5 percent, *** sgnfcant at 1 percent.

25 Table 4: Alternatve specfcaton for economc sze and smlarty Model 1 Model 2 Model 3 Model 4 Model 5 Log maxmum of exporter and mporter GDP *** *** *** *** *** Log mnmum of exporter and mporter GDP Absolute log dfference n blateral labor ratos: DRLFAC *** *** *** *** *** Log dfference n dfferentated and homogeneous goods transport costs: ln(t X )-ln(t Y ) *** *** *** Log 1+blateral homogeneous transport costs: ln(t Y ) *** *** Interacton term: DRLFACµ[ln(t X )-ln(t Y )] *** *** Interacton term: DRLFACµln(t Y ) ** ** Constant *** *** *** *** *** Observatons Between R Fgures below coeffcents are standard errors. Two-taled t-tests: * sgnfcant at 10 percent, ** sgnfcant at 5 percent, *** sgnfcant at 1 percent.

26 Table 5a: Margnal effect of trade costs Trade cost dfference Total trade costs Model 3 n Table 4 Model 4 n Table 4 Lowest decle of DRLFAC (0.124) *** ** Mean of DRLFAC (0.946) *** ** Hghest decle of DRLFAC (2.157) *** ** The dependent varable s the logstc transformaton of GLI. Margnal effects refer to the transformed GLI. Fgures below margnal effects are standard errors. Two-taled t-tests: * sgnfcant at 10 percent, ** sgnfcant at 5 percent, *** sgnfcant at 1 percent. Table 5b: Margnal effect of trade costs Trade cost dfference Total trade costs Model 5 n Table 4 Lowest decle of DRLFAC (0.124) *** ** Mean of DRLFAC (0.946) *** ** Hghest decle of DRLFAC (2.157) *** ** The dependent varable s the logstc transformaton of GLI. Margnal effects refer to the transformed GLI. Fgures below margnal effects are standard errors. Two-taled t-tests: * sgnfcant at 10 percent, ** sgnfcant at 5 percent, *** sgnfcant at 1 percent.

27 Table 6: Senstvty analyss Model 1 - based Model 3 - based Model 4 - based Model 5 - based RLFAC RLFAC RLFAC RLFAC ln(t X )-ln(t Y ) ln(t Y ) ln(t Y ) (ln(t X )-ln(t Y )) ln(t X )-ln(t Y ) ln(t Y ) ln(t X )-ln(t Y ) ln(t Y ) Influental observatons: Excludng outlers a) * *** *** *** *** *** *** *** Medan regresson *** ** *** * *** ** ** * Excludng smallest country-par *** *** *** * *** ** * *** Excludng largest country-par *** *** *** * *** *** * *** Jackknfe analyss: Jackknfe analyss (mnmum ln(t X )-ln(t Y ) coeff.) *** *** *** ** Jackknfe analyss (maxmum ln(t X )-ln(t Y ) coeff.) * *** *** *** Jackknfe analyss (mnmum ln(t Y ) coeff.) ** *** *** * Jackknfe analyss (maxmum ln(t Y ) coeff.) *** *** *** * Jackknfe analyss (mnmum RLFACµ[ln(t X )-ln(t Y )] coeff.) *** ** ** * Jackknfe analyss (maxmum RLFACµ[ln(t X )-ln(t Y )] coeff.) *** *** Jackknfe analyss (mnmum RLFACµln(t Y ) coeff.) *** ** * *** Jackknfe analyss (maxmum RLFACµln(t Y ) coeff.) *** * ** * Trade mbalance adjusted GLI: Trade mbalance adjusted GLI as LHS varable ** *** *** ** *** *** *** Fgures below coeffcents are standard errors. Two-taled t-tests: * sgnfcant at 10 percent, ** sgnfcant at 5 percent, *** sgnfcant at 1 percent. - a) We follow Belsley et al. (1980) n defnng outlers as observatons wth absolute resduals larger than two standard errors of the regresson.