epub WU Institutional Repository

Transcription

1 epub WU Insttutonal Repostory Harald Badnger and Frtz Breuss Trade and productvty. An ndustry perspectve. Paper Orgnal Ctaton: Badnger, Harald and Breuss, Frtz (2005) Trade and productvty. An ndustry perspectve. EI Worng Papers / Europansttut, 66. Europansttut, WU Venna Unversty of Economcs and Busness, Venna. Ths verson s avalable at: Avalable n epub WU : March 2005 epub WU, the nsttutonal repostory of the WU Venna Unversty of Economcs and Busness, s provded by the Unversty Lbrary and the IT-Servces. The am s to enable open access to the scholarly output of the WU.

2 EUROPAINSTITUT WIRTSCHAFTSUNIVERSITÄT WIEN EUROPAINSTITUT UNIVERSITY OF ECONOMICS AND BUSINESS ADMINISTRATION VIENNA Worng Papers EI Worng Paper No. 66 HARALD BADINGER, FRITZ BREUSS Trade and Productvty: An Industry Perspectve March 2005 Download: Althanstraße 39-45, A Wen / Venna Österrech / Austra Tel.: ++43 / 1 / / 4135, 4134, 4133 Fax.: ++43 / 1 / / 758, 756 e-mal: europafragen2@wu-wen.ac.at

3 Impressum: De EI Worng Papers snd Dsussonspapere von MtarbeterInnen und Gästen des Europansttutes an der Wrtschaftsunverstät Wen, de dazu denen sollen, neue Forschungsergebnsse m Fachres zur Dsusson zu stellen. De Worng Papers geben ncht notwendgerwese de offzelle Menung des Insttuts weder. Se snd gegen enen Unostenbetrag von 7,20 am Insttut erhältlch. Kommentare snd an de jewelgen AutorInnen zu rchten. Medennhaber, Egentümer Herausgeber und Verleger: Europansttut der Wrtschaftsunverstät Wen, Althanstraße 39 45, A 1090 Wen; Für den Inhalt verantwortlch: Unv.-Prof. Dr. Stefan Grller, Althanstraße 39 45, A 1090 Wen. Nachdruc nur auszugswese und mt genauer Quellenangabe gestattet.

4 Trade and Productvty: An Industry Perspectve 3 Trade and Productvty: An Industry Perspectve * Harald Badnger Europansttut, Wrtschaftsunverstät Wen Emal: harald.badnger@wu.wen.ac.at Frtz Breuss Europansttut, Wrtschaftsunverstät Wen, and Austran Insttute of Economc Research (WIFO), Venna Emal: frtz.breuss@wu-wen.ac.at March 2005 Abstract: We use a sample of 14 OECD countres and 15 manufacturng ndustres to test for the effect of trade on productvty. Endogenety concerns are accounted for usng the geographcal component of trade as nstrument as suggested by Franel and Romer (1999). Our results are n lne wth prevous studes: Trade ncreases productvty. What s puzzlng, however, s the sze of the effect: An ncrease n the export rato by one percentage pont ncreases productvty n manufacturng by 0.6 percent on average. Ths s less than half of the effect obtaned n prevous studes. We dscuss lely explanatons for ths dscrepancy. JEL Classfcaton: F14, F43, L60 Keywords: Trade, Productvty, Manufacturng * We wsh to than Davd Romer for very helpful comments on an earler draft.

5 EI Worng Paper No I. Introducton The relaton between trade and productvty has always been at the heart of nternatonal economcs. Theoretcal arguments can be traced bac at least to Adam Smth s famous dctum that the dvson of labour s lmted through the sze of the maret. The lterature has emphaszed several addtonal channels va whch trade may nfluence productvty: the explotaton of ncreasng returns from larger marets (e.g. Balassa, 1961), the transmsson of nternatonal technology spllovers (e.g. Coe and Helpman, 1995), the exchange of deas through travel and communcaton, as well as the pro-compettve effect of nternatonal trade (e.g. Bhagwat, 1965). Numerous emprcal studes have confrmed a postve correlaton between trade and ncome (see Lewer and Van den Berg (2003) for a comprehensve survey), but endogenety concerns and the absence of convncng nstruments have cast doubts on whether these observed correlaton actually reflects a causal relatonshp. Franel and Romer (1999) argue that geography affects ncome only va trade and suggest usng the geographcal component of aggregate trade as nstrument. Usng data from a large cross-secton of countres referrng to the year 1985, they fnd that the least squares estmates are not nvaldated by the results of the nstrumental varable estmaton. In contrast, the estmated effect of trade on productvty even ncreases f endogenety s accounted for, though the estmates turn out only moderately sgnfcant. These results were bascally confrmed by Irwn and Tervö (2002) for alternatve reference years from the twenteth century, rangng from 1913 to Nevertheless, the academc debate s yet unsettled: Rodr et al. (2002) fnd that trade s nsgnfcant once nsttutonal qualty s controlled for and argue that the postve effect of trade n prevous studes may smply capture omtted nsttutonal characterstcs. Also, Irwn and Tervö (2002) note that the effect of trade on productvty s not robust to the ncluson of dstance from the equator, a proxy for Western nfluence accordng to Hall and Jones (2002). Another

6 Trade and Productvty: An Industry Perspectve 5 crtque of the geographcal nstrument suggested by Franel and Romer (1999) was put forward by Durlauf (2001). Part of these concerns on the relatonshp between trade and ncome have been addressed n a recent paper by Alcala and Cccone (2004). They advocate the use of real rather than the nomnal openness, whch may be a dstorted measure of openness as a result of the trade-related Balassa-Samuelson effect, and fnd trade to be a sgnfcant determnant of productvty even when nsttutons are controlled for. Ths paper taes an alternatve approach to addressng the concerns on the relatonshp between trade and productvty. We follow the approach by the Franel and Romer (1999), but rather than ncludng proxes for nsttutonal characterstcs or lattude (as control for Western nfluence) we focus on a sample of Western, nsttutonally largely homogenous countres (14 OECD countres) and use ndustry data (15 manufacturng ndustres). Provdng an ndustry perspectve on the relatonshp between trade and productvty s not only of nterest n tself; t also allows to control for country-specfc effects that are nvarant over ndustres, provdng an addtonal robustness chec aganst the possble omsson of (unobservable) nsttutonal characterstcs. The results suggest that trade has a statstcally sgnfcant and robust effect on productvty n manufacturng. Our nstrument, the geographcal component of trade, s of hgh qualty yeldng comparably precse and hghly sgnfcant estmates. The sze of the effect, however, appears to be farly low compared wth prevous estmates. We dscuss lely explanatons for ths dfference, rasng several ponts that deserve further attenton. The remander of the paper s organzed as follows. Secton II sets up the emprcal model. Secton III constructs the nstrument for trade. Secton IV presents the estmates of the relaton between trade and productvty. Secton V summarzes the results and concludes.

7 EI Worng Paper No II. Trade and Productvty: The Emprcal Model Trade may affect productvty n varous ways: through specalzaton and the explotaton of economes of scale from larger marets, by ncreasng the ncentve to nnovate as a result of enhanced competton, and through nowledge spllovers by exposure to new goods and technologes. The logc of these arguments s not confned to nternatonal trade but apples equally to wthn-country trade. In lne wth Franel and Romer (1999), equaton (1) sets up the hypothess that economc nteractons of country wth other countres (nternatonal trade) and economc nteractons wthn the country (wthn-country trade) have a (postve) effect on productvty: ln y = α + βt + γw + ε. (1) Here y s GDP per worer, T s nternatonal trade, and W s wthn-country trade. Equaton (1) cannot be estmated by least squares snce trade (T ) s lely to be endogenous. Internatonal and wthn-country trade are n turn modelled as functons of proxmty to other countres (P ) and country sze (S ) respectvely. T = ψ + φ + δ, (2) P W = η + λ + ν. (3) S The problem that wthn-country trade s not observable s addressed by substtutng (3) nto (1), yeldng ln y = ( α + γη) + βt + γλs + ( γν + ε ), or equvalently (4) ln y = µ + ϕt + τs + u (5) Under the dentfyng assumptons that the geographcal varables P and S are uncorrelated wth the composte error term γν + ε ), equaton (4) can be estmated consstently usng S ( and P as nstruments for T. We use two measures of sze, populaton (N) and area (A), such that equaton (4) becomes

8 Trade and Productvty: An Industry Perspectve 7 ln y µ A + u (6) = + ϕt + τ1n + τ 2 Ths basc setup follows Franel and Romer (1999). Our fnal model dffers n several respects. Frst, Franel and Romer (1999) use a large sample of 150 (and 98) countres, ncludng many developng countres. We focus on a sample 14 OECD countres (AUT, BEL, DEU, DNK, ESP, FIN, FRA, GBR, GRC, ITA, NLD, NOR, SWE, USA). Besdes the avalablty of more relable data, focussng on countres wth smlar nsttutons maes t unlely that our trade varable captures unobserved nsttutonal heterogenety, a pont promnently rased by Rodr et al. (2002). Smlarly, Durlauf s (2003) crtque of the geography-based nstrument by Franel and Romer (1999) s less relevant for our sample. 1 Second, we use ndustry data comprsng 15 manufacturng ndustres (See Table A1 n the Appendx for an overvew). That s, equaton (6) becomes ln y µ A + u, (7) = + ϕt + τ1n + τ 2 where denotes ndustry. Productvty (y ) s now measured n terms of value added per worer. Whle the nstrument P becomes ndustry-specfc as well (snce t s derved from ndustry-specfc gravty estmates as wll be outlned below) the sze measures N and A are stll country-specfc rather than ndustry-specfc. Snce ndustres also delver ntermedates to other ndustres, t not the sze of the own sector but that of the whole economy whch s the relevant determnant of wthn-country trade. A further advantage of usng ndustry data s that we have several observatons per county, whch enables us to nclude country dummes as an addtonal robustness chec. 1 The pont rased by Durlauf s that the error term n (6) or here n (7) captures omtted varables such as taxes and democracy. Area, so the argument, s related to mltary expendtures, whch are n turn correlated wth tax levels and democracy. Whle ths may n fact be a serous concern n a large sample of countres ncludng many developng and non-democratc countres, ths argument vanshes for our sample of hgh-ncome countres wth smlar democratc nsttutons.

9 EI Worng Paper No As most common measure for openness at the ndustry level we use the export share of producton. 2 Snce no ndustry-specfc deflators are avalable we have to use nomnal openness rather than real openness as advocated by Alcala and Cccone (2004). For our sample, whch ncludes manly countres wth a smlar level of development, ths s no major drawbac, snce the trade related Balassa-Samuelson effect s less relevant here. Equaton (7), whch relates ndustry-specfc productvty to ndustry-specfc openness, stll reflects the man channels va whch the effects of trade on productvty materalze: ntra-ndustry specalzaton, ntra-ndustry Economes of Scale, technology-spllovers through exchange of smlar goods and the effects of ncreased competton between frms producng smlar goods. Snce our productvty measure s value added per worer, ncreases n productvty n ndustry do not necessarly ncrease productvty n ndustry j, even f uses the output of j as ntermedate good. Stll there are channels that may operate across ndustres: Specalzaton accordng to comparatve advantage would be reflected n an ncrease n nter-ndustry rather than ntra-ndustry trade; and (external) Economes of Scales may also be of nter-ndustry type (Balassa, 1961, p. 131). 3 By rulng out these nter-ndustry effects, the hypothess expressed by equaton (7) s somewhat narrower than the correspondng model at the aggregate level (6). 2 Imports are less preferable snce a large mport share may also reflect ntensve use of mported ntermedates, not necessarly a hgh exposure to nternatonal trade n fnal goods. We do not use alternatve measures such as mport penetraton whch have a dfferent nterpretaton and would mae our results less comparable wth prevous studes that use openness. 3 Part of these nter-ndustry effects s transmtted through nteractons wthn the country ( wthn-country trade) and should thus be captured by the coeffcent of the wthn-country varable. Ths s another ratonale for ncludng aggregate rather then ndustry sze measures n equaton (7).

10 Trade and Productvty: An Industry Perspectve 9 III. Constructon of the Instrument a) The geographcal gravty model For the nstrument P to be vald s must not be correlated wth the error term n (7); to be useful t must also be strongly correlated wth trade. Franel and Romer (1999) argue that geography s an mportant determnant not only of blateral but also of overall trade. At the same tme t s dffcult to thn of any channel other than nternatonal trade va whch geographcal characterstcs may affect ncome. Ths s the ratonale for constructng the nstrument P as geographcal component of aggregate trade ( Tˆ ). Ths geographcal component of trade s calculated from the predcted values of the followng gravty model, whose regressor matrx (X) ncludes geographcal varables only: lnt j = α X ' = α 0 j + ϑ = 1 j + α d j + α 2 N + α 3 A + α 4 N j + α 5 A j + α ( LL 6 + LL ) + α CB j 7 j + ϑ, (8) j where Tj s the rato of exports from county to country j to country s producton n ndustry, d j s dstance, N s populaton, A s area, LL s a dummy for landloced countres, and CB s a common-border dummy. 4 Our sample comprses 14 countres () and 15 manufacturng ndustres (); we have blateral trade data by ndustry wth 44 partner countres (j, j ) coverng some 90 percent of trade on average. Trade data are averages of the perod 1995 to A detaled descrpton of the data s gven n the Appendx. b) Estmaton results Equaton (8) s estmated separately for each of the 15 ndustres. Table 1 summarzes the results. 4 Franel and Romer (1999) also nclude nteractons between the common border dummy and all other varables; n our estmaton these nteractons all turned out nsgnfcant; moreover, they nduced collnearty problems such that we omt these nteractons from the begnnng.

11 EI Worng Paper No < Table 1 here > As expected dstance has a large negatve effect on trade (defned as rato of exports to producton). The elastcty of trade wth respect to dstance ranges from to ; for ndustres wth a larger weght/value rato (e.g. 23: Coe, refned petroleum products and nuclear fuel) the effect s more pronounced than for ndustres producng more sophstcated goods (e.g : Electrcal and optcal equpment). Trade s strongly ncreasng n country j s populaton wth an average elastcty of and decreasng n both country s and country j s area. The mpact of country s populaton s nsgnfcant or negatve for most ndustres. In lne wth prevous studes we fnd that, all else beng equal, landloced countres trade consderably less (some 60 percent on average) and that a countres sharng a border trade more (some 140 percent). Fnally, and most mportantly, our regressons confrm that geographcal varables are an mportant determnant of nternatonal trade. The values of the R 2 of our regressons range from to c) Implcatons for aggregate trade and the qualty of the nstrument To obtan our nstrument Tˆ, that s the geographcal component of aggregate (rather than blateral) trade by ndustry, the predcted values for the blateral export ratos are aggregated as follows: Tˆ = J j= 1 a' Xj θ e, (9) where a s the estmate of α from (8). To obtan consstent predcted values for the levels of Tj from the estmates n log form, a correcton factor θ s requred. Under normalty θ s j equal to E[ e ϑ ] = 2 ( ˆ /2) e σ, where 2 ˆ σ s a consstent estmator of the varance of ϑ j. To avod mang dstrbutonal assumptons we follow the approach suggested by Wooldrdge (2003, p.

12 Trade and Productvty: An Industry Perspectve ff.) and estmate θ from a regresson of T j on j e a' X through the orgn. Usng (9) we calculate ndustry-specfc trade shares for each of the 14 countres; the summaton n (9) runs not only over the (44) countres for whch we have blateral trade data, but over all countres for whch the varables n X are avalable (addtonal 179 countres). 5 Our nstrument Tˆ j must fulfl two propertes. Frst, t must be uncorrelated wth the error term n (7), an assumpton made here for theoretcal reasons. Second, the nstrument must also be relevant. The ssue of wea nstruments has receved consderable attenton n the recent lterature (Stager and Stoc, 1997; Stoc, Wrght, and Yogo, 2002; Andrews, Morera, and Stoc, 2004). Two stages least squares wth wea nstruments may yeld strongly based estmates and tests wth large sze dstortons. Hence t wll be mportant to chec the qualty of the nstrument. < Fgure 1 here > The scatter plot of the constructed aganst the actual trade share n Fgure 1 suggests a strong assocaton between our nstrument and the endogenous varable. The qualty of the nstruments can also be judged more formally. Stoc and Yogo (2004) wor out a defnton of wea nstruments (based on bas and sze dstorton) and develop a test for wea nstruments. For our case of one endogenous regressor, the test statstc amounts to the F- statstc on excludng the nstrument n the frst stage regresson. Stoc and Yogo (2004) provde crtcal values dependng on the maxmum tolerable bas and sze dstorton. < Table 2 here > 5 As mentoned above, however, more than 90 percent of trade s covered by the countres for whch blateral trade data are avalable.

13 EI Worng Paper No The frst column n Table (2), whch shows the results from regressng the actual on the constructed trade share and ndustry-specfc constants (Table 2 gves only the average value), confrms the strong assocaton llustrated n Fgure 1. In the two-stage least squares estmaton of equaton (7), however, N and A have to be ncluded as nstruments anyway. Hence our constructed trade share s useful only nsofar, as t contans nformaton about the endogenous varable (the trade share) that goes beyond that contaned n country sze (N and A). To put t dfferently: what s relevant s the strength of the partal correlaton between the actual and the constructed trade share, controllng for country sze. As can be seen from columns (2) and (3) of Table 2, addng the constructed trade share to the model ncludng only populaton and area (as well as ndustry-specfc constants) ncreases the R 2 from (column (2)) to (column (3)). Moreover, the F-test of the excludng restrcton for the constructed trade share n column (3), whch s the frst stage regresson n the two-stages least squares estmaton of equaton (7), amounts to Ths value by far exceeds the crtcal values provded by Stoc and Yogo (2004). Hence the hypothess that our two stages least squares estmates are flawed by wea nstruments s strongly rejected. IV. Trade and Productvty: Estmaton Results a) Basc results Havng constructed the nstrument and verfed ts qualty we can now turn to the estmaton of trade s effect on productvty usng equaton (7). Agan, the data are averages over the perod 1995 to 2000; a detaled descrpton s gven n the Appendx. The frst column of Table 3 shows the results of the OLS estmaton. Trade turns out as both statstcally and economcally sgnfcant determnant of productvty; ts t-value s 6.34 and the coeffcent mples that a one-percentage-pont ncrease n trade (the export rato) ncreases productvty by percent. The regresson also confrms a postve relatonshp between country sze

14 Trade and Productvty: An Industry Perspectve 13 and productvty; ncreasng both area and populaton by one percent rases productvty by percent. Turnng to the IV-estmates n column (2) we observe only a mnor change n the parameter estmates; n contrast there s a huge ncrease n the effect of trade n the Franel and Romer (1999) study. Moreover, due to the hgh qualty of our nstrument the frst stage F-statstc on the excluded nstrument amounts to 83.5 our estmates are stll very precse: the standard error of the coeffcent of trade only doubles (whle t quadruples n the Franel and Romer (1999) analyss), leavng the coeffcent on trade sgnfcant at the one per cent level wth a t-value of 3.05 (whle t remans only margnally sgnfcant n the Franel and Romer analyss). 6 So far we have allowed for ndustry-specfc dfferences only by lettng the ntercept vary across ndustres. Ths s a mnmum requrement, as a Wald-test for restrctng the constants to be equal across ndustres s strongly rejected. Ths s not surprsng, gven the huge varaton n productvty across ndustres whch s due to very dfferent producton technologes and cannot be expected to be explaned by varatons n natonal and nternatonal trade. 7 < Table 3 here > 6 Throughout, we account for the fact that the nstrument depends on the parameters of the blateral gravty equatons usng the delta method. That s, the covarance matrx of b (the vector contanng the coeffcents of equaton (7)), s gven by Var [ b( a)], that s the usual two stages least squares varance-covarance matrx for gven a (whch s the staced vector of coeffcents a from the ndustry-specfc gravty models) plus ( b ( a)/ a ) Ωˆ ( b( a)/ a ), where ˆ 2 1 Ω = ˆ ( σ X X) s the varance-covarance matrx of a. The summaton over ndustres results from the bloc-dagonalty of the regressor matrx n the staced gravty model wth ndustry-specfc coeffcents. Ths correcton has only a moderate effect on the precson of the parameter estmates: On average the standard errors ncrease by less than 10 percent. 7 We note that usng the log of the export rato yelds de facto dentcal results. Typcally, the sze of the coeffcent s halved (as expected snce the average export-rato s some 50 percent).

15 EI Worng Paper No b) Testng for cross-ndustry heterogenety In a next step we nvestgate whether there are also sgnfcant dfferences n the effect of trade on productvty across ndustres. We nclude nteracton terms between the trade varable and (-1) ndustry dummes; the parameters of the nteractons terms thus have the nterpretaton of devatons from an arbtrary benchmar ndustry, whose parameter s that of the unweghted trade varable. The test for parameter homogenety can then be carred out wth a smple Wald-test that the parameters of all nteracton terms are zero. The detaled results are gven n Table A2 n the Appendx. The Wald test for parameter homogenety s only slghtly rejected at the 10 percent level (see bottom of Table A2). Tang a closer loo at the nteracton terms t becomes evdent that the sgnfcant test statstc can be traced bac to two ndustres: 23 (coe, refned petroleum products and nuclear fuel) and 34 (motor vehcles, tralers and sem-tralers). The effect of trade on productvty n these two ndustres s ( ) and ( ) respectvely. As a result of farly large standard errors, a Wald-test that these two coeffcents are equal to zero cannot be rejected, nether separately where the test statstcs amount to 1.16 (p-value: 0.282) and 1.49 (p-value: 0.222) respectvely, nor when tested jontly (test statstc: 2.23, p-value: 0.328). Ths s not mplausble: Both ndustres are largely domnated by a few large multnatonal companes. Hence the trade varable reflects a large share of ntra-frm trade and s thus a less proper measure of openness here. As a consequence we exclude these two ndustres from our sample. Repeatng the heterogenety test (see Table A1 n the appendx, column (2)), only one of the nteracton terms s sgnfcant. However, the zero restrcton for ths ndustry s now rejected. Moreover, the Wald-test for parameter homogenety cannot be rejected any more wth a value of (p-value: 0.148). Hence we proceed wth the reduced sample model, excludng the two ndustres 23 and 34. Ths mples that our sample stll covers 92.3 percent of manufacturng. The correspondng OLS and two stages least squares estmates of the homogenous model are

16 Trade and Productvty: An Industry Perspectve 15 gven n columns (3) and (4) of Table 3. The effect of trade on productvty n column (3) s now somewhat larger as expected, snce the effect s averaged down by the two ndustres where the effect s close to zero n columns (1) and (2). 8 It suggests tself to test for cross-country heterogenety as well; unfortunately, ths leads to multcollnearty problems of the nteractons of the countres dummes wth populaton (N) and area (A), whch are country-specfc varables as well. The only way to allow for more heterogenety across countres s the ncluson of selected country dummes, a pont that wll be taen up n the subsequent robustness chec. c) Robustness Before relatng our results to prevous estmates, we now provde some robustness checs of our results. We proceed wth the reduced (homogenous) sample. 9 The frst two columns n Table 3 reproduce the estmates of the reduced model to facltate the comparson. As a frst robustness chec we exclude obvous outlers, defned as observatons where the standardzed resduals exceed the value of two. Ths reduces our sample by 11 observatons. The correspondng results are gven n columns (3) and (4) of Table 4. The standard error falls consderably (by 30 percent) and the ft n terms of the R 2 ncreases from to a remarable The man results are not affected; n fact the coeffcent of trade ncreases to < Table 4 here > 8 Snce the OLS estmates are potentally flawed by endogenety of the trade varable, we focussed on the two stages least squares results n testng for heterogenety. The prce s larger standard errors, of course. Hence the nterpretaton should not be overstressed. We cannot rule out heterogenety; t s rather that at our level of precson we cannot detect sgnfcant dfferences. From an econometrc perspectve, however, t s vald to proceed wth the homogenety assumpton. 9 The results are very smlar, f the full sample s used.

17 EI Worng Paper No In a next step we nclude country dummes to control for remanng nsttutonal heterogenety; snce populaton (N) and area (A) n equaton (7) are country-specfc as well, ncludng a full set of country dummes s not possble due to perfect collnearty. As compromse we nclude only 11 of the 14 country dummes. Ths means the remanng countres are assumed to have the same ntercept; accordngly one should choose the most homogenous group of three countres. To avod mang ad hoc assumptons about the smlarty of countres, we estmated (7) ncludng all possble combnatons of 11 of the 14 country dummes. It turns out that, whle the coeffcents of area and populaton vary consderably across the estmates, the effect on the coeffcent of trade s very smlar across the specfcatons. Columns (5) and (6) of Table 4 show one example, whch may be regarded as representatve. Here three Nordc countres (Denmar, Fnland, and Sweden) have been grouped together; for all other countres a separate dummy was ncluded. As a consequence, the coeffcent of trade falls to but remans sgnfcant at the fve percent level. If outlers are excluded as before, the coeffcent ncreases to agan and turns out sgnfcant at the 1 percent level. We conclude our results are not drven by unobserved country-specfc heterogenety. Fnally, we chec the senstvty of the results wth respect to the constructon of the nstrument. In lne wth Franel and Romer (1999) we use alternatve specfcatons of the gravty model excludng varables whch mght be regarded as endogenous n the very long run. In partcular, we omt the common border dummy, the landloced dummy, area and populaton respectvely from the gravty equaton and re-estmate (7) wth the so constructed nstrument. The results of the two stages least squares estmates are gven n the last four columns of Table 4. As expected the F-statstc of the frst stage regresson falls, when less nformaton s used n the constructon of the nstrument. Nevertheless, the F-statstcs are hgh enough to reject the hypothess of wea nstruments. The coeffcent of trade vares somewhat across the specfcatons from to but always remans sgnfcant.

18 Trade and Productvty: An Industry Perspectve 17 d) Relaton to prevous estmates Our estmates suggest that an ncrease n openness (n terms of the export rato) by onepercentage pont ncreases productvty by approxmately two thrd of a percent on average; the IV estmates range from 0.39 to How do these estmates of trade s effect on productvty compare wth prevous results n the lterature? The most comprehensve survey on the relaton between trade and growth s Lewer and Van den Berg (2003). The studes surveyed do not exactly match our setup; they are prmarly specfed n growth rates, do not use trade relatve to GDP and f specfed n ntensve form at all use per capta values; nevertheless the results are suggestve. Lewer and Van den Berg (2003) fnd that many emprcal studes yeld surprsngly consstent results: a one percentage pont ncrease n growth of exports s a assocated wth a one-ffth percentage pont ncrease n economc growth. The results by Franel and Romer (1999) are not drectly comparable; they fnd that openness (defned as mports plus exports as share of GDP as measure) ncreases ncome by 0.5 to 2 percent; the estmates of Irwn and Tervö (2002) vary somewhat over the samples but are of a smlar sze. For a rough comparson, we can translate the estmates of the Franel and Romer (1999) study nto elastctes by evaluaton them at the sample mean of openness (73.4 per cent): ths yelds an average value of Alcala and Cccone (2004), who use the log of real openness, obtan much elastctes between 0.91 and 1.49 (and, smlarly, elastctes around one f nomnal openness s used). Translatng our estmates nto elastctes by evaluatng them at the sample mean of the export rato (43.5 per cent, see Table A1 n the Appendx) we obtan an average value of 0.37 for manufacturng (whch maes up only roughly one ffth of total value added). As we argued n secton II, the ndustry-specfc specfcaton (7) narrows the effects captured by the coeffcent of the trade varable, rulng out effects that materalze across ndustres (nterndustry productvty spllovers wthn manufacturng and from manufacturng to servces). In order to roughly account for ths fact we assume that trade n manufacturng goods affects

19 EI Worng Paper No productvty n servces equally 10 ; even then the mpled aggregate effect of 0.37 s stll less than half of the Franel and Romer (1999) estmates. Hence, the dfferent setup of the hypotheses can only partally account for ths large gap. There are two further explanatons for ths dscrepancy n the sze of the estmated effect. Frst, t could be due to the choce of our sample of countres. If there are decreasng returns to economc (trade) ntegraton, the small effect of our estmates could be due to the fact that ours sample conssts of hghly ntegrated countres wth comparably large trade shares. One possble reason for such decreasng returns s that trade may have a larger effect on productvty for countres whch are more dstant from the technology fronter, snce trade facltates the catchng-up process by transmttng technologcal now-how for mtaton. Ths effect dmnshes as countres are movng closer to the technology fronter. Usng aggregate data, the cross secton of our sample would reduce to fourteen countres, whch s too small to yeld sensble estmates. We can, however, consder whether our sample of countres has a sgnfcantly dfferent coeffcent n the Franel and Romer (1999) sample. Table A3 n the Appendx gves the estmaton results. Columns (1) and (2) reproduce the Franel and Romer (1999) estmates; columns (3) and (4) show the results when a dummy for our group of 14 countres (D OECD ) s ncluded along wth an nteracton of the trade rato and D OECD, whch measures the devaton of trade s productvty effect n our sample from the average. 11 Whle the devaton s n fact negatve, mplyng a substantally lower effects of trade on productvty, the dfference s nsgnfcant. The same holds true f the restrcted sample of Ths assumpton requres some justfcaton: As Franel and Romer argue t s not the lteral shppng of goods between countres that rases ncome. Rather, trade may be vewed as a proxy for the many ways n whch nteractons between countres rase ncome. From ths perspectve, one would expect that any mpact of trade n goods on productvty n manufacturng would also be assocated wth an mpact on productvty n servces. 11 Snce the 14 OECD countres may have a hgher productvty for reasons other than ther trade share (agan, nsttutons) ncludng a separate constant s a mnmum requrement for the heterogenety tests.

20 Trade and Productvty: An Industry Perspectve 19 countres s used as reference (see columns (5) and (6)). Hence, there s no strong support for the presumpton of decreasng returns to trade, but n lght of the fact that the estmates are rather mprecse, the results are somewhat nconclusve. Second, the effect of trade n productvty on trade n prevous studes s lely to be based upwards. Most studes use aggregate productvty but trade n goods as explanatory varable. Servces mae up more than two thrd of GDP and trade n servces has rsen consderably over the last years. Snce trade n goods s lely to be hghly correlated wth trade n servces, the coeffcent could also capture productvty effects of trade n servces. For studes, whch use many developng countres (where servce trade s neglgble) or use tme perods where servces have been less mportant (such as 1985 n the Franel and Romer paper), ths s an unlely explanaton. Even today and for developed countres, where trade n servces mae up only lttle more than 20 percent of total trade ths may agan provde a partal explanaton at best. Anyway, consderng the effect of trade on productvty n servce sectors s a worthwhle lne for future research. A pror, none of these explanatons can be favoured on grounds of theory or exstng studes. In lght of the recent evdence, the ln between trade and productvty appears to be farly robustly establshed. Nevertheless, the sze and the transmsson channels of the effect deserve further attenton. V. Summary and Conclusons Ths paper complements prevous studes on the productvty effects of trade, provdng an analyss at the ndustry-level for a sample of 14 OECD countres. Followng Franel and Romer (1999) we use the geographcal component of aggregate trade, whch s constructed from the predcted values of a purely geographcal gravty model, as nstrument for trade to deal wth the lely endogenety of a country s (ndustry s) openness.

21 EI Worng Paper No The effect of trade on productvty turns out to have a statstcally and economcally sgnfcant effect on productvty: On average, an ncrease n openness (n terms of the export to producton rato) by one percentage pont ncreases the productvty n manufacturng by 0.6 percent. Snce we use a sample of nsttutonally largely homogenous countres, our results are unlely to be drven by unobserved nsttutonal heterogenety. Ths s also confrmed by a robustness chec ncludng country dummes. Comparng the OLS wth the IV estmates, the OLS estmates appear to be downward based (due to measurement error, snce trade s a poor proxy for nternatonal nteractons) rather than upward based (due to reverse causalty). Ths s n lne wth Franel and Romer (1999) and Irwn and Tervö (2002), but the dfference between our OLS and IV estmates s much smaller than that n prevous studes. One puzzlng result s the sze of the estmated effect on aggregate productvty, whch amounts to less than half of prevous estmates. We offer three explanatons all of whch deserve further nvestgaton: ) Usng ndustry data our coeffcent of trade does not capture effects of trade that operate across ndustres such as external Economes of Scale of nterndustry type wthn manufacturng. ) Usng a sample of developed countres wth hgh trade ntegraton may result n low estmates f there are decreasng returns to the gans from trade. ) The sze of prevous estmates may be upward-based, f the trade (n goods) varable also captures productvty effects of trade n servces (whch s lely to be hghly correlated wth trade n goods). The contrbuton of each of these arguments s unclear. Provdng a convncng explanaton for the large dscrepancy between our estmates and that of prevous studes s an mportant further step to gan a deeper understandng of the economc sgnfcance of trade n enhancng productvty and the channels va whch these effects materalze.

22 Trade and Productvty: An Industry Perspectve 21 References Alcala, F. and Cccone, A. (2004). Trade and Productvty. Quarterly Journal of Economcs, 119(2), Andrews, D., Morera, M. and Stoc, J.H. (2004). Optmal nvarant tests for nstrumental varables regresson. NBER Techncal Worng Paper. Balassa, B. (1961). The Theory of Economc Integraton. London: Allen & Unwn. Bhagwat, J.N. (1965). On the equvalence of tarffs and quotas. In: Baldwn, R.E. et al. (eds.), Trade, Growth, and the Balance of Payments: Essays n honor of Gottfred Haberler, Amsterdam: North Holland, Clar, G., Gauler, G., Mayer, T. and Zgnago, S. (2004). Notes on CEPII s dstances measures. Pars, www-cep.fr. Coe, D.T. and Helpman, E. (1995). Internatonal R&D spllovers. European Economc Revew, 39(5), Durlauf, S. (2001). Manfesto for a growth econometrcs. Journal of Econometrcs, 100(1), Franel, J.A. and Romer, D. (1999). Does trade cause growth? Amercan Economc Revew, 89(3), Hall, R.E. and Jones, C.I. (2000). Why do some countres produce so much more output per worer than others? Quarterly Journal of Economcs, 114, Irwn, D.A. and Tervö, M. (2002). Does trade rase ncome? Evdence from the twenteth century. Journal of Internatonal Economcs, 58, Lewer, J.J. and Van den Berg, H. (2003). How large s nternatonal trade s effect on economc growth. Journal of Economc Surveys, 17(3),

23 EI Worng Paper No Rodr, D., Subramanran, A., Trebb, F. (2002). Insttutons rule: the prmacy of nsttutons over geopraphy and ntegraton n Economc Development. NBER Worng Paper, No Stager, D. and J.H. Stoc (1997). Instrumental varables regresson wth wea nstruments. Econometrca, 65, Stoc, J.H., Wrght, J.H., and Yogo, M. (2002). A survey of wea nstruments and wea dentfcaton n generalzed method of moments, Journal of Busness and Economc Statstcs, 20, Wooldrdge, J. (2003). Introductory Econometrcs, 3 rd edton, Prentce Hall.

24 Trade and Productvty: An Industry Perspectve 23 Appendx Data descrpton All data are averages over the perod Cross-country dmenson comprses 14 countres (AUT, BEL, DEU, DNK, ESP, FIN, FRA, GBR, GRC, ITA, NLD, NOR, SWE, USA); dmenson of partner countres j (j ) comprses the 223 countres contaned n the CEPII dataset; for the estmaton of the gravty equaton (8) the dmenson j reduces to the 44 partner countres (j ) for whch blateral trade data by ndustry are avalable. T = /, where T j export rato; j X j PROD country j n sector and OECD Structural Analyss (STAN) Database. X j s blateral exports from country to PROD s the producton of country n sector. Source: y d j labour productvty; y = VA / EMPL, where VA s real valued added of country n sector n 1995$ (base year 1995, converted nto $ wth average PPPs exchange rate over the perod ) and EMPL s total employment n ndustry of country. Source: OECD Structural Analyss (STAN) Database. smple dstance between country and country j (smple dstance dst ). Source: CEPII (Clar et al., 2004). N populaton of country n 1000 persons. Source: Unted Natons: Demographc Yearboo. A area of country n square lometres. Source: CEPII (Clar et al., 2004). LL dummy varable tang a value of one f country s landloced and zero otherwse. Source: CEPII (Clar et al., 2004). CB j dummy varable tang a value of one f countres and j share a border. Source: CEPII (Clar et al., 2004). < Table A1 here > < Table A2 here > < Table A3 here >

25 EI Worng Paper No Table 1. Geography and Blateral Trade: Estmaton Results for Model (8) by Industry Dependent varable: ln T j ndustry constant *** *** *** *** *** (0.849) (0.914) (1.138) (0.948) (1.416) (0.750) (0.808) (0.861) ln dstance *** *** *** *** *** *** *** *** (0.059) (0.063) (0.079) (0.066) (0.101) (0.053) (0.056) (0.060) ln populaton *** ** *** (country ) (0.062) (0.066) (0.082) (0.069) (0.118) (0.062) (0.059) (0.062) ln area *** *** *** *** (country ) (0.063) (0.068) (0.085) (0.071) (0.113) (0.060) (0.060) (0.064) ln populaton *** *** *** *** *** *** *** *** (country j) (0.045) (0.048) (0.060) (0.050) (0.077) (0.041) (0.043) (0.045) ln area *** ** ** (country j) (0.035) (0.037) (0.047) (0.039) (0.060) (0.031) (0.033) (0.035) landloced *** *** *** ** *** *** *** *** (0.120) (0.129) (0.160) (0.134) (0.203) (0.108) (0.114) (0.121) common border *** *** *** *** *** *** *** *** (0.249) (0.268) (0.333) (0.278) (0.422) (0.224) (0.237) (0.252) SE of regresson R observatons

26 Trade and Productvty: An Industry Perspectve 25 Table 1 (contnued). Geography and Blateral Trade: Estmaton Results for Model (8) by Industry Dependent varable: ln T j ndustry constant ** ** *** *** (0.939) (0.766) (0.685) (0.786) (0.972) (1.251) (1.000) ln dstance *** *** *** *** *** *** *** (0.065) (0.053) (0.047) (0.054) (0.067) (0.087) (0.069) ln populaton ** ** ** ** ** (country ) (0.068) (0.056) (0.050) (0.057) (0.071) (0.091) (0.073) ln area *** ** ** (country ) (0.070) (0.057) (0.051) (0.059) (0.073) (0.093) (0.075) ln populaton *** *** *** *** *** *** *** (country j) (0.049) (0.040) (0.036) (0.041) (0.051) (0.066) (0.053) ln area ** *** ** ** (country j) (0.038) (0.031) (0.028) (0.032) (0.040) (0.051) (0.041) landloced *** *** *** *** *** ** (0.132) (0.108) (0.097) (0.111) (0.137) (0.177) (0.141) common border *** *** *** *** *** *** *** (0.275) (0.225) (0.201) (0.230) (0.285) (0.366) (0.293) SE of regresson R observatons Notes: Standard errors are n parenthess. *** sgnfcant at 1 percent, ** 5 percent, * 10 percent.

27 EI Worng Paper No Table 2. Qualty of Instrument: The Relevance of Geography for Aggregate Trade dependent varable: exports (T ) (1) (2) (3) constant 1) constructed trade share *** *** (0.071) (0.095) ln populaton *** (1.747) (1.544) ln area *** (0.655) (0.569) SE of regresson R observatons 2) Notes: 1) average of ndustry-specfc constants. 2) data for Norway (23,24) are mssng. *** sgnfcant at 1 percent, ** 5 percent, * 10 percent. Table 3. Trade and Productvty: Estmaton Results for Model (7) dependent varable: productvty (y ) full sample reduced sample (excl. 23, 34) LS IV LS IV constant 1) export rato *** *** *** *** (0.086) (0.172) (0.078) (0.144) ln populaton *** ** *** *** (0.021) (0.021) (0.019) (0.019) ln area *** ** *** *** (0.021) (0.027) (0.019) (0.025) SE of regresson R observatons Frst stage F-Test Notes: 1) average of ndustry-specfc constants. *** sgnfcant at 1 percent, ** 5 percent, * 10 percent. R 2 n IV regressons calculated as squared correlaton between dependent varable and predcted values.

28 Trade and Productvty: An Industry Perspectve 27 Table 4. Trade and Productvty: Robustness dependent varable: exports baselne outlers country dummes alternatve gravty models omttng 2) LS IV LS IV LS IV CB j LL +LL j N, N j A, A j constant 1) export rato *** *** *** *** *** ** *** *** *** *** (0.078) (0.144) (0.064) (0.113) (0.073) (0.179) (0.164) (0.166) (0.175) (0.187) Ln populaton *** *** *** *** *** ** *** *** (0.019) (0.019) (0.013) (0.013) (0.123) (0.130) (0.019) (0.019) (0.019) (0.019) Ln area *** *** *** *** *** *** *** *** ** *** (0.019) (0.025) (0.014) (0.017) (0.030) (0.034) (0.026) (0.027) (0.028) (0.029) SE of regresson R observatons Frst stage F-Test Notes: 1) average of ndustry-specfc constants. *** sgnfcant at 1 percent, ** 5 percent, * 10 percent. 2) IV-estmates, when the respectve varables are omtted from the gravty equaton for the constructon of the nstruments. The correspondng LS regresson s that gven n the frst column.

29 EI Worng Paper No Table A1. Overvew of Manufacturng Industres ISIC Rev3 ndustry share n value added 1) export rato 2) Food products, beverages and tobacco Textles, textle products, leather and footwear Wood and products of wood and cor Pulp, paper, paper products, prntng and publshng Coe, refned petroleum products and nuclear fuel Chemcals and chemcal products Rubber and plastcs products Other non-metallc mneral products Basc metals Fabrcated metal products, except machnery and equpment Machnery and equpment, n.e.c Electrcal and optcal equpment Motor vehcles, tralers and sem-tralers Other transport equpment Manufacturng nec Notes: 1) Share n total manufacturng value added n percent. 2) Rato of exports to producton n percent. All values are averages of the perod and the 14 countres.

30 Trade and Productvty: An Industry Perspectve 29 Table A2. Trade and productvty: Heterogenety tests dependent varable: exports full sample reduced sample (excl. 23, 34) constant 1) export rato ** *** (0.523) (0.444) ln populaton ** *** (0.023) (0.019) ln area *** *** (0.032) (0.029) Interacton of export rato wth (0.534) (0.449) (0.780) (0.659) (0.762) (0.638) ** (1.037) (0.550) (0.462) (0.576) (0.483) (1.689) (1.431) (0.641) (0.539) (1.780) (1.513) (0.579) (0.487) ** (0.527) (0.445) ** (0.624) (0.836) (0.704) (0.580) (0.487) SE of regresson R Observatons Wald-test for parameter homogenety * Notes: 1) average of ndustry-specfc constants. *** sgnfcant at 1 percent, ** 5 percent, * 10 percent.

31 EI Worng Paper No Table A3. Heterogenety Tests n Franel and Romer (1999) sample dependent varable: productvty (y ) Baselne: Franel and Romer (1999) Heterogenety Test Full Sample (150) Heterogenety Test Reduced Sample (98) LS IV LS IV LS IV constant *** *** *** *** *** *** (0.332) (0.912) (0.306) (0.971) (0.430) (1.502) trade *** ** *** ** ** (0.332) (0.912) (0.229) (0.865) (0.284) (1.238) Ln populaton * ** (0.064) (0.088) (0.061) (0.087) (0.090) (0.129) Ln area (0.055) (0.098) (0.051) (0.088) (0.068) (0.144) D OECD *** *** *** *** (0.596) (0.701) (0.617) (0.741) D OECD trade (0.776) (0.977) (0.801) (1.054) SE of regresson R observatons Notes: Estmates based on the data of Franel and Romer (1999), provded n the Appendx. *** sgnfcant at 1 percent, ** 5 percent, * 10 percent. The standard errors of the IV estmates have not been corrected to account for the varance of the gravty estmates; hence, the heterogenety tests (slghtly) tend to overreject.

32 Trade and Productvty: An Industry Perspectve 31 Fgure 1. Actual versus Constructed Trade Share 200 Actual trade share n percent Constructed trade share n percent

33 EI Worng Paper No EI Worng Papers 1 Gerhard Fn, A Schedule of Hope for the New Europe, Otober Gerhard Fn, Jutta Gumpold, Österrechsche Behlfen m europäschen Wrtschaftsraum (EWR), Otober Gerhard Fn, Mcroeconomc Issues of Integraton, November Frtz Breuss, Herausforderungen für de österrechsche Wrtschaftspolt und de Sozalpartnerschaft n der Wrtschafts- und Währungsunon, November Gerhard Fn, Alexander Petsche, Central European Economc Polcy Issues, July Gerhard Fn, Alexander Petsche, Antdumpng n Österrech vor und nach der Ostöffnung, November Frtz Breuss, Karl Stennger, Reducng the Greenhouse Effect n Austra: A General Equlbrum Evaluaton of CO2-Polcy-Optons, March Franz-Lothar Altmann, Wladmr Andreff, Gerhard Fn, Future Expanson of the European Unon n Central Europe, Aprl Gabrele Tondl, Can EU's Coheson Polcy Acheve Convergence?, Aprl Jutta Gumpold, Natonale bzw. gesamtwrtschaftlche Effete von Behlfen - nsbesondere Exportbehlfen, Aprl Gerhard Fn, Martn Opptz, Kostensenungspotentale der Wener Wrtschaft - Salenerträge und Kostendruc, August Alexander Petsche, De Verfassung Ungarns m Lchte enes EU-Betrtts, September Mchael Sora, De Europäsche Unon m Internet, September Frtz Breuss, Jean Tesche, A General Equlbrum Analyss of East-West Mgraton: The Case of Austra-Hungary, January Alexander Petsche, Integratonsentwclung und Europaabommen EU - Ungarn, Jul Jutta Gumpold, De Ausfuhrförderung n der EU, Jun Jutta Gumpold, Internatonale Rahmenregelungen zur Ausfuhrförderung, Jun Frtz Breuss, Austra's Approach towards the European Unon, Aprl Gabrele Tondl, Neue Impulse für de österrechsche Regonalpolt durch de EU- Struturfonds, Ma Grller, Droutsas, Falner, Forgó, Klatzer, Mayer, Nentwch, Regerungsonferenz 1996: Ausgangspostonen, Jun Stefan Grller, En Staat ohne Vol? Zur Zuunft der Europäschen Unon, Otober Mchael Sora, Der EU-Info-Broer en datenbangestütztes Europanformatonssystem m World Wde Web über de KMU-Förderprogramme der Europäschen Kommsson, November Katrn Forgó, Dfferenzerte Integraton, November 1996.