Endowments, Specialization, and Policyroie_916 913..923

Olivier Cadot and Yuliya Shakurova*

AbstractThe paper explores the relationship between industry shares in production and their determinants includingfactor endowments, technology, and government policies, in a GDP–function framework. We use a newinternational panel dataset on production and trade compiled by the World Bank.As an intermediate step wecalculate Hicks-neutral productivity indices that vary across industries, time, and countries. We find thatown-TFP is robustly associated with industry shares across time and countries and that, after correcting forthese productivity differences, output shares are related to factor endowments (Rybczynski effects) in aplausible way. Once Rybczynski effects are controlled for, we find little evidence of demand-side policies(import tariffs) affecting the allocation of resources; we find, however, more role for supply-side policies as therelative size of capital-intensive industries is positively associated with infrastructure–capital endowments.

1. Introduction

What determines a country’s production pattern across sectors? Trade theory directlysuggests a number of potential determinants: factor endowments, demand-side policies(tariffs and subsidies), and supply-side policies (e.g. the provision of infrastructure andpublic capital). But then, which of these determinants have proved across time andcountries to be robustly associated with cross-sectoral production patterns, and canthey be influenced by policy?

In a perfect world, production patterns would surely best be left for markets todetermine. However, as noted by Rodrik (2007, pp. 2–3):

“The market failures that provide a role for industrial policy are thebread-and-butter of development economists. They are widely perceived tobe pervasive, even if systematic evidence is sketchy and hard to come by.Informational and political problems in administering industrial policy arelegion—but in that respect too industrial policy is no different from manyother areas of policy. Moreover, most governments do carry out variousforms of industrial policy already, even if they call it by other names (‘exportfacilitation,’ ‘promotion of foreign investment,’ ‘free-trade zones,’ etc.). Con-sequently, it is far more productive for the discussion to focus on howindustrial policy should be carried out than on whether it should be carriedit out at all.”

Rodrik’s paper is representative of a revival in academic and political interest inindustrial policy. Our opening question is thus, from a policy perspective, one of

* Cadot: The World Bank, University of Lausanne, Faculty of Business, CERDI, CEPR, and CEPREMAP,Internef-Dorigny, CH-1015 Lausanne, Switzerland. Tel: +41-21-692-34-63; Fax: +41-21-692-34-95; E-mail:ocadot@worldbank.org. Shakurova: University of Lausanne, Faculty of Business, Internef-Dorigny, CH-1015Lausanne, Switzerland. E-mail: juliya.shakurova@unil.ch.This paper is part of a World Bank research project.We thank an anonymous referee, Alessandro Nicita, Claudio Sfreddo, Rafael Lalive, Jean Imbs, PascalSt-Amour, and participants in the Department of Economics Research Days, University of Lausanne, foruseful suggestions. Special thanks to Claudio Sfreddo for his help in the calculation of the public infrastruc-ture index, and to Alessandro Nicita for his advice in the early stages of the paper.

Review of International Economics, 18(5), 913–923, 2010DOI:10.1111/j.1467-9396.2010.00916.x

© 2010 Blackwell Publishing Ltd

interest. At a general level, the Heckscher–Ohlin–Vanek (HOV) model suggests thatthe best way for a government to influence a country’s pattern of specialization is tocreate the right policy environment for the appropriate factors of production to beaccumulated. Put differently, under the HOV assumptions, whereas demand-side poli-cies such as trade barriers and sectoral promotion are doomed—because they areunlikely to lead to anything but capture by special interests—supply-side policies canbe justified. However, there is currently a gap between this type of loose presumptionand the operational policy advice that is typically hoped for by governments.

At a theoretical level, specialization is linked to the “factor content of trade,” as anindustry that has a large share in GDP is likely to be an export one, unless theeconomy is severely distorted. So the relationship between production patterns andendowments is not independent of the relationship between trade and endowments.A large body of literature has been devoted to testing the HOV model’s basic pre-diction for the factor content of trade. The results were, at first, a long series ofsetbacks. Although Leamer (1984) solved Leontief’s paradox by showing that the testwas not robust to large trade imbalances, Trefler (1995) confirmed that the measuredfactor content of international trade did not reflect differences in trade endowments,a finding he called “missing trade.” Based on his work, the subsequent literature(Harrigan, 1997; Davis and Weinstein, 2001, DW (2001) hereafter) used unobservedor partially observed technology differences across countries and industries toexplain the observed deviations of the factor content of trade from the predictions ofthe HOV model. Together with other adjustments—for example, for unobservedtrade barriers through the use of gravity-predicted trade flows instead of observedones—as in DW (2001) the modifications to the model suggested by the empiricalliterature, which consisted of assuming various forms of technology differences,turned out to close much of the gap between theory and measurement. The mostrecent paper in this tradition—and closest to ours—is Schaur et al. (2008) publishedin this journal. The debate is, by now, largely settled.

Surprisingly, the normative side of the question (the policy debate about what typeof trade intervention, if any, is appropriate) has evolved in parallel with the positive sidebut without much interaction between the two. Early conceptions in terms-of-tradepolicy involved a basic laissez-faire presumption with exceptions only for infant indus-tries and in other very particular cases. Schools of thought with a more activist stand,such as Latin America’s ECLAC, drew on altogether different theoretical premises, andthe import-substitution policies that were based on those premises largely failed to liveup to their promises. The idea of proactive trade and industrial policies was brieflyrevived, under the generic name of “strategic trade policy,” with the advent of intra-industry trade models in the 1980s. Essentially, the argument was that under imperfectcompetition, beggar-thy-neighbor policies could be welfare-enhancing from a nationalpoint of view although welfare-reducing from a global one. The vogue of game-theoretic arguments as guides for strategic trade policy was, however, short-lived, asmany of those arguments proved to be not robust to slight changes in the assumptions(e.g. moving from Cournot to Bertrand competition) and US concerns about lowproductivity growth and large trade deficits with Japan and Europe, which had con-tributed to the appeal of proactive policy prescriptions, vanished in the 1990s. One ofthe few recent papers relating policy or institutional issues to the factor endowments–specialization nexus is Belloc (2009) which looks at the relationship betweenunionization and comparative advantage.

The objective of this paper is to shed light on this issue using a traditional GDP–function approach but applied on a dataset that has only recently been made available

914 Olivier Cadot and Yuliya Shakurova

© 2010 Blackwell Publishing Ltd

and that is arguably very well suited to the question. The dataset is the World Bank’sTrade and Production Database (Nicita and Olarreaga, 2006, NO (2006) hereafter),which contains data on production, trade and endowments at the ISIC three-digit level.We estimate Rybczynski derivatives for the database’s 100 countries and 27 sectorsusing aggregate capital/labor ratios constructed from investment and employment datacontained in the database.

Results are encouraging. The baseline model yields positive and significant shareeffects for own-TFP parameters and sorts industries by capital intensiveness in aplausible way. Concerning tariff protection, however, it suggests very few effects.

2. Background

Let R(qp, v) be a representative country’s GDP function, solving

R p f v vv

j jj

jj

n

jj

n

jθ θ⋅( ) = ( ) =⎧

⎨⎩

⎫⎬⎭= =

∑ ∑p v v, max . . ,1 1

s t (1)

where q is an n ¥ n diagonal matrix of Hicks-neutral country- and sector-specific technology parameters, p is an n ¥ 1 vector of commodity prices, and v is anm ¥ 1 vector of factor endowments. Following Kohli (1978) and (inter alia) Harrigan(1997), for estimation purposes we assume a translog form (a flexible functional formin which elasticities are not constrained to be constant):

ln ln ln

ln ln

ln

R p v

p p

v

j j jj

k kk

ij i i j jji

kl k

= + +

+

+

∑ ∑

∑∑

α α θ β

γ θ θ

δ

0

12

12

lln

ln ln .

v

p vlk

ik i i kki

1∑∑∑∑+ φ θ (2)

Like any revenue function (Dixit and Norman, 1980) the GDP function is homo-geneous of degree one in prices and endowments, which implies the followinghomogeneity restrictions:

α β γ δ φ φjj

kk

iji

klk

iki

ikk

∑ ∑ ∑ ∑ ∑ ∑= = = = = =1 0; . (3)

Sectoral shares can be obtained by differentiating (2) with respect to ln pj:

∂∂

lnln

ln ln .Rp

p vj

j ij i ii

jk kk

= + +∑ ∑α γ θ φ (4)

Homogeneity restrictions (3) imply that γ γi jN

ij1 2= − ∑ = , so (4) can be rewritten as

∂∂

lnln

ln ln ln

ln

Rp

p p vj

j jl ij i ij

n

jk kk

j ij

= + + +

= −

=∑ ∑α γ θ γ θ φ

α γ θ

1 12

1 pp p v

pp

i

n

ij i ii

n

jk kk

j iji i

i

n

j

12 2

1 12

= =

=

∑ ∑ ∑

∑

+ +

= + +

γ θ φ

α γ θθ

φ

ln ln

ln kk kk

vln .∑ (5)

ENDOWMENTS, SPECIALIZATION, AND POLICY 915

© 2010 Blackwell Publishing Ltd

Similarly using (3) to write φ φjl kM

jk= − ∑ =2 we have

∂∂

lnln

ln ln ln .Rp

pp

vvj

j iji

i

n

iji

i

n

jkk

k

= + + += =∑ ∑ ∑α γ γ θ

θφ

12 12 1

(6)

Finally, noting that

∂∂

∂∂

lnln

,Rp

pR

Rp

p YY

sj

j

j

j jj= = ≡ (7)

where sj is sector j’s share in GDP, we have

spp

vv

j j iji

i

n

iji

i

n

jkk

k

= + + += =∑ ∑ ∑α γ γ θ

θφln ln ln .

12 12 1(8)

3. Estimation

Empirically testing an expression like (8) involves substantial prior work to generateproxies for right-hand-side variables.This section describes our data-construction strat-egy. Essential differences between our work and Harrigan’s come from the availabilityof a new dataset compiled by the World Bank and from our use of tariffs to approxi-mate international price differences (whereas he simply assumed free trade). Includ-ing tariffs in the equation not only makes it possible to do away with a dubiousassumption—universal free trade—but it also makes it possible to assess the extent towhich trade protection seems to affect the allocation of productive resources. In orderto bring more of a policy flavor to our exercise, we also include measures of publicinfrastructure capital. If infrastructure is complementary with private capital, animprovement in its provision should shift resources toward capital-intensive sectors.

The Data

Data sources The primary source of our data is the revised version of the WorldBank’s Trade and Production Database (TPDB, described in detail in NO, 2006). TheTPDB covers 100 countries and 28 manufacturing sectors at the ISIC-revision 2 three-digit level, and includes data on trade, production, and protection. All data are incurrent US dollars. Descriptive statistics are given in Table 1. Two variables must becalculated: capital stocks and total factor productivity (TFP), taken as a proxy for thetechnology parameter q .

Constructing the capital stock Capital stocks for each manufacturing sector andcountry were calculated from the TPDB’s sectoral investment data (GFCF) using thePerpetual Inventory Method (PIM).1 Several variants of the method exist, dependingon how capital consumption is modeled, and are described in detail in OECD (2001).Geometric depreciation over an infinite lifetime, used by Statistics Canada, assumesthat capital assets lose value most rapidly in their first years. Hyperbolic depreciation,used, inter alia, by the US Bureau of Labor Statistics, assumes instead that capital’sproductive value is a concave function of time, until a finite end of service life.2 Thisassumption is better suited to productivity analysis as it reflects the machinery’s capa-city to deliver productive services (e.g. by not breaking down) rather than its market(or accounting) value (OECD, 2001).

916 Olivier Cadot and Yuliya Shakurova

© 2010 Blackwell Publishing Ltd

Formally, let It be the flow of gross investment at time t (observed) and Kt the netcapital stock at time t (to be estimated). Let also L be the capital stock’s average servicelife and t the date of its purchase.The “age–efficiency” adjustment factor, dt,t is given by

δτ β τ τ

τt

L t L t t L,

,=

− − −( )[ ] − − −( )[ ] − <{ 1 1

0

if

otherwise,(9)

where b is the slope coefficient (the US Bureau of Labor Statistics (BLS) assumesb = 0.5 for machinery and b = 0.75 for construction; we used b = 0.625).3 The efficiency–age value of the capital stock at time t (t = 0 in the sample’s initial year) is then

K K It t t

t

= +=

−

∑δ δ τ ττ

, , .0 01

1

(10)

Practical estimation of (10) requires an initial estimate for K0, whose influence on latervalues of Kt vanishes over time. We set K0 at k times value-added, where k has to bechosen as a parameter (see Table 1) and use a service life of 23 years, corresponding tothe BLS’s average across industries (see also Annex 3 of OECD, 2001, which listsestimates of machinery service lives from various national statistical offices).All valuesin the TPDB are in current dollars, so we deflated them using the US producer priceindex from NBER–CES Manufacturing Industry Database, assuming that purchasingpower parity (PPP) holds in the long run.

TFP The productivity parameters were calculated relative to the US using an indexdue to Caves et al. (1982) and described in some detail in Harrigan (1997).4 That is, weassumed that for each industry, cross-country technology differences are Hicks-neutral.For Cobb–Douglas technologies, the index’s formula is

θα α

α αjtc jt

cjtc

jtc

jtUS

jtUS

jtUS

y L K

y L K

j j

j j=

( ) ( )⎡⎣ ⎤⎦( ) ( )⎡⎣

−

−

1

1 ⎤⎤⎦, (11)

where α j is a geometric average of the share of labor in all countries in industry j,calculated using employment and wage data given in the TPDB using

α == =

∏ ∏W ycjt

ccjt

c1

100100

1

100100 ,

Table 1. Summary Statistics

No. of obs. Mean Std. dev. Min. Max.

GDP share 54,972 0.4224045 0.7244915 4.38e-06 18.7972Tariff 56,166 14.49044 20.42839 0 349.1Public capital 75,180 -0.0432877 0.3733612 -0.476144 0.955331TFP 32,209 0.56282 0.3448961 0.0005478 2.49599Capital ind. 45,052 113.9718 2334.378 0.0015464 93,627.8Capital 73,892 53.61715 50.69451 0.413685 231.569Land 73,472 0.0006888 0.000762 4.63e-07 0.0067401Primary educ. 60,900 0.196377 0.1506267 0.0042019 1.15718Secondary educ. 60,900 0.1597826 0.1330986 0.0015446 0.695144Postsec. educ. 60,900 0.0811598 0.0715105 0.0007608 0.577095

ENDOWMENTS, SPECIALIZATION, AND POLICY 917

© 2010 Blackwell Publishing Ltd

where Wcjt is the wage bill and yc

jt is value-added in the TPDB. The resulting indexvalues are noisy, which is to be expected as measurement errors leading to values ofcapital or labor near zero send the ratio in (11) to infinity.As a reality check, one wouldexpect national averages of the productivity parameters to be somewhat correlatedwith GDP per capita across countries. A simple pooled OLS regression of θ θt

ctUS on

GDP GDPtc

tUS gives a coefficient of 0.58 significant at the 1% (robust t-stat. 7.36).

R2 = 0.48, reflecting the index’s noise. All in all, the crude plausibility test of an overallpositive relationship between average technology and GDP per capita is passed.

Estimating Rybczynski Elasticities

We estimated stochastic versions of (8) at the industry level (27 industries meaning 27equations) with time and country fixed effects and measurement–error correction on apanel of 100 countries and 25 years, using 3SLS.5 We included only own productivityand own tariffs so as to save on degrees of freedom. Cross-industry productivity effects(spillovers) would have been interesting to identify, but preliminary estimation exer-cises showed them to be all insignificant. Similarly for tariffs: cross-industry crowding-out effects would have been interesting to track, but they all proved insignificant. Weestimated the industry equations simultaneously for groups of similar industries inorder to control for cross-equation correlation in the error terms.6

Letting the superscript c stand for countries and dropping industry indices, we havethus

s PC

KL

TL

ct c t ct ct ct

ct

ct

ct

cth

= + + + +

+ + +

η δ α α τ α θ

α α α

1 2 3

4 5

ln ln

ln ln lnn ,,HLtype ct

cthct

=∑ +

6

8

ε (12)

where hc is a country fixed effect, dt is a time effect, PCct is relative country public capitalindex, tct is tariff, q ct is an industry productivity index, Kct /Lct is country c’s relativecapital endowment, Tct /Lct is its relative land endowment, and Htype,ct /Lct is humancapital per worker.7

Two potential sources of measurement error on the right-hand side must be takencare of. One is persistent errors due to differences in data collection across countries.The other, a classical measurement error, may be generated in the calculation of theproductivity index. As shown by Harrigan (1997), the measurement error’s persistentcomponent (the first) is controlled for by the country and time effects included in theequation. By contrast, the classical measurement error (the second) requires instru-mentation if one is to get consistent estimates. Assuming that measurement errors areuncorrelated across countries,8 we instrument productivity in country c by the averagevalue of productivity in all other countries (industry by industry). That is, q jct is instru-mented by 1 1C k c

Cjkt−( )∑ ≠ θ , where j = 1, . . . , N indexes industries (equations). Then,

these instruments are used for correcting measurement error bias for each industry forcountry c at time t.

Finally, political-economy models suggest that tariffs are likely to be endogenous toproduction levels; for instance, the common-agency model (Grossman and Helpman,1994) makes tariffs an increasing function of the ratio of domestic production toimports. Alternatively, tariffs can be applied to protect infant industries, in which casethey would be decreasing in the level of domestic production.9 So we do not knowwhich way the bias would go but what is clear is that tariffs are unlikely to be exogenousto production levels and hence to industry shares in GDP. In order to control for this

918 Olivier Cadot and Yuliya Shakurova

© 2010 Blackwell Publishing Ltd

likely endogeneity, we instrument tariffs the same way we instrument productivities, byusing, in each industry-specific equation, the average tariff applied in that industry byall other countries.

4. Results

Three-stage least squares estimation results are shown in Table 2 for three-digit indus-tries. In all cases, the dependent variable is the share of industry j’s value-added in totalvalue-added in the manufacturing sector (as reported in the TPDB). Beverages, leather,wood and paper, and glass are omitted to save space. In spite of the data’s noisiness,results are largely as expected.

Own-TFP coefficients are all positive and significant at the 1% level. As for indus-try shares, according to the Rybczynski theorem, in a 2 ¥ 2 setting, at constant relativegood prices, an increase in a country’s endowment of a factor will cause an increasein output of the good which uses that factor intensively, and a decrease in the outputof the other good. In terms of (12), this means that a4 should be positive for capital-intensive industries and negative for labor-intensive ones—see Schaur et al. (2008)for a discussion of the n ¥ n case. By and large, signs (when significant) are plausible.Textile (under 3SLS) and apparel are negatively correlated with capital endowments.Chemicals (“industrial” and “other”), petroleum, steel and other metallurgical indus-tries, machinery, transport, and scientific equipment are all positively correlated withcapital endowments. One puzzling result is the positive correlation of food productswith capital endowments.10 Unlike Harrigan (1997), using the World Bank’s largedatabase, we find that most manufacturing industries are capital-intensive relative tothe omitted category (services and nontradables), which somehow accords betterwith intuition. Land endowments are negative or insignificant in most cases, possiblyreflecting a higher share of agriculture in the GDP for countries with high land/laborendowments. Human capital endowments, broken down by level of education, showno particular pattern except for postsecondary education which is correlated (signifi-cant coefficients with high values) with the shares of capital-intensive and high-techindustries (fabricated metal products, machinery, electrical machinery, transportequipment). The only counterintuitive case here—but it is very counterintuitive—isscientific equipment, for which we have no particular explanation. The simple corre-lation between the share of scientific equipment and levels of postsecondaryeducation is, however, positive.11

A more formal reality check on estimated Rybczynski effects consists of plotting ourRybczynski elasticities with respect to capital endowments against industry capitalintensities (averaged across countries in the sample).The resulting scatterplot is shownin Figure 1. The cloud of points is upward-sloping, as expected, with machinery, electricmachinery, and transport being outliers. The share of these three sectors in GDP thusrises faster with capital endowments than warranted by their own capital-intensity,suggesting that they are also intensive in something else that is correlated with capitalendowments—technology and education being obvious suspects.12

Tariffs are associated with higher shares for food, tobacco, wearing apparel, andfabricated metal products. They have negative and significant effects for machinery,transport equipment, and scientific instruments, possibly reflecting infant-industry pro-tection. Public infrastructure capital has a positive coefficient in the equations formachinery, electrical machinery, and transport equipment, reflecting the complemen-tarity of infrastructure with private capital.

ENDOWMENTS, SPECIALIZATION, AND POLICY 919

© 2010 Blackwell Publishing Ltd

Tabl

e2.

Sect

orSh

are

Reg

ress

ions

,Cou

ntry

–Tim

eF

ixed

Eff

ects

,3SL

SE

stim

ator

Food

prod

ucts

Text

iles

App

arel

Indu

str.

chem

ical

Oth

erch

emic

als

Pet

rol.

Non

met

allic

Iron

,st

eel

Non

ferr

ous

Fab.

met

alN

onel

ect.

mac

h.E

lect

.m

ach.

Tra

nspo

rteq

uip.

Scie

nt.

equi

p.

TF

P1.

140*

*0.

349*

*0.

172*

*0.

394*

*0.

534*

*1.

133*

*0.

361*

*0.

349*

*0.

133*

*0.

686*

*0.

958*

*1.

816*

*0.

642*

*0.

276*

*[0

.118

][0

.019

5][0

.009

87]

[0.0

201]

[0.0

242]

[0.1

14]

[0.0

158]

[0.0

223]

[0.0

122]

[0.0

369]

[0.0

540]

[0.0

812]

[0.0

322]

[0.0

162]

Tari

ff0.

0893

-0.0

0048

0.04

53**

0.02

090.

0075

40.

0502

0.01

49-0

.023

50.

0238

*-0

.009

4-0

.135

*-0

.011

9-0

.202

**-0

.018

[0.0

866]

[0.0

0788

][0

.005

22]

[0.0

262]

[0.0

177]

[0.1

99]

[0.0

134]

[0.0

226]

[0.0

118]

[0.0

353]

[0.0

619]

[0.1

07]

[0.0

379]

[0.0

297]

Pub

.cap

.-0

.033

2-0

.042

3*-0

.048

0**

-0.0

482*

*0.

0029

1-0

.470

*0.

0309

*-0

.007

50.

0179

-0.0

231

0.16

4**

0.11

20.

160*

*-0

.016

5[0

.045

3][0

.018

1][0

.011

8][0

.014

6][0

.016

4][0

.215

][0

.013

2][0

.019

6][0

.011

5][0

.026

2][0

.049

1][0

.067

5][0

.032

7][0

.019

1]C

apit

al0.

499*

*-0

.090

6**

-0.0

844*

*0.

327*

*0.

332*

*1.

173*

*0.

148*

*0.

137*

*0.

0977

**0.

298*

*0.

738*

*0.

972*

*0.

622*

*0.

152*

*[0

.083

4][0

.029

9][0

.019

7][0

.024

6][0

.026

0][0

.312

][0

.018

3][0

.030

3][0

.017

9][0

.038

2][0

.065

8][0

.092

8][0

.044

0][0

.025

5]L

and

-0.2

76-0

.144

-0.1

62**

-0.2

22**

0.13

1*-1

.17

-0.1

87**

-0.3

91**

0.12

5**

-0.0

046

-0.1

410.

312

-0.4

08**

0.00

35[0

.160

][0

.073

8][0

.049

1][0

.054

8][0

.058

6][0

.611

][0

.043

4][0

.059

0][0

.034

8][0

.086

5][0

.174

][0

.228

][0

.111

][0

.067

7]P

rim

ed.

-0.2

45*

0.12

0**

0.11

8**

0.00

747

-0.0

703*

-0.5

78*

0.01

490.

147*

*0.

0139

-0.0

238

-0.5

65**

-0.4

69**

0.11

7-0

.182

**[0

.116

][0

.030

8][0

.020

8][0

.030

8][0

.032

3][0

.282

][0

.024

8][0

.037

4][0

.022

8][0

.054

1][0

.105

][0

.139

][0

.070

3][0

.039

9]Se

c.ed

.-0

.174

-0.0

682

0.15

5**

0.02

07-0

.039

4-0

.446

0.05

20*

0.24

7**

-0.0

324

-0.0

282

-0.4

59**

-0.5

61**

0.07

48-0

.237

**[0

.143

][0

.039

8][0

.025

9][0

.032

9][0

.037

0][0

.301

][0

.026

5][0

.038

9][0

.022

5][0

.053

9][0

.104

][0

.141

][0

.070

9][0

.039

4]Po

st-s

ec.

0.42

6**

0.21

8**

0.19

4**

0.18

8**

-0.0

349

-1.9

73**

0.18

6**

0.22

1**

0.00

658

0.39

8**

0.62

8**

0.35

6**

0.34

2**

-0.0

987*

[0.1

03]

[0.0

377]

[0.0

251]

[0.0

368]

[0.0

395]

[0.6

37]

[0.0

289]

[0.0

531]

[0.0

309]

[0.0

524]

[0.1

02]

[0.1

37]

[0.0

679]

[0.0

388]

No.

ofob

s.58

260

860

870

770

727

558

163

463

460

660

660

660

660

6

Not

es:

Stan

dard

erro

rsin

brac

kets

.*Si

gnifi

cant

at10

%;*

*si

gnifi

cant

at5%

;***

sign

ifica

ntat

1%.

920 Olivier Cadot and Yuliya Shakurova

© 2010 Blackwell Publishing Ltd

5. Concluding Remarks

Results so far should be treated cautiously, if only because the countries in oursample do not necessarily belong to the same diversification cone.A given industry cantherefore have different factor intensities in different countries, implying that therelationship between endowments and intensities is fundamentally heterogeneous(Schott, 2004).With a sufficiently small number of cones, an alternative to our approachwould be to use switching-regime techniques. We leave this for future research.

This said, the results are broadly consistent with a Heckscher–Ohlin perspective ontrade, confirming that production patterns are significantly influenced by factor endow-ments and “technology” (the latter taken in a broad sense that can include managementquality, since it is measured as TFP).Tariffs, by contrast, appear to play but a minor rolein the allocation of resources, suggesting that reliance on demand-side policies tochannel resources to whatever industries are perceived by public authorities as “theones to be in” is unlikely to have any long-term effects—without even talking about theinefficiencies involved in the process.

The obvious policy implication is that policies aimed at influencing a country’spattern on specialization should be essentially supply-side policies to encourage theaccumulation of the right factors and horizontal policies to encourage the adoption oftechnological and managerial advances.

References

Belloc, Marianna,“International Specialization and Labor Unions: Evidence from OECD Coun-tries,” Review of International Economics 17 (2009):34–50.

Caves, Douglas W., Laurits R. Christensen, and W. Erwin Diewert, “Multilateral Comparisons ofOutput, Input, and Productivity using Superlative Index Numbers,” Economic Journal 92(1982):73–86.

Capital/Labor ratio

Para

met

er e

stim

ates

10

0.5

1

1.5

0

20

wearingsfootwear

scientific equipmentfurniture

textiles leather

rubberwood

potterytobaccoglass

plastic

food

machinery

electric machinery

petroleum ref.

transportfabricated metal prod.

printing oth. chem.ind. chem.

paperiron, steelnonferrous metalsnonmetallic

petroleum, coal products

beverages

10050 200 400

Figure 1. Estimated Rybczynski Elasticities against Factor Intensities: Capital /LaborRatio

ENDOWMENTS, SPECIALIZATION, AND POLICY 921

© 2010 Blackwell Publishing Ltd

Davis, Donald R. and David E. Weinstein, “An Account of Global Factor Trade,” AmericanEconomic Review 91 (2001):1423–53.

Dixit, Avinash K. and Victor Norman, Theory of International Trade, Cambridge: CambridgeUniversity Press (1980).

Goldberg, Pinelopi and Giovanni Maggi, “Protection for Sale: An Empirical Investigation,”American Economic Review 89 (1999):1135–55.

Grossman, Gene M. and Elhanan Helpman, “Protection for Sale,” American Economic Review84 (1994):833–50.

Harrigan, James, “Technology, Factor Supplies, and International Specialization: Estimating theNeoclassical Model,” American Economic Review 87 (1997):475–94.

Kohli, Ulrich R.,“A Gross National Product Function and the Derived Demand for Imports andSupply of Exports,” Canadian Journal of Economics 11 (1978):167–82.

Leamer, Edward E., Source of International Comparative Advantage: Theory and Evidence,Cambridge, MA: MIT Press (1984).

Nicita, Alessandro and Marcelo Olarreaga, “Trade, Production and Protection 1976–2004,”World Bank Economic Review 21 (2006):165–71.

OECD, Measuring Capital: OECD Manual, Paris: Organisation of Economic Co-operation andDevelopment (2001).

Rodrik, Dani, “Political Economy of Trade Policy,” in G. Grossman and K. Rogoff (eds), Hand-book of International Economics, Vol. III, Amsterdam: North-Holland (1995).

———, “Normalizing Industrial Policy,” manuscript, Harvard University (2007).Schaur, Georg, Chong Xiang, and Anya Savikhin, “Factor Uses and the Pattern of Specializa-

tion,” Review of International Economics 16 (2008):368–82.Schott, Peter, “Across-Product versus Within-Product Specialization in International Trade,”

Quarterly Journal of Economics 119 (2004):647–78.Trefler, Daniel, “The Case of the Missing Trade and Other Mysteries,” American Economic

Review 85 (1995):1029–46.

Notes

1. Capital stock variables are available from the Penn World Tables Mark 5.6 only up to 1992.Mark 6.1 of the PWT is incomplete precisely for those variables.2. The service life of machinery is variable and has been estimated in various ways, e.g. byobserving discard decisions over a sample to infer the hazard rate of a distribution of discardtimes assumed to be of the Weibull type. Other distributions (e.g. log-normal, Winfrey) are alsosometimes used. The average service life is then taken as the distribution’s mean. A simpleaverage of estimated service lives of machinery in the manufacturing sector from StatisticsCanada gives 13.6 years, against 14.4 years using the US’s Bureau of Economic Analysis data (seeAnnex 3 of OECD, 2001); revised data by the US Department of Labor gives on average 22 yearsof service life.3. Bureau of Labor Statistics, BLS Handbook of Methods, ch. 11, “Industry Productivity Meas-ures,” available at www.bls.gov.4. An alternative would be to estimate industry- and country-specific TFP values asresiduals from estimated production functions. This alternative route has the drawbackthat TFP equations would have only 29 observations each, as they would have to be estimatedon country- and industry-specific time series. We leave this alternative route to futureresearch.5. The nominal sample size of 2900 observations is, however, drastically reduced by missingvalues.6. Such correlation is implied by the symmetry restrictions and can also be the result of shockson the omitted service sector.A Breusch–Pagan test rejects independence of the equations at the1% level.

922 Olivier Cadot and Yuliya Shakurova

© 2010 Blackwell Publishing Ltd

7. All regressions are performed in levels. A panel unit-root test rejects the null of a unit root atthe 5% level. We also estimated a partial-adjustment version of (12) by GMM but with limitedsuccess, the sample size being too small for GMM to be efficient.8. Idiosyncratic measurement errors, like their persistent component.9. The common-agency model’s prediction that tariffs should rise with domestic production hasbeen a subject of controversy (see Rodrik, 1995), since it seems to imply that industries in whicha country has a comparative disadvantage (hence small production) are the least likely to beprotected, a rather counterintuitive prediction. However, Goldberg and Maggi (1999) showedthat that prediction is upheld by the data if politically organized sectors are treated distinctlyfrom nonorganized ones.10. Textiles can also be considered a dubious case, as upstream textile operations (weaving,dyeing, and cutting) are rather capital-intensive.11. We are grateful to a referee for attracting our attention to this.12. All three sector shares indeed have high coefficients on postsecondary education and own-TFP relative to other sectors.

ENDOWMENTS, SPECIALIZATION, AND POLICY 923

© 2010 Blackwell Publishing Ltd