Journal of International Economics 71 (2007) 389 409

www.elsevier.com/locate/econbase

Technology, endowments, and the factor content of

bilateral trade
Huiwen Lai a , Susan Chun Zhu b,
a
Economics Group, Wachovia Corporation, 301 South College Street, DC-7, Charlotte, NC 28288, USA
b
Department of Economics, Michigan State University, Marshall-Adams Hall, East Lansing, MI 48824, USA

Received 25 July 2005; received in revised form 12 July 2006; accepted 12 July 2006

Abstract

We derive testable restrictions relating the factor content of bilateral trade to bilateral differences in
technology and endowments. This departs from the HeckscherOhlinVanek theorem which compares the
factor content of net trade with factor abundance. We test the theoretical restrictions using a unique dataset
that covers 41 developed and developing countries with disparate endowments and technology. We find
evidence supporting the predictions. In addition: (1) The factor content predictions perform best for country
pairs with larger endowment differences, and (2) for trade between capital-abundant countries, Ricardian
international technology differences matter more than HeckscherOhlin factor endowment differences.
2006 Elsevier B.V. All rights reserved.

Keywords: HeckscherOhlin model; Factor content of bilateral trade; Ricardian technology differences

JEL classification: F1

Empirical research on the HeckscherOhlin model, the centerpiece of traditional trade

theories, has largely focused on its generalization, the HeckscherOhlinVanek (HOV) theorem.
The HOV theorem compares the factor content of net trade with factor abundance and predicts
that a capital-abundant country should export capital services. Empirically, the HOV theorem that
maintains strict assumptions of identical technology, factor price equalization and identical
homothetic preferences has been rejected repeatedly (e.g., Maskus, 1985; Brecher and Choudhri,

We are indebted to James Harrigan, Pravin Krishna, Dan Trefler, and two anonymous referees for very helpful
comments and suggestions. We also benefit from seminar presentations at the workshop on Recent Advances in
International Economics at the City University of Hong Kong, University of Windsor, and University of Kiel.
* Corresponding author. Tel.: +1 517 355 9647; fax: +1 517 432 1068.
E-mail address: [email protected] (S.C. Zhu).
0022-1996/$ - see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.jinteco.2006.07.002
390 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

1988; Staiger, 1988). In recent major contributions, Trefler (1993, 1995) and Davis and Weinstein
(2001) amend the traditional model by relaxing those restrictive assumptions and provide
evidence supporting the modified HOV theorem.1
The objective of this paper is to test the HeckscherOhlin model. Deviating from the traditional
approach, we do not examine the HOV prediction. Instead, we follow the empirical approach taken
by Choi and Krishna (2004) and test the theoretical predictions that relate the factor content of
bilateral trade to bilateral differences in technology and endowments. Our theory is built on the
work by Brecher and Choudhri (1982), Helpman (1984) and Staiger (1986). Following these
authors, we consider a trade equilibrium in which factor prices are allowed to differ across
countries. In the absence of factor price equalization one can predict the factor content of trade
from post-trade data without imposing any restrictions on preferences, and this can be done not
only for every country's net trade vector but also for bilateral trade patterns. Choi and Krishna
(2004) are the first to note the implications of those relaxed assumptions for HeckscherOhlin
testing. Using a sample of 8 OECD countries and assuming identical technology across countries,
they test the theoretical predictions in Helpman (1984) and find strong evidence supporting the
theory.
In this paper we extend Choi and Krishna (2004) in two important directions. First, we expand
the sample substantially to include 41 developed and developing countries with sufficiently
disparate factor abundance and productivity. The large sample variation in factor abundance is
essential for testing the theory that emphasizes endowment differences. Second, because countries
in our sample are at very different technology levels, we incorporate international technology
differences into our empirical analysis. In particular, we allow technology differences to be
country- and industry-specific, i.e., Ricardian technology differences. As shown by Harrigan
(1997), Ricardian technology differences are an important determinant of specialization. However,
in the literature on HOV testing, the effective factor content of trade is not well defined when there
are non-uniform technology differences across sectors. In contrast, we will show that it is
straightforward to incorporate Ricardian technology differences into our framework.2
We focus on the following empirical hypotheses. First, on average, a country imports the
content of those factors that are cheaper in its trading partner and exports the content of those
factors that are more expensive for its trading partner. It implies that in the trade equilibrium the
exporter's actual unit cost of production (i.e., using the actual exporter's factor prices and factor
usage) cannot be greater than the importer's hypothetical unit cost of production (i.e., using the
importer's factor prices and exporter's factor usage). In the presence of international technology
differences, both factor prices and factor usage should be expressed in productivity-equivalent
units. Inspired by Debaere (2003), we derive our second hypothesis which relates the factor
content of bilateral trade to relative factor abundance. It says that exports by capital-abundant
countries embody a higher capitallabor ratio than the exports by labor-abundant countries.
However, differing from Debaere (2003) who examines the relationship between relative factor
abundance and trade in factor services from the HOV perspective, we focus on the factor content of
bilateral trade rather than the factor content of net trade. Also note that the second hypothesis has

1
Trefler (1995) reports that the traditional HOV model is rejected in favor of a modification that allows for
international technology differences and Armington home bias in consumption. Davis and Weinstein (2001) show that
the HOV model, when modified to allow Hicks-neutral technology differences, factor price differences, the existence of
nontraded goods and trade costs, is consistent with data from ten OECD countries.
2
Extending Helpman (1984), Choi and Krishna (2004) derived theoretical restrictions when there exist Ricardian
technology differences. However, their empirical analysis assumes identical technology across countries. This
assumption is reasonable in their context because their sample includes countries with fairly similar technology.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 391

not been examined by Choi and Krishna (2004). Finally, it is worth pointing out that the two
hypotheses represent complimentary views about the relationship between the factor content of
bilateral trade and bilateral differences in technology and factor abundance. The first hypothesis
looks at the relationship in absolute terms, while the second one is in relative terms.
We find that the empirical hypotheses are confirmed by the vast majority of country pairs in
our sample. This result is robust to various alternative specifications. Compared to Choi and
Krishna (2004), we find even stronger evidence supporting the model, which confirms their
expectation: While it is unwise to speculate out of sample, this raises the expectation that the
theory would hold with even greater success outside of the OECD countries we are working
with, where factor price differences may be expected to be even larger (page 905). We also find
that the model performs better for country pairs with substantially different endowments. This
result accords well with the finding by Debaere (2003) that the HOV model works remarkably
well for country pairs with very different capitallabor ratios. Furthermore, we find that Ricardian
technology differences appear to play a bigger role than endowment differences in determining
trade between capital-abundant countries.
The paper is organized as follows. Section 1 lays out the theoretical framework. Section 2
derives the empirical hypotheses. Section 3 details how productivity differences are measured.
Empirical results are presented in Sections 45. Section 6 draws the conclusions.

1. Theory

In this section we derive theoretical restrictions on the factor content of bilateral trade, factor
prices, and Ricardian technology differences in the trade equilibrium. The basic setup follows
Staiger (1986). Final goods are produced using primary factors (e.g., labor, capital) and inter-
mediate inputs. Production exhibits constant returns to scale. All product markets are perfectly
competitive. There are no barriers to trade. As will become clear, the theory can be extended easily
to include nontraded intermediate inputs. And the assumption whether intermediate inputs are
freely traded or non-traded does not make any difference empirically.
Extending Staiger (1986), we allow technology to differ across countries and industries. For
simplicity, we assume that technology differences are factor-augmenting and Hicks-neutral.3 We
also assume that the requirement for intermediate inputs is identical across countries. As argued
by Davis et al. (1997), a car may be produced with varying degrees of substitution between capital
and labor across countries. Yet the same car may require a certain amount of steel, rubber and
other intermediate inputs. However, the assumption of identical requirement for intermediates is
not needed when intermediate goods are nontraded.
Let g index goods, and c index countries. Let gc be the production function for good g in
country c. Let dgc be the vector of factors needed directly to produce one unit of good g in
country c. By definition, gc (dgc) = 1. To simplify notation, in gc we suppress the requirement
for intermediate inputs because it is identical across countries. Let gc denote productivity of
industry g in country c. With factor-augmenting and Hicks-neutral technology differences, gc
(dgc) = g(gcdgc) for some internationally common production functions g (see Trefler, 1995).
Because gc(dgc) = 1, g(gcdgc) = 1. Thus, a larger gc indicates fewer inputs per unit of output or
greater productivity.

3
It is straightforward to extend the theory to allow for factor-bias technology differences. However, the lack of
measures of factor-bias technology differences prevents us from pursuing any empirical test for this interesting case.
392 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

Let Tgcc be the volume of gross exports of good g from country c to country c. When
intermediates are freely traded, all countries face the same price of intermediate inputs and have
identical requirements for intermediate inputs, implying that the cost of intermediate inputs must
be equal across countries. Let pgI be the cost of intermediates used to produce one unit of good g.
Let wc be the vector of factor prices in country c. With constant returns-to-scale technology, the
per-unit cost of producing g in country c is given by wcdgc + pgI. Perfect competition implies zero
profits on exports of g from country c to country c. Hence,
pg wc dgc pIg 1
where pg is the world price of good g.
For importing country c, unit profits on good g must be nonpositive:
pg Vwc Vdgc V pIg : 2

With constants returns-to-scale technology and no firm heterogeneity within industries, Eq. (2)
holds for all firms in industry g in country c. Zero profits obtain only when country c produces
good g. Combining Eqs. (1) and (2) we obtain
wc dgc Vwc Vdgc V: 3
pgI cancels because, being costlessly tradable, intermediates are not a source of comparative
advantage.4
However, direct factor requirements may differ across countries, i.e., dgc dgc. The gap in
factor usage arises from international differences in both technology and factor prices. With Hicks-
neutral differences in factor efficiency, if country c and country c had the same factor prices,
country c would need (gc / gc)dgc directly to produce one unit of good g. (Recall that a higher
gc indicates greater productivity.) For example, for industry g, if workers in country c are twice
as productive as workers in country c (i.e., gc / gc = 2), country c would need just half of the
workers that are required by country c to produce the same amount of output. However, if country
c and country c face different factor prices, although (gc / gc)dgc is a feasible way for country c
to produce one unit of g, it may not be optimal. Country c can reduce production cost via factor
substitution. The optimal bundle of factors is given by dgc. Therefore, cost minimization implies

wc Vdgc VVwc Vkgc =kgc Vdgc : 4

Combining inequalities (3)(4) yields wcdgc wc(gc / gc)dgc or
wc wc V
dgc V dgc : 5
kgc kgc V

Inequality (5) describes the restriction on factor prices, direct factor requirements and Ricardian
technology differences for industry g in the trade equilibrium. Because all the variables are
observable in the post-trade equilibrium, inequality (5) is empirically testable. It is worth pointing

4
If intermediate inputs are traded with costs, country c and country c may purchase intermediates at different prices.
Let pIgc and pIgc be the per-unit cost of intermediates paid by country c and country c, respectively. Then pIg in Eq. (1) is
replaced with pIgc while pIg in inequality (2) is replaced with pIgc. If pIgc is sufficiently smaller than pIgc, it is possible that
wcdgc wcdgc is compatible with Eq. (1) and inequality (2). In this case, the restriction in inequality (5) may be violated.
Therefore, admitting both trade barriers and intermediate inputs poses theoretical difficulties. This point has been noted
by Deardorff (1979), Brecher and Choudhri (1982) and Staiger (1986). We remind the reader that there is also no
substitution possibility between intermediate inputs and primary factors.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 393

out that the derivation does not require the productivity-adjusted factor prices to be equalized
internationally i.e., wc / gc = wc / gc.
To derive the national-level restrictions, we aggregate inequality (5) over g using Tgcc as the
weight. By giving a larger weight to industries with a higher trade volume, we can capture the
V
effects of international specialization. Defining Tgcc dgcTgcc (i.e., the vector of factors required
directly to produce Tgcc), industry aggregation yields


wc wc V
Rg kgc TgcV
cV V0: 6
kgc kgc V
By symmetry,


wc V wc
Rg kgc VTgcc
V
VV0: 7
kgc V kgc

As Brecher and Choudri (1982, footnote 10) point out, when bilateral trade is not balanced
between country c and country c, it is not appropriate to compare the factor content of total
exports. Thus, in the following analysis we will examine the factor content of one million dollars'
V
worth of exports from each country: We divide Tgcc by the value (in millions of dollars) of gross
V
exports from country c to country c, and Tgcc by the value (in millions of dollars) of gross
exports from country c to country c.
Several points are worth mentioning about inequalities (6) and (7). First, the derivations allow
for differences in factor prices. Factor price equalization is likely to break down when countries
differ in technology or have sufficiently different factor endowments.5 On the other hand, the
derivations do not require that factor prices must be different between country pairs. As is clear
from the weak inequalities, the theoretical restrictions still hold when there exist factor price
equalization and identical technology. However, we expect that inequalities (6)(7) would
perform better the larger are the within country-pair differences in endowments.
Second, with factor-augmenting technology differences, both factor prices and the factor
content of bilateral trade should be transformed into productivity-equivalent units. Specifically,
for industry g, the productivity-adjusted factor costs in country c are wc / gc and the productivity-
adjusted factor content of per-unit exports from country c to c is gcTgcc V
. This type of
productivity transformation has been used by Trefler (1993, 1995).
Third, the restrictions are imposed on the relationship between the post-trade bilateral factor-
price differentials and the direct factor content of bilateral trade. As stressed by Staiger (1986),
with free trade in intermediate goods, inequalities (6) and (7) should not be applied to the indirect
factor content of bilateral trade.

1.1. Nontraded intermediates

Now we briefly discuss the case where intermediate goods are not traded. In this case,
intermediate inputs must come from domestic sources. Let Igc be the vector of factor content of

5
With sufficiently disparate endowments, countries specialize in the particular subset of goods most suited to their mix
of endowments. Using industry-level data, Schott (2003) finds strong empirical evidence supporting this multiple-cone
equilibrium against the overly restrictive one size fits all equilibrium of the HeckscherOhlin model. Debaere and
Demiroglu (2003) provide further evidence for the existence of multiple-cone equilibrium in which developed OECD
countries and less developed ones belong to different diversification cones.
394 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

intermediate inputs used to produce one unit of good g. Hence, Eq. (1) becomes pg = wc(dgc + Igc) and
inequality (2) becomes pg wc(dgc + Igc). They imply that wc(dgc + Igc) wc(dgc + Igc), which is
analogous to inequality (3). The national-level restrictions in the presence of nontraded intermediates
are analogous to inequalities (6) and (7), except that both direct and indirect factor requirements
should be included to calculate the factor content of bilateral trade. Note that when intermediates are
nontraded, the assumption of identical requirement for intermediates is not needed.

2. Empirical hypotheses

The empirical hypotheses are derived based on inequalities (6)(7). Combining the two
inequalities yields


wc V wc
Rg kgc Tgc
V
cV kgc VTgcc Vz0:
V
8
kgc V kgc

Inequality (8) is the productivity-adjusted version of equation (16) in Helpman (1984). Both
factor prices and the factor content of bilateral trade are expressed in productivity-equivalent units to
take into account Ricardian technology differences. It says that on average country c is a net
importer from country c of the content of those factors that are cheaper in c than in c and vice versa.
To test the theory, we rewrite inequality (8) as

Rg wc V=kgc Vkgc Tgc

V
cV Rg wc =kgc kgc VTgcc V
V
hRicardian
c cV u z1 9
V wc Tc cV
V V
wc VTcc
P V P V
where TcVcV u g Tgc cV and Tcc Vu
V
g Tgcc V. cc
Ricardian
represents a ratio of the importer's hypothe-
tical unit cost of production (i.e., using the importer's factor prices and exporter's factor usage) to
the exporter's actual unit cost of production (i.e., using the actual exporter's factor prices and
factor usage). Because in the trade equilibrium the unit cost of production in the importing
country cannot be lower than that in the exporting country, the cost ratio represented by cc Ricardian

should not be less than 1. In addition, the magnitude of cc Ricardian

can reveal the extent to which the
theory is confirmed or violated in the sense that cc Ricardian
1 is a measure of excess costs that
would be incurred by not specializing.
If international technology
P V differences P are uniform across sectors (i.e., gc = c and gc = c
for all g), using TcVcV u g Tgc cV and T V
u
cc V
V
g Tgcc V, inequality (9) reduces to

wc V=kc Vkc TcVcV wc =kc kc VTcc

V
V
hUniform
c cV u z1: 10
V
wc VTcc V wc T V
c cV

If technology is identical across countries (i.e., c = c for all c and c), inequality (9) can be
further simplified as6

wc VTcVcV wc Tcc
V
V
hIdentical
c cV u z1: 11
wc VTcc V wc TcVcV
V

6
If wc = wc, Identical
cc = 1. Thus, the hypothesis in inequality (11) allows factor prices to be equal between country pairs.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 395

Choi and Krishna (2004) tested inequality (11) using data from 8 OECD countries and find
strong evidence supporting the theory.7 They also derived theoretical restrictions similar to
inequalities (9)(10). But they did not test these restrictions empirically. In contrast, inequality (9)
is the focal point of our empirical analysis.
It is worth noting that inequalities (9)(11) can be applied to many factors. However, due to
data constraints, in the empirical analysis we include capital K and aggregate labor L as primary
factors. As will be shown in Section 4.4, for the 8 OECD countries examined by Choi and Krishna
(2004), our results on cc
Identical
1 are very similar to their estimates which include capital, skilled
and unskilled labor as primary factors. Hence, we do not expect that our results would be sensitive
to factor aggregation.
Inspired by Debaere (2003), we also derive a relationship between the factor content of
bilateral trade and relative factor abundance. It is straightforward to derive



w LcV wcL K
Tcc K
V Tc cV
gc cV u L z0 12
wcKV wKc L
Tcc V Tc cV

where wcf ( f = K, L) is the price of factor f in country c and Tcc f

is the amount of factor f required
8
directly to produce gross exports from country c to c. Inequality (12) implies that if w cL /
K
w c N w cL / w cK , then T cc
K L
/ T cc T cc
K L
/ T cc . That is, if country c has a higher wagerental ratio
than country c, the capitallabor ratio embodied in country c's exports to c cannot be lower
than the capitallabor ratio embodied in country c's exports to c.
As noted by Debaere (2003), when considering relative factor abundance, Hicks-neutral and
factor augmenting productivity differences do not matter. Hence, the hypothesis cc 0 is robust
to Hicks-neutral technology differences. This property allows us to compute cc without first
imputing technology parameters. Thus, we can avoid the potential problem of measurement error
in the estimated technology parameters.

3. Measuring productivity gc

To test the hypothesis ccRicardian

1, we first compute the industry-level total factor
productivity (TFP) to measure Ricardian technology differences gc. Following Caves et al.
(1982), Harrigan (1997), Keller (2002) and Griffith et al. (2004), we calculate the multi-lateral
TFP index in order to make the measure internationally comparable. TFP calculations require

7
To be precise, inequality (11) slightly differs from Choi and Krishna's hypothesis. Inequality (10) in Choi and Krishna
(2004) does not normalize the factor content of trade by the value of gross exports.
8
For the case with two factors, K and L, inequality (5) can be rewritten as wcLdgc
L
+ wcKdgc
K
(wcL /gc)gcdgc
L
+ (wcK/gc)
gcdgc
K f
, where dgc ( f = K, L) is the amount of factor f required directly to produce one unit of good g in country c.
Multiplying both sides by Tgcc and using Tgcc f
dgc
f
Tgcc we obtain

wcL Tgc cV wc Tgc cV Vwc V=kgc Vkgc Tgc cV wc V=kgc Vkgc Tgc cV : 13
L K K L L K K

By symmetry,

wcLVTgcc V wc VTgcc VVwc =kgc kgc VTgcc V wc =kgc kgc VTgcc V: 14

L K K L L K K

Because both sides of inequalities (13) and (14) are positive, we multiply both sides of them. Then summing the result over g, we
obtain wcLTcc
L K K
wcTcc +wcKTccwcTcc wcLTcc
K L L L K K
wc Tcc +wKcTcc
K L L
wc Tcc. Dividing both sides by wcKTccwc Tcc gives (wcL/ wcK wcL / wcK)
K K K

L
(Tcc K
/ Tcc Tcc
L K
/ Tcc) 0. Since (Tcc
L K
/ Tcc Tcc
L K
/ Tcc K
) and (Tcc L
/ Tcc Tcc
K L
/ Tcc) have the same sign, inequality (12) follows.
396 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

real, internationally comparable data on outputs, inputs of primary factors, and intermediate
inputs. At the industry level, data exist only for capital and aggregate labor, not for intermediate
inputs. So we calculate the value-added TFP indexes. Therefore, the TFP estimates used in this
paper should be viewed as approximations to the true TFP measures.
Because our empirical analysis is cross sectional (for the year 1997), to simplify notation, we
will suppress year t as an argument unless it is necessary. Let Zgc denote value added in industry g
in country c, Lgc labor inputs, Kgc capital inputs, gc labor cost share, and N the number of coun-
P P P
tries in the sample. Define ln Zgc uRc ln Zgc =N , ln Lg uRc lnLgc =N , ln Kg uRc ln Kgc =N and
gc [gc + (cgc) / N] / 2. Then the multi-lateral TFP index for industry g in country c can be
calculated as
lnTFP ulnZ ln Z f a lnL ln L 1f
P P P
gc gc g gc gc g a lnK ln K :
gc gc g 15
This TFP index is superlative, meaning that it is exact for the flexible translog function. It is
also transitive: TFPgc / TFPgc = (TFPgc / TFPg,US) / (TFPgc / TFPg,US) where TFPg,US and TFPgc
denote the industry-level TFP for the United States and country c, respectively. Hence, the
choice of the base country is inconsequential. Without loss of generality, the United States is
chosen as the base country and thus the estimates of TFPgc are expressed relative to TFPg,US.
Following Keller (2002) and Griffith et al. (2004), we take into account cross-country
differences in labor and capital utilization. To adjust labor inputs, we multiply labor employment
by average annual hours worked per person in employment. To adjust capital inputs, we multiply
capital stock series by an estimate of capacity utilization. Capacity utilization is estimated as
follows. The actual usage of capital inputs may fluctuate over economic cycles: during down
turns capital may not be fully used while during booms it may be over used. We thus measure
gct; where Qgct is the actual output level in industry g in
capacity utilization for year t as Qgct / Q

country c in year t, and Qgct is predicted from the regression Qgct = gc + tc + gct where gc denote
the industrycountry dummy variables, tc is the country-specific time trend, and gct is the error
term. The regression is done for the period 19852000 when all the key variables are available
(see the Appendix for more details). It is worth pointing out that adjusting capital utilization does
not make any difference empirically.
To test the hypothesis cc
Uniform
1, we first calculate the national-level TFP to measure the
country-specific productivity differences c. To obtain the national-level TFP, we aggregate the
industry-level TFP using the industry's share of value added as the weight.
Because international technology differences play an essential role in our analysis, it is
important to know whether our results are robust to alternative productivity measures. We
f
construct an alternative measure based on direct factor requirements. Let dgc ( f = K, L) be the
amount of factor f needed directly to produce one unit of good g in country c. Modifying the
estimating equation (P4) in Davis and Weinstein (2001) we run the regression In d gc f
= gc + fg + f
ln(Kc / Lc) + fgc, where gc captures the country- and industry-specific productivity differences,
fg provides average estimates of the factor requirements across countries, Kc / Lc is the capital
labor ratio in country c, and fgc is the error term. The inclusion of ln(Kc / Lc) is motivated by the
fact that if factor prices are not equal and countries are in different diversification cones, input
coefficients may vary according to country capital abundance. The predicted factor requirements
in each country differ from the average factor requirements fg depending on the Ricardian
technology differences gc, and depending on the effect of factor abundance f ln(Kc / Lc). Based
on the estimated gc, gc can be imputed as exp( gc). Without loss of generality, the United
States is chosen as the numeraire, i.e., g,US 0. Hence, the productivity estimates for other
countries are expressed relative to the U.S. level.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 397

The summary statistics on the industry-level productivity estimates are presented in Appendix
Table A.1. Column 1 gives the geometric average of TFPgc with country shares of value added as the
weight. As expected, nearly all industries have average TFP estimates below 1 (the U.S. level).
Column 2 displays large cross-country variations in productivity for all industries. The table also
shows the correlation of the TFP indices with GDP per capita (column 3) and the alternative
productivity measure exp( gc) (column 4). Except for a few industries, the TFP indices are strongly
correlated with GPD per capita, confirming our expectation that high-income countries should have
higher productivity. The TFP estimates are also strongly and positively correlated with exp( gc) for
the majority of industries. See the Appendix for more details about data sources and measurement.

4. Exporter's actual production cost versus importer's hypothetical production cost

In this section we focus on the hypotheses in inequalities (9)(11). They mean that in the trade
equilibrium, the importer's hypothetical unit cost of production (i.e., using the importer's factor
prices and exporter's factor usage) cannot be lower than the exporter's actual unit cost of
production (i.e., using the actual exporter's factor prices and factor usage).

4.1. Labor-abundant and capital-abundant country groups

We divide the 41 countries in our sample into two groups based on the similarity of wage
rental ratios (wcL / wcK) and capitallabor endowment ratios (Kc / Lc).9 The labor-abundant group
has 19 countries and the capital-abundant group has 22 countries. On average, the wagerental
and capitallabor ratios for the capital-abundant group are three times as high as those for the
labor-abundant group. The large sample variation in endowments is essential for testing the
HeckscherOhlin model that emphasizes the role of endowment differences.

4.2. Sign tests

We first examine the hypothesis cc Ricardian

1 in inequality (9) where Ricardian technology
differences, gc and gc, are measured by the industry-level TFP. The results are listed in column
1 of Table 1. Rows 13 report the sign statistics which are defined as the percentage of country
pairs that satisfy cc
Ricardian
1. For 418 (= 22 19) pairs involving one capital-abundant country
and one labor-abundant country, cc Ricardian
1 is satisfied for 96% of the time (see row 1). The p-
value of the sign test is below 0.01 which means that the probability of having cc Ricardian
1 for
more than 96% of the time is less than 1%. Thus, the hypothesis cc Ricardian
1 performs
remarkably well for pairs of capital-abundant and labor-abundant countries. Rows 23 of column
1 show that ccRicardian
1 is satisfied for more than 80% of the pairs involving only capital-
abundant or labor-abundant countries.10 Again, the hypothesis cc Ricardian
1 cannot be rejected at
the 1% significance level.
Column 2 of Table 1 gives the results for cc Uniform
1 where international technology dif-
ferences are assumed to be uniform across sectors. The technology parameters, c and c, are

9
On a more technical level, our country clustering is based on the Euclidean distances computed from Kc / Lc and wcL /
wcK. The cluster centers are based on least-squares estimation and are the means of the observations assigned to each
cluster when the algorithm is run to complete convergence. See Anderberg (1973) for more details.
10
Because Ricardian
cc = Ricardian
cc , the number of non-duplicate pairs involving only capital-abundant countries is 231
(=22 21 / 2). Similarly, the number of non-duplicate pairs of only labor-abundant countries is 171 (=19 18 / 2).
398 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

Table 1
Exporter's actual production cost versus importer's hypothetical production cost
Ricardian
cc 1 Uniform
cc 1 Identical
cc 1 Obs
(1) (2) (3)
Sign test
1. Pairs of capital-abundant and labor-abundant countries 0.96 0.96 0.99 418
(b0.01) (b0.01) (b0.01)
2. Pairs of capital-abundant countries 0.82 0.68 0.73 231
(b0.01) (b0.01) (b0.01)
3. Pairs of labor-abundant countries 0.85 0.85 0.81 171
(b0.01) (b0.01) (b0.01)

Probit regressions
4. |Kc / Lc Kc / Lc| 33.94 32.37 33.28 820
Standard error 4.65 3.71 4.27
Log-likelihood 240.27 298.28 270.18
P
5. f K;L wcf wcfV 2 V f 2 0.79 0.71 3.41 820
Standard error 0.17 0.14 0.47
Log-likelihood 257.20 318.70 237.71

OLS regressions
6. |Kc / Lc Kc / Lc| 13.28 13.75 19.19 820
Standard error 0.82 0.81 1.02
R2 0.28 0.30 0.37
P
7. f K;L wcf wcfV2 V f 2 0.39 0.39 0.80 820
Standard error 0.03 0.03 0.03
R2 0.21 0.20 0.57
Notes: This table examines the hypotheses in inequalities (9)(11) that compare the importer's hypothetical unit production
cost with the exporter's actual unit production cost. The sample includes 22 capital-abundant countries and 19 labor-abundant
countries. In the probit regressions the dependent variable is a dummy variable that equals one if the hypothesis is satisfied and
zero otherwise. In the OLS regressions the dependent variable is log Ricardian
cc , log Uniform
cc and log Identical
cc , respectively, in
columns 13. P In rows 4 and 6, |Kc /Lc Kc /Lc| is the difference in capitallabor ratios between country c and country c. In
rows 5 and 7, f K;L wcf wcfV2 V f 2 is an alternative measure of country-pair differences in endowments, where wcf is the
price of factor f in country c, and V f is the sample average of actor endowment (capital K or labor L) relative to GDP. The
standard errors are robust to heteroskedasticity. In parentheses are p-values of the sign test.

measured by the national-level TFP. Column 3 reports the results for cc Identical
1 under the
assumption of identical technology across countries. Columns 23 show that for all country pairs,
both hypotheses cannot be rejected at the 1% level. For pairs involving capital-abundant and labor
abundant countries, both hypotheses are satisfied at a higher rate than for the pairs of only capital-
abundant or labor-abundant countries. Therefore, the sign tests show strong support for the theory.
The model performs even better for pairs of countries with more disparate endowment differences.
Interestingly, we also find that for the pairs that consist of only capital-abundant countries, the
sign statistic for cc
Ricardian
1 is 82% which is higher than the success rate for both cc Uniform
1
and cc
Identical
1. The result suggests that allowing for Ricardian technology differences may help
one explain trade in factor services between capital-abundant countries.

4.3. Model performance and endowment differences

Now we present more direct evidence supporting the view that the HeckscherOhlin model
performs better for country pairs with larger endowment differences (Evenett and Keller, 2002).
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 399

We directly measure endowment differences between country pairs in order to avoid the potential
problem of arbitrary country grouping. Endowment differences are measured in two ways. The
first measure is |Kc / Lc Kc / Lc| which is the difference in capitallabor ratios between country c
and country c. It has been used by P Debaere (2003). Following Choi and Krishna (2004), we
construct the second measure as f K;L wcf wcf V2 V f 2 , where V f is the sample average of
factor endowment f (capital K or labor L) relative to GDP. Using V f as the weight handles the
problem of different measurement units of wage (wcL) and rental (wcK). The correlation between the
two measures of endowment differences is 0.59 (p b 0.01).
We run probit regressions on endowment differences. The dependent variable is a dummy
variable that equals one if the hypothesis is satisfied, and zero otherwise. The regression results
are given in rows 45 of Table 1.11 In all cases the coefficients on endowment differences are
positive and statistically significant, which indicates that the hypotheses are more likely to be
satisfied for country pairs with bigger endowment differences. This result accords well with
Debaere (2003) who finds that the HOV model performs well for countries with very different
capitallabor ratios.
Because the 's represent a ratio of the importer's hypothetical unit cost to the exporter's
actual unit cost, the magnitude of the 's is meaningful. We thus regress the logarithm of on
endowment differences using the ordinary least squares (OLS) estimation. Rows 67 show that in
all cases the estimated coefficients on endowment differences are significantly positive,
suggesting that the cost ratio increases in bilateral endowment differences. It further implies that
the probability of 1 is likely to be higher for country pairs with more disparate endowments.
Further evidence is given in Figs. 1 and 2. Panel A of Fig. 1 displays the percentage of country
pairs that satisfy the hypothesis cc Ricardian
1 by percentile of endowment differences measured
by |Kc / Lc Kc / Lc|. For country pairs with very small endowment differences, e.g., below the
10th percentile of the distribution, cc Ricardian
1 is satisfied for less than 70% of the time. In
contrast, for country pairs with very large endowment differences, e.g., above the 90th percentile,
cc
Ricardian
1 is confirmed for 100% of the time. Panel B reveals a similar pattern for the
hypothesis ccIdentical
1. Country pairs in the upper end of the distribution satisfy the hypothesis at
much higher rates than country pairs in the lower end, indicating that the model performs better
for country pairs with larger endowment differences. We obtain similar P patterns for cc
Uniform
1.
Results are also similar when endowment differences are measured by f K; L wc wcf V2 V f 2. f

To save space, we do not show those plots in the paper.

Fig. 2 plots log cc Ricardian
and log cc
Identical
against |Kc / Lc Kc / Lc|. Evidently, the OLS esti-
mates shown in Table 1 are not driven by outliers. In addition, the plots reveal that for country pairs
with larger endowment differences, log cc Identical
are more likely to be bigger than log cc Ricardian
. It
implies that ignoring Ricardian technology differences is likely to overstate the cost ratio and
thus model performance for country pairs with more disparate endowments. This is consistent
with the sign statistics as shown in row 1 of Table 1: cc Identical
1 has a higher sign statistic than
cc
Ricardian
1 for the pairs involving capital-abundant and labor-abundant countries. In the
following robustness analysis we will focus on cc Ricardian
1.

4.4. Robustness

In the previous analysis we have assumed that intermediate inputs are traded without any costs.
Now we allow intermediates to be nontraded. Under this assumption, both direct and indirect

11
Because cc = cc, with 41 countries in the sample, the number of non-duplicate country pairs is 820 (=41 40 / 2).
400 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

Fig. 1. Performance of Ricardian

cc 1 and Identical
cc 1 by endowment differences.

factor requirements should be used to compute the factor content of bilateral trade (see Section 1).
The results for cc
Ricardian
1 with nontraded intermediates are presented in column 2 of Table 2. For
comparison, column 1 of Table 2 lists the baseline result carried over from column 1 of Table 1. For
pairs of capital-abundant and labor-abundant countries, the sign statistic is 97% which is almost
identical to the baseline result. For pairs of only capital-abundant or labor-abundant countries, the
sign statistics are slightly higher than the baseline case. Both the probit and OLS regressions yield
significantly positive coefficients on endowment differences, suggesting that the model performs
better for country pairs with bigger endowment differences. Therefore, our results are not sensitive
to the assumption whether intermediates inputs are freely traded or nontraded.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 401

Fig. 2. The magnitude of Ricardian

cc and Identical
cc by endowment differences.

Our second robustness check is on the rental rate of capital. The inclusion of many developing
countries in our sample prevents us from constructing the return to capital in the same way as
done by Choi and Krishna (2004). As described in the Appendix, our data include information
on wcK / wUS
K
rather than wcK . To overcome this obstacle, in the previous analysis we have set
K
wUS = 16.5% based on the Capital II measure of U.S. rental rate in Choi and Krishna (2004). Then
wcK is imputed as (wcK / wUS
K
) 16.5%. In order to examine whether our results are sensitive to this
K
imputation, we now choose wUS = 8% based on Choi and Krishna's Capital I measure of U.S.
rental rate and impute wc as (wcK / wUS
K K
) 8%. The results are given in column 3 of Table 2. It is
clear that our results are little changed.
Since international technology differences play an essential role in our analysis, it is important
to see whether our results are robust to alternative measures of productivity differences. In
402 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

Table 2
Robustness of Ricardian
cc 1
Baseline Nontraded wKUS = 8% gc = exp gc = lagged Ricardian
gcc
intermediates ( gc) TFP 1
(1) (2) (3) (4) (5) (6)
Sign test
1. Pairs of capital-abundant and labor- 0.96 0.97 0.95 0.87 0.95 0.88
abundant countries (b0.01) (b0.01) (b0.01) (b0.01) (b0.01) (b0.01)
2. Pairs of capital-abundant countries 0.82 0.84 0.83 0.71 0.81 0.70
(b0.01) (b0.01) (b0.01) (b0.01) (b0.01) (b0.01)
3. Pairs of labor-abundant countries 0.85 0.90 0.82 0.71 0.85 0.80
(b0.01) (b0.01) (b0.01) (b0.01) (b0.01) (b0.01)

Probit regressions
4. |Kc / Lc Kc / Lc| 33.94 25.10 34.00 22.74 32.38 17.13
Standard error 4.65 4.71 4.56 3.04 4.63 0.85
Log-likelihood
P 240.27 209.57 254.65 395.00 247.75 9152.30
5. f K;L wcf wcf V2 V f 2 0.79 1.32 0.70 0.61 0.79 0.41
Standard error 0.17 0.29 0.15 0.11 0.16 0.03
Log-likelihood 257.20 203.17 275.84 405.64 263.28 9345.49

OLS regressions
6. |Kc / Lc Kc / Lc| 13.28 12.15 15.39 6.14 13.24 12.42
Standard error 0.82 0.82 0.93 0.37 0.82 0.31
2
R
P 0.28 0.25 0.30 0.26 0.29 0.20
7. f K;L wcf wcf V2 V f 2 0.39 0.36 0.45 0.19 0.42 0.35
Standard error 0.03 0.03 0.03 0.01 0.03 0.01
R2 0.21 0.18 0.22 0.19 0.26 0.13
Notes: The sample includes 22 capital-abundant countries and 19 labor-abundant countries. Column 1 reports the baseline
results for Ricardian
cc > 1 carried over from column 1 of Table 1, in which Ricardian
cc is computed using the 1997 industry-level
TFP and under the assumption of freely traded intermediates and wUSK = 16.5%. In columns 45, Ricardian technology
differences gc are measured by exp( gc) and the 1996 industry-level TFP, respectively. Column 6 examines the industry-
level hypothesis Ricardian
gcc N 1 in inequality (16) for 24 ISIC industries. In the probit regressions the dependent variable is a
dummy variable that equals one if the hypothesis is satisfied and zero otherwise. In the OLS regressions the dependent
variable is log Ricardian
cc in columns 15, and log Ricardian
gcc in column 6. See the notes to Table 1 for a description of other
variables. The standard errors are robust to heteroskedasticity. In parentheses are p-values of the sign test.

column 4 of Table 2 we replace the industry-level TFP measure with exp( gc), which is
estimated based on direct factor requirements (see Section 3). A comparison of columns 1 and 4
reveals that our results are somewhat sensitive to how productivity is measured. However, the sign
tests in rows 13 show that the hypothesis cc
Ricardian
1 cannot be rejected for all country pairs at
the 1% level. The regression results in rows 47 show that for all specifications coefficients on
endowment differences are significantly positive. Thus, our main conclusions hold.
For a further robustness check, we use lagged industry-level TFP estimates rather than
contemporaneous ones (for the year 1997) used in the previous analysis. We find almost no change
in our results. For example, column 5 of Table 2 displays the results when the 1996 industry-level
TFP estimates are used. We also experimented with the industry-level TFP estimates that do not
make any adjustments to labor and capital inputs. Our results are little changed. To save space, we
do not report them in the table.
So far we have focused on the aggregate-level results. An alternative way to test the theory is to
examine the industry-level restriction on factor requirements, factor prices and productivity
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 403

differences. Based on inequality (5), it is straightforward to derive the industry-level counterpart

of inequality (9) as
wc V=kgc Vkgc dgc wc =kgc kgc Vdgc V
hRicardian
gc cV u z1: 16
wc Vdgc V wc dgc
Differing from inequality (9), however, inequality (16) is not directly related to international
specialization. The industry-level results are shown in column 6 of Table 2. Because larger industries
tend to be measured more accurately than smaller industries, both the sign tests and regressions are
weighted by the industry shares of output in order to give a bigger weight to larger industries. It is
evident that the industry-level results are consistent with the aggregate-level estimates.
Finally, we examine model performance for a sample of 8 OECD countries examined by Choi
and Krishna (2004).12 Compared to our sample, the variation among the 8 countries in
endowments, factor prices and productivity levels is relatively small. On the other hand, because
we expand the sample substantially to include many developing countries, our data include
capital and aggregate labor. In contrast, the analysis of Choi and Krishna (2004) deals with capital
and disaggregated labor (skilled and unskilled). We interpret the labor cost in our empirical
analysis as an average cost of skilled and unskilled labor. Choi and Krishna (2004) tested the
hypothesis ccIdentical
1. Although our data sources and measurements differ from theirs, our
estimates of cc
Identical
are fairly similar to those reported in Choi and Krishna (2004). In particular,
for the majority of country pairs that exclude Korea, the estimated cc Identical
are slightly above 1. In
contrast, for country pairs involving Korea, the estimated cc Identical
are well above 1. The
hypothesis cc
Identical
1 holds for 86% of 28 country pairs, which is comparable to the 79%86%
in Choi and Krishna (2004, Tables 45).13 When we allow for Ricardian technology differences,
we find that ccRicardian
1 holds for 82% of the time (p b 0.01).
To summarize, we find that the hypothesis cc Ricardian
1 cannot be rejected for all country
pairs. This result is robust to alternative assumptions of the model and alternative measures of
productivity differences. We also find that the model gains the strongest support from country
pairs that differ substantially in endowments. Thus, our results confirm the expectation of Choi
and Krishna (2004) that the theory would hold with even greater success outside of the OECD
countries where factor price differences are even larger.

5. Relative factor abundance and the factor content of bilateral trade

Now we turn to the hypothesis cc 0 in inequality (12). It relates the factor content of bilateral
trade to bilateral differences in relative factor abundance. The hypothesis implies that if country c
has a higher wagerental ratio than country c, country c's exports to c embody a higher capital
labor ratio than country c's exports to c. Thus, cc Ricardian
1, cc 0 is not affected by Ricardian
technology differences which assume that productivity differs across industries but not between
factors within each industry.
Column 1 of Table 3 lists the results for cc 0. The sign statistics are summarized in rows 13.
Notably, for pairs of only capital-abundant countries, cc 0 is confirmed for just 54% of the time, a
success rate that is not much better than a coin toss (Trefler, 1995). The p-value is 0.15. So cc 0 can
be rejected for pairs involving only capital-abundant countries at the 5% level. Because the hypothesis

12
The 8 countries are Canada, Denmark, France, Germany, Korea, the Netherlands, the United Kingdom and the United States.
13
We compare our results with Tables 4 and 5 in Choi and Krishna (2004) because our calculations use their Capital II
measure of U.S. rental rate.
404 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

cc 0 relates the factor content of bilateral trade to endowment ratio differences, the result indicates
that trade between capital-abundant countries cannot be fully explained by endowment differences.
On the other hand, as mentioned above, allowing for Ricardian technology differences improves
model performance for the pairs of only capital-abundant countries (see row 2 of Table 1). Thus we
conclude that Ricardian technology differences might be a more important factor than endowment
differences in determining the trade patterns between capital-abundant countries.
For pairs involving capital-abundant and labor-abundant countries, cc 0 is satisfied
for 89% of the time (see row 1). The p-value of the sign test is below 0.01. Thus, the hypothesis
cc 0 performs remarkably well for country pairs with sufficient disparate endowments. This is
consistent with the result for cc Ricardian
1.
The industry-level counterpart of cc 0 can be expressed as

L  K !
K
wc V wcL dgc V dgc
ggc cV u K K L
L z0: 17
wc V wc dgc V dgc

f
Recall that dgc ( f = K, L) gives the amount of factor f needed directly to produce one unit of
good g in country c. Inequality (17) implies that capital-abundant countries use more capital-
intensive techniques than labor-abundant countries in each industry g. The results for cc 0 are
given in column 2 of Table 3. The sign tests are weighted by the industry shares of output in order
to give a bigger weight to larger industries. It is clear that the industry-level results are fairly
similar to those at the aggregate level. In particular, the hypothesis gcc 0 is satisfied for just
half of the pairs involving only capital-abundant countries (see row 2). In contrast, gcc 0 is
confirmed at a much higher rate for pairs of capital-abundant and labor-abundant countries.

Table 3
Relative factor abundance and the factor content of bilateral trade
cc 0 gcc 0
(1) Obs (2) Obs
Sign test
1. Pairs of capital-abundant and labor abundant countries 0.89 418 0.77 10,032
(b0.01) (b0.01)
2. Pairs of capital-abundant countries 0.54 231 0.50 5544
(0.15) (0.35)
3. Pairs of labor-abundant countries 0.63 171 0.56 4104
(b0.01) (0.01)

Probit regressions
4. |Kc / Lc Kc/ Lc| 35.59 820 20.86 19,680
Standard error 3.22 0.75
Log-likelihood
P 410.21 12,007.62
5. f K; L wcf wcf V2 V f 2 0.73 820 0.05 19,680
Standard error 0.12 0.03
Log-likelihood 448.38 12,370.22
Notes: This table examines the hypotheses in inequalities (12) and (17) that relate the factor content of bilateral trade to
bilateral differences in relative factor abundance. The sample includes 22 capital-abundant and 19 labor-abundant
countries. Column 1 reports the aggregate-level results for the 41 countries. Column 2 gives the industry-level results
which involve 24 ISIC industries. The dependent variable in the probit regressions is a dummy variable that equals one if
the hypothesis is satisfied and zero otherwise. See the notes to Table 1 for a description of other variables. The standard
errors are robust to heteroskedasticity. In parentheses p-values of the sign test.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 405

In order to illustrate how model performance varies with bilateral endowment differences, rows 45
of Table 3 present the results from probit regressions on endowment differences.14 The dependent
variable is a dummy variable that equals one if the hypothesis is satisfied P and zero otherwise.
Endowment differences are measured by |Kc /Lc Kc /Lc| in row 4 and by f K; L wcf wcf V2 V f 2 in
row 5. For all specifications the coefficients on endowment differences are positive and statistically
significant, indicating that the hypotheses are more likely to be satisfied for country pairs with bigger
endowment differences. Thus, our results are consistent with the finding by Debaere (2003) that the
HOV model performs well for countries with very different capitallabor ratios.

6. Conclusions

In this paper we incorporated Ricardian technology differences into Staiger (1986) and derived
testable restrictions that relate the factor content of bilateral trade to bilateral differences in
technology and endowments. This departs from the traditional HOV theorem that compares the
factor content of net trade with factor abundance. Following Helpman (1984) and Staiger (1986),
we considered a trade equilibrium in which factor prices are allowed to differ across countries. In
the absence of factor price equalization one can predict the factor content of trade from post-trade
data without restricting preferences, and this can be done for bilateral trade patterns. So less
restrictive assumptions are required than the traditional HOV model.
We studied the following empirical hypotheses. First, the exporter's actual unit cost of
production cannot be greater than the importer's hypothetical unit cost of production. This
hypothesis was first examined by Choi and Krishna (2004) under the assumption of identical
technology across countries. Our second hypothesis relates the factor content of trade to relative
factor abundance. It says that exports by capital-abundant countries embody a higher capital
labor ratio than the exports by labor-abundant countries. The two hypotheses are complementary.
The first one looks at the relationship between factor abundance and the factor content of bilateral
trade in absolute terms, while the second one is in relative terms.
In contrast to previous studies that have been confined to developed countries, our empirical
analysis exploited a unique dataset that covers 41 countries with sufficiently disparate endowments
and technology. We find that the empirical hypotheses are confirmed by the majority of country
pairs in our sample. This result is robust to various alternative specifications. Thus, our empirical
study confirms the finding in Choi and Krishna (2004). In addition, the HeckscherOhlin model
performs even better for country pairs with substantially different endowments. This is consistent
with Debaere (2003) who finds that the HOV model performs remarkably well for countries with
very different capitallabor ratios. Finally, we find some evidence suggesting that for trade
between capital-abundant countries, Ricardo matters more than HeckscherOhlin.
Appendix A. Data sources and measurement

All data pertain to 1997 unless otherwise stated.

A.1. Factor Endowment and Factor Price Data

Capital. Capital stock is constructed as follows. We use the latest capital stock data from the
Penn World Table 5.6 (PWT 5.6) and update the data to 1997 by applying Leamer's (1984) double
14
Although the sign of cc and gcc is economically meaningful, the magnitude of them is not. Thus we do not regress
their values on endowment differences as done for the 's.
406 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

declining balance method to investment. The real gross domestic investment series come from the
Penn World Table 6.1 (PWT 6.1). Let Kc,t0 be capital stock in country c in year t0 (the latest year
available) from PWT 5.6 (in 1985 international prices).15 Let kct be the investment series for year t
from PWT 6.1 (in 1996 international prices).16 Let PIt0PWT5.6 and PIt0PWT6.1 be the price level of
investment for year t0 from PWT 5.6 and PWT 6.1, respectively. Assuming a typical asset life
of 15 years, the depreciation rate is = 13.3%. Then country c's capital stock Kc at the
beginning of 1997 (in 1996 international prices) is defined as

X
1996
Kc u1r1996t0 Kc;t0 PItPWT6:1
0
=PItPWT5:6
0
1r1996t kct :
tt0 1

Data on wcK are the PWT 6.1 investment index which is expressed relative to U.S. rental rate
K
wUS . Under the assumption that interest rates and depreciation rates are equalized across
countries, wcK / wUS
K
is equal to the ratio of user costs of capital in country c relative to the United
States.
Labor. Data on industry labor employment Lgc are from the OECD STAN database for
OECD countries, the UNIDO data base for manufacturing in non-OECD countries andPfrom the
ILO for non-manufacturing in non-OECD countries. The endowment of labor, Lc u g Lgc , is
scaled so that it sums to the PWT 6.1 workforce totals in 1997. Data on wcL are calculated as total
payroll divided by the total labor employment.

A.2. Technology and International Trade Data

Data on inputoutput tables are from GTAP (version 5). Direct usage of capital by industry is
generated by assuming that industry capital stocks are proportional to industry payments to
capital. This will be the case in steady state under the assumption of constant depreciation rates.
Data on capital payments are from the GTAP inputoutput accounts. Direct usage of labor by
industry is calculated as industry labor employment divided by industry output. Industry output is
from GTAP. In order to match the classification of industries in data on factor usage with those in
inputoutput tables, we aggregated industries up to 24 ISIC (rev. 2) industries.
Data on international trade flows are from GTAP (version 5).

A.3. TFP Measures

Calculations of the TFP indices in Eq. (15) require real, internationally comparable data on
value added, labor and capital inputs.
OECD Countries. Data come from the OECD STAN database. Value added are deflated using
the value added deflators available from the STAN database. To obtain the labor input measure, the
number of employees is adjusted by the average annual hours worked per person in employment from
the STAN. Capital stock series are constructed using a 15-year double declining balance method
applied to deflated gross fixed capital formation. The country- and industry-specific investment
deflators are derived from the STAN database. The capital stock measure is also adjusted for cyclical

15
Kc,t0 KAPWc,t0 RGDPCHc,t0 POPc,t0 /RGDPWc,t0 where KAPWc is country c's capital per worker, RGDPWc is c's real GDP
per worker using the chain index, RGDPCHc is c's real GDP per capita using the chain index, and POPc is c's population.
16
Kct RGDPLct KIct POPct where RGDPLc is country c's real GDP per capita using the Laspeyre index, KIc is c's
share of real gross domestic investment in RGDPLc and POPc is c's population.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 407

differences in capacity utilization. Since the STAN data only go back to 1970, with a typical asset life
of 15 years, 1985 is the earliest year for which the capital stock measure is available. Thus, capacity
utilization is estimated based on industry gross output over the period 19852000 for which all the key
variables are available. Industry gross output come from the STAN database.
Non-OECD Countries. Data are taken from the UNIDO database. Value added are deflated using
the GDP price index from the PWT 6.1. The average annual hours worked per person are derived by
multiplying the ILO average weekly hours worked per person by 52 weeks. Capital stock series are
constructed using a 15-year double declining balance method applied to deflated gross fixed capital
formation. The investment deflators are taken from the PWT 6.1. Since the UNIDO data only cover the
manufacturing industries, we impute the non-manufacturing TFP as the average manufacturing TFP.

A.4. Countries

The 41 countries are divided into two groups according to the capitallabor ratio and wage
rental ratio. The capital-abundant group has 22 countries including Australia, Austria, Belgium,
Canada, Denmark, Finland, France, Germany, Hong Kong, Ireland, Italy, Japan, the Netherlands,
New Zealand, Singapore, Korea, Spain, Sweden, Switzerland, Taiwan, the United Kingdom, and
the United States. The labor-abundant group has 19 countries including Argentina, Brazil, Chile,
China, Colombia, Greece, Hungary, Indonesia, Malaysia, Mexico, Peru, the Philippines, Poland,
Portugal, Sri Lanka, Thailand, Turkey, Uruguay, and Venezuela.

A.5. Industries

The 24 ISIC industries are: 110130 (Agriculture, Hunting, Forestry and fishing); 200 (Mining
and quarrying); 311 + 312 (Food); 313 + 314 (Beverages, Tobacco); 321 (Textiles); 322 (Apparel);
323 + 324 (Leather products, Footwear); 331 + 332 (Wood products, Furniture); 341 + 342 (Paper
products, Printing and publishing); 353 + 354 (Petroleum refineries, Misc. petro and coal pro-
ducts); 351 + 352 + 355 + 356 (Industrial chemicals, Other chemicals, Rubber products, Plastic
products); 361 + 362 + 369 (Pottery, Glass, Other non-metallic mineral products); 371 (Iron
and steel); 372 (Non-ferrous metals); 381 (Fabricated metal products); 384 (Transport equipment);
382 + 383 + 385 (Non-electrical machinery, Electric machinery, Instruments); 390 (Misc.
manufacturing); 400 (Electricity, gas, and water); 500 (Construction); 600 (Wholesale and retail
trade and restaurants and hotels); 700 (Transport, storage and communication); 800 (Financing,
insurance, real estate and business services); and 900 (Community, social and personal services).

Table A1
Summary Statistics on the Industry-level TFPgc
Industry description Average Std dev. Corr Corr(TFPgc, exp
(TFPgc, GDPc) ( gc))
(1) (2) (3) (4)
Agriculture, Hunting, Forestry and fishing 0.83 0.27 0.21 0.21
Mining and quarrying 1.01 0.41 0.33 0.31
Food 0.78 0.29 0.42 0.32
Beverages, Tobacco 0.64 0.28 0.11 0.18
Textiles 0.77 0.30 0.73 0.66
Apparel 0.73 0.31 0.76 0.63
(continued on next page)
408 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409

Appendix
Table A (continued
A1 (continued ) )
Industry description Average Std dev. Corr Corr(TFPgc, exp
(TFPgc, GDPc) ( gc))
(1) (2) (3) (4)
Leather products, Footwear 0.68 0.31 0.70 0.44
Wood products, Furniture 0.77 0.26 0.74 0.65
Paper products, Printing and publishing 0.82 0.24 0.57 0.44
Industrial chemicals, Other chemicals, Rubber products, 0.73 0.22 0.53 0.43
Plastic products
Petroleum refineries, Misc. petro and coal products 0.93 0.35 0.32 0.21
Pottery, Glass, Other non-metallic mineral products 0.74 0.28 0.53 0.74
Iron and steel 0.72 0.29 0.41 0.41
Non-ferrous metals 0.82 0.38 0.18 0.21
Fabricated metal products 0.74 0.24 0.64 0.62
Non-electrical machinery, Electric machinery, Instruments 0.93 0.35 0.65 0.64
Transport equipment 0.75 0.23 0.52 0.56
Misc. manufacturing 0.72 0.24 0.75 0.42
Electricity, gas, and water 0.74 0.22 0.48 0.43
Construction 0.72 0.24 0.62 0.67
Wholesale and retail trade and restaurants and hotels 0.82 0.21 0.54 0.55
Transport, storage and communication 0.74 0.20 0.55 0.55
Financing, insurance, real estate and business services 0.61 0.19 0.44 0.35
Community, social and personal services 0.87 0.27 0.62 0.56
Notes : The TFPgc are relative to the U.S. industry-level TFP. Column 1 reports the geometric average of TFPgc across
countries with country shares of value added as the weight. Column 2 gives the standard deviations of TFPgc across
countries. Columns 34 report the correlation of TFPgc with GDP per capita and exp( gc), respectively. and
indicate the 1% and 5% significance level, respectively.

References

Anderberg, M.R., 1973. Cluster Analysis for Applications. Academic Press, New York.
Brecher, R.A., Choudhri, E.U., 1982. The factor content of international trade without factorprice equalization. Journal of
International Economics 12, 277283.
Brecher, R.A., Choudhri, E.U., 1988. The factor content of consumption in Canada and the United States: a two-country
test of the HeckscherOhlinVanek model. In: Feenstra, R.C. (Ed.), Empirical Methods for International Trade. MIT
Press, Cambridge, MA, pp. 517.
Caves, D.W., Christensen, L.R., Diewert, W.E., 1982. Multilateral comparisons of output, input, and productivity using
superlative index numbers. Economic Journal 92, 7386.
Choi, Y., Krishna, P., 2004. The factor content of bilateral trade: an empirical test. Journal of Political Economy 112,
887914.
Davis, D.R., Weinstein, D.E., 2001. An account of global factor trade. American Economic Review 91, 14231453.
Davis, D.R., Weinstein, D.E., Bradford, S.C., Shimpo, K., 1997. Using international and Japanese regional data to
determine when the factor abundance theory of trade works. American Economic Review 87, 421446.
Deardorff, A.V., 1979. Weak links in the chain of comparative advantage. Journal of International Economics 9, 197209.
Debaere, P., 2003. Relative factor abundance and trade. Journal of Political Economy 111, 589610.
Debaere, P., Demiroglu, U., 2003. On the similarity of country endowments. Journal of International Economics 59, 101136.
Evenett, S.J., Keller, W., 2002. On theories explaining the success of the gravity equation. Journal of Political Economy
110, 281316.
Griffith, R., Redding, S., Van Reenen, J., 2004. Mapping the two faces of R&D: productivity growth in a panel of OECD
industries. Review of Economics and Statistics 86, 883895.
Harrigan, J., 1997. Technology, factor supplies, and international specialization: estimating the neoclassical model.
American Economic Review 87, 475494.
Helpman, E., 1984. The factor content of foreign trade. Economic Journal 94, 8494.
Keller, W., 2002. Geographic localization of international technology diffusion. American Economic Review 92, 120142.
 H. Lai, S.C. Zhu / Journal of International Economics 71 (2007) 389409 409

Leamer, E.E., 1984. Sources of International Comparative Advantage: Theory and Evidence. MIT Press, Cambridge, MA.
Maskus, K.E., 1985. A test of the HeckscherOhlinVanek theorem: the Leontief Commonplace. Journal of International
Economics 19, 201212.
Schott, P.K., 2003. One size fits all? HeckscherOhlin specification in global production. American Economic Review 93,
686708.
Staiger, R.W., 1986. Measurement of the factor content of foreign trade with traded intermediate goods. Journal of
International Economics 21, 361368.
Staiger, R.W., 1988. A specification test of the HeckscherOhlin theory. Journal of International Economics 25, 129141.
Trefler, D., 1993. International factor price differences: Leontief was right! Journal of Political Economy 101, 961987.
Trefler, D., 1995. The case of the missing trade and other mysteries. American Economic Review 85, 10291046.