Trade Partner Diversification and Growth: How

Trade Links Matter∗

Ali Sina Önder† Hakan Yilmazkuday ‡

Abstract

We analyze the effects of a country’s export connections on its income growth

using Trade Partner Diversification (TPD) measures that capture the country’s
relative importance in the international trade network. On top of the standard
trade openness measures, TPD measures are shown to enter growth regressions
positively and significantly, where one standard deviation increase in TPD is
associated with a 1 to 1.5 percentage point increase in the annual growth rate.
Threshold analyses show that TPD measures are positively and significantly cor-
related with growth in countries that have low financial depth, high inflation,
low levels of human capital, or high trade openness.

Keywords: Trade; Economic Growth; Export Networks; Thresholds; Cross-

Country Analysis; Trade Partner Diversification

JEL Classification: F13, G20, O19

∗
We thank William Lastrapes, Hartmut Egger and an anonymous referee for helpful comments.
The usual disclaimer applies.
†
University of Bayreuth, Dept. of Economics, RW1, Universitätsstrasse 30, 95447 Bayreuth,
Germany; Tel:+49-921-556085; Fax: +49-921-556081. Email: [email protected]
‡
(corresponding author) Florida International University, Dept. of Economics, 11200 SW
8th Street, 33199 Miami FL, USA; Tel: +1-305-348-2316; Fax: +1-305-348-1524; Email: hyil-
[email protected]
1 Introduction
Trade openness can support growth by providing access into large markets, low-cost
intermediate inputs, and higher technologies. Empirical analyses of Frankel and Romer
(1999), Irwin and Terviö (2002), and Wacziarg and Welch (2008) provide evidence for
a robust, statistically significant, and positive effect of international trade on income
growth. Nevertheless, measures for a country’s trade openness and trade volume usu-
ally do not differentiate between trade partners so that two countries with the same
trade volume but trading with different sets of countries are identical in the eyes of
these measures.1 We investigate in this paper whether and in what ways a country’s
connectivity in international trade matters in addition to its trade openness (measured
by residual openness, to be defined below); hence, we ask: Does the way a country is
connected within the web of international trade explain this country’s per capita in-
come growth above and beyond what can solely be explained by conventional measures
of trade openness, and moreover, how does this connectivity correlate with growth in
countries at different stages of economic and financial development?
We measure a country’s connectivity by evaluating the overall position of this
country in the web of international trade. This requires paying attention not only to
trade partners of this country but also to trade partners of its trade partners, because
a country’s connectivity depends on quantity as well as quality of its trade partners.
Although a country may have fewer trade partners than some other countries, it may
still be deemed better connected than others if its trade partners are better connected
than those of other countries. Similarly, between two countries that have the same
number of trade partners, the one that has better connected trade partners is also
deemed better connected than the other country. We employ conventional network
centrality measures to capture the connectivity of a country and will refer to these
centrality measures as trade partner diversification (TPD) measures throughout this
paper.
Using a cross-country panel data set for 83 countries, we first analyze whether

1
In order to have a better idea about the progress of such measures over time, consider Kali et
al. (2007) who show that the share of total trade in the GDP for the average economy went from
58.3 percent in 1970 to 88.5 percent in 2003, while the average number of trading partners more than
doubled as it went from 46.4 to 93.9 in that same period.
1
TPD has any role in explaining growth, given the degree of trade openness together
with other control variables. Such an investigation requires a measure for TPD, for
which we consider three alternative definitions, all obtained by using international bi-
lateral trade data. The baseline regression analysis, which is robust to endogeneity
concerns, shows that TPD enters growth equations positively and significantly on top
of the control variables, including the financial system (measured by financial depth),
price stability (measured by inflation), human capital (measured by secondary educa-
tion level), government size (measured by government expenditure as a percentage of
GDP), per capita income level, and trade openness (measured by residual openness).
Therefore, on average across all countries in the sample, there is a positive role of TPD
on growth. A country’s measures of TPD increase as this country gains access to more
and/or better export markets—better in the sense that these export markets are well
connected with the rest of the world.
Herzer (2013) finds, in addition to a positive and significant effect of trade on in-
come, a large heterogeneity in this effect across countries. In order to take into account
that countries having different macroeconomic conditions may be affected asymmet-
rically by their respective positions in the web of international trade regarding their
growth, we investigate nonlinear effects of countries’ positions in the international
trade network using a threshold analysis. In particular, this methodology enables us
to observe how coefficients of network measures change continuously over the whole
range of other explanatory variables so that we do not need to consider discrete cat-
egories or interaction terms and we are able to prevent any potential problems of
over-representation and biases created as a result thereof.
In our threshold investigation, we consider alternative sets of countries distin-
guished with respect to their financial system (measured by financial depth), price sta-
bility (measured by inflation), human capital (measured by secondary education level),
government size (measured by government expenditure as a percentage of GDP), per
capita income level, and trade openness (measured by residual openness). Our thresh-
old analysis results show that TPD is positively and significantly correlated with the
growth of countries that lack a developed financial system, price stability, or advanced
human capital, after controlling for the degree of trade openness. Therefore, a plausible
interpretation is that the TPD provides channels to hedge against a lack of financial

2
development or human capital as well as against macroeconomic volatility borne by
high inflation. This result is important especially for developing economies where, on
average, financial depth is low, inflation is high, and human capital is low. Developing
(or underdeveloped) countries that experience higher growth rates than their counter-
parts are those that are well connected to export markets or have export partners that
are themselves well connected. Moreover, TPD positively and significantly correlates
with the growth rate in countries that have large trade openness, which implies the
importance of TPD in creating channels for a country to avoid vulnerability borne
by the vulnerability of its trading partners. Finally, TPD correlates positively and
significantly with growth irrespective of government size or per capita income level.
The rest of this paper is organized as follows: the next section provides a brief
discussion of the underlying mechanisms and a review of the related literature. Section
3 introduces the measures of TPD. Section 4 introduces the trade data and presents
descriptive statistics of the TPD measures. Results obtained from growth regressions
and threshold analysis are shown and discussed in Section 5. We draw our conclusions
in the last section.

2 How Trade Partner Diversification Works and

Relations to the Existing Literature
Our paper follows the line of literature investigating the relationship between inter-
national trade and growth, using trade network indicators for the role and strength
of a country in international trade. Kali and Reyes (2007) and Kali, Méndes, and
Reyes (2007) provide the first set of very intriguing results, using network measures.
Kali and Reyes (2007) use degree centrality and importance measures to capture a
country’s international economic integration in addition to its trade openness. They
use rankings of countries according to these measures in their regression analysis and
show that a country’s ranking in either of these network measures is significantly cor-
related with its growth rate even after controlling for conventional measures of trade
openness: e.g., a 10-unit increase in a country’s rank in degree centrality is correlated
with a 1.11 percentage point increase in income of that country. Moreover, they argue
that the network position of a country substitutes for physical capital by making up
3
for the lack of technology and at the same time complements human capital. Kali,
Méndez, and Reyes (2007) investigate how degree centrality (what they refer to as the
number of trading partners) and trade concentration (as measured by the Herfindahl-
Hirschmann Index) correlate with growth in rich and poor countries. They find that
the number of trading partners and trade concentration are positively correlated with
growth across all countries; however, the former turns out to be more important for
rich countries whereas the latter is more important for poor countries.
A related line of research has mainly focused on the relationship between eco-
nomic growth and the diversification of goods exported. Kali et al. (2013) provide a
powerful analysis using network density and proximity measures linking all interna-
tionally traded products and products traded by a country. They unravel interesting
characteristics of the growth acceleration process: interactions of agglomeration ex-
ternalities and proximity of products are shown to determine the likelihood of higher
growth. Their findings also support those of Cadot et al. (2011) who show that
a country travels along diversification cones as its income grows.2 This line of lit-
erature focuses on sector-specific links and hence offers insights for possible effects
of industry/goods-specific shocks on growth. However, as shown by Costello (1993)
and Karadimitropoulou and León-Ledesma (2013) from a business cycle perspective
through variance decomposition analyses, fluctuations in growth rates are dominated
by country-specific shocks rather than industry/goods-specific shocks. Thus, our anal-
ysis using measures of TPD bridges an important gap between network structure of
international trade and economic growth by focusing on country-specific rather than
sector-specific links, which constitute an important transmission mechanism of global
shocks to country-specific growth rates.
In our analysis, we employ degree centrality of a country in the international net-
work, as Kali and Reyes (2007) and Kali, Méndez, and Reyes (2007) do,3 and in addi-
tion, we employ measures of closeness centrality and eigenvector centrality. Whereas
Kali and Reyes (2007) use rank of a country according to its degree centrality and
importance, we prefer using numerical values of centrality measures in our regression
and threshold analyses. Similar to these two papers, we show that network mea-
2
See Cadot et al. (2013) for an excellent survey on diversification of export goods and growth.
3
Kali et al. (2007) cover years between 1980 and 2000 in their empirical analysis, and Kali and
Reyes (2007) focus on the period from 1987 to 1998.
4
sures are positively and significantly correlated with growth even after controlling for
trade openness. In addition to it, we show that this canonical finding still holds after
controlling for finance, inflation, and government spending of countries. We further
advance this line of literature by providing an investigation of nonlinear relationships
between a country’s position in the international trade network and its growth rate,
using threshold analysis with respect to various explanatory variables of the growth
equation.
In the theoretical literature, the positive effects of trade on growth are well estab-
lished. One channel through which these positive effects work is the transmission and
creation of knowledge, as shown in studies by Romer (1990), Grossman and Helpmann
(1993), and Rivera-Batiz and Romer (1991). In particular, countries that are involved
in trade are exposed to new ideas, designs, and technologies, together with innova-
tive managerial decisions. Accordingly, when a country has a larger trade network,
such knowledge spillovers would increase, where size of a country’s trade network is
measured by the number of direct or indirect (more than one geodesic distance apart)
trade partners. While the direct partnership would have an impact on growth through
the trade of final goods, the indirect partnership would have an impact through the
trade of intermediate goods or re-exports.
Another channel for an exporting country is by getting exposed to different poten-
tial buyers because different buyers may have alternative consumer tastes, government
regulations, or climate. Accordingly, an exporter country would invest more in local
research and development strategies, such as innovation, brand recognition, and patent
registrations. As an exporting country gets involved in a larger trade network, such
destination-specific requirements, either for final goods or intermediate goods, would
result in a wider range of research and development, which, in turn, would positively
affect the overall productivity of the exporter country. Another channel is through
the competition in the destination markets (see Vickers and Yarrow, 1991; Bourbakri
and Cosset, 1998). As an exporter country gets involved in a larger trade network,
the complexity of the competition would increase for both direct trade partnerships
(through final goods) and indirect trade partnerships (through intermediate goods).
The magnitudes of the above-mentioned channels are, however, subject to the
changes in the macroeconomic environment or shocks, especially exchange rate fluc-

5
tuations between the exporting country and (both direct and indirect) trade partners.
Accordingly, this paper is also connected to the literature based on strategic hedging
for exchange rate volatility through diversification (see Hitt et al., 2006; Hoskisson and
Hitt, 1990; Ito, 1997; Meyer, 2006). In particular, consider two extreme countries with
the same level of trade openness, the first with only one direct trade partner and the
second with many direct and indirect trade partners. The export of the first country
would highly depend on economic volatilities (mostly reflected in the exchange rates)
in the unique destination country. Hence, in the case of a crisis in the destination
country, the first country would need additional resources in order to cope with such
a reduction in its exports. If the first country has a developed financial system, low
inflation (i.e., low uncertainty in its domestic economy), or high human capital, the
economic loss due to the reduction in exports can be compensated by additional credits
(due to the developed financial system) that can be provided without any uncertainty
(due to the low inflation) in order to sustain and promote economic growth. However,
if the first country does not have such characteristics, its growth would be affected
adversely in this case. Consider the second country in the example above having a
similar problem in one of the countries that it is exporting to (i.e., one of the direct
trade partners) or that its trading partner is exporting to (i.e., one of the indirect trade
partners). Even without any developed financial system, low inflation, or high human
capital, the second country would be affected much less by the economic volatility in
any of its trade partners.
In light of the above example, an expected result of a cross-country analysis (as
we also run in this paper) would be that trade networks compensate for low levels
of financial depth, high levels of inflation, and low levels of human capital. The
existing literature also supports this expectation. For instance, Aghion et al. (2009)
show that exchange rate volatility can stunt growth of a country when its financial
development is low. Similarly, Gylfason (1999) suggests that inflation hurts growth
through lower ratios of exports to output;4 and Kali and Reyes (2007) show that a

4
These effects are also subject to firm-level strategies, such as sequential export entry (see Al-
bornoz et al., 2013). In particular, many firms enter new international markets to increase their sales
through exports by accepting to pay high sunk costs. However, such costs are worthwhile to pay
when they expand their exports to alternative countries through the trade network. Therefore, if
the first destination market that is being entered (i.e., the central/hub country through direct trade
6
country’s integration to the international economy may complement its human capital,
however, significance of such complementarity depends on the particular integration
measure being employed in their analysis.
Accordingly, a country’s TPD in international trade can be used as a proxy for two
important macroeconomic characteristics of this country: first, a country that enjoys a
high quantity and/or high quality of trading partners may find it easier to substitute for
financial development by hedging, diversifying, and pooling risk arising from exchange
rate volatility because exchange rate risk is distributed among its trading partners.5
Moreover, high and variable inflation may create considerable uncertainty about future
prices, interest rates, and exchange rates, which, in turn, increases the overall risk of
business among trade partners due to the possibility of a devaluation and vulnerability
to speculative attacks.6 TPDs of trade partners show how such risks are distributed
among trade partners because the possibility of a bilateral depreciation (with respect
to one currency) is much higher than the possibility of a multilateral depreciation (with
respect to a basket of currencies consisting of the currencies of the trade partners).
Second, good connectivity can compensate for international shocks that may arise
due to having higher degrees of trade openness.7 As shown by Svaleryd and Vlachos
(2002), trade openness increases income volatility, and thus better diversification of
risk is needed. A country with a high score of TPD is likely to find it easier to achieve
such diversification.

3 Measuring Trade Partner Diversification

In this section, we explain how TPD measures are obtained, we discuss their various
aspects, and we provide interpretations with the help of a simple example. The col-

partnership) is already involved in a good trade network due to its trade partners (i.e., if the source
country has a good indirect trade partnership), the exporting firm/country would benefit more from
it.
5
Bailliu et al. (2003) have shown that exchange rate risk is reduced when exports are diversified
across markets. See Levine (1997) for the relationship between finance and trade.
6
See Rousseau and Wachtel (2002) for a discussion on the channels through which inflation may
negatively affect growth.
7
See Rodrik (1998) who has emphasized the role of international fluctuations (imported through
trade openness) in growth.
7
lection of bilateral trade links between countries yields the basic structure of the web
of international trade. Links between pairs of countries at time t are captured by the
adjacency matrix A(t) where an element a(t)ij is defined as follows:


1, if there is export of goods from country i to country j

a(t)ij =
0, if there is no export of goods from country i to country j


We employ export degree, export closeness, and export eigenvector (centrality) to

measure a country’s TPD. Suppose there are n countries at time t in the international
trade network. The fraction of country i’s existing export links to the total number
of countries (excluding itself) is the export degree of country i, and we denote it by
ED(t)i such that

1 X
ED(t)i = a(t)ij
n−1 j

Export degree takes on values from zero to one: ED(t)i = 1, if country i is exporting
to every possible country in the international trade network at time t, and ED(t)i = 0,
if country i does not export anywhere.
The distance between any two countries i and j in the international trade network
is the geodesic distance measured by the length of the shortest path connecting the
two countries; if country i is exporting to country j, then the distance from i to j is
one. If country i does not export to country h but it exports to country j, and country
j exports to country h, then the distance from country i to country h is two.8 Such
linkages are especially important in order to consider the effects of intermediate-input
trade, which is mostly discussed in the literature under the title of ”Trade in Value-
Added and Global Value Chains,” which is also given utmost consideration by World

8
If countries i and j are not connected by any collection of edges, then the distance is assumed
to be infinite, but it remains a practical issue as to how big ”infinity” should be. For our empirical
analysis we take the infinite distance to be the equivalent of a geodesic of 10. It is important to
emphasize that this specific value does not affect our growth regression results in the next section.
We check its robustness by using two alternative distances to capture infinity, namely 5 and 100.
Comparing coefficient estimates using 5, 10, and 100 we find that these are qualitatively the same
and quantitatively very similar. Details of estimations using 5 and 100 as the infinite geodesic distance
are available upon request.
8
Trade Organization.9 All such bilateral distances in the international trade network at
time t are collected in the distance matrix D(t) where an element d(t)ij is the distance
from country i to country j following export links in the international trade network.
Export closeness of country i at time t is calculated by

n−1
EC(t)i = P
j d(t)ij

and it takes on values from zero to one, increasing as a country gets closer to the rest of
the world. Export closeness adds valuable information on top of what we already know
from the export degree of a country. Not only do we care whether there is direct trade
between country i and the rest of the world, we also care for the geodesic distance from
country i to countries that are not its direct trade partners. Thus, export closeness
gives a country credit for having access to the export market of a country that itself
has access to other export markets that the former country does not have.
The third measure we employ in our analysis is the export eigenvector of countries
in the international trade network, as developed by Bonacich (1987). The export
eigenvector captures the idea that the centrality of a country must be proportional to
the centrality of countries where it exports to. This is also referred to as the prestige
or influence (Newman, 2010). According to this measure, not only the total number of
direct links is important but also the prestige of those links counts; for a high score of
the export eigenvector, a country should be connected to countries that are also well
connected and thus have high prestige. Countries that are exporting to trade partners
that have a high score of the export eigenvector (hence high prestige) are in a better
position to diversify and minimize country-specific risks affecting international trade
because their trade partners are in a better position to do so as well.
Next, we provide a simple example for the international trade network, where we
apply these three centrality measures and discuss what they capture. We further show
how these measures evolve as new trade links are established. Suppose the world
consists of eight countries and exports flow between these countries as depicted in

9
The related theoretical and empirical studies are surveyed in a recent study Johnson (2014) who
shows that intermediate inputs (i.e., the part of exports that has not been produced in the source coun-
try) range between 10 percent (Russia) and 50 percent (Taiwan). Please see https://www.wto.org/
for more details.
9
Figure 1.
1 3 6

4 7

2
5
8
Figure 1: Export flows between eight countries

The direction of an arrow from one country to another shows the direction of
export flow between these countries, e.g., country 1 is exporting to countries 3 and
4, but country 1 imports only from country 4. There are two blocks that are not
interconnected: countries 1 to 5 make up one block and countries 6, 7, and 8 make up
the other. Since these two blocks are not interconnected, the distance between them
is infinite. We calculate our TPD measures for these eight countries and present the
results in Table 1.
Table 1: Network centralities as TPD measures

Export
Country Export Degree Export Closeness Eigenvector
1 0.29 0.194 0.45
2 0.14 0.179 0.29
3 0.14 0.175 0.16
4 0.43 0.2 0.64
5 0.29 0.194 0.52
6 0.14 0.132 0
7 0.29 0.135 0
8 0 0.1 0

Export degree captures the overall export market access of a country: country 4
exports to 43% of all countries, while country 8 is not exporting anywhere. Countries
1, 5, and 7 have the same export degree, but according to the export closeness, country
7 is further away from the rest of the world than countries 1 and 5. This follows from
the positioning of country 7; it exports to countries that are not connected to any
10
other country besides country 7.
Export eigenvector is increasing in a country’s export degree and the connectivity
of its trade partners. Countries 2 and 3 have an identical export degree and very
close export closeness. When it comes to the export eigenvector, country 2 has a
better measure, which reveals that countries that import from country 2 are themselves
exporting either to a lot of countries or to key exporters. Although countries 1 and
5 have the same export degree and closeness, country 5 is a more important player
in the international trade network than country 1 according to the export eigenvector
measure. Countries 6, 7, and 8 have a zero export eigenvector because they are not
connected to any important exporter.10
Next, we illustrate how new export links change our TPD measures. Suppose that
country 2 starts exporting to country 8, and after a while country 8 starts exporting
to country 2. These new export links are depicted in Figure 2 panels (a) and (b).

(a) (b)

1 3 6 1 3 6

4 7 4 7

2 2
5 5
8 8

Figure 2: Export flows between eight countries, including exports from country 2 to
country 8

10
It is possible to avoid obtaining a zero export eigenvector by assuming a base prestige for each
country. This is especially helpful if the network consists of many small components. This is not ob-
served in international trade data so that a strictly positive base prestige would not have a meaningful
effect on our empirical analysis.
11
 Table 2: Revised TPD measures

Export
Country Export Degree Export Closeness Eigenvector
(a) (b) (a) (b) (a) (b)
1 0.29 0.29 0.24 0.24 0.45 0.41
2 0.29 0.29 0.23 0.23 0.29 0.37
3 0.14 0.14 0.218 0.219 0.16 0.19
4 0.43 0.43 0.25 0.25 0.64 0.58
5 0.29 0.29 0.25 0.25 0.52 0.5
6 0.14 0.14 0.132 0.26 0 0.07
7 0.29 0.29 0.134 0.33 0 0.14
8 0 0.14 0.1 0.21 0 0.19

Table 2 lists our revised TPD measures for each country based on their position
in the international trade network. Each measure has two columns (a) and (b), corre-
sponding to networks depicted in panels (a) and (b), respectively, of Figure 2. When
country 2 starts exporting to country 8, this directly affects the export degree of coun-
try 2. Moreover, closeness of most countries is being affected positively. It is worth
mentioning that the new export link does not affect the eigenvector measure of any
country. This is simply because the exporter, country 2 in this case, is not getting any
prestige from exporting to country 8 because country 8 itself has no prestige.
When country 8 starts exporting to country 2, this leads to several curious changes
in eigenvector centralities: country 2’s import demand generates prestige for country 8.
Since country 8 now has prestige, i.e., a non-zero export eigenvector, this also leads to
non-zero export eigenvectors for countries 6 and 7. Changes in the export eigenvector
are very sensitive to the case that a low prestige country starts exporting, especially
to a prestigious trade partner or equivalently if this trade partner is a member of a
large trading block.

12
4 Trade Data and International Trade Networks
Our trade data set contains bilateral trade flows between 83 countries from 1965 to
2004; this is a subset of the trade data set of Rose and Spiegel (2011) that covers the
annual real FOB exports.11 The list of countries is given in the note under Table 3.
In this section, we describe how the TPD measures evolve over time for a selected
group of countries. It is not the aim of this paper to provide a full discussion of the
international trade network12 . Our aim for this section is to provide insight to how
the TPD measures evolve over time for different countries because patterns of these
measures are important for the subsequent econometric analysis and interpretation of
our results in the next section. Arribas et al. (2009) document an overall increasing
trend in direct trade connections between countries and also find that indirect or higher
order connections reveal distinct and varying patterns across countries. In the context
of our analysis, export degree captures direct export links between countries, whereas
export closeness and export eigenvector are measures that capture the strength and
influence of such indirect connections.
In this section, we briefly describe the positions of the United States, Brazil, and
Israel in the international trade network based on their export connections. Being a
rich and large country, the US has been one of the most important players in world
trade. Brazil and Israel are developing countries that differ by the characteristics of
their trade in goods; Brazil has long been an exporter of raw materials and agricultural
products, whereas Israeli exports rely on capital-intensive manufactures. A common
pattern for export degree and export closeness is that both display an increasing trend
over time for most countries in our data set. The export eigenvector reveals an in-
teresting pattern. As discussed in the previous section, well-connected countries lose
network prestige measured in the export eigenvector as the rest of the trade network
establishes further links to outsiders. This trend is obvious for the US. Since Brazil
and Israel rose later in world trade, their trade network centrality measures reveal
different patterns.

11
We focus on this subset to be consistent with our growth data set, which we document below.
12
See De Benedictis and Tajoli (2011) for a recent and detailed description of the international
trade network with implications on trade structure. Also see Serrano and Boguña (2003) for an
analysis of complex network properties observed in the international trade network.
13
 Export Degree Export Closeness

1
Closeness1_F
Degree_F
.9

.9
.8

.8
.7

.7
1960 1970 1980 1990 2000 1960 1970 1980 1990 2000
year year

Export Eigenvector
.16
Eigen_UW_D
.12 .14
.1

1960 1970 1980 1990 2000

year

Figure 3: Trade Network Centralities for the United States, 1965-2004

The US is one of the most important players in international trade, and this im-
portance clearly manifests itself in its TPD measures shown in Figure 3. The US was
exporting to about 75% of the countries that are contained in our data for 1965, and
to more than 95% of countries in 2004. Connecting to different trade partners and
hence establishing more links necessarily cuts down the overall distance of the US to
other countries, which leads to a higher centrality measure over years. The decrease
in the export eigenvector of the US in the 1980s captures the fact that the world trade
network kept growing in this decade and export markets of the US lost their relative
importance and prestige, which in turn led to the US losing its importance and prestige
in the international trade network measured by the export eigenvector.
Trade network measures for Brazil reveal an even more interesting story, as shown in
Figure 4. Brazil’s export markets made up 40% of countries that were actively trading
in 1965, and in 2004, Brazil’s exports reached about 90% of countries. Although the
trend of the degree centrality measure has been increasing and is thus similar to that of
the US in that respect, its relative change from 40% to more than 90% is impressive.
The increase in Brazil’s export eigenvector during the 1960s and 1970s shows that
Brazil gained access to prestigious markets in these years.
14
 Export Degree Export Closeness

1
Closeness1_F
.9
Degree_F
.8

.8
.6

.7
.4

.6
1960 1970 1980 1990 2000 1960 1970 1980 1990 2000
year year

Export Eigenvector
.11 .12 .13 .14
Eigen_UW_D
.1

1960 1970 1980 1990 2000

year

Figure 4: Trade Network Centralities for Brazil, 1965-2004

Export Degree Export Closeness

.4 .45 .5 .55 .6 .65

.75
Closeness1_F
Degree_F

.65 .6.7

1960 1970 1980 1990 2000 1960 1970 1980 1990 2000
year year

Export Eigenvector
.08 .1 .12.14
Eigen_UW_D
.06

1960 1970 1980 1990 2000

year

Figure 5: Trade Network Centralities for Israel, 1965-2004

15
 Export market access of Israeli goods experienced a downward trend from 1970
until the late 1980s, as revealed by the degree centrality (see Figure 5). During the
same period, the export eigenvector for Israel decreased as well. The decrease in the
export eigenvector turned out to be strong because the downward trend is fed by two
factors simultaneously: first, the world trade network is becoming more dense over
time, and this creates downward pressure on a country’s export eigenvector, even if a
country keeps its share of access to the world’s export markets from previous periods.
Second, even if the world trade network remained stable except for one country that
loses its existing export links, this country is bound to lose its prestige as captured by
the export eigenvector.13 In the Israeli case, the world trade network grows more dense
(most countries establish new export links and their export degree increases), and the
export degree of Israel decreases at the same time; thus, the two forces mentioned
above work in the same direction and lead to an even stronger decrease in its export
eigenvector.

5 Growth Data and Baseline Growth Regressions

An almost standard empirical framework has emerged since Barro (1991) and Levine
and Renelt (1992) introduced cross-country regression as an empirical representation
of the Solow growth model because new growth theories suggest that trade policy
affects growth through its impact on technological change (i.e., Solow’s residual).14
The growth data set was constructed for 83 countries covering the period 1965-2004 as
a panel of country observations from the World Bank’s World Development Indicators
and International Monetary Fund’s Direction of Trade CD-ROM. The list of countries
is the same as the one that we have for the trade data, which can be found in the note
under Table 3.
We present descriptive statistics of variables in our analysis and their correlations in

13
One possibility to see whether a country’s export eigenvector is declining due to an expanded
world trade or a reduction in trading relationships is to take a closer look at the co-movements of the
TPD measures in our dataset. In 1,056 observations (that is 33% of all observations) the decrease in
export eigenvector is accompanied by a decrease in export degree and closeness as well, and in 516
(16%) we observe that the export eigenvector decreases whereas export degree and closeness increase.
14
See, for example, Grossman and Helpman (1993) and Harrison (1996).
16
 Table 3: Descriptive Statistics, 1965-2004, 83 Countries

Per capita Log of

income Finance initial Log of
growth Trade (M3, Inflation SEC Gov't initial Export Export Export
Variable (%) Openness % GDP) (%) (%) (% GDP) GDP Degree Closeness Eigenvector

Mean 1.78 53.68 45.59 12.75 3.65 14.44 7.65 48.85 65.84 8.56
Maximum 11.66 113.20 184.03 86.25 4.96 40.59 10.73 97.81 97.85 16.70
Minimum -9.27 8.92 4.15 0.49 0.00 4.36 4.97 0.00 10.00 0.00
Standard deviation 2.68 10.43 28.48 13.48 0.84 5.26 1.48 25.78 14.09 3.49
Coefficient of Variation 1.51 0.19 0.62 1.06 0.23 0.36 0.19 0.53 0.21 0.41

Correlations
Per capita income growth (%) 1.00
Trade Openness 0.16 1.00
Finance (M3, % GDP) 0.25 0.05 1.00

17
Inflation (%) -0.20 0.05 -0.25 1.00
Log of initial SEC (%) 0.19 0.00 0.53 0.01 1.00
Government (% GDP) -0.04 -0.01 0.33 -0.16 0.33 1.00
Log of initial GDP 0.18 0.00 0.52 -0.11 0.71 0.43 1.00
Export Degree 0.25 0.03 0.54 -0.16 0.60 0.27 0.61 1.00
Export Closeness 0.21 0.04 0.49 -0.16 0.55 0.24 0.55 0.95 1.00
Export Eigenvector 0.34 0.00 0.43 -0.10 0.49 0.23 0.59 0.88 0.80 1.00

Source: Authors' analysis based on data sources discussed in the text.

Notes: The list of 83 countries is as follows: Algeria, Argentina, Australia, Austria, Bangladesh, Barbados, Belgium, Bolivia, Brazil, Cameroon, Canada, Central African Republic,
Chile, Colombia, Costa Rica, Cote d'Ivoire, Denmark, Dominican Republic, Ecuador, Egypt, El Salvador, Fiji, Finland, France, Gambia, Ghana, Greece, Guatemala, Guyana,
Haiti, Honduras, Iceland, India, Indonesia, Iran, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kenya, Korea Rep., Luxembourg, Malawi, Malaysia, Malta, Mauritius, Mexico,
Morocco, Nepal, Netherlands, New Zealand, Nicaragua, Niger, Nigeria, Norway, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Portugal, Rwanda, Senegal,
Sierra Leone, South Africa, Spain, Sri Lanka, Sudan, Sweden, Switzerland, Syria, Thailand, Togo, Trinidad and Tobago, Turkey, United Kingdom, United States, Uruguay,
Venezuela, Zimbabwe.
Table 3. As is evident, TPD measures are positively correlated with per capita income
growth and with each other; however, they are not correlated (either negatively or
positively) with trade openness. A possible explanation for the low correlation between
trade openness and TPD measures is the following: since TPD measures are based
on unweighted trade networks and don’t depend on the volume or scale of exports
whereas trade openness does, it is possible that two countries have similar positions in
the international trade network and yet their trade volumes are completely different.
Hence TPD measures are expected to have significant effects on growth on top of
the trade openness measure, which is also positively correlated with per capita income
growth. Nevertheless, these are just descriptive statistics, and next we employ a formal
econometric analysis to show the effects of TPD measures.
Following Barro (1991) and Levine and Renelt (1992), the baseline growth equa-
tions include a standard set of explanatory variables that provide robust and widely
accepted proxies for growth determinants. The dependent variable is the growth rate
of real per capita income averaged over 5-year periods from 1965 to 2004. On top of
the TPD measures15 , the baseline regression analysis includes standard explanatory
variables, such as log initial per capita GDP, log initial secondary enrollment rate
(SEC), the ratio of liquid liabilities (i.e., M3) to GDP, inflation rate, trade openness,
and government size. In order to consider the effects of trade policies (rather than the
volume of trade that is endogenous to within-country variations in per capita GDP),
consistent with earlier studies, such as by Balassa (1985), Leamer (1988), Syrquin and
Chenery (1989), and Edwards (1992), who have considered the deviation of actual
from predicted trade flows in their growth regressions, we measure trade openness
as residual openness that is obtained as residual from the regression where exports
plus imports over GDP are regressed on per capita income, country-fixed effects, and
time-fixed effects.16
As discussed in detail by Kali et al. (2007), although income growth may affect
the change in trade network, there is no reason to believe that growth affects the
level of trade network. Nevertheless, we use two-stage least squares (TSLS) to control

15
TPD measures in raw form take values between zero and one. We scale these measures by 100
when we use them in our regression analysis.
16
We also considered the raw trade openness measure of exports plus imports over GDP; the results
were virtually the same. Such results are available upon request.
18
for any potential endogeneity, where initial values of the corresponding variables (i.e.,
trade network, finance, inflation, trade openness, and government size) are used as
instruments in the first stage for each 5-year period. The usage of initial values as
instruments is based on the assumption that they are exogenous, since they are prede-
termined. However, since such an assumption may not be valid as discussed in studies
such as by Brock and Durlauf (2001), we support our investigation with state-of-
the-art tests of under-identification, weak-identification, and weak-instrument-robust
inference. Moreover, since ordinary least squares (OLS) is more efficient when variables
are exogenous, we also perform a standard endogeneity test.
Time-fixed effects for 5-year periods are also included to control for shocks with
common growth effects across countries. For robustness, country-fixed effects are also
considered in order to capture any country-specific characteristics that are constant
over time, such as their geographic location or historical experience/institutions.17
Accordingly, regression results are shown in Table 4 where country-fixed effects
are excluded in the benchmark case. As is evident, all TPD measures are positive
and significant at the 1% level, with or without the trade openness measure included
in the regressions. All other control variables have their expected signs, although
inflation is the only variable that is insignificant at any considered level. We present
regression results without TPD measures in column (1), and as can be seen, signs and
significances of other explanatory variables do not change when TPD measures are
included in the analysis. When country-fixed effects are also included for robustness,
as shown in Table 5, all TPD measures are again positive and significant at the 1%
level, although some of the control variables have changed their estimated signs and/or
significance. In sum, robust to the inclusion of country-fixed effects, trade openness,
and other control variables, TPD measures are significant and positive in all considered
cases.
When we perform endogeneity tests by using the null hypothesis that variables are
exogenous, we obtain the results (in terms of p-values) given in Table 4 and Table

17
In order to make a comparison with the existing literature, Kali et al. (2007) have also considered
IV regressions where they have instrumented trade networks with country-specific measures, such as
physical access to international waters or tropical climate. Since these are mostly time-invariant
country-specific variables, our methodology involving country-fixed effects is comparable to the one
in Kali et al. (2007).
19
Table 4: Instrumental variables growth regressions, 1965-2004, 83 Countries
Dependent Variable: Growth of Per Capita Income (%)
Variables (1) (2) (3) (4) (5) (6) (7)

Export Degree 0.0215*** 0.0203***

(0.00622) (0.00612)

Export Closeness 0.0396*** 0.0367***

(0.0111) (0.0109)

Export Eigenvector 0.161*** 0.159***

(0.0449) (0.0441)

Trade Openness 0.0326** 0.0327** 0.0324** 0.0343**

(0.0117) (0.0115) (0.0115) (0.0115)

Finance (M3, % GDP) 0.0208*** 0.0181*** 0.0184*** 0.0186*** 0.0170*** 0.0174*** 0.0172***
(0.00490) (0.00504) (0.00502) (0.00499) (0.00497) (0.00494) (0.00491)

Inflation (%) -0.0168 -0.00847 -0.00883 -0.0106 -0.0132 -0.0136 -0.0154

(0.0109) (0.0111) (0.0111) (0.0110) (0.0109) (0.0109) (0.0107)

Government (% GDP) -0.0728** -0.0667** -0.0667** -0.0649** -0.0685** -0.0685** -0.0666**

(0.0243) (0.0244) (0.0244) (0.0243) (0.0240) (0.0240) (0.0239)

Log of initial GDP -0.191 -0.330** -0.327** -0.307* -0.321** -0.316** -0.304*
(0.117) (0.124) (0.123) (0.121) (0.122) (0.121) (0.119)

Log of initial SEC (%) 0.999*** 0.839*** 0.880*** 0.770*** 0.872*** 0.912*** 0.799***
(0.211) (0.216) (0.214) (0.219) (0.213) (0.211) (0.215)

F-test (Endogeneity) 0.0821 0.0395 0.0463 0.0318 0.1122 0.1381 0.0957

R-bar Squared 0.265 0.259 0.259 0.266 0.282 0.281 0.290

Source: Authors' analysis based on data sources discussed in the text.

Notes: +, *, ** and *** indicate significance at the 10%, 5%, 1% and 0.1% levels, respectively. Standard errors are in parentheses.
Growth rates are five-year averages. Network measures are five-year averages; their initial values in each five-year period are used as
instruments for the corresponding five-year averages. All equations also include time fixed effects that are not shown. Estimation is by
two-stage least squares. The sample size in each equation is 464. Endogeneity (p-value) shows p-values of the regression-based
endogeneity test results regarding the null hypothesis that variables are exogenous.

20
Table 5: Instrumental variables growth regressions with country-fixed effects, 1965-2004, 83 Countries
Dependent Variable: Growth of Per Capita Income (%)
Variables (1) (2) (3) (4) (5) (6) (7)

Export Degree 0.0585*** 0.0518***

(0.0142) (0.0139)

Export Closeness 0.100*** 0.0893***

(0.0203) (0.0199)

Export Eigenvector 0.293*** 0.280**

(0.0887) (0.0859)

Trade Openness 0.0351*** 0.0344*** 0.0340*** 0.0373***

(0.00952) (0.00941) (0.00938) (0.00937)

Finance (M3, % GDP) -0.0159 -0.00897 -0.0131 -0.00716 -0.0143 -0.0179+ -0.0129
(0.0108) (0.0108) (0.0108) (0.0109) (0.0107) (0.0106) (0.0107)

Inflation (%) -0.0120 -0.00533 -0.00678 -0.00408 -0.0146 -0.0158 -0.0145

(0.0139) (0.0142) (0.0142) (0.0142) (0.0137) (0.0137) (0.0137)

Government (% GDP) -0.0914* -0.0808+ -0.0956* -0.0640 -0.0863* -0.0994* -0.0703+

(0.0423) (0.0430) (0.0427) (0.0439) (0.0418) (0.0415) (0.0425)

Log of initial GDP -2.564*** -3.503*** -3.326*** -2.863*** -3.346*** -3.193*** -2.790***
(0.522) (0.567) (0.544) (0.535) (0.554) (0.531) (0.520)

Log of initial SEC (%) -0.993** -1.118** -1.124** -1.314*** -1.052** -1.058** -1.239***
(0.371) (0.377) (0.375) (0.387) (0.366) (0.365) (0.375)

Endogeneity (p-value) 0.2207 0.1566 0.0655 0.2424 0.3630 0.2525 0.4747

R-bar Squared 0.420 0.403 0.409 0.399 0.434 0.439 0.434

Source: Authors' analysis based on data sources discussed in the text.

Notes: +, *, ** and *** indicate significance at the 10%, 5%, 1% and 0.1% levels, respectively. Standard errors are in parentheses.
Growth rates are five-year averages. Network measures are five-year averages; their initial values in each five-year period are used as
instruments for the corresponding five-year averages. All equations also include country fixed effects and time fixed effects that are
not shown. Estimation is by two-stage least squares. The sample size in each equation is 464. Endogeneity (p-value) shows p-values of
the regression-based endogeneity test results regarding the null hypothesis that variables are exogenous.

21
5, which are mixed. Since OLS is more efficient when variables are exogenous, in
order to give the reader a better insight, we also replicated the results in Table 4
and Table 5 by using OLS in Appendix Table A.1 and Table A.2, respectively. As is
evident, the estimated coefficients are virtually the same qualitatively and very similar
quantitatively when OLS is used instead of TSLS.
We run further state-of-the-art tests of under-identification, weak-identification,
and weak-instrument-robust inference in order to test the validity of our instruments
and document their results in Appendix Table A.3 and Table A.4. Regarding under-
identification, we use the Kleibergen-Paap rk LM test. The null hypothesis that the
estimated equation is under-identified is rejected for all the TSLS regressions that we
have run at the 1 percent significance level.
Regarding weak-identification, we use the Cragg-Donald Wald F -test and the
Kleibergen-Paap Wald rk F -test; both have the null hypothesis that the estimated
equation is weakly identified. The corresponding test statistic is based on the rejec-
tion rate r that is to be tolerated if the true rejection rate is the standard 5 percent.
In this context, weak instruments are defined as instruments that lead to a rejection
rate of r when the true rejection rate is 5 percent. The results are given in Appendix
Table A.3 and Table A.4, which we compare against the critical value tables in Stock
and Yogo (2005). The rejection rate r of 10 percent (with a critical value lower than
16.87) for all the TSLS regressions that we have run at the 5 percent significance level
implies that our estimations do not suffer from a weak identification problem.
Regarding weak-instrument-robust inference, we employ the Anderson-Rubin Wald
test (based on both F -statistic and chi-square statistic) in order to check whether the
estimated coefficients of the endogenous variables are compatible with the our sample,
independent of the strength of our instruments. The corresponding null hypothesis is
that the estimated coefficients of the endogenous variables are jointly equal to zero.
We reject the null hypothesis for all the TSLS regressions that we have run at the
1 percent significance level, independent of the statistic considered, and we conclude
that our estimated coefficients are compatible with the data used, independent of the
strength of instruments used. Hence the empirical results in Tables 4-5 are robust to
the consideration of any endogeneity problem, since all the employed statistical tests
suggest evidence for the validity of the instruments used in TSLS regressions.

22
 In order to gain a better understanding of the correlation between TPD measures
and income growth rate, consider Table 6 where we document the magnitude of the
change in a country’s annual growth rate associated with one standard deviation in-
crease in a given network measure. Using network measure coefficients of Table 4
between columns (2) and (4), we find that one standard deviation increase in a net-
work measure is associated with a 0.56 percentage point increase in the annual growth
rate, which corresponds to 0.21 of one unit standard deviation in the growth rate. Co-
efficients of network measures after controlling for trade openness (shown in columns
(5) to (8) of Table 4) reveal that one standard deviation increase in network measures
is associated with about a 0.52 - 0.56 percentage point increase in the annual growth
rate, corresponding to about 0.2 of one unit standard deviation in the growth rate.
Inclusion of country-fixed effects leads to higher coefficient estimates for network mea-
sures (shown in Table 5), and the associated increase in the annual growth rate with
one standard deviation increase in network measures turns out to be about 1 to 1.5
percentage points, corresponding to 0.37 to 0.56 of one unit standard deviation in the
growth rate. One standard deviation increase in trade openness, on the other hand,
is associated with about a 0.34 - 0.39 percentage point increase in the annual growth
rate. This shows the importance of having access to well-connected export markets
and improving a country’s trade networks is not being captured by trade openness
measure.
The time it takes to reach one standard deviation increase in export degree takes
on average 22 years, but the variation across countries is also large, e.g. Malawi and
New Zealand reach this milestone in 1971 whereas Greece and Guyana reach it in 1980,
Turkey in 1982, Chile and Ghana in 1992, Rwanda reaches it in 2002. One standard
deviation increase in export closeness and eigenvector take on average 23 years and
11 years, respectively. While 61 countries in our sample have been able to increase
their export degree or export closeness by at least one standard deviation during the
sample period, 23 countries were able to increase their export eigenvector by at least
one standard deviation.

Threshold Analysis. It is important to note that all of the results shown so far
represent on-average effects across countries. However, we would also like to know how

23
Table 6: Change in Growth in Response to One Standard Deviation Increase in Trade Network Measures
By how many percentage points does the annual growth rate change when controlled for:
ALL: Finance (M3),
ALL, country fixed
Inflation, Human ALL and country fixed
ALL and trade openness effects and trade
Capital (SEC), Gov. effects
openness
spending, initial GDP

Export Degree 0.567 0.516 1.495 1.341

Export Closeness 0.564 0.521 1.409 1.254

Export Eigenvector 0.562 0.555 1.023 0.977

Source: Standard deviations of export degree, export closeness, and export eigenvector are reported in table 3. These are multiplied by
estimated coefficients taken from regression results in tables 4 and 5.

different country characteristics would change the effects of trade networks on growth
across countries, which we achieve next. This is in line with Herzer (2013) who shows
that effects of trade on income have a large heterogeneity across countries. Therefore,
a given level of connectivity in the international trade network may also have very
different effects on countries that have very different underlying macroeconomic con-
ditions. Accordingly, in order to have a systematic explanation of such heterogeneity,
we would like to know for which countries larger and/or better trade networks are
associated with higher income growth. For example, in the related literature, Aghion
et al. (2009) show that exchange rate volatility can stunt growth of a country when
its financial development is low. Since macroeconomic volatilities, especially exchange
rate volatility, constitute an important channel through which trade networks may
affect income growth (as discussed in Section 2 in detail), one would expect superior
trade network connections to be more important for a country where financial de-
velopment is low. Similarly, Gylfason (1999) suggests that inflation adversely affects
growth through lower ratios of exports to output; therefore, one would expect trade
networks to be more important for countries experiencing high inflation. Finally, since
trade openness increases income volatility according to Svaleryd and Vlachos (2002),
one would expect trade networks to be more important for countries with high trade
openness because high trade openness comes with an increased risk of being exposed
to external shocks, and being connected to larger and/or better export markets may
provide good diversification of risk.
24
 Accordingly, in order to account for the heterogeneity across countries regarding the
effects of trade networks on income growth, we consider thresholds of the right-hand
side variables in the growth regressions. Following Yilmazkuday (2011), rolling-window
TSLS regressions are employed with a constant window size of 200 after ordering the
data according to the threshold variable.18 For instance, if the inflation thresholds were
of interest, all the observations (i.e., the pooled sample of 5-year average data from
all countries) are sorted in the order of the lowest to the highest inflation rates; the
first regression is run with the first 200 observations of the sorted data set, the second
regression by moving the 200 window toward higher inflation rates by one observation,
and so on. The selection of a constant window size is important for comparison of
coefficient estimates across the windows, while the selection of a window size of 200 is
important to ensure a fair distribution across the power of regressions and the degree
of nonlinearity.19
For a consistent inference across linear and nonlinear frameworks, the rolling-
window regressions use the specifications in columns 4-6 of Table 5. The rolling-window
regression results are given in Figures 6 to 11, where the x-axes show the median of
the threshold variable in 200 sample windows (i.e., the variable according to which all
the observations have been sorted). The y-axes of the figures in the left panel show the
coefficient estimates of the TPD measures. The bold solid lines show the coefficient
estimates, and the dashed lines the 10-percent confidence intervals. The dashed lines
in the right panels of figures 6 to 11 show the mean of R-bar squared values. As is
evident in Figures 6 to 9, the coefficient estimates of the TPD measures are positive

18
Earlier studies that have also used rolling-window regressions in the context of long-run growth
are Rousseau and Wachtel (2002,2011) and Yilmazkuday (2013). The use of rolling windows is
well established and has been considered in a variety of other contexts, including international and
monetary economics, exchange rates, and inflation modeling as well as returns predictability (e.g.,
Cushman, 1988; Guidolin et al., 2013; Meese and Rogoff, 1988; and Swanson, 1988), while Giacommi
and White (2006) provide general support for the use of rolling-window regressions.
19
It is important to emphasize that having a rolling-window regression corresponds to having
different residual characteristics across different windows of 200 observations. This is to our benefit
since it results in cluster-robust standard errors, where clusters depend on the level of the variable
of interest. Moreover, the main advantage of using rolling-window regressions is allowing the data
to speak, whereas other methods such as using interaction terms have the underlying assumptions of
linearity (of the derivatives) regarding the effects of TPD measures on long-run growth.
25
Figure 6: Effects of Trade Partner Diversification on Growth: Thresholds in Finance

Figure 7: Effects of Trade Partner Diversification on Growth: Thresholds in Inflation

26
Figure 8: Effects of Trade Partner Diversification on Growth: Thresholds in Human
Capital

and significant in subsets of countries where financial development is low, inflation is

high, human capital is low, and trade openness is high.20 Hence, positive effects of
TPD are more evident for such countries because from two countries that have, e.g.,
very similar low levels of financial development, the one with better TPD (that is, the
one with access to more/better export markets) has a significantly higher growth rate
than the other one. A brief look in our data reveals that many countries with low
financial development and high TPD maintain this combination over time, i.e. the
relationship of low financial development and high TPD seems to stay the same over
time. This points out that high TPD substitutes for low financial development rather
than supporting a country’s financial development growth. Figures 10 and 11 reveal
that such effects are relatively stable across alternative levels of government size and
level of income (log-initial GDP).
The rolling-window regression results have important implications for the stages of
economic development. Correlation between a country’s position in the international
trade network is positively and significantly correlated with the annual growth rate
in countries that display traits of developing or underdeveloped countries (low finan-
20
We observe several changes in the estimated coefficients based on the threshold variable; it is
simply implied that certain country characteristics play important roles in the determination of TPD
effects on long-run growth, supporting our nonlinear investigation.
27
Figure 9: Effects of Trade Partner Diversification on Growth: Thresholds in Trade
Openness

Figure 10: Effects of Trade Partner Diversification on Growth: Thresholds in Govern-

ment

28
Figure 11: Effects of Trade Partner Diversification on Growth: Thresholds in Initial
GDP

cial development and human capital) or experience macroeconomic volatility in the

domestic economy (high inflation). In particular, Figures 6 to 8 imply that a superior
position in the international trade network, which means having many trade partners
and/or having trade partners that themselves have superior positions in the trade net-
work, compensates countries in their early stages of development for their low levels
of financial depth, high levels of inflation, and low levels of human capital.
The results also yield an important policy implication for countries with high levels
of trade openness: TPD measures and growth are positively and significantly corre-
lated in countries that are exposed to external shocks due to their high trade openness.
Hence, of two similar highly open countries, the one with a superior position in the
international trade network has a significantly higher growth rate than the other one.
TPD compensates countries that have a high degree of trade openness for their vulner-
ability to international trade shocks through having access to superior export markets,
thereby providing better diversification of risk.

6 Conclusion
The relationship between trade and growth has been an important area of ongoing
research and policy discussion. We contribute to this literature by showing that a
29
country’s position in the international trade network enters per capita income growth
regressions positively and significantly on top of trade openness and other standard
control variables. In technical terms, when all other country characteristics are con-
trolled for, one standard deviation increase in network measures is associated with
about a 1 to 1.5 percentage point increase in the annual growth rate, depending on
the trade network measure used; this corresponds to about 0.37 to 0.56 of one unit
standard deviation of the growth rate. Moreover, from a welfare perspective, having
international trade connections with a larger and/or better connected set of countries
is positively and significantly correlated with higher rates of growth, ceteris paribus,
and should thus be considered important for achieving higher growth and development
levels.
We use network centrality measures to capture the position (and hence the impor-
tance) of a country in the web of international trade and call it the TPD of this country.
Since the TPD of a country increases based on quantity and quality of its export part-
ners, TPD yields an overall comparison between countries as to how diversified and
well connected they are (and their trade partners are) with the rest of the world. The
positive and significant effect of TPD on growth can best be understood once its po-
tentially important role in coping with uncertainty borne by country-specific shocks is
understood: being linked to diversified and well-connected trade partners provides an
important insurance against country-specific shocks that might disturb international
trade and thus affect growth.
Using threshold analyses, we also show that countries’ position and importance
in the international trade network can compensate for their low levels of financial
depth, high levels of inflation, and low levels of human capital. This result is especially
important for countries in their early stages of development where, on average, financial
depth is low, inflation is high, and human capital is low; therefore, gaining access into
more and/or better connected export markets turns out to be crucial. The results also
show that trade networks are effective for income growth through better diversification
of risk, especially for countries that are more vulnerable to external shocks due to their
high levels of trade openness.

30
 References

Albornoz, F., Pardo, H.F.C., Corcos, G., Ornelas, E., 2012. Sequential exporting. Journal
of International Economics. 88(1), 17-31.
Aghion, P., Bacchetta, P., Ranciere, R., Rogoff, K., 2009. Exchange rate volatility and
productivity growth: The role of financial development. Journal of Monetary Eco-
nomics. 56(4), 494-513.
Arribas, I., Pérez, F., Tortosa-Ausina, E., 2009. Measuring globalization of international
trade: theory and evidence. World Development. 37(1), 127-145.
Bailliu, J., Lafrance, R., Perrault, J.F., 2003. Does exchange rate policy matter for
growth? International Finance. 6(3), 381-414.
Balassa, B., 1985. Exports, policy choices, and economic growth in developing countries
after the 1973 oil shock. Journal of Development Economics. 18, 23-35.
Barro, R. J., 1991. Economic growth in a cross section of countries. Quarterly Journal
of Economics. 106, 407-443.
Bonacich, P., 1987. Power and centrality: a family of measures. American Journal of
Sociology. 92(5), 1170-1182.
Bourbakri, N., Cosset, J.-C., 1998. The financial and operating performance of newly
privatized firms: evidence from developing countries. Journal of Finance. 53(3), 1081-
1110.
Brock, W.A., Durlauf, S.N., 2001. Growth empirics and reality. The World Bank Eco-
nomic Review. 15(2), 229-272.
Cadot, O., Carrére, C., Strauss-Kahn, V., 2011. Export diversification: what’s behind
the hump? The Review of Economics and Statistics. 93(2), 590-605.
Cadot, O., Carrére, C., Strauss-Kahn, V., 2013. Trade diversification, income, and
growth: what do we know? Journal of Economic Surveys. 27(4), 790-812.
Costello, D., 1993. A cross-country, cross industry comparison of productivity growth.
Journal of Political Economy. 101, 206-222.
Cushman, D.O., 1988. U.S. bilateral trade flows and exchange risk during the floating
period. Journal of International Economics. 24, 317-330.
De Benedictis, L., Tajoli, L., 2011. The world trade network. The World Economy.
34(8), 1417-1454.

31
Edwards, S., 1992. Trade orientation, distortions, and growth in developing countries.
Journal of Development Economics. 39, 31-57.
Frankel, J.A., Romer, D., 1999. Does trade cause growth? American Economic Review.
89(3), 379-399.
Giacomini, R., White, H., 2006. Tests of conditional predictive ability. Econometrica.
74, 1545-1578.
Grossman, G.M., Helpman, E., 1993. Innovation and growth in the global economy. MIT
Press, Cambridge.
Guidolin, M., McMillan, D.G., Wohar, M.E., 2013. Time varying stock return predictabil-
ity: evidence from US sectors. Finance Research Letters. 10, 34-40.
Gylfason, T., 1999. Exports, inflation and growth. World Development. 27, 1031-1057.
Harrison, A., 1996. Openness and growth: a time series, cross-country analysis for de-
veloping countries. Journal of Development Economics. 48, 419-447.
Herzer, D., 2013. Cross-country heterogeneity and the trade-income relationship. World
Development. 44, 194-211.
Hitt, M.A., Tihanyi, L., Miller, T., Connelly, B., 2006. International diversification:
antecedents, outcomes, and moderators. Journal of Management. 32(6), 831-867.
Hoskisson, R.E., Hitt, M.A., 1990. Antecedents and performance outcomes of diversifica-
tion: a review and critique of theoretical perspectives. Journal of Management. 16(2),
461-509.
Irwin, D.A., Terviö, M., 2002. Does trade raise income? Evidence from the twentieth
century. Journal of International Economics. 58(1), 1-18.
Ito, K., 1997. Domestic competitive position and export strategy of Japanese manufac-
turing firms: 1971-1985. Management Science. 43(5), 610-622.
Johnson, R.C., 2014. Five facts about value-added exports and implications for macroe-
conomics and trade research. Journal of Economic Perspectives. 28(2), 119-142.
Kali, R., Méndez, F., Reyes, J., 2007. Trade structure and economic growth. Journal of
International Trade and Economic Development. 16(2), 245-269.
Kali, R., Reyes, J., 2007. The architecture of globalization: a network approach to
international economic integration. Journal of International Business Studies. 38(4),
595-620.
Kali, R., Reyes, J., McGee, J., Shirrell, S., 2013. Growth networks. Journal of Develop-

32
 ment Economics. 101, 216-227.
Karadimitropoulou, A., León-Ledesma, M., 2013. World, country, and sector factors in
international business cycles. Journal of Economic Dynamics and Control. 37(12),
2913-2927.
Leamer, E.E., 1988. Measures of openness, in: Baldwin, R.E. (Ed.), Trade policy issues
and empirical analysis. The University of Chicago Press, Chicago, pp.147-204.
Levine, R., 1997. Financial development and economic growth: views and agenda. Jour-
nal of Economic Literature. 35, 688-726.
Levine, R., Renelt, D., 1992. A sensitivity analysis of cross-country growth regressions.
American Economic Review. 82, 942-963.
Meese, R.A., Rogoff, K., 1988. Was it real? The exchange rate-interest differential
relation over the modern floating-rate period. Journal of Finance. 43, 933-948.
Meyer, K.E., 2006. Global focusing: from domestic conglomerates to global specialists.
Journal of Management Studies. 43(5), 1109-1144.
Newman, M.E.J., 2010. Networks: An introduction. Oxford University Press.
Rivera-Batiz, L., Romer, P.M., 1991. Economic integration and endogenous growth.
Quarterly Journal of Economics. 106(2), 531-555.
Rodrik, D., 1998. Why do more open economies have bigger governments? Journal of
Political Economy. 106(5), 997-1032.
Romer, P., 1990. Endogenous technological change. Journal of Political Economy. 98(5),
2.
Rose, A., Spiegel, M., 2011. The olympic effect. The Economic Journal. 121(553) 652-
677.
Rousseau, P.L., Wachtel, P., 2002. Inflation thresholds and the finance-growth nexus.
Journal of International Money and Finance. 21, 777-793.
Rousseau, P.L., Wachtel, P., 2011. What is happening to the impact of financial deepen-
ing on economic growth? Economic Inquiry. 49(1), 276-288.
Serrano, M.A., Boguña, M., 2003. Topology of the world trade web. Physical Review E.
68, 015101.
Stock, J.H., Yogo, M., 2005. Testing for weak instruments in linear IV regression, in:
Andrews, D.W.K., Stock, J.H. (Eds.), Identification and Inference for Econometric
Models: Essays in Honor of Thomas Rothenberg. Cambridge University Press, pp.80-

33
 108.
Svaleryd, H., Vlachos, J., 2002. Markets for risk and openness to trade: how are they
related? Journal of International Economics. 57, 369-395.
Swanson, N.R., 1988. Money and output viewed through a rolling window. Journal of
Monetary Economics. 41, 455-474.
Syrquin, M., Chenery, H., 1989. Three decades of industrialization. The World Bank
Economic Review. 3(2), 145-181.
Vickers, J., Yarrow, G., 1991. Economic perspectives on privatization. Journal of Eco-
nomic Perspectives. 5(2), 111-132.
Wacziarg, R., Welch, K.H., 2008. Trade liberalization and growth: new evidence. The
World Bank Economic Review. 22(2), 187-231.
Yilmazkuday, H., 2011. Thresholds in the finance-growth nexus: a cross-country analysis.
The World Bank Economic Review. 25(2), 278-295.
Yilmazkuday, H., 2013. Inflation thresholds and growth. International Economic Journal.
27(1), 1-10.

34
Table A.1: OLS growth regressions, 1965-2004, 83 Countries
Dependent Variable: Growth of Per Capita Income (%)
Variables (1) (2) (3) (4) (5) (6) (7)

Export Degree 0.0202*** 0.0196***

(0.00580) (0.00572)

Export Closeness 0.0357*** 0.0339***

(0.0101) (0.0100)

Export Eigenvector 0.167*** 0.169***

(0.0416) (0.0410)

Trade Openness 0.0395*** 0.0384*** 0.0378*** 0.0400***

(0.0102) (0.0101) (0.0101) (0.0101)

Finance (M3, % GDP) 0.0188*** 0.0165*** 0.0170*** 0.0167*** 0.0151** 0.0157** 0.0151**
(0.00479) (0.00491) (0.00488) (0.00485) (0.00485) (0.00482) (0.00479)

Inflation (%) -0.0337*** -0.0283** -0.0287** -0.0298*** -0.0308*** -0.0311*** -0.0323***

(0.00879) (0.00884) (0.00882) (0.00877) (0.00873) (0.00873) (0.00865)

Government (% GDP) -0.0829*** -0.0797*** -0.0797*** -0.0775*** -0.0791*** -0.0792*** -0.0767***

(0.0234) (0.0235) (0.0234) (0.0234) (0.0231) (0.0231) (0.0230)

Log of initial GDP -0.201+ -0.336** -0.328** -0.324** -0.328** -0.318** -0.322**
(0.118) (0.124) (0.123) (0.121) (0.122) (0.121) (0.119)

Log of initial SEC (%) 1.071*** 0.927*** 0.968*** 0.840*** 0.951*** 0.992*** 0.858***
(0.212) (0.215) (0.213) (0.218) (0.212) (0.211) (0.214)

R-bar Squared 0.271 0.267 0.268 0.273 0.288 0.288 0.297

Source: Authors' analysis based on data sources discussed in the text.

Notes: +, *, ** and *** indicate significance at the 10%, 5%, 1% and 0.1% levels, respectively. Standard errors are in parentheses.
Growth rates are five-year averages. Network measures are five-year averages; their initial values in each five-year period are used as
instruments for the corresponding five-year averages. All equations also include time fixed effects that are not shown. Estimation is by
OLS. The sample size in each equation is 464.

35
Table A.2: OLS growth regressions with country-fixed effects, 1965-2004, 83 Countries
Dependent Variable: Growth of Per Capita Income (%)
Variables (1) (2) (3) (4) (5) (6) (7)

Export Degree 0.0432*** 0.0383**

(0.0118) (0.0116)

Export Closeness 0.0737*** 0.0660***

(0.0173) (0.0170)

Export Eigenvector 0.235** 0.231**

(0.0754) (0.0733)

Trade Openness 0.0435*** 0.0406*** 0.0398*** 0.0433***

(0.00924) (0.00916) (0.00912) (0.00913)

Finance (M3, % GDP) -0.0125 -0.00280 -0.00495 -0.00110 -0.0113 -0.0131 -0.0101
(0.0108) (0.0108) (0.0107) (0.0109) (0.0107) (0.0106) (0.0107)

Inflation (%) -0.0251* -0.0220* -0.0234* -0.0212+ -0.0260* -0.0271* -0.0256*

(0.0109) (0.0110) (0.0110) (0.0111) (0.0108) (0.0107) (0.0108)

Government (% GDP) -0.0948* -0.0908* -0.100* -0.0785+ -0.0899* -0.0986* -0.0771+

(0.0409) (0.0414) (0.0412) (0.0420) (0.0404) (0.0402) (0.0408)

Log of initial GDP -2.698*** -3.490*** -3.378*** -3.034*** -3.270*** -3.179*** -2.875***
(0.568) (0.600) (0.583) (0.579) (0.587) (0.571) (0.564)

Log of initial SEC (%) -0.894* -0.966* -0.966* -1.127** -0.946* -0.947* -1.110**
(0.405) (0.410) (0.407) (0.418) (0.400) (0.398) (0.406)

R-bar Squared 0.424 0.411 0.418 0.405 0.439 0.445 0.437

Source: Authors' analysis based on data sources discussed in the text.

Notes: +, *, ** and *** indicate significance at the 10%, 5%, 1% and 0.1% levels, respectively. Standard errors are in parentheses.
Growth rates are five-year averages. Network measures are five-year averages; their initial values in each five-year period are used as
instruments for the corresponding five-year averages. All equations also include country fixed effects and time fixed effects that are
not shown. Estimation is by OLS. The sample size in each equation is 464.

36
 Table A.3: Validity of Instruments in TSLS regressions in Table 4

Table 4 Table 4 Table 4 Table 4 Table 4 Table 4 Table 4

Column (1) Column (2) Column (3) Column (4) Column (5) Column (6) Column (7)

Under-identification tests

Kleibergen-Paap rk LM statistic Chi-sq(1)= 33.86 Chi-sq(1)=32.87 Chi-sq(1)=33.21 Chi-sq(1)=33.44 Chi-sq(1)=33.02 Chi-sq(1)=33.28 Chi-sq(1)=33.75
[0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000]

Weak identification tests

Cragg-Donald Wald F statistic 194.74 183.65 184.00 188.46 150.23 150.62 153.96

37
Kleibergen-Paap Wald rk F statistic 13.79 38.50 38.95 39.42 32.10 32.49 32.88

Weak-instrument-robust inference

Anderson-Rubin Wald test F(4,450)=7.98 F(4,450)=9.67 F(4,450)=9.81 F(4,450)=9.81 F(4,449)= 8.47 F(5,449)=8.47 F(5,449)=8.96
[0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000]

Anderson-Rubin Wald test Chi-sq(4)=32.90 Chi-sq(4)=39.88 Chi-sq(4)=40.47 Chi-sq(4)=40.46 Chi-sq(5)=43.78 Chi-sq(5)=43.74 Chi-sq(5)=46.29
[0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000]

p values in brackets.

Under-identification tests, Ho: matrix of reduced form coefficients has rank=K1-1 (under-identified); Ha: matrix has rank=K1 (identified).
Weak identification test, Ho: equation is weakly identified.
Weak-instrument-robust inference, Tests of joint significance of endogenous regressors B1 in main equation, Ho: B1=0 and over-identifying restrictions are valid.
 Table A.4: Validity of Instruments in TSLS regressions in Table 5

Table 5 Table 5 Table 5 Table 5 Table 5 Table 5 Table 5

Column (1) Column (2) Column (3) Column (4) Column (5) Column (6) Column (7)

Under-identification test

Kleibergen-Paap rk LM statistic Chi-sq(1)=35.41 Chi-sq(1)=37.81 Chi-sq(1)=36.36 Chi-sq(1)=36.55 Chi-sq(1)=37.42 Chi-sq(1)=35.68 Chi-sq(1)=35.91

[0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000] [0.0000]

Weak identification tests

Cragg-Donald Wald F statistic 90.09 86.06 86.78 86.92 71.20 71.83 72.08

38
Kleibergen-Paap Wald rk F statistic 21.91 22.27 20.84 21.11 19.11 17.72 17.92

Weak-instrument-robust inference

Anderson-Rubin Wald test F(4,372)=3.32 F(4,372)=4.00 F(4,372)=4.22 F(4,372)=2.57 F(5,371)= 5.71 F(5,371)=5.89 F(5,371)=4.97
[0.0129] [0.0034] [0.0023] [0.0378] [0.0000] [0.0000] [0.0002]

Anderson-Rubin Wald test Chi-sq(4)=16.57 Chi-sq(4)=19.95 Chi-sq(4)=21.08 Chi-sq(4)=12.81 Chi-sq(5)=35.72 Chi-sq(5)=36.86 Chi-sq(5)=31.07
[0.0023] [0.0005] [0.0003] [0.0122] [0.0000] [0.0000] [0.0000]

p values in brackets.

Under-identification tests, Ho: matrix of reduced form coefficients has rank=K1-1 (under-identified); Ha: matrix has rank=K1 (identified).
Weak identification test, Ho: equation is weakly identified.
Weak-instrument-robust inference, Tests of joint significance of endogenous regressors B1 in main equation, Ho: B1=0 and over-identifying restrictions are valid.