Modeling the International-Trade Network:

A Gravity Approach
Marco Due
nas

Giorgio Fagiolo

December 2011

Abstract
This paper investigates whether the gravity model (GM) can explain the statistical properties
of the International Trade Network (ITN). We fit data on international-trade flows with a
GM specification using alternative fitting techniques and we employ GM estimates to build
a weighted predicted ITN, whose topological properties are compared to observed ones. Furthermore, we propose an estimation strategy to predict the binary ITN with a GM. We find
that the GM successfully replicates the weighted-network structure of the ITN, only if one
fixes its binary architecture equal to the observed one. Conversely, the GM performs very
badly when asked to predict the presence of a link, or the level of the trade flow it carries,
whenever the binary structure must be simultaneously estimated.
Keywords: International Trade Network; Gravity Equation; Weighted Network Analysis;
Topological Properties; Econophysics.
JEL Classification: F10, D85.

Institute of Economics, SantAnna School of Advanced Studies, Pisa, Italy. Mail address: SantAnna School of
Advanced Studies, Pisa, Italy.Email: m.duenasesterling@sssup.it

Corresponding Author. Institute of Economics, SantAnna School of Advanced Studies, Pisa, Italy. Mail
address: SantAnna School of Advanced Studies, Piazza Martiri della Libert`a 33, I-56127 Pisa, Italy. Tel: +39-050883282. Fax: +39-050-883344. Email: giorgio.fagiolo@sssup.it

Introduction

The International Trade Network (ITN), aka World-Trade Web (WTW) or World Trade Network
(WTN), is defined as the graph representing in each year the web of bilateral-trade relationships
between countries in the World. The statistical properties of the ITN, and their evolution over
time, have been recently received a lot of attention in a number of contributions.1
Understanding the topology of the ITN is important for two related reasons. First, trade
is one of the most important channels of interaction among countries (Helliwell and Padmore,
1985; Krugman, 1995; Galvandatilde et al., 2007; Forbes, 2002). The knowledge of macroeconomic
phenomena such as economic globalization and internationalization, the spreading of international
crises, and the transmission of economic shocks, may be improved by looking at international-trade
patterns in a holistic framework, where indirect as well as direct linkages between countries are
explicitly taken into consideration (Fagiolo, 2010).2 Second, ITN topological properties can help
to statistically explain macroeconomics dynamics. For example, Kali et al. (2007) and Kali and
Reyes (2010) have shown that country position in the trade network has substantial implications for
economic growth and a good potential for predicting episodes of financial contagion. Furthermore,
Reyes et al. (2010) suggest that country centrality in the ITN may help to account for the evolution
of international economic integration better than what standard statistics, like openness to trade,
do.
The statistical properties of the ITN, in its undirected/directed or binary/weighted characterizations, have been extensively studied and today we know a great deal about the topological
architecture of the web of international-trade flows. For example, Serrano and Bogu
na (2003) and
Garlaschelli and Loffredo (2004) show that the binary-directed representation of the ITN exhibits
a disassortative pattern: countries with many trade partners (i.e., high node degree) are on average
connected with countries with few partners (i.e., low average nearest-neighbor degree). Furthermore, partners of well connected countries are less interconnected than those of poorly connected
ones, implying some hierarchical arrangements. Remarkably, Garlaschelli and Loffredo (2005)
show that this evidence is quite stable over time. This casts some doubts on whether economic
integration (globalization) has really increased in the last 30 years. Furthermore, node-degrees
appear to be very skewed, implying the coexistence of few countries with many partners and many
countries with only a few partners.
These issues are taken up in more detail in a few subsequent studies adopting a weighted1

See for example Li et al. (2003); Serrano and Bogu

na (2003); Garlaschelli and Loffredo (2004, 2005); Garlaschelli
et al. (2007); Serrano et al. (2007); Bhattacharya et al. (2007, 2008); Fagiolo et al. (2008, 2009); Reyes et al. (2008);
Fagiolo et al. (2010); Fagiolo (2010); Barigozzi, Fagiolo and Garlaschelli (2010); Barigozzi, Fagiolo and Mangioni
(2010); De Benedictis and Tajoli (2011).
2
For example, Abeysinghe and Forbes (2005) show that bilateral trade can only explain a small fraction of the
impact that an economic shock originating in a given country can have on another one, which is not among its
direct-trade partners. Similarly, Dees and Saint-Guilhem (2011) report that countries that do not trade very much
with the U.S. are largely influenced by its dominance over other trade partners linked with the U.S..

network approach to the study of the ITN. The motivation is that a binary approach, by treating all
relationship equally, might dramatically underestimate the impact of trade-linkage heterogeneity.
This seems indeed to be the case: Fagiolo et al. (2008, 2009, 2010) find that the statistical properties
of the ITN viewed as a weighted undirected network crucially differ from those exhibited by its
binary counterpart. For example, the strength distribution is highly right-skewed, indicating that
a few intense trade connections co-exist with a majority of low-intensity ones. This confirms the
results obtained by Bhattacharya et al. (2007) and Bhattacharya et al. (2008), who find that the
size of the group of countries controlling half of the worlds trade has decreased in the last decade.
Furthermore, weighted-network analyses show that the ITN architecture has been extremely stable
in the 1981-2000 period and highlights some interesting regularities (Fagiolo et al., 2009). For
example, countries holding many trade partners and/or very intense trade relationships are also
the richest and most globally central; they typically trade with many partners, but very intensively
with only a few of them, which turn out to be very connected themselves; and form few but
intensive-trade clusters (i.e., triangular trade patterns).
Most of existing network literature on the ITN, however, has been focusing on a purely empirical
quest for statistical properties, largely neglecting the issue of exploring whether theoretical models
are able to explain why the ITN is shaped the way it is.3
This paper is a preliminary attempt to fill this gap. We extend the work in Fagiolo (2010)
to ask whether the gravity model (GM) can provide a satisfactory theoretical benchmark able to
reproduce the observed architecture of the ITN across time. The GM (van Bergeijk and Brakman,
2010) aims at explaining international-trade bilateral flows using an equation obtained as the
equilibrium prediction of a large family of micro-founded models of trade (more on that in Section
2). The term gravity comes about because the predicted relation between trade flows and
explanatory variables is similar to Newtons formula: the magnitude of aggregated trade flows
between a pair of countries is proportional to the product of country sizes (e.g. the masses, as
proxied by country GDPs) and inversely proportional to their geographic distance (interpreted
as proxies of trade-resistance factors, e.g. tariffs). From an econometric perspective, the original
model-driven prediction can be augmented with a set of country-specific explanatory variables
(e.g., population, area, land-locking effects, etc.), as well as with a set of bilateral variables (i.e.,
geographical contiguity, common language and religion, colony relation, bilateral trade agreements,
etc.). The GM can be fitted to the data using different econometric techniques, ranging from
simple ordinary least squares (OLS) applied to the log-linearized equation, to two-stage Poisson
estimations, employed to correctly deal with the large number of zero trade flows characterizing
the data. Overall, the GM is very successful: independent on the technique employed, it typically
achieves a very high goodness of fit, e.g. in terms of R-squared coefficients.
3

See Bhattacharya et al. (2008) and Garlaschelli and Loffredo (2004) for exceptions. See also Squartini et al.
(2011a,b) for an alternative approach employing null random models that are able to predict whether observed
properties of the ITN are statistically meaningful or simply the result of constrained randomness.

Motivated by the well-known empirical success of the GM, we fit data on bilateral trade flows
to build a GM-predicted weighted-directed representation of the ITN, which we then compare to
the observed one, constructed using original bilateral-flow data. We employ both a static and a
dynamic approach. In the static approach, we assume that a GM holds in each subsequent year
and we estimate a series of predicted ITN snapshots. In the dynamic approach, we control for
time dummies in the estimation to account for change over time and get a unique predicted ITN
from the unbalanced panel of predicted flows. In both cases, we end up with a prediction for
the expected bilateral-trade flow occurring between any two countries in a given year, and for the
probability that a binary link is in place. We complement this information with standard errors
of predicted values, so as to evaluate the precision of GM-based estimated quantities.
In a nutshell, our results suggest that the GM well predicts weighted ITN properties only
when the binary structure is kept fixed, equal to the observed one. Conversely, the performance
of the GM is very poor when asked to predict ITN weighted properties together with its binary
architecture, or when one employs a GM specification to estimate the presence of a link only.
The rest of the paper is organized as follows. Section 2 discusses the gravity model and presents
data and related methodologies. Our main results are reported in Section 3. Finally, Section 4
concludes and flags some of the challenges facing ITN modeling in the future.

Data and Methodology

2.1

Bilateral Trade-Flow Data

We use international-trade data taken from Subramanian and Wei (2003), which contains aggregate
bilateral imports reported by the IMF Direction of Trade Statistics, measured in U.S. dollars and
deflated by U.S. Consumer Price Index at 1982-83 prices. We focus on seven unbalanced crosssections for the years 1970 to 2000, with a five-year lag. Let wij (t) be exports from country i to
country j in year t and let N (t) the correspondent number of countries reporting at least a positive
flow.
Table 1 summarizes some descriptive statistics. The number of participating countries and
average per-country trade both increase over time. Entry of new countries in the database may
be possibly caused either by the availability of new data or by the actual entry of the country in
international-trade markets. New trade links, however, seem to increase more than quadratically
with the number of participating countries in the last part of the sample, as testified by the rising
density.4 Note also that the number and percentage of countries making up 50% of total trade
seem to remain stable across the years, hinting to a stable core of top traders. Conversely, the
4

Defined as the ratio between L(t) (existing trade partnerships) and N (t) [N (t) 1] (all possible trade partnerships).

percentage of countries controlling 90% of total world trade has substantially decreased. The
concentration process going on in the ITN, despite globalization and international integration, is
confirmed also by the decrease in both the number and percentage of flows making up a certain
share of total trade.
Given wij (t) and N (t), we build weight matrices for the correspondent observed trade networks.
More precisely:
Definition 1 (Observed Weighted ITN) The observed weighted International Trade Network
in a given year t is represented by a weighted-directed graph, where the nodes are the N (t) countries
and link weights are fully characterized by the N (t) N (t) asymmetric matrix W (t), with entries
wij (t), i.e. exports from country i to country j.
Similarly, one can define the observed binary ITN, where links represent import-export partnerships, as:
Definition 2 (Observed Binary ITN) The observed binary International Trade Network in a
given year t is represented by a binary-directed graph, where the nodes are the N (t) countries and
binary links are fully characterized by the N (t) N (t) asymmetric adjacency matrix A(t), with
entries aij (t) = 1 if and only if wij (t) > 0, i.e. exports from country i to country j are strictly
positive.
This database has been studied from a binary/weighted network perspective in De Benedictis
and Tajoli (2011). They show that international integration in trade has been increasing over time,
but it is still far from being fully accomplished. Indeed, a strong heterogeneity in the profiles of
across-country trade partnerships does emerge. This has important implications for both the role
of regional trade agreements (i.e., the WTO) and the interplay between extensive and intensive
margins of trade (Felbermayr and Kohler, 2006).
In this paper, we take an alternative approach. We characterize the topological properties of
the observed ITN and we compare them to the properties displayed by the gravity-based predicted
ITN, which we define in the next sub-sections.

2.2

Gravity-Model Specifications

The GM, independently proposed by Tinbergen (1962) and Poyhonen (1963), is the workhorse
model to explain bilateral trade flows among countries as a function of import and export market
sizes (i.e., GDP) and trade-resistance factors, proxied by geographical distance. The GM derives
its name from the functional form linking trade to size and distance, which resembles the expression
for the attraction force between two bodies derived by Isaac Newton in classical mechanics. Thus,

in analogy with the physics law, it is expected that trade flows increase with the product of some
power of country sizes and decrease with some power of geographical distance.
This empirically-inspired law has been found to be consistent with a number of theoretical
foundations (Anderson, 1979; Bergstrand, 1985; Deardorff, 1998; Anderson and van Wincoop,
2003). In other words, many possibly-conflicting micro-foundations can generate has their equilibrium outcome some gravity-like relation between trade, market sizes and trade-resistance terms.5
For example, a gravity-like equation can be derived in trade specialization models, monopolisticcompetition frameworks with intra-industry trade, or Hecksher-Ohlin models (see Fratianni, 2009;
De Benedictis and Taglioni, 2011, for comprehensive surveys).
Notwithstanding the preferred micro-founded explanation, modern empirical interpretations of
the gravity expression generalize the original idea including in the formulation a list of additional
explanatory variables, covering aspects related to geography, culture, bilateral trade agreements,
among others. In Table 2 we report the list of explanatory variables that, following existing
literature (see, e.g., Glick and Rose, 2001; Rose and Spiegel, 2002), we employ in our exercises. GM
explanatory variables can be typically grouped in country- or link-specific ones. The former include,
in addition to GDP, other country-size proxies like population and geographical area, as well as
geographically-related aspects controlling for land-locking effects and continent membership. The
latter instead include relational variables characterizing bilateral relationships, as geographical
contiguity, colonial ties, regional trade agreements, commonalities in language, colonial history,
religion, and currency. Together, these factors have been shown to successfully explain, in a way
or in the other, international-trade flows in gravity-equation econometric exercises (van Bergeijk
and Brakman, 2010).
The most general GM specification that we employ in what follows then reads:
wij (t) = 0 Yi (t)
exp

H
X
h=1

Yj (t)2 dij3

h Dijh (t) +

"

K
Y

k=1
L
X

Xik (t)

1k

Xjk (t)

2k

(1)

(1l Zil + 2l Zjl ) ij (t),

l=1

where t is the year (t = 1950, 1955, . . . , 2000); wij (t) are export flows from the observed weighted
ITN; i, j = 1, ..., N (t), i 6= j; Yh (t) is year-t GDP of country h = i, j (i=exporter; j=importer); dij
is geographical distance; Xh (t), h = i, j, are additional country-size effects (area and population);
Dij is a vector of bilateral-relationship variables (contiguity, common language, past and current
colonial ties, common religion, common currency, a dummy to control if both countries share a
generalized system of preferences, and a regional trade agreement flag); Zi and Zj are country5

This has led Deardorff (1998) to argue that just about any plausible model of trade would yield something
very like the gravity equation. See also Evenett and Keller (2002).

specific dummies (controlling for land-locking effects and continent membership); finally, ij (t) are
the errors (whose mean conditional to explanatory variables obeys E[ij (t)|] = 1).
Two remarks are in order. First, note that standard GM specifications assume that any pair
of countries trade. In other words, zero trade flows, which are quite frequent in trade data either
because of missing values or because the two countries are not trade partners (see Table 1), are
ruled out from the analysis by the non-linear functional form employed. Therefore, one either loglinearizes Eq. (1) (and excludes observed zero-trade flows) or explicitly deals with over-estimation
errors coming from positive GDPs and other country and bilateral variables. In other words, the
standard GM specification is not suited to address the issue why any pair of countries that were
previously not trading start to trade at some point, or why existing trade relationships terminate.
We shall get back to this point in Section 3.2.
Second, and more importantly, we employ a GM specification that slightly differ from Anderson and van Wincoop (2003) one, which is one of the most commonly used in GM exercises.
Anderson and van Wincoop (2003) introduce multilateral resistance terms and importer-exporter
fixed effects. Formally, that approach considers that the constant term of the equation (1) must be
generalized to a set of importer and exporter dummies. One important implication is that countrysize effects are captured by country dummies. This means that characteristics of exporters and
importers cannot be generalized (Santos Silva and Tenreyro, 2006). In any case, all our results
are robust to Anderson and van Wincoops specification. We have therefore chosen to retain the
traditional specification because of its more immediate empirical interpretation.

2.3

Estimation

Estimation of Eq. (1) is not easy. A straightforward approach consists in log-linearizing the GM
specification and apply standard OLS techniques to estimate parameters and obtain predicted
values. The existing empirical literature on GM has largely employed this approach (Cf. for
example Glick and Rose, 2001; Rose and Spiegel, 2002).
However, a series of more recent contributions highlighted the risk of biases in estimation
induced by OLS applied to log-linear specifications. The main sources of bias come from the treatment of zero-valued flows (Santos Silva and Tenreyro, 2006; Linders and de Groot, 2006; Burger
et al., 2009), non-linearity and heteroscedasticity (Santos Silva and Tenreyro, 2006), endogeneity
and omitted-term (Baldwin and Taglioni, 2006). In particular, the issue of zero-flow treatment is
particularly relevant to our analysis. Indeed, log-linearizing the GM equation and applying OLS
estimation implies using non-zero trade flows only in the estimation. In network terms, this means
that we are keeping the observed binary structure constant (i.e. we are conditioning on adjacency
matrices A(t)). This is a serious issue if one wants to estimate the presence of a link together with
its weight.
To properly account for all these potential difficulties, we can resort to count-data analysis
7

(Long, 1997) and fit to the data Poisson pseudo-maximum likelihood models (PMML), either
in their standard formulation (Santos Silva and Tenreyro, 2006) or in zero-inflated specifications
(Linders and de Groot, 2006). In a nutshell, PPML models allow to estimate Eq. (1) in its original
non-linear form, thus avoiding possible correlation between errors and regressors. PPML models
use a Poisson distribution to model simultaneously the probability of a zero flow and of a positive
(integer) flow. However, it has been noticed that, in the case of international trade, zero flows occur
much more frequently than a plain Poisson model would predict (Burger et al., 2009), cf. also Table
1. This has led to the family of zero-inflated (ZI) models (Winkelmann, 2008). The underlying idea
is to model the presence of zeros and positive values as a two-stage process. In this way one treats
differently the process of presence-absence of trade partnerships from link-weight determination.
In the first stage, one estimates zero-flow probabilities using a standard logit model, and employing
a series of regressors that often coincide with those used in the standard GM formulation. In the
second stage, conditionally to having non-zero flows, one estimates the magnitude of trade-flow
values using either a Poisson (ZIP) or a negative-binomial (ZINB) distribution. Notice that in
the second stage there is a non-zero probability of having a zero flow, as the process governing
link-weight value determination may attach a zero flow independently on what the first process
has done.
To double check our results, we have applied a full range of models to estimate Eq. (1). In
particular, we have employed standard OLS, PPML, ZIP and ZINB approaches. Furthermore, in
order to control for dynamic effects, we have estimated Eq. 1 using both a cross-section perspective
(i.e., fitting a separate model for each of the 7 waves we end up with in our database) and an
unbalanced panel-data approach (i.e., adding time dummies and estimating once and for all the
entire data set). We have also controlled for country fixed effects as suggested in Baldwin and
Taglioni (2006). Our results turn out to be very robust to all these alternatives. Therefore in this
paper, to avoid redundancy, we report only results from three sets of models (OLS, PPML and
ZIP), where a sequence of independent cross sections is estimated without country fixed effects.6
By doing so, we are able to compare a setup where the binary structure of the ITN is kept fixed
(OLS) with two alternative setups (PPML and ZIP) where instead one estimates the probability
that a link is in place or not, correcting or not for the zero-inflation effect, i.e. when adjacency
matrices are endogenous.
Table 3 presents estimation results for year 2000 (similar results hold also for the remaining
years) for OLS, PPML and ZIP. Note that, by and large, both signs and orders of magnitude of
estimated coefficients do not change with the estimation technique employed. All coefficients have
6

Indeed, ZINB estimates turn out to be very similar to ZIP ones. No dramatic differences are detected between
cross-section and panel-data analyses. Similarly, the introduction of country fixed effects do not alter our results
below in any crucial ways. Note also that we employ the same set of regressors in both stages of ZIP and ZINB
estimates, as listed in Table 2. Reducing the set of regressors in the first stage does not dramatically change our
main results. The whole set of estimation results is available from the authors upon request.

the expected signs, although there are some relevant differences in estimated-coefficient values
across methods. For example, OLS estimates for GDP and distance coefficients are quite different
from PPML and ZIP ones. The geographical distance coefficient in OLS is negative and stronger
than in PPML/ZIP. This may depend on the fact that OLS estimates are computed on a smaller
sample (the number of positive observations is almost one half of the whole sample size). OLS
differ from PPML/ZIP ones also as far as the importer/exporter GDP elasticity is concerned. In
general, GDP elasticities tend to be larger than in other studies as we explicitly consider population
and area as additional size effects (entering with a negative sign). This hints to a relevant effect
played by per-capita GDP. Furthermore, variables as contiguity, common language, and regional
trade agreements enhance trade. In contrast, all variables related to colony relations, common
religion and common currency are statistically significant under the OLS models but not so much
for PPML.7
Note also that in (first-stage) logit estimation of the ZIP method, GDP (resp. distance) negatively (resp. positively) affect the probability of having unlinked countries, as expected. Conversely,
distance or land-locking effects enhance the probability of missing links. Contiguity coefficient is
instead positive: after controlling for geographical distance, sharing a border does not influence
the emergence of bilateral trade. This is however a result that does not hold robustly over all
cross-sections, where contiguity does not affect significantly the estimated probability.
Finally, all diagnostic statistics indicate that the estimated models are well-specified (Wooldridge,
2001) and achieve a quite good (pseudo) R2 .

2.4

The Predicted Weighted ITN

As long as Yi (t), dij and Xik (t) are strictly positive for all (i, j) and t, one can rewrite Eq. (1)8 as:
wij = exp{xij M }ij ,

(2)

where xij are logged country-specific and bilateral explanatory variables, and is the vector of all
parameters to estimate. Let M be estimated parameters with model M {OLS, P P M L, ZIP }.
In the OLS case, therefore, one can straightforwardly define a linear prediction for the log of
non-zero flows as:
OLS

ijOLS = log[wij
] = xij OLS ,
(3)
Note that we prefer to use logs of non-zero trade-flows to avoid over-dispersion issues. This means
that when comparing observed and OLS-predicted ITN properties we will always refer to logs of
7

Whenever a variable resulted not significant we decided to keep it among the regressors anyway to preserve
comparison between estimation techniques.
8
From now on, we suppress time labels for the sake of notational convenience and we refer to a cross-section
sequence of estimations.

2
non-zero flows to define link-weights. The variance of the prediction (
OLS
) equals the variance of
the model (i.e., the sum of squared residuals) divided by the degrees of freedom.
As far as PPML specification is concerned, the probability of observing a certain trade flow
is estimated using a Poisson model with expected value equal to the exponential of the linear
prediction. Therefore, predicted flows (in levels) read:
P P ML
wij
= exp{xij P P M L }.

(4)

Being a Poisson model, the variance of predictions equals their expected value.
Finally, in the ZIP case, one first estimates the probability ij that a link (i, j) is zero. This
is done by fitting a Logit model taking as dependent variable the elements aij of the adjacency
matrices, and as explanatory variables those usually employed to fit a GM equation. Next, for all
and only active trade links, one fits wij with the PPML model above.9 The overall probability of
observing a given trade-flow level is described by combination of the PPML and the logit processes.
As a result, the predicted bilateral flow (in levels) is defined as:
ZIP
wij
= (1 ij ) exp{xij ZIP } = (1 ij )
ij ,

(5)

whereas the variance of the prediction is given by:

ZIP
|xij ) = ij (1 ij )[1 + ij
ij ].
V ar(wij

(6)

As a by-product, one can also estimate the overall probability of a zero flow (i.e. of an absent
link), which reads ij + (1 ij )
ij .
Given any of the foregoing predictions for bilateral-trade flows, we then define:
Definition 3 (Predicted Weighted ITN) The predicted weighted International Trade Network,
for each given cross-section t, is represented by a weighted-directed graph, where the nodes are counM
M , with entries wij
tries and link weights are fully characterized by the asymmetric matrix W
and
10
M {OLS, P P M L, ZIP }.
As far as the binary predicted ITN is concerned, two remarks are in order. First, since in the
OLS case, as already mentioned, the predicted binary ITN coincides with the observed one, there
is no need to address any binary analysis at all. Second, note that by construction estimated trade
flows from both PPML and ZIP models are always strictly positive (although in some cases very
small). This means that one always ends up with an estimated full binary ITN, which impairs any
9

Therefore, the second stage of a ZIP model takes as given the underlying binary structure as in OLS estimation.
As mentioned, in the OLS case we shall compute weighted network statistics on the logged predicted matrix,
OLS
whose generic element is
ij
.
10

10

statistical comparison with the observed binary ITN. Section 3.2 discusses these points in more
detail.

2.5

Network Statistics and Confidence Intervals

We study the extent to which the architecture of the observed ITN over time can be explained by
the GM employing a set of standard topological properties (i.e., network statistics), see Fagiolo
et al. (2009) for a discussion. As Table 4 shows, we focus on three families of properties. First, total
node-degree and total node-strength, measure, for binary and weighted networks respectively, the
number of node partners and total trade intensity. In a directed network, one can also distinguish
between node in-degree/in-strength (i.e., number of markets a country imports from, and total
imports) and node out-degree/out-strength (i.e., number of markets a country exports to, and
total exports).
Second, total average nearest-neighbor degree (ANND) and strength (ANNS) compute, respectively, the average number of trade partners and total trade value of trade partners of a given node.
This gives us an idea of how much a country is connected with other very well-connected countries. ANND and ANNS statistics can be disaggregated so as to account for both import/export
partnerships of a country, and import/export partnerships of its partners. More precisely, one can
compute four different measures of average nearest-neighbor degree/strength, obtained by coupling the two ways in which a node A can be a partner of a given target country B (importer
or exporter) and the two ways in which the partners of A may be related to it (as exporters or
importers). Finally, we consider clustering coefficients (CCs), see Fagiolo (2007) for a discussion.
In the binary case, a node overall CC returns the likelihood that any two trade partners of that
node are themselves partners. In the weighted case, these likelihoods are computed taking into
account link weights to proxy how strong are the edges of the triangles that are formed in the
neighborhood of a node. Again, in the directed case one can disaggregate total node CC according
to the four different shapes that directed triangular motifs can exhibit.11
We are interested not only in node average and standard deviation of such statistics over time,
but also in the way node statistics correlate, and how such correlation patterns evolve across the
years. In particular, we focus on correlation between node degrees (resp., strengths) and ANND
(resp., ANNS). This gives us information on the assortativity/disassortativity nature of the ITN.
We are also interested in correlation between ND/NS and clustering, to understand the extent to
which more and better connected countries trade with partners that trade a lot between them.
We also derive the variance of predicted network statistics, so as to build their confidence intervals and evaluate the precision of GM estimates. The variance of population averages of a few pre11

These are labelled cycle (if i exports to j, who exports to h, who exports to i), in (if both j and h, who are
trade partners, exports to i), out (if both j and h, who are trade partners, imports from i) and mid (if i imports
from h and exports to j, and j and h are trade partners).

11

dicted network statistics can be computed analytically. For instance, given M {OLS, P P M L, ZIP },
the predicted population-average of node out-strength (N SOU T ) reads:
M
1 XX M
d
N
S out =
wij ,
N i

(7)

jIi

M
d
S out equals the sum
where Ii are the predicted export partners of country i. The variance of N
M
M
of variances of the elements wij
divided by N 2 , as by construction the predictions wij
have zero
covariance. Then, for the PPML model we get:
P P ML
1 X X P P ML
1 dP P M L
d
S out ) = 2
S out ,
V ar(N
wij
= N
N i
N

(8)

jIi

as it is expected from the Poisson nature of the model. In the case of a ZIP estimation, using Eq.
(6), one instead gets:
ZIP
1 XX
d
S out ) = 2

ij (1 ij )(1 +
ij ij ).
V ar(N
N i

(9)

jIi

2
In the OLS case the variance of wij equals
OLS
for each observation. Hence:
2
OLS

2 X

ols
(N 1)
d
S out ) = OLS
a
=
,
V ar(N
ij
2
N
N
ij

(10)

where is the observed density of the network.

More generally, whenever an analytical expression cannot be obtained, we perform simulations to estimate the variance of predicted network-statistics. To that end, for each model
M {OLS, P P M L, ZIP }, we generate 10,000 replications of the ITN using the data-generation
process implied by the model M to proxy the second moment of the distribution of predicted
statistics. For example, in the OLS model the generic entry of the sampled ITN weight matrices in each run can be simply drawn from a normal distribution with mean
ijOLS and variance
2

OLS
, see Eq. (3). In a PPML model, instead, wij are drawn from their corresponding Poisw

son distribution: Prob{wij } =

ijij eij /wij !, where
ij = exp(xij P P M L ). Finally, simulation of
ZIP-generated trade flows proceeds in two stages. First, we simulate the binary structure using a
Bernoulli process, where each link is in place with probability (1 ij ). Second, we superimpose on
this simulated binary structure a weight matrix whose generic entries are sampled from a Poisson
distribution with parameter
ij = exp{(xij P P M L )}.12
12

This simulation technique allows the Bernoulli-generated binary structure to converge to the correspondent
logit predictions from the first stage of the ZIP. In small samples, however, a bias is introduced. This in general
implies an overestimation of the performance of the model. Notice also that in each simulated instance the adjacency

12

Results

This Section explores the question whether the statistical properties of the predicted ITN are
similar to those observed in the real-world ITN. We start with basic (non-directed) weighted
statistics (total NS, ANNS and clustering). Next, we discuss results related to directed weighted
measures (e.g., in and out strength, etc.). Finally, we focus on the binary ITN.

3.1

Weighted Statistics

We begin to study population averages of total node strength:

M
1 X dM
1 XX M
d
N S i,tot =
N
S tot =
w ,
N i
N i j ij

(11)

where N is the number of countries in the target cross section. Note that N Sitot measures total
country trade. Therefore its population average equals total world trade divided by the number
of countries. Figure 1 reports predicted values with confidence bands against observed ones across
years. It is easy to see that all three methods perfectly match observed values, with very narrow
prediction errors. This is not surprising, as the very purpose of the GM is to predict bilateral
trade flows, and NS are just linear combinations of them. Therefore one expects the GM to be
well equipped to predict linear transformations of total world trade.
The picture substantially changes when we turn to higher-order statistics like ANNS and WCC,
which involve link weights that are two steps away from the origin node. As Figures 2 and
3 indicate, OLS predictions are quite successful in replicating average total ANNS and, to a
lesser extent, average total clustering.13 More precisely, the OLS-predicted ITN tends to slightly
overestimate observed average total ANNS and to underestimate observed average total WCC.
Nevertheless, predicted values are very close to (and in many cases within) error bands. Note also
that the precision of GM estimates is very high, as the narrow 95% error bars suggest. Conversely,
both PPML and ZIP largely underestimate both ANNS and WCC, although they are able to
correctly get the time trend. In addition, PPML predictions are more precise than ZIP ones.
The reason for this mismatch lies in the way the three estimation techniques work in reproducing
the binary structure. Recall from Table 4 that weighted-network statistics as ANNS and WCC
are in fact a mix of link weights and node degrees. To correctly reproduce such network properties
any predictor of link weights must also correctly reproduce the underlying binary topology. Using
OLS means fitting the GM over positive link weights only, i.e. the observed binary topology is
preserved. PPML and ZIP employ instead all possible country pairs in the regression. As a result,
matrix changes and so do binary topological properties.
13
Clustering coefficients are computed without rescaling link weights in the unit interval in order not to bias the
analysis with network-dependent rescaling factors (Fagiolo, 2007; Saram
aki et al., 2007). Therefore, the range of
WCC is not within [0, 1].

13

they both completely destroy the underlying binary structure and obtain a full predicted binary
network, where all links are in place. This is because all predicted bilateral flows and positive-link
probabilities are strictly positive (although sometimes very small).14 Note that density in ITN
ranges from 0.40 to 0.50 (see Table 1), meaning that slightly less than a half of possible trade
relationships are present. In fact, both PPML and ZIP predict very small link-probabilities for
unconnected countries, and consequently very weak link weights. This implies that predicted total
country trade (i.e., total node strength) is not dramatically far from the observed one, although in
principle it may be slightly overestimated. Conversely, badly predicting the binary structure results
in remarkably-smaller predictions for average ANNS and WCC (as compared to observed ones),
because one highly overestimates node degrees, appearing at the denominator of both statistics.
To further explore this issue, we perform two-sample Kolmogorov-Smirnov (K-S) tests to compare predicted vs. observed node-statistic distributions. More precisely, given any of the three
statistics of interest (total node strength, average nearest-neighbor strength and weighted clustering), we test the null hypothesis that predicted and observed statistics come from the same
distribution. The results in Table 5 confirm the message coming from population averages. Consistently over the years, OLS predictions are able to generate network-statistic distributions very
similar to those observed in the observed ITN. In contrast, PPML and ZIP can reproduce total
strength, but they hardly replicate second-order topology measures like ANNS and WCC.
Another fundamental set of stylized facts characterizing the evolution of the ITN concerns the
way in which different network statistics correlate. Figures 4 and 5 show observed vs. predicted
correlation patterns between, respectively, total ANNS and NS, and total WCC and NS. Note that
OLS are able to correctly predict the existing disassortativity emerging between total country trade
and average trade of the partners of a node. Conversely, both PPML and ZIP strongly overestimate
the magnitude of such negative correlation.15 OLS estimates are also able to match strengthclustering correlation, even if with less precision than before, as confidence intervals suggest. Again,
neither PPML nor ZIP can reproduce the positive strength-clustering correlation characterizing
the original weight matrices (in levels), as the underestimation bias appears to be persistently
high.
Correlation results are in line with recent findings by Squartini et al. (2011a,b), who show that
higher-order weighted properties in the ITN cannot be reproduced by any random model that
takes as given the observed strength sequence (but does not control for the underlying binary
structure). Here we show that a satisfactory replication of ITN properties can be achieved only
if one fixes the binary structure and attributes link weights using a GM. As soon as the binary
14

This interpretation is confirmed by an additional fitting exercise where we employ a PPML estimation technique
performed on positive link weights only. In this case, binary-restricted PPML fits are able to reproduce quite
successfully all statistics, in line with what happens for OLS. In fact, restricted PPML predictions do not add much
more precision in the estimates compared to OLS.
15
The difference between observed correlations in the OLS vs PPML/ZIP cases must be attributed to the fact
that in OLS we compute all network statistics on logged weight matrices.

14

structure is badly reproduced, one also looses the possibility to correctly recover weighted-network
patterns, primarily because most of weighted-network statistics are inherently dependent on the
binary representation.
So far, we have been studying the performance of GM predictions for weighted undirected
statistics. In fact, total strength, ANNS and clustering all neglect the directed nature of trade
flows and ensuing asymmetries, as they do not discriminate between in and out links (i.e., import
and export flows). To check if the foregoing results also apply in the case of weighted-network
directed statistics, which instead take fully into account trade-flow directionality, we have studied
predicted vs. observed values of population averages of such statistics and their correlation. We
have focused on in- and out-strength, and the breakdown in four directed statistics of ANNS and
WCC (see Table 4). In all cases, all results obtained above hold. In particular, OLS can easily
reproduce all versions of average disaggregated ANNS, while it slightly underestimates the average
of all directed clustering coefficients. Both PPML and ZIP fail to capture average ANNS and
WCC. Incidentally, according to K-S tests, all three methods are able to fully explain both in- and
out-strength distributions, but only OLS predictions get higher-order statistic distributions right.
All correlations16 are correctly predicted by OLS (with a slight overestimation as far as clusteringstrength relationships are concerned), whereas both PPML and ZIP always fail in replicating
correlations among directed statistics. Once again, the ability to predict the binary (directed)
structure of the ITN becomes crucial: despite the fact that the three methods correctly replicate
the correlation between in- and out-strength, only the OLS (by construction) exploits a perfect
prediction of the binary structure, and therefore results in a good approximation of the patterns
characterizing weighted statistics.

3.2

Binary Statistics

Our weighted-network exercises show that the GM can provide a quite satisfactorily picture of ITN
properties only if one restricts the estimation to strictly-positive trade flows, i.e. if the observed
binary structure is taken as given. The fact that binary trade links play a crucial role in explaining
ITN weighted topology indicates that any GM model aiming at endogenously estimating binary
links must somewhat take into account the discrete nature of the binary ITN and try to obtain a
more accurate estimation of the exact location of the zeros in trade matrices.
But is the GM able to correctly predict the binary structure of the ITN? In other words, can
one employ the independent variables traditionally used in GM equations to predict whether a
trade link exists or not? To address this issue, we employ the most natural candidate model for
16

Among all possible correlations of directed statistics with node in- and out-strength we have selected only
those economically more relevant. For example, we have focused on the correlation coefficient between AN N S out,in
and N S out (and not that between AN N S out,in and N S in ) because one is much more interested in understanding
whether a country that exports more, in turn exports to countries that imports more, rather than knowing whether
a country that imports more, in turn exports to countries that imports more.

15

estimating the probability that a given link is present, i.e. a logit specification. More specifically,
for any cross-section t, we estimate:
P rob{aij = 1|xij } =

exp{xij }
= (xij ; )
1 + exp{xij }

(12)

Notice that Eq. (12) is exactly the functional form that we fit in the first stage of the ZIP
estimation. Therefore, we can employ first-stage estimates for a zero flow ij from the ZIP model
whose generic entry ij = 1 ij represents the
and build the predicted probability matrices ,
estimated probability of observing a directed link from country i to country j in that year.17
we
Of course, as already mentioned, ij > 0 for all links ij. Therefore, if we just employ ,
will end up with a full-network for the predicted binary ITN, as it happens for PPML and ZIP
procedures. This impairs all subsequent analyses, as they strongly depend on a correct estimation
of the binary structure.
In what follows, we propose a strategy to employ the GM to predict the binary ITN. To get a
reasonable prediction for the binary ITN, we proceed in two ways. First, for each year, we take the
and we delete all the links associated to predicted probabilities smaller than the observed
matrix
ITN density. This generates a predicted binary ITN with a density approximately similar (for
numerical reasons) to the observed one.18 This leads to the following:
Definition 4 (Density-Induced Predicted Binary ITN) The density-induced predicted binary
International Trade Network, for each given cross-section t, is represented by a binary-directed

graph, where the nodes are countries and the adjacency asymmetric matrix A()
has entries
a
ij () = 1 if and only if ij > , where is observed ITN density in year t.
Second, we exploit a simulation-based procedure that, instead of using density-matching thresh More precisely, in each year, we generate a sample of M
olds, fully exploits the information in .
independent adjacency matrices Am = {
am
m
ij }, for m = 1, . . . , M where in each sample a
ij is drawn
from a Bernoulli distribution with parameter ij , independently across all pairs ij. More formally:
Definition 5 (Bernoulli Predicted Binary ITN) The Bernoulli predicted binary International
Trade Network, for each given cross-section t, is represented by a distribution of M binary-directed
17

We have also experimented with the matrix of predicted probabilities coming from the full ZIP estimation,
where each element equals ij + (1 ij )
ij , without noticing any dramatic changes in the results we present below.
18
We also performed two alternative threshold-based exercises. In the first one, we delete all the links associated to predicted probabilities smaller than a given threshold s, which is chosen so as to approximately match the
empirically-observed ITN density. In the second one, the optimal threshold is chosen so as to minimize the Manhattan distance between the observed adjacency matrix and the predicted binary one, where the latter is defined, for
each given threshold s, as the binary matrix where a link is in place if and only if ij < s. Both procedures lead to
very similar optimal thresholds, which are in turn very close to the density we get by straightforwardly setting the
threshold equal to . Therefore, we present here only results for the case where the threshold is equal to observed
density.

16

graphs, where for each graph the nodes are countries and the adjacency asymmetric matrix Am ,
m = 1, . . . , M , has entries a
m
ij that are drawn independently from a Bernoulli distribution with

parameter ij .
In our exercises, we set M = 10, 000 and we employ simulated predicted matrices to compute
Monte-Carlo standard deviations for all statistics of interest (of course with the first method we
do not have any source of variation in predicted values, therefore no error bars can be computed).
Note also that density-induced predicted binary ITN preserves almost exactly observed density,
while the Bernoulli predicted binary ITN preserves that quantity only on average.
Our main results are reported in Figures 6 and 7, where we plot observed binary statistics vs.
predicted ones, using density-induced and Bernoulli procedures (see Definitions 4 and 5 above).19
To begin with, note that both density-induced and Bernoulli predictions are quite successful in
tracking average total ND. Actually, Bernoulli-predicted binary ITN can exactly replicate, on
average, that statistics. Conversely, density-induced predictions slightly deviate from observed
values.20 This is not surprising, as it means that both models are able to get a good proxy of
observed density.21
The fact that a Logit estimation is on average able to predict observed density explains why
a ZIP model, which employs the very same Logit specification in its first stage, predicts very well
average total NS. For that statistics is an average over all existing links and it is not so much
affected by where these links are actually located. This is not true of ANNS and WCC, which in
fact are badly reproduced by a ZIP model because require a more precise knowledge of where links
are placed.
A similar problem arises in the binary ITN for both density-induced and Bernoulli predictions:
the former persistently overestimates observed average ANND and BCC, whereas the latter persistently underestimates them. Again, this hints to an inherent inability of the GM to well predict
the presence a link.
Things seem to improve a bit when we move to correlation structure. Density-induced predictions are able to well capture binary disassortativity in the last part of the sample, but they only
partially get right clustering-degree correlation. Conversely, Bernoulli-based predictions seem to
perform quite satisfactorily in both cases: although on average observed correlations are rarely
replicated, the inherent variability of this procedure allows one to conclude that there exists a
sufficiently large number of simulations where predicted correlations are very similar to observed
ones.

19

We focus here only on undirected measures. All main results hold also for directed network statistics.
a
This is entirely due to numerical problems, as often it is impossible to purge from the predicted matrix
number of links so as to perfectly match observed density.
21
Average total ND is indeed equal to L/N , where L is the number of links, and hence equals (N 1).
20

17

Concluding Remarks

In this paper, we have studied whether a gravity model (GM), the work-horse theoretical reference
in international trade, can explain the statistical properties of the international-trade network.
Our exercises show that the GM does a very good job in replicating the weighted-network
structure of the ITN only if one fixes its binary architecture equal to the observed one. More
generally, the GM performs very badly when asked to predict the presence of a link, or the level
of the trade flow it carries, whenever the binary structure must be simultaneously estimated.
Therefore, the GM turns out to be a good model for estimating trade flows, but not to explain
why a link in the ITN gets formed and persists over time. In other words, knowing country-specific
variables (country GDP, etc.) and country bilateral interactions (bordering conditions, belonging
to the same RTA, etc.) is not enough to predict the presence of a link. However, conditional on the
information that a link exists, such variables can well predict how much trade that link actually
carries.
Notice that these results are largely independent on which variables are actually entering the
gravity equation we fit to the data. In the foregoing exercises, we have used a standard specification
where many of the most-employed GM variables enter the regression. We have also tried and
augment the equation with other explanatory variables that resulted statistically not significant,
but can nevertheless improve the percentage of explained trade-flow variance, without observing
any dramatic increase in the goodness of fit of ITN network statistics.
In order to better explain the topological properties of the ITN many alternative strategies
may be pursued. First, one may consider to augment a GM specification with network-related
variables. It may be indeed argued that if standard economic variables entering in the GM are not
enough to explain link formation, perhaps this is because the presence of a link between any two
countries might be actually explained by the very local structure of the network (e.g., degrees of
the two countries, etc.). Of course this introduces some endogeneity to the problem, because the
presence of a link in turn affects local network properties. By properly dealing with endogeneity
issues in estimation, one can hope to better explain the binary structure of the ITN.
Second, one might borrow social-network statistical methodologies currently employed to model
the evolution of directed graphs over time as continuous-time Markov processes (Snijders, 2005).
For example, one may envisage setups where each single node chooses its outgoing link (i.e. whether
to export to another country or not) based on a myopic optimization of some objective function,
where the latter may be the result of many firm-level decisions within the origin country.
Finally, one may think to explore international-trade models where the decision of a firm located
in country A to export goods to country B, which possibly never imported products from A before,
is rooted in a more detailed micro-foundation. This may require to blend together two strands of
literature, one on the role of heterogeneous firms in international trade (Melitz, 2003; Bernard et al.,

18

2007) and the other on models of trade network formation based on simple aggregate dynamics
(Garlaschelli and Loffredo, 2004; Bhattacharya et al., 2008; Riccaboni and Schiavo, 2010).

References
Abeysinghe, T. and Forbes, K., 2005. Trade linkages and output-multiplier effects: a structural
VAR approach with a focus on Asia. Review of International Economics. 13, 356375.
Anderson, J. E., 1979. A Theoretical Foundation for the Gravity Equation. American Economic
Review. 69, 10616.
Anderson, J. E. and van Wincoop, E., 2003. Gravity with Gravitas: A Solution to the Border
Puzzle. American Economic Review. 93, 170192.
Baldwin, R. and Taglioni, D., 2006. Gravity for Dummies and Dummies for Gravity Equations.
NBER Working Papers 12516. National Bureau of Economic Research, Inc.
Barigozzi, M., Fagiolo, G. and Garlaschelli, D., 2010. Multinetwork of international trade: A
commodity-specific analysis. Physical Review E. 81, 046104.
Barigozzi, M., Fagiolo, G. and Mangioni, G., 2010. Identifying the Community Structure of the
International-Trade Multi Network. Physica A. 390, 20512066.
Bergstrand, J. H., 1985. The Gravity Equation in International Trade: Some Microeconomic Foundations and Empirical Evidence. The Review of Economics and Statistics. 67, 47481.
Bernard, A. B., Jensen, J. B., Redding, S. J. and Schott, P. K., 2007. Firms in International Trade.
Journal of Economic Perspectives. 21, 105130.
Bhattacharya, K., Mukherjee, G. and Manna, S., 2007. The International Trade Network. In Econophysics of Markets and Business Networks, A. Chatterjee and B. Chakrabarti (eds). SpringerVerlag, Milan, Italy.
Bhattacharya, K., Mukherjee, G., Saramaki, J., Kaski, K. and Manna, S., 2008. The International
Trade Network: Weighted Network Analysis and Modeling. Journal of Statistical Mechanics:
Theory Exp. A. 2, P02002.
Burger, M., Oort, F. v. and Linders, G., 2009. On the Specification of the Gravity Model of Trade:
Zeros, Excess Zeros and Zero-Inflated Estimation. Research Paper ERS-2009-003-ORG Revision.
Erasmus Research Institute of Management (ERIM).
De Benedictis, L. and Taglioni, D., 2011. The Gravity Model in International Trade. In The Trade
Impact of European Union Preferential Policies, L. De Benedictis and L. Salvatici (eds). Springer.
De Benedictis, L. and Tajoli, L., 2011. The World Trade Network. The World Economy. 34, 1417
1454.

19

Deardorff, A., 1998. Determinants of Bilateral Trade: Does Gravity Work in a Neoclassical World?.
In The Regionalization of the World Economy. National Bureau of Economic Research, Inc.
pp. 732.
Dees, S. and Saint-Guilhem, A., 2011. The role of the United States in the global economy and its
evolution over time. Empirical Economics. 41, 573591.
Evenett, S. J. and Keller, W., 2002. On Theories Explaining the Success of the Gravity Equation.
Journal of Political Economy. 110, 281316.
Fagiolo, G., 2007. Clustering in Complex Directed Networks. Physical Review E. 76, 026107.
Fagiolo, G., 2010. The international-trade network: gravity equations and topological properties.
Journal of Economic Interaction and Coordination. 5, 125.
Fagiolo, G., Schiavo, S. and Reyes, J., 2008. On the topological properties of the world trade web:
A weighted network analysis. Physica A. 387, 38683873.
Fagiolo, G., Schiavo, S. and Reyes, J., 2009. World-trade web: Topological properties, dynamics,
and evolution. Physical Review E. 79, 036115.
Fagiolo, G., Schiavo, S. and Reyes, J., 2010. The Evolution of the World Trade Web: A WeightedNetwork Approach. Journal of Evolutionary Economics. 20, 479514.
Felbermayr, G. J. and Kohler, W., 2006. Exploring the Intensive and Extensive Margins of World
Trade. Review of World Economics (Weltwirtschaftliches Archiv). 142, 642674.
Forbes, K., 2002. Are trade linkages important determinants of country vulnerability to crises?. In
Preventing Currency Crises in Emerging Markets, E. Sebastian and F. Jeffrey (eds). Chicago,
University of Chicago Press.
Fratianni, M., 2009. The Gravity Model in International Trade. In The Oxford Handbook of International Business, A. M. Rugman (ed). Oxford University Press, Oxford, U.K.
Galvandatilde, A. B., Artis, M. and Marcellino, M., 2007. The transmission mechanism in a
changing world. Journal of Applied Econometrics. 22, 3961.
Garlaschelli, D., Di Matteo, T., Aste, T., Caldarelli, G. and Loffredo, M., 2007. Interplay between
topology and dynamics in the World Trade Web. The European Physical Journal B. 57, 1434
6028.
Garlaschelli, D. and Loffredo, M., 2004. Fitness-Dependent Topological Properties of the World
Trade Web. Physical Review Letters. 93, 188701.
Garlaschelli, D. and Loffredo, M., 2005. Structure and evolution of the world trade network. Physica
A. 355, 13844.
Glick, R. and Rose, A. K., 2001. Does a Currency Union Affect Trade? The Time Series Evidence.
NBER Working Papers 8396. National Bureau of Economic Research, Inc.

20

Helliwell, J. F. and Padmore, T., 1985. Empirical Studies of Macroeconomic Interdependence. In

Handbook of International Economies, R. Jones and P. Kenen (eds). Elsevier Science Publishers
B.V.
Kali, R., Mendez, F. and Reyes, J., 2007. Trade structure and economic growth. Journal of International Trade & Economic Development. 16, 245269.
Kali, R. and Reyes, J., 2010. Financial Contagion On The International Trade Network. Economic
Inquiry. 48, 10721101.
Krugman, P., 1995. Growing World Trade: Causes and Consequences. Brookings Papers on Economic Activity. 26, 327377.
Li, X., Jin, Y. Y. and Chen, G., 2003. Complexity and synchronization of the World trade Web.
Physica A: Statistical Mechanics and its Applications. 328, 28796.
Linders, G.-J. M. and de Groot, H. L., 2006. Estimation of the Gravity Equation in the Presence
of Zero Flows. Tinbergen Institute Discussion Papers 06-072/3. Tinbergen Institute.
Long, J., 1997. Regression models for categorical and limited dependent variables. Advanced quantitative techniques in the social sciences. Sage Publications, London.
Melitz, M. J., 2003. The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry
Productivity. Econometrica. 71, 16951725.
Poyhonen, P., 1963. A Tentative Model for the Volume of Trade Between Countries.
Weltwirtschaftliches Archiv. 90, 9399.
Reyes, J., Schiavo, S. and Fagiolo, G., 2008. Assessing the evolution of international economic integration using random-walk betweenness centrality: The cases of East Asia and Latin America.
Advances in Complex Systems. 11, 685702.
Reyes, J., Schiavo, S. and Fagiolo, G., 2010. Using complex networks analysis to assess the evolution
of international economic integration: The cases of East Asia and Latin America. Journal of
International Trade & Economic Development. 19, 215239.
Riccaboni, M. and Schiavo, S., 2010. Structure and growth of weighted networks. New Journal of
Physics. 12, 023003.
Rose, A. K. and Spiegel, M. M., 2002. A gravity model of sovereign lending: trade, default and
credit. Working Papers in Applied Economic Theory 200209. Federal Reserve Bank of San
Francisco.
Santos Silva, J. M. C. and Tenreyro, S., 2006. The Log of Gravity. The Review of Economics and
Statistics. 88, 641658.
Saramaki, J., Kivela, M., Onnela, J.-P., Kaski, K. and Kertesz, J., 2007. Generalizations of the
clustering coefficient to weighted complex networks. Physical Review E. 75, 027105.
Serrano, A. and Bogu
na, M., 2003. Topology of the World Trade Web. Physical Review E.
68, 015101(R).
21

Serrano, A., Bogu

na, M. and Vespignani, A., 2007. Patterns of dominant flows in the world trade
web. Journal of Economic Interaction and Coordination. 2, 111124.
Snijders, T. A. B., 2005. Models for Longitudinal Network Data. In Models and Methods in Social
Network Analysis, P. Carrington, J. Scott and S. Wasserman (eds). Cambridge University Press.
pp. 215247.
Squartini, T., Fagiolo, G. and Garlaschelli, D., 2011a. Randomizing world trade. I. A binary
network analysis. Physical Review E. 84, 046117.
Squartini, T., Fagiolo, G. and Garlaschelli, D., 2011b. Randomizing world trade. II. A weighted
network analysis. Physical Review E. 84, 046118.
Subramanian, A. and Wei, S.-J., 2003. The WTO Promotes Trade, Strongly But Unevenly. Working Paper 10024. National Bureau of Economic Research.
Tinbergen, J., 1962. Shaping the World Economy: Suggestions for an International Economic
Policy. New York.
van Bergeijk, P. and Brakman, S. (eds), 2010. The Gravity Model in International Trade. Cambridge University Press, Cambridge.
Winkelmann, R., 2008. Econometric Analysis of Count Data. Springer Verlag, New York.
Wooldridge, J. M., 2001. Econometric Analysis of Cross Section and Panel Data. Boston, The MIT
Press.

22

23

Countries (No.)
Trade Flows (No.)
Density
Average Trade
Countries making up 50% of trade
Flows making up 50% of trade
Countries making up 90% of trade
Flows making up 90% of trade
% Countries making up 50% of trade
% Flows making up 50% of trade
% Countries making up 90% of trade
% Flows making up 90% of trade

1970
1975
129
135
6583
7618
0.40
0.42
51.03
56.43
7
8
73
99
39
39
794
900
5.43% 5.93%
1.11% 1.30%
30.23% 28.89%
12.06% 11.81%

1980
142
8162
0.41
57.48
7
90
38
894
4.93%
1.10%
26.76%
10.95%

1985
148
9108
0.42
61.54
7
73
37
871
4.73%
0.80%
25.00%
9.56%

1990
145
10289
0.49
70.96
7
69
31
749
4.83%
0.67%
21.38%
7.28%

Table 1: Subramanian and Wei (2003) Database. Summary statistics.

1995
2000
157
154
12138
11828
0.50
0.50
77.31
79.81
8
8
74
79
32
33
826
855
5.10% 5.19%
0.61% 0.67%
20.38% 21.43%
6.81% 7.23%

Label
W
Y
area
pop
d

Related to
Link
Country
Country
Country
Link

landl

Country

continent

Country

contig

Link

comlang off

Link

comcol

Link

colony

Link

curcol

Link

comrelig

Link

comcur

Link

gsp

Link

rta

Link

Description
Imports in U.S. Dollars
Gross-domestic product
Country area in Km2
Country population
distance between two countries, based on bilateral distances between the largest
cities of those two countries,
weighted by the share of the
city in the overall countrys
population
Dummy variable equal to 1 for
landlocked Countries
Categorical variable indicating the continent of the country
Contiguity dummy equal to 1
if two countries share a common border
Dummy equal to 1 if both
countries share a common official language
Dummy equal to 1 if both
countries have had a common
colonizer
Dummy equal to 1 if both
countries have ever had a colonial link
Dummy equal to 1 if both
countries are currently in a
colonial relationship
Percentage in which both
countries share religions
Dummy equal to 1 if both
countries have a currency
unions
Dummy equal to 1 if both
countries share a generalized
system of preferences
Dummy variable equal to 1 if
both countries involved in regional, bilateral or preferential trade agreements

Source
Subramanian and Wei (2003)
Subramanian and Wei (2003)
Subramanian and Wei (2003)
Subramanian and Wei (2003)
CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)
CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)
CEPII (http://www.cepii.fr/)

CEPII (http://www.cepii.fr/)

WTO (http://www.wto.org/)

Table 2: Variables employed in the gravity-model estimation.

24

Regressor
ln gdp i
ln gdp j
ln dist ij
ln area i
ln area j
ln pop i
ln pop j
landl ci
landl cj
continent i
continent j
contig
comlang off
comcol
colony
curcol
comrelig
comcur
gsp
rta
cons
No. Obs
F or Wald chi2
Prob > F or chi2
R2 or Pseudo R2
Vuong Z
Prob > Z

(1 )
(2 )
(3 )
(11 )
(21 )
(12 )
(22 )
(11 )
(21 )
(12 )
(22 )
(1 )
(2 )
(3 )
(4 )
(5 )
(6 )
(7 )
(8 )
(9 )
()

OLS(W>0)
1.415***(0.023)
1.323***(0.021)
-1.034***(0.023)
-0.068***(0.013)
-0.108***(0.012)
-0.402***(0.025)
-0.42***(0.025)
-0.456***(0.051)
-0.472***(0.046)
0.004(0.02)
-0.05**(0.019)
0.823***(0.112)
0.637***(0.055)
0.785***(0.084)
1.091***(0.088)
-2.334(1.923)
0.266***(0.066)
0.554***(0.111)
0.484***(0.047)
0.338***(0.053)
-20.666***(0.394)
11828
1423
0
0.68
-

PPML
1.302***(0.049)
1.697***(0.047)
-0.725***(0.033)
-0.097***(0.022)
-0.14***(0.035)
-0.401***(0.07)
-0.773***(0.057)
-0.509***(0.092)
-0.451***(0.124)
-0.16***(0.038)
-0.257***(0.042)
0.572***(0.113)
0.407***(0.084)
0.399(0.276)
-0.252**(0.093)
0.156(0.737)
-0.094(0.109)
-0.139(0.107)
0.349***(0.107)
0.204**(0.078)
-21.693***(0.852)
23562
15029
0
0.93
-

ZIP-Poisson
1.278***(0.048)
1.65***(0.047)
-0.721***(0.033)
-0.09***(0.022)
-0.135***(0.034)
-0.393***(0.068)
-0.744***(0.056)
-0.48***(0.091)
-0.426***(0.123)
-0.153***(0.038)
-0.259***(0.041)
0.622***(0.114)
0.376***(0.083)
0.41(0.274)
-0.226**(0.092)
0.345(0.718)
-0.148(0.109)
-0.136(0.107)
0.303***(0.106)
0.181**(0.078)
-20.8***(0.868)
11828
14083
0
0.92
82.76
0

ZIP-Logit
-0.954***(0.025)
-0.961***(0.023)
0.533***(0.033)
0.107***(0.013)
0.15***(0.013)
0.206***(0.026)
0.25***(0.026)
0.517***(0.047)
0.579***(0.047)
0.043*(0.019)
-0.086***(0.018)
0.996***(0.216)
-0.728***(0.061)
-0.26**(0.077)
0.476*(0.239)
2.081(1.291)
-0.676***(0.081)
-1.493***(0.18)
-2.015***(0.116)
-0.672***(0.078)
21.636***(0.487)
23562
5691.15
0
0.43

Table 3: GM estimation. Year: 2000. ZIP-Poisson: second stage of the ZIP estimation process.
ZIP-Logit: first stage of the ZIP estimation process. Note: i = exporter; j = importer.

25

4.5

1700
Observed
OLS

1400
1300

2.5
2

1100

1.5

1975

1980

1985

1990

1995

2000

3
2.5
2
1.5
1

1
1970

Year

Observed
ZIP

3.5

1200

x 10

3.5

Average NStot

Average NStot

Observed
PPML

1500

1000
1970

4.5

Average NStot

1600

x 10

1975

1980

1985

1990

1995

2000

0.5
1970

1975

1980

1985

1990

1995

2000

Year

Year

Figure 1: Observed vs. GM-predicted average node total strength. 95% confidence bands are
displayed as error bars around predicted values. Note: in the OLS plot node strength is computed
using log of trade flows.

10

2400
Observed
OLS

Average ANNStot

Average ANNStot

Observed
ZIP
8

2100
2000
1900
1800

7
6
5
4

1700

1600

2
1975

1980

1985

1990

1995

1
1970

2000

Year

x 10

Observed
PPML

2200

1500
1970

10

Average ANNStot

2300

x 10

1975

1980

1985

1990

1995

0
1970

2000

1975

1980

1985

1990

1995

2000

Year

Year

Figure 2: Observed vs. GM-predicted average total ANNS. 95% confidence bands are displayed
as error bars around predicted values. Note: in the OLS plot node strength is computed using log
of trade flows.

Observed
OLS

8.2

Average WCCtot

Average WCCtot

8.4

4.5

8
7.8
7.6

4.5

3.5

3.5

3
2.5
2

7.4
7.2
1970

x 10

Average WCCtot

8.6

1980

1985

Year

1990

1995

2000

1
1970

3
2.5
2

Observed
PPML

1.5

1975

x 10

1975

1980

1985

Year

1990

1995

2000

Observed
ZIP

1.5
1
1970

1975

1980

1985

1990

1995

2000

Year

Figure 3: Observed vs. GM-predicted average total WCC. 95% confidence bands are displayed as
error bars around predicted values. Note: in the OLS plot node strength is computed using log of
trade flows.

26

Topological
Properties
Degrees/Strengths

ANND/ANNS

27
Clustering

Binary
P
N Diin = kiin = P
j aji
out
out
N Di = ki = j aij
N Diout = kitot =Pkiin + kiout

aji kjin
kin
P i
aji kout
AN N Din,out = j kin j
P i in
aij k
AN N Dout,in = jkout j
Pi
aij kout
AN N Dout,out = j kout j
i
P
(aij +aji )ktot
AN N Dtot = j ktot j
i
P P
cyc
j
k aij ajk aki
BCCi = kin kout k
i
P i Pi
j
k aik aji ajk
mid
BCCi = kin kout k
i
P Pi i
a aji a
BCCiin = jkink(kinki1) jk
P iP i
aik ajk aij
BCCiout = kjoutk(kout
P iP i 1)
(aij +aji )(a +a )(a +a )
BCCitot = j k  tot tot jk kj ki ik
2 ki (ki 1)2ki

AN N Din,in =

Weighted P
N Siin = sin
i = P
j wji
out
out
N Si = si = j wij
out
sin
N Sitot = stot
i + si
i =P

aji sin
j
kiin
P
aji sout
AN N Sin,out = j kin j
P i in
aij s
AN N Sout,in = jkout j
Pi
aij sout
AN N Sout,out = jkout j
i
P
(aij +aji )stot
AN N S tot = j ktot j
P P i1/3 1/3 1/3
cyc
k wij wjk wki
W CCi = j kin
kout ki
P Pi i 1/3 1/3
1/3
k wik wjk wji
W CCimid = j kin
out

ki ki
P P i 1/3
1/3 1/3
w
wji wki
j
k
jk
W CCiin =
in
in
k (ki 1)
P Pi
1/3 1/3 1/3
j
k wik wij wjk
W CCiout =
out
k (kiout 1)
P Pi 1/3
1/3
1/3
1/3
1/3
1/3
(w +wji )(wjk +wkj )(wki +wik )

W CCitot = j k ij  tot
2 ki (kitot 1)2ki

AN N Sin,in =

Table 4: Binary and weighted topological properties of the ITN. Note: Time labels are suppressed for notational convenience.

OLS
1970
0.05 (1.00)
0.05 (1.00)
0.05 (0.99)

1975
0.03 (1.00)
0.05 (0.99)
0.05 (0.99)

1980
0.04 (1.00)
0.04 (1.00)
0.04 (1.00)

1985
0.03 (1.00)
0.04 (1.00)
0.07 (0.88)

1990
0.04 (1.00)
0.03 (1.00)
0.05 (0.99)

1995
0.03 (1.00)
0.03 (1.00)
0.05 (0.98)

2000
0.03 (1.00)
0.05 (1.00)
0.06 (0.95)

NS
ANNStot
WCCtot

1970
0.06 (0.96)
0.98 (0.00)
0.45 (0.00)

1975
0.10 (0.44)
0.99 (0.00)
0.36 (0.00)

PPML
1980
1985
0.11 (0.31) 0.11 (0.27)
0.99 (0.00) 0.97 (0.00)
0.32 (0.00) 0.38 (0.00)

1990
0.08 (0.78)
0.94 (0.00)
0.32 (0.00)

1995
0.08 (0.64)
0.94 (0.00)
0.29 (0.00)

2000
0.08 (0.72)
0.94 (0.00)
0.34 (0.00)

NStot
ANNStot
WCCtot

1970
0.06 (0.96)
0.95 (0.00)
0.48 (0.00)

1975
0.07 (0.92)
0.93 (0.00)
0.41 (0.00)

1980
0.08 (0.77)
0.94 (0.00)
0.38 (0.00)

1990
0.06 (0.94)
0.83 (0.00)
0.33 (0.00)

1995
0.08 (0.73)
0.85 (0.00)
0.32 (0.00)

2000
0.05 (0.98)
0.83 (0.00)
0.36 (0.00)

NStot
ANNStot
WCCtot

tot

ZIP
1985
0.09 (0.6)
0.87 (0.00)
0.40 (0.00)

Table 5: Two-sample Kolmogorov-Smirnov test statistics. Null hypothesis: predicted and observed
node-statistics sequences come from the same distribution. P-values in parentheses.
((%

((#
((#%
((&
((&%
((

(!
!#

((

((

((%
(()

:;</=>
??

((#
((&
((

((%
7

((+

515

01//234515673345158

(()%

((+

515

8=2

(()

./0123/3451123/36

89:+;+<

01//234 67334 8

((%%

!#%

!&

!&%

!%

(!
9
!#

((%
(()

:;</=>
?

((#
((&
((

!#%

!&

!&%

+,

!%

(!
9
!#

!#%

!&

,./

!&%

!%

,./

0.9

0.9

0.9

tot

0.8

0.7

0.6

0.5
1970

Observed
OLS
1975

1980

1985

Year

1990

1995

2000

0.8

0.7
Observed

0.6

0.5
1970

PPML

1975

1980

1985

Year

1990

1995

2000

Corr(NStot,WCCtot)

Corr(NStot,WCCtot)

tot

Corr(NS ,WCC )

Figure 4: Observed vs. GM-predicted correlation between total node strenght and total ANNS.
95% confidence bands are displayed as error bars around predicted values. Note: in the OLS plot
node strength and ANNS are computed using log of trade flows.

0.8

0.7
Observed

0.6

0.5
1970

ZIP
1975

1980

1985

1990

1995

2000

Year

Figure 5: Observed vs. GM-predicted correlation between total node strenght and total weighted
clustering coefficient (WCC). 95% confidence bands are displayed as error bars around predicted
values.
28

160

240

150

1
0.95

220

130

Observed

120

200

180

Observed

160

Bernoulli

0.8
0.75
0.7
Observed
Density
Bernoulli

0.6

Bernoulli

140

0.85

0.65

Density

Density
110

Average BCCtot

tot

Average ANND

Average ND

tot

0.9
140

0.55
100
1970

1975

1980

1985

1990

1995

2000

120
1970

1975

1980

Year

1985

1990

1995

2000

0.5
1970

1975

1980

Year

1985

1990

1995

2000

Year

Figure 6: Observed vs. GM-predicted average statistics in the binary ITN. Logit estimation. Density: average statistics in the Density-Induced Predicted Binary ITN (see Definition 4). Bernoulli:
average statistics in the Bernoulli Predicted Binary ITN (see Definition 5). 95% confidence bands
are displayed as error bars around predicted Bernoulli values.

((%
((%%
(()
(()%
((#
((#%
((&
((&%

(()
(()%
((#
((#%
((&
((&%

((

((

((%

((%

(!

(!

!#

!#%

!&

!&%
+,

!%

89:+;+<
2+=:>3?
6+=/>

((%%

./0123/3456..3/37

3/3

3/3

./012 456112 7

((%

89:+;+<
2+=:>3?
+=/>

!#

!#%

!&

!&%
+,

!%

Figure 7: Observed vs. GM-predicted correlation between statistics in the binary ITN. Logit estimation. Density: average statistics in the Density-Induced Predicted Binary ITN (see Definition
4). Bernoulli: average statistics in the Bernoulli Predicted Binary ITN (see Definition 5). 95%
confidence bands are displayed as error bars around predicted Bernoulli values.

29