Exploring the Intensive and Extensive Margins

of World Trade

Gabriel J. Felbermayr and Wilhelm Kohler

University of Tübingen

Abstract: World trade evolves at two margins. Where a bilateral trading relation-
ship already exists it may increase through time (intensive margin). But trade may
also increase if a trading bilateral relationship is newly established between coun-
tries that have not traded with each other in the past (extensive margin). We pro-
vide an empirical dissection of post–World War II growth in manufacturing world
trade along these two margins. We propose a “corner-solutions version” of the
gravity model to explain movements on both margins. A Tobit estimation of this
model resolves the so-called “distance puzzle”. It also finds more convincing evi-
dence than recent literature that WTO-membership enhances trade. JEL no. F12,
F15
Keywords: Bilateral trade; globalization; gravity model

1 Introduction

Despite the unquestionable increase in economic globalization, world trade

still covers a surprisingly small part of the world. In 1950, almost 52 percent
of the potential number of bilateral trade relationships did not report any
manufacturing trade at all. By 1997 the share of bilateral trade relationships
that were actually utilized was still no more than 58 percent.1 Globalization
thus evolves along two major margins. At the intensive margin, established
bilateral trade relationships change their trade volume, while at the extensive

Remark: We wish to thank Rudi Winter-Ebmer and participants of the CESifo Venice Sum-
mer Institute Workshop on “Dissecting Globalization”, July 2004 and at the joint Mu-
nich/Tübingen Internationalization Workshop at the IAW 2005, for stimulating discus-
sions. Thanks are also due to Alan Deardorff, Michael Funke, Keith Head, Andy Rose,
Farid Toubal, and John Whalley, as well as to an anonymous referee for helpful comments.
Marcel Smolka has provided valuable editorial assistance. Please address correspondence to
Wilhelm Kohler, Eberhard Karls University Tübingen, Nauklerstrasse 47, D-72074 Tübin-
gen, Germany; e-mail: [Link]@[Link]
1 These numbers are from a data set to be described in detail below. Admittedly, the ratios
would look much different if trading relationships were weighted by GDPs or population.
The most appropriate weight to use is the amount of trade involved. We shall return to
this in Section 3.

© 2006 The Kiel Institute DOI: 10.1007/s10290-006-0087-3

 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 643

margin new trade relationships are established, or existing ones abandoned.

It is somewhat surprising that systematic treatments of this two-fold mar-
gin in the empirical literature have only just begun to emerge, Feenstra
and Rose (1997) being a notable exception.2 In gravity studies of trade,
the usual approach is to restrict attention to those country pairs for which
strictly positive trade flows are observed. This seems inadequate for two
reasons. First, it ignores an important part of the “action” across time. And
secondly, given the coexistence of the two margins, the exact interpreta-
tion of estimates obtained with this procedure is questionable, as are their
statistical properties.
The purpose of this paper is two-fold. First, we introduce a vintage-
accounting framework to quantify the importance of the extensive and the
intensive margin in the growth of world trade since World War II. In doing
so, we separate two forms in which the extensive margin may arise: through
the formation of new countries, and through first-time establishment of
trade relationships between preexisting countries. The second purpose is
to reformulate the gravity equation to take explicit account of the dual
margin of world trade. The extensive margin will appear in what we call
a corner solution of our generalized gravity model. We discuss econometric
implications of this model and then present a consistent estimation of this
model, based on a comprehensive panel data set for post–World War II
world trade in manufactures.
Among other things, our approach allows us to readdress the so-called
“puzzling persistence of distance”, i.e., recent econometric evidence from
the gravity equation suggesting that the elasticity of bilateral trade with re-
spect to distance has increased (in absolute value) over time. This evidence,
surveyed by Disdier and Head (2004), seems at odds with received wisdom
and extraneous evidence that highlights improvements in transport and
communication as a key force of globalization. Buch et al. (2004) argue
against the notion of a “distance puzzle” in which globalization may work
as much through affecting the intercept of the gravity equation, as through
a change in the estimated distance elasticity. However, many things other
than transport technology may affect the intercept. At least, an increasing
(absolute value of the) distance elasticity constitutes a puzzle in which it
implies that technical progress has been biased toward short distances, con-
trolling for potentially confounding factors such as membership in regional

2See also Wang and Winters (1992), Evenett and Venables (2002), and Haveman and
Hummels (2004).
644 Review of World Economics 2006, Vol. 142 (4)

trade agreements. This fact is difficult to reconcile with the largely undis-
puted fact that the past five decades have witnessed enormous progress in
air and sea transport and long-distance communication.
We argue that the distance puzzle may simply reflect a mis-specification
of the gravity equation that arises from inadequate treatment of the afore-
mentioned dual margin of world trade. More specifically, we show that in
the corner-solutions version of the gravity model the usual linear estimator
as well as the nonlinear least squares estimator imply a mis-specification
bias that shows up in the form of the distance puzzle. More importantly, the
conventional approach of estimating the gravity equation does not disen-
tangle the extensive and the intensive margins of world trade. As a result, the
coefficient estimates are devoid of a clear theoretical interpretation. Based
on our corner-solutions model, we achieve such a disentangling by means
of a Tobit estimation approach. In addition to resolving the distance puzzle,
our results also indicate that WTO membership has been more conducive to
trade than would appear from previous evidence presented by Rose (2004).
Our paper is closely related to Helpman, Melitz and Rubinstein, hence-
forth called HMR (2004). However, there are several important differences.
First, our novel vintage-accounting framework allows for a complete decom-
position of the growth of world trade into movements on the intensive and
the extensive margins, including an important dissection of the extensive
margin into what we call the extensive margin proper, and a “pseudo-
extensive” margin which arises from the emergence of new countries. We
thus provide a richer dissection of world trade than do HMR. Secondly, while
HMR treat zeros in their trade data as missing values, applying a Heckman
sample selection procedure, we take the zeros at face value. In particular,
we argue that observed zeros contain valuable information which should be
exploited for efficient estimation, and we do so using a Tobit approach to es-
timate our corner-solutions model.3 We do acknowledge that the zeros may
also reflect mis-reporting and mis-measurement, particularly that of small
and poor countries. But when confronted with zeros in the trade data, sim-
ply treating them as missing values is hardly less arbitrary than our choice of
treating them as a corner solution. While our approach poses econometric
problems in that the log of zero is not defined, it does allow us to extract
3 While formally closely related to the Tobit model, the Heckman procedure addresses the
problem of nonrandom selection of some country pairs with positive or zero trade vol-
umes into the sample. The corner-solutions model, in contrast, starts from the assumption
that all relevant data is observed, and that the problem is not one of sample selection but
rather of how to deal with the censoring of the dependent variable at zero.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 645

more information from the data, particularly relating to the role of distance
and other variables affecting the extensive margin of world trade. Finally, in
contrast to our paper, HMR do not address the time-varying nature of the
distance coefficient.
The importance of extensive versus intensive margins in understanding
the evolution of international trade volumes has been recognized in several
recent studies. Fundamentally, trade growth can be decomposed in several
ways, depending on the researcher’s interest and the nature of data avail-
able. In the present study, we use aggregate data with countries as units
of observation, distinguishing between changes in the number of active
bilateral trade relationships (extensive margin), and the growth of trade
volumes in existing relationships (intensive margin). Bernard et al. (2006)
focus on firm-level data and decompose the growth in U.S. exports into
entry and exit, respectively, of firms into and out of exporting (extensive
margin), and changes in foreign sales that exporting firms achieve (intensive
margin). In a similar vein, Hummels and Klenow (2005) draw on highly
disaggregated product-level data to distinguish between the variety dimen-
sion of U.S. trade growth (extensive margin), the quality (price) and the
quantity dimension (intensive margins). Another popular way of looking
at the evolution of world trade is to distinguish between trade in final and
intermediate goods, as emphasized by Yi (2003). In this context, the inten-
sive margin is related to changes in trade volumes, based on a given pattern
of vertical specialization, while the extensive margin addresses the question
of whether a certain input is traded internationally or not. In each of these
cases, trade costs play an important role in the story. Bernard et al. (2006)
emphasize that reductions in trade costs may boost industry productivity,
even without intra-firm productivity gains, through entry and exit at the
extensive margin alone. Hummels and Klenow (2005) point out that adverse
terms-of-trade effects are much less likely if a country’s trade grows at the ex-
tensive product margin, increasing the range of products traded rather than
the volume of trade within a given range. And finally, Yi (2003) demonstrates
that an increase in vertical specialization at the extensive margin of inter-
mediate inputs trade may explain the large and nonlinear responsiveness
of trade volumes with respect to reductions in trade barriers. Interestingly,
the studies cited above refrain from using corner-solutions models in their
econometric analyses, looking at each margin separately. Hence, this pa-
per provides a methodological contribution that extends beyond the role
of distance as a trade barrier on the country-level extensive margin that
we focus on.
646 Review of World Economics 2006, Vol. 142 (4)

The paper is structured as follows. Section 2 discusses the methodolog-

ical problem underlying the distance puzzle and gives intuitive reasons why
a consistent incorporation of the dual margin of world trade should be an
important part of the solution. Section 3 provides an in-depth dissection
of world trade growth from 1950 to 1997, documenting the relative impor-
tance of the extensive and the intensive margins of globalization. Section 4
introduces the corner-solutions gravity model and discusses the relevant
econometric issues leading to the Tobit approach. Section 5 presents empir-
ical results from a Tobit estimation, and Section 6 summarizes the results
and their implications for future research.

2 A Methodological View on the Distance Puzzle

Fundamentally, the distance puzzle may be seen as a reflection of missing

trade, meaning that observed trade through time increasingly falls short
of what early estimates of the gravity model would predict, based on the
evolution of time-varying determinants of trade, like income. Provided
these estimates are accepted on face value, missing trade gets picked up by
an increasing barrier effect attributed to time-invariant distance. If there is
a common trend in missing trade and income, then the income coefficient
might pick up part of it, if allowed to be time-variant and estimated jointly
with the distance coefficient, thus taking some of the explanatory burden
off the distance variable. However, if we still estimate an increasing role
of distance through time, and if we have sufficient extraneous evidence on
a falling time trend of transport costs, then the simple verdict is that gravity
theory performs badly, and increasingly so over time. The challenge then
is to find the “missing factors” reflected in seemingly missing trade. Any
reformulation of the theory that takes the explanatory burden from missing
trade off the distance variable would be seen as an improvement.
Why should we expect the dual margin of trade to play an important
role in this attempt? The intuition runs as follows. Gravity theory maintains
that trade is the result of mass attraction and resistance from geographi-
cal distance. If attraction in some cases is not strong enough to generate
trade at all, then ignoring such cases altogether implies that we systemati-
cally overestimate the force of attraction, or—equivalently—underestimate
the trade-inhibiting force of distance. Now suppose that the prevalence of
such “zero trade” cases falls through time, say because of technological
improvements in transport and communication. Then, the bias toward
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 647

overestimation of attraction and underestimation of distance falls through

time. The distance puzzle may thus be a reflection of an ever smaller under-
estimation of the distance barrier, as world trade expands on the extensive
margin. However, it is not clear a priori to what extent the bias (and its
change through time) arises with the income coefficient, or the distance
coefficient. The answer requires empirical analysis of the dual margin.4
There are, of course, other factors that might play a role. An argument
often made relates to trade liberalization. For large distances, tariffs are
generally a relatively small part of overall trade costs. If tariffs are equal
for all trade relationships to start with, and if they are reduced by an equal
(absolute) amount, then—other things equal—the percentage reduction in
the destination price is larger for low-distance trade. Tariff liberalization will
then have a disproportionately large impact on short-distance trade. In the
gravity model, this might be picked up by an increasing role of distance as
a trade-inhibiting factor through time. However, this route of explanation
hinges on an equal level of tariffs to start with, and on an equal amount of
tariff liberalization across all trade relationships, which seems questionable.
In any case the effect is easily controlled for in the empirical analysis pursued
in this paper.
A further point relates to the degree of product differentiation and
substitutability. If goods from different locations are distant substitutes,
then increasing distance might not reduce trade by much. If they become
closer substitutes, then distance plays an ever increasing role, until—in
the limit with perfect substitution—trade in any good occurs only with the
closest location. The distance puzzle could thus be due to a long-run decline
in the importance of product differentiation in traded goods. However, this
explanation seems questionable, since it would appear at least as plausible
that the role of product differentiation in trade has increased, rather than
declined over recent decades. Moreover, if valid, the argument implies that
the empirical performance of the gravity equation would fall through time.
If anything, the empirical literature points to the opposite.
A final argument relates to foreign direct investment (FDI). If trade
and FDI are substitutes, and if during the past decades FDI has system-
atically replaced long distance trade more than short distance trade, then
4 There is an additional motivation for looking at the extensive margin which has to do
with welfare. If trade is based on product differentiation along the lines of the Dixit–
Stiglitz approach, as suggested by gravity theory (see below), then expansion of world
trade on the extensive margin seems particularly important from a welfare perspective,
since it increases the degree of product differentiation.
648 Review of World Economics 2006, Vol. 142 (4)

this effect might get picked up by a rising estimated distance coefficient

in a gravity equation. However, the relationship between trade and FDI
is far from clear cut, both theoretically and empirically. The same holds
true for the role of distance in FDI. While there is no doubt that distance
and trade costs play a role for FDI, particularly through the well-known
proximity-concentration trade-off, it is not at all clear why FDI should
be more attractive for long-distance markets. Market-size and monitoring
costs play a role as well. Moreover, FDI may be used as a platform to serve
other markets, in which case it is complementary to trade (Neary 2005).
Empirically, distance appears to play an ambiguous role for FDI, which
would cast additional doubt on whether it is a promising candidate for
resolving the distance puzzle.5
Overall, our priors are that the distance puzzle establishes a convincing
case for exploring the dual margin in world trade, and to extend the gravity
model in such a way that it consistently captures simultaneous movements
on both margins through time. We therefore move on to a detailed empirical
account of such movements, followed by a reformulation of the gravity
equation, including a discussion of the econometric issues arising from
the dual margin, and empirical panel estimation, using a Tobit estimation
approach.

3 Decomposing the World Trade Expansion

Available data sources treat any country pair for which there is no reported
trade as a missing observation. This must be distinguished from explicit
recording of zero trade. Unfortunately, data based on explicit reporting
of zero trade are not available. However, an extensive scrutiny of the IMF
Direction of Trade Statistics (DoTS) by Gleditsch (2002), based on a detailed
comparison with other data (e.g., UN COMTRADE and WTO), shows that
80 percent of all observations coded as missing do in fact represent zeros. In
this paper, we therefore rely on the DoTS and, following Coe et al. (2002)
and Brun et al. (2004), we treat missing observations as zero trade. However,
while the evidence provided by Gleditch does provide some reassurance, this
strategy admittedly involves a measurement problem. We shall return to it
when discussing econometric issues below.

5 The role of distance for FDI has been examined, for instance, by Markusen (2002: ch.10)
and by Egger and Pfaffermayr (2004), with ambiguous results.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 649

Import data are usually more reliable than export data since imports
constitute a tax base, while no comparable incentive for correct reporting
exists on the export side. This is also reflected by a larger number of country
pairs with positive trade, if bilateral trade data are constructed from import
data alone. Better reliability has prompted some researchers to construct
trade data from imports alone; see Coe et al. (2002) and Brun et al. (2004).
But imports are evaluated c.i.f., including costs of transport and insurance.
Hence, using such data in a gravity regression yields an inconsistent esti-
mator for the distance coefficient, as distance will be correlated with the
error term. Using export data may involve a cost in terms of larger errors,
but avoids this correlation and thus yields consistent estimates. In this pa-
per, we follow Rose (2004) in using an average of c.i.f. import and f.o.b.
export values, in order to obtain a maximum number of observations.6
In the sequel, Tij is defined as the sum of bilateral exports and imports
recorded by countries i and j, divided by 4.7 The data cover 1950–1997 trade
in manufactures and are in constant U.S. dollars, based on the U.S. CPI
(1983 = 100).
We define the vintage of a trade relationship as the earliest time at which
trade may occur between a specific pair of two countries, based on a) their
independent existence and b) their principal openness. A country is judged
open if it reports trade with at least one other country. Of course, an open
country need not trade with all other open economies. We use Nt,h to denote
the number of trade relationships of vintage h that are active at time t. This
will typically be lower than the number of potential vintage-h trade relation-
ships at time t, which
is denoted by Vt,h . Total

world trade at time t may thus
be written as Tt ≡ th=t0 T t,h Nt,h ≡ Tt th=t nt,h ζt,h Vt , where nt,h ≡ Nt,h
0 Vt,h
is defined as

the share of active trading relationships within vintage h, and
t
ζt,h ≡ Vt,h h=t0 Vt,h is defined as the share of vintage h in the potential
number of trading relationships. T t,h is the average bilateral trade volume
based on trading relationships of vintage h, and T t is the average volume

t 
t
 

tat time t across all vintages: T t ≡ h=t0 T t,h
of bilateral trade h=t0 Nt,h .
Finally, Vt ≡ h=t0 Vt,h denotes the overall number of potential trading
relationships, and t0 denotes the “beginning of time”. Obviously, h ≤ t.

6 We use the data kindly made available by Andy Rose on his website. See [Link]
[Link]/arose/[Link]. Choosing a sample period from 1950 to 1997 avoids
incomplete recordings at the beginning and toward the end.
7 Where only 3 (or 2) observations on bilateral trade are available, their sum is divided
by 3 (or 2).
650 Review of World Economics 2006, Vol. 142 (4)

We now call

t−1  
∆Tint,t ≡ h=t0 T t−1,h Nt−1,h
t,h Nt−1,h − T (1)
a movement of world trade on the intensive margin, where preexisting
relationships vary in trade volumes. Accordingly, variations in the number
of active trading relationships,


∆Nt ≡ t−1 h=t0 (Nt,h − Nt−1,h ) + nt,t ∆Vt = ∆Nx,t + nt,t Vt,t , (2)

are called movements on the extensive margin of world trade, where ∆Vt ≡
Vt − Vt−1 is the change in the number of potential trade relationships, due
to the formation of new countries or disappearance of old ones. If no old
countries disappear, then ∆Vt ≡ Vt,t , as assumed in the above equation.
Equation (2) separates two types of extensive margins. At the extensive mar-
gin proper, denoted by ∆Nx,t , the utilization of potential trade relationships
between preexisting countries and vintages h ≤ t − 1 changes from t − 1 to
t. By way of contrast, ∆Vt captures the first-time emergence of new trading
partners, and disappearance of existing ones. This is certainly exogenous to
gravity theory. We therefore call ∆Vt the pseudo-extensive margin of world
trade. But the extent to which new potential trade relationships become

term nt,t , is treated as endogenous below.
active at once, i.e., the
Defining Nt ≡ th=t0 nth ζth Vt , world trade growth may now be decom-
posed as follows:
t Nt−1 + ∆Nt T
∆Tt ≡ ∆T t ≡ ∆T t Nt−1
(3)
t ,
+ (∆Nx,t + nt,t ∆Vt )T
where ∆Nt is taken from (2). With this decomposition, changes on the
extensive margin are “weighted” by end-of-period average trade volumes,
while changes on the intensive margin are “weighted” by beginning-of-
period numbers of trade relationships. The decomposition could equally
well be defined with opposite weighting. Indeed, one could construct more
complex weighting schemes that maintain the identity between observed
trade growth and the component terms. However, the component terms are
used only for descriptive purposes and do not enter the estimation procedure
below. Hence, the weighting chosen in (3), while in some sense arbitrary,
does not give rise to a measurement bias. But in this descriptive section it is
still important to be aware of the arbitrariness of the decomposition chosen,
and to duly recognize its implications when interpreting the numbers; see
below.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 651

Figure 1: The Extensive Margin of Merchandize World Trade

(Looking at Vintages up to 1950 / 1960 / 1980 / 1993 / 1997)

Figures 1 through 3 provide descriptive evidence on the role that these

margins have played in post–World War II evolution of world trade in
manufactures. Figure 1 highlights the extensive margin, looking at the in-
creasing utilization over time t of potential trading relationships, singling
out

1950five different groups of vintages.
The left-most horizontal line gives
1950
h=t0 Vt,h , while the line below gives h=t0 Nt,h for t up to 1959. Note the
difference between t0 (“beginning of time”) and the beginning of the sample
period

60 which is
1950. The corresponding lines for 1960 ≤ t ≤ 1979 depict
60
V
h=t0 t,h and h=t0 Nt,h , respectively, and analogously for vintages up to
1980 and 1993 further to the right.8 These lines do not fully trace out the
evolution of world trade at the extensive margin. Indeed, they do not even
look at trade as such, but simply count trading relationships for four cumu-
lative groups of vintages, in order to illustrate the type of movement that
occurs at the extensive margin. The jumps in the horizontal lines indicate

8 These points in time have been chosen primarily to illustrate the definitions of the ex-
tensive and the pseudo-extensive margin. A more continuous way of tracing the pseudo-
extensive margin in the evolution of world trade is presented in Figures 3a and 3b below.
652 Review of World Economics 2006, Vol. 142 (4)

movements on the pseudo-extensive margin when moving from one group

to the next, while the gap between the two lines at any point in time reveals
the extent to which potential trading relationships of the respective groups
of vintages have not yet become active.
Thus, of the potential trading relationships based on the group of vin-
tages up to 1950, about 40 percent had not yet become active by 1950 (the
numbers are in natural logs). By 1959, the active relationships within this
group have grown by about 20 percent, narrowing the gap of unutilized rela-
tionships to 30 percent. Including new vintages emerging between 1950 and
1959, mainly from colonies becoming independent, the maximum number
of trading relationships increases by about 65 percent, as indicated by the
second horizontal line starting at 1960. However, the number of active re-
lationships increases only by about 40 percent (comparing the solid lines
between 1959 and 1960). Hence, the gap of unutilized relationships within
this larger group of vintages (comparing the dashed and solid line) is still
43 percent in 1960, but is reduced substantially in the next two decades.
Analogous interpretations hold for the subsequent groups of vintages, in-
cluding 1960 ≤ h ≤ 1979 and 1980 ≤ h ≤ 1992. Notice the relatively large
jump at the sample end of 1997, reflecting the emergence of new indepen-
dent states from disintegration of the former Soviet Union and Yugoslavia.9
Obviously, these are institutional shocks driven by forces other than those
of the intensive margin. Our statistical model therefore does not purport
to explain ∆Vt . On the other hand, the number of new vintages that are
immediately active, nt,t , is treated as an endogenous variable.10 In terms of
Figure 1, what we want to explore is the time-varying role of distance in
explaining the gap between the horizontal and the solid line. Disintegration
of countries is also likely to increase the share of low-distance trade relation-
ships. But, as will be seen below, such a change in the sample composition
does not, in and of itself, generate any estimation problem regarding the
distance coefficient.
Figure 1 clearly shows that during our sample period there was signifi-
cant change at both types of extensive margins, and there is still much room
for further movements, even at the end of the sample period in 1997. It does
not, however, tell us how much trade was involved in movements at the ex-
tensive margin. One would probably expect the amount of trade generated
9 It should be emphasized that COMECON countries did not belong to the sample before
opening up and disintegration.
10 Referring to the world trade matrix, ∆V > 0 amounts to adding new rows and
t
columns, while nt,t gives the share of new cells emerging in “filled form”.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 653

Figure 2: Growth of Merchandize World Trade

(Vintage Distribution of Cumulative Growth 1950–1997)

at this margin to be rather small.11 Figure 2 therefore moves to a somewhat

more comprehensive perspective by asking a very simple question: What is
the contribution of vintages up to H toward the overall cumulative growth
in world trade between 1950 and 1997? The bars indicate the frequency dis-


1950
tribution, plotting the ratio of H 
h=t0 T 1997,h N1997,h −

h=t0 T 1950,h N1950,h

1997
1950
to h=t0 T 1997,h N1997,h − h=t T 1950,h N1950,h on the left-hand vertical axis,
0
with H going from 1950 to 1997 on the horizontal axis. Using the right-
hand axis, the line depicts 
the growth rate for trade of vintages up to H,

H

i.e., h=t0 T 1997,h N1997,h 1950 
T 1950,h N1950,h − 1 between 1950 and 1997

H h=t0 
1950

of vintages up to H, i.e., h=t0 T 1997,h N1997,h 
h=t0 T 1950,h N1950,h − 1.
Figure 2 is perhaps best understood by looking at extreme cases. If all
growth had happened entirely at the intensive margin, then the distribution
function would be degenerate, with all mass concentrated at vintages up

11 The working paper version of this paper Felbermayr and Kohler (2004) presents a view
on the amount of trade generated through movements at the extensive margin, relative to
the intensive margin, for the five vintage groupings depicted in Figure 1.
654 Review of World Economics 2006, Vol. 142 (4)

to H = 1950. If growth had evolved in a completely symmetric way at the

extensive margin only, then we would have a linear increase of the bars,
and a straight line for the growth rates. Vertical jumps at interior points
indicate movements at the extensive margin (proper plus pseudo), while
flat segments indicate prevalence of the intensive margin. For example,
based on vintages up to 1950 the growth rate of world trade was about
1.100 percent, contributing about 60 percent to cumulative growth during
the entire sample period. Including vintages 1950–1960, the growth rate
increases to about 1.420 percent, contributing a further 20 percent to overall
cumulative growth. Cumulative growth of trade from 1950 to 1997 involves
both, episodes where the contribution was more important on the extensive
margin (late 1950s and early 1960s, as well as the 1990s) and an interim
period dominated by the intensive margin.
Finally, Figures 3a and 3b depict a decomposition of world trade
growth into its constituent parts according to equation (3) above. Plot-
ting t = 1959, ..., 1997 on the horizontal axis, Figure 3a reveals how the
cumulative difference of actual trade from the initial volume for 1950,
indicated by bars, may be decomposed into changes at the respective mar-

Figure 3a: Decomposing the Evolution of Merchandize World Trade

(Cumulative Difference to 1950, in Trillion Real Dollars)
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 655

Figure 3b: Decomposing Yearly Growth of Merchandize World Trade

(Yearly Growth Rates 1950–1997)

gins, indicated by lines. The intensive margin line depicts (T t − T 1950 )

1950
h=t0 N1950,h , while the line for the extensive margin proper plots

1950
1950 
t
T h=t0 Nt,h − h=t0 N1950,h , and the pseudo-extensive margin is plot-

t
1950 
ted as Tt h=t0 Nt,h − h=t0 N1950,h . By definition, the three lines add up
to the bars. Figure 3b plots the yearly growth rates. Extreme values of yearly
growth tend to be dominated by the intensive margin, the exception being
the period from the late 1950s to the late 1960s, as well as the early 1990s.
But even for the 1970s and 1980s, the extensive margin did play a role, as
evidenced by the difference between the bar-values and the solid line.
Figure 3a indicates a surprisingly large contribution of the extensive
margin, relative to Figure 2. One should, however, be aware of the specific
decomposition chosen in (3). For the extensive margin, the contribution to
the growth of trade is weighted by the current average trade volume  Tt , while
the intensive margin receives base-period

weights from the 1950 number
of active trading relationships, i.e., 1950h=t0 1950,h . For instance, from 1950
N
to 1980 world trade has increased by about 0.93 trillion (real 1983) dollars.
If all new trading relationships emerging and becoming active between
1950 and 1980 had been trading on the average 1980 level, T 1980 , then this
margin alone (pseudo-extensive margin) would imply an increase by as
656 Review of World Economics 2006, Vol. 142 (4)

much as about 0.65 trillion real dollars. Taking only vintages up to 1950 and
the increase in their utilization (extensive margin proper) adds a further
increase by 0.14 trillion real dollars. About that same increase is observed
on the intensive margin, where we look at the trading relationships that had
been active already in 1950, assuming that they have increased their trading
volume according to T 1980 − T 1950 .
Our analysis clearly suggests that the extensive margin is important
quantitatively, both along the cross-sectional and the time-series dimension
of the data. First, even in the last year of the time span under investigation,
only about 58 percent of potential bilateral trading relationships are actu-
ally utilized. Second, about 40 percent of world trade growth from 1950
to 1997 comes from movements at the extensive margin. Admittedly, this
result is “contaminated” by the pseudo-extensive margin where new coun-
tries emerge, due to decolonialization and other fundamental changes in
the political environment. But even from 1970 to 1997, where the pseudo
extensive margin arguably plays a minor role in our data, about 15 percent
of total world trade growth is attributable to the extensive margin. The
empirical importance of the extensive margin calls for a proper model-
ing of the dual margin in gravity-type investigations of world trade. This
relates to both, the conceptual framework and the econometric implica-
tions. In the following sections of the paper, we undertake such a modeling
effort.

4 Modelling the Dual Margin of Globalization

Established trade theory does not go very far in explaining movements on

the dual margin of world trade. In a large class of models, the focus lies
squarely on countries’ overall trade. As noted recently by Deardorff (2004),
the comparisons are mostly done globally, and not by pairs of countries, thus
leaving bilateral trade undetermined. This seems justifiable on the grounds
that bilateral trade of a country is largely irrelevant for its level of welfare,
but in many respects bilateral trade is important.12
Common sense suggests that bilateral trade is importantly determined
by trade costs related to geographic distance and transport. Traditional trade

12 A case in point is the trade potential that arises if a country opens up to world trade
in the process of systemic change. Prominent examples are Eastern European countries in
the 1990s.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 657

theory almost entirely neglects all such costs, the important exception, of
course, being the gravity approach where distance as a trade-inhibiting
force combines with economic mass of two countries to explain bilateral
trade. The gravity force arises, whenever trade is based on perfect special-
ization, meaning that each good (with a sufficiently narrow definition) is
produced in only one country, but consumed everywhere (say due to identi-
cal preferences). Such a case arises, almost by definition, in a love-for-variety
product differentiation framework with increasing returns to scale, which
is the usual theoretical justification of the gravity equation going back to
Anderson (1979). It also arises with comparative-advantage-based trade,
provided domestic trade is treated on an equal footing with foreign trade,
or if the number of goods and countries is very large, relative to the number
of factors; see Deardorff (1998).
The gravity equation derived using models of monopolistic competition
assumes that trade costs are of the iceberg type and that there is no fixed
cost of entering a certain market. With CES preferences, this model cannot
account for zero trade. By way of contrast, our corner-solution approach
to gravity importantly hinges on zero trade as an equilibrium outcome.
Theoretical foundations of the gravity equation that allow for zero trade
have been provided by Feenstra et al. (2001), Eaton and Kortum (2002),
Haveman and Hummels (2004), and more recently Helpman et al. (2004).
In this paper, we abstain from providing an explicit structural model of
trade leading to an estimable gravity equation that allows for zero bilateral
trade volumes. For our purposes it suffices to work with an reinterpretation
of the standard gravity equation as derived, e.g., by Baier and Bergstrand
(2001) and reviewed by Feenstra (2004).
Suppose that the bilateral trade potential is determined according to the
conventional gravity equation. It is, thus, positive regardless of the magni-
tude of real trade costs. Our key assumption in this paper is that, for reasons
external to each exporting firm, positive trade between two countries arises
only if their bilateral trade potential exceeds some dyad-specific threshold
value Tijt . One may think of various explanations for the existence of such
a threshold value, the principal line of reasoning is as follows. Maintaining
trade relationships requires certain infrastructure and institutions that fa-
cilitate an efficient flow of information and exchange of ideas, in addition
to the best-practice transport of goods. Such institutions need to be present
both, in the export and import country. While the private costs of informa-
tion flows and transport are captured in the conventional gravity equation
determining the trade potential, the underlying infrastructure and insti-
658 Review of World Economics 2006, Vol. 142 (4)

tutions often exhibit characteristics of public goods. In other words, they

involve fixed costs (largely independent of the magnitude of potential trade)
that are borne by the public sector. Important examples are international
payment systems, legal agreements, consulates, or the activities of the re-
spective chambers of commerce. The assumption here is that individual
firms will not be able to trade in the absence of services provided by such
institutions. A further key assumption is that the benefits accruing to both
countries from such institutions depends on how much trade they are likely
to support once established. And this is where the gravity-determined trade
potential comes into play. Rational governments will invest into these insti-
tutions only if the expected benefit from a materializing trade potential in
present value terms exceeds the investment cost. These will, in turn, deter-
mine the threshold value T ijt that the trade potential needs to surpass for
governments to invest into the required institutions.
For the present purpose, we need not provide an explicit model deter-
mining the threshold value T ijt , but it would seem natural to assume that
it depends on geographical barriers such as distance, and on the two coun-
tries’ GDP levels. It is thus specific to each dyad. Armed with these ideas, we
may now modify the standard gravity equation by postulating the following
system of equations
 
∗ Yit Yjt θijt 1−σ
Tijt = , (4)
pσi Ȳt Pjt
∗
Tijt  
Tijt = ∗
max 0, Tijt ijt ,
−T (5)
∗
Tijt ijt
−T
where the first expression explains the trade potential Tij∗ . It is taken from
Baier and Bergstrand (2001: 9–10), where it is derived as an approximation
to a gravity equation that includes multilateral trade resistance, proposed
by Anderson and van Wincoop (2003).13 The equation is based on product
differentiation on the firm level, with σ > 1 as the elasticity of substitution

13 Our gravity equation is log-linear in distance. This is a common feature in the applied
literature. However, it is worth noting that the Anderson and van Wincoop (2003) version
of the gravity model is highly nonlinear in distance (and other covariates). Differentiat-
ing their system of equations, one finds that the elasticity of trade with respect to distance
depends on the Helpman (1987) similarity index. Potentially, changes in this index could
also explain the time pattern of the elasticity of distance. However, empirically, the index
exhibits very little time variance and thus does not seem a good candidate explanation for
the distance puzzle. Therefore, we do not pursue this argument any further.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 659

between varieties. While Yi and Yj are total GDPs of the two countries i
and j, Ȳ denotes world-GDP, θij > 1 indicates iceberg trade costs and pi are
country i’s mill (ex factory) prices and Pj is country j’s exact price index,
based on love-of-variety preferences. Equation (5) introduces the threshold
∗
value to relate the trade potential Tijt to actual trade Tijt , in line with the
reasoning proposed above. Trade will materialize in its full potential if it
exceeds the threshold level Tijt , otherwise bilateral trade is zero. We call this
the corner solutions of the gravity model.
Specifying iceberg-type trade costs determined by geographical and
cultural distance, as well as membership in trade agreements, we may write
β γ
θijt = Θijt Dδij Bij Lij Aαijt , (6)

where Dij is bilateral distance, Bij is a dummy for a common border, Lij
is a dummy for common language, and Aijt denotes joint membership in
trade agreements such as the WTO or the European Union. Importantly,
Θij denotes the overall level of technology pertaining to transport or com-
munication. Substituting and taking logs gives the following equation for
the bilateral trade potential
∗
 
ln Tijt = (1 − σ) Θijt + ln Yit Yjt + (1 − σ)
 
× δ ln Dij + β ln Bij + γ ln Lij + α ln Aijt (7)
+ φi + φj + φt + uijt .

Note that this specification includes separate sets of dummies for countries
i and j as well as time dummies to control for the unobserved multilateral
resistance indexes proposed by Anderson and van Wincoop (2003). Techni-
cal progress in transport or communication technologies may affect both,
the constant (1 − σ) Θij and the coefficients δ and β. However, the effect of
distance on transport costs, δ, is not identified in (7). Following established
literature, we make the identifying assumption that σ is constant over time.
In the subsequent discussion, we shall refer to the right-hand-side variables
in (7) collectively as Xijt .
As we have shown in Section 3 above, our data set features a large number
zero trade cases, whereby the number of cases with positive trade increases
over time. Broadening our perspective, we now need to distinguish three
types of events. The first is the formation of new countries. In terms of the
world trade matrix, this adds new rows and columns. Although we have
discussed this under the pseudo-extensive margin above (denoted by ∆Vt
660 Review of World Economics 2006, Vol. 142 (4)

in Section 3 above), it is driven by forces entirely unrelated to the gravity

model, and indeed unrelated to economics. This margin is therefore treated
as exogenous in our statistical model. The second type of event occurs when
changes in gravity-type variables cause empty cells to be filled over time.
This is the extensive margin proper, and it includes the question of whether
a new cell enters in “filled form” (denoted by nt,t in Section 3 above), which
is therefore treated as an endogenous variable in our model. And finally,
there is the traditional case where a cell entry increases or falls in magnitude,
which is the intensive margin.
A separate issue relates to measurement. As we have noted above, avail-
able data sets do not include explicit reporting of zero trade, but simply
code all country pairs not reporting positive trade as missing observa-
tions. This precludes a multistep procedure, where the first step would
explain whether a given country pair is a missing observation, followed
by a separate explanation of whether a given observation involves positive
or zero trade, and a final step explaining the extent of trade if positive.
The first step is not only negated by lack of data, it would also feature ex-
planatory factors for missing observations that are unrelated to the gravity
equation and not of immediate interest for the present purpose. There-
fore, our strategy, supported to some extent at least also by empirical evi-
dence (see Section 3 above), is to directly treat missing observations as zero
trade.
Zero trade entries in the data pose a problem when taking logarithms.
One way to deal with this problem is to estimate the gravity equation in
semilogarithmic form, with Tijt instead of ln Tijt as the dependent variable.
This is the strategy followed by Eaton and Tamura (1994). However, the-
ory suggests ln Tijt as the correct left-hand variable. In line with this, the
literature so far clearly shows that the semilog equation performs consid-
erably worse than the log-log one.14 Hence, we follow Eichengreen  and

Irwin (1995, 1997), in transforming the dependent variable to ln 1 + Tijt ,
which may be justified  by the following logic of interpretation. If Tijt is
large, then ln 1 + Tijt is approximately equal to ln Tijt , thus validating the
common practice of interpreting
  estimated coefficients as elasticities. If Tijt
is very small, then ln 1 + Tijt is approximately equal to Tijt , in which
case coefficients have to be interpreted as semielasticities. Arguably, either

14 We have experimented with the semilog specification. Qualitatively, the main patterns
in the interesting regression parameters do not change. But many coefficients are estimated
imprecisely and with elasticities that take implausible magnitudes.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 661

of the two strategies is somewhat arbitrary, making the results dependent

on the unit of measurement.  To check the robustness, we shall express
the dependent variable as ln a + Tijt , experimenting with different values
for a > 0.15
There is a key statistical problem that follows from our corner-solutions
version of the gravity equation. This is that the conditional mean of actual
trade ln Tijt cannot be linear in Xijt , because there is positive probability mass
at Tijt = 1, or ln Tijt = 0. Addressing this problem by means of nonlinear
least squares (NLS), as in Santos Silva and Tenreyro (2003) or Coe et al.
(2002),  poses several problems. First, since ln Tijt includes corner outcomes,
ln Tijt Xijt is likely to be heteroskedastic, which renders NLS inefficient.16
Using weighted NLS requires  an arbitrary
 choice of a specific model for the
conditional variance, var ln Tijt Xijt , and would thus seem questionable;
see Wooldridge (2002: 518 ff). More importantly,   the coefficients obtained
by NLS estimation of a model for E ln Tijt Xijt are difficult  to interpret.
They do not tell us anything about distribution of ln Tijt Xijt , other than
its mean. In particular, they would not allow us to empirically identify
the intensive and extensive margins of world trade.  In theunderlying dis-
tribution, the extensive margin  is given
 by Pr ln Tijt ≥ 0 Xijt , while the
intensive margin is given by E ln Tijt Xijt , ln Tijt > 0 .
To deal with these difficulties, we treat our estimation equation as
a corner-solutions model in the sense of Wooldridge (2002), which is a spe-
cial case of the censored regression. In the following, we call Xijt , which
stands for all exogenous explanatory variables in (7), the “the gravity force”.
The underlying statistical model is written as
 ∗

ln Tijt − Xijt β σ ∼ N[0, 1] , (8)

where β is a vector of gravity-related parameters, and N[0, 1] denotes the

standard-normal distribution. In other words, the mean of trade poten-
∗
tial ln Tijt , conditional on gravity forces Xijt , is equal to Xijt β, with variance
equal to σ 2 .

15 Helpman et al. (2004) circumvent these difficulties by interpreting all observations with
Tijt = 0 as missing observation. However, this comes at the cost of sacrificing all informa-
tion present in zero trade data, which clearly runs counter to our corner-solutions inter-
pretation of the gravity equation.   
16 Heteroskedasticity arises because at the “corner”, i.e., where E ln T X
ijt ijt = 0, we ob-
serve only one-sided
  deviations.
 More generally, for values of explanatory variables leading
to a lower E ln Tijt Xijt = 0, the variance of the error term is smaller.
662 Review of World Economics 2006, Vol. 142 (4)

The relationships between the parameters β and the two margins men-
tioned above are as follows. The overall effect is given by
E(ln Tijt |Xijt ) = Φ(Xijt β /σ)[Xijt β + σλ (z)] , with (9)
∂ E(ln Tijt |Xijt )/∂Xijt = β Φ(Xijt β /σ) ,
r r
(10)

where z ≡ Xijt β σ, and a superscript r denotes a specific explanatory vari-
able r. The term λ(z) is the inverse Mills ratio, λ(z) =: φ(z)/Φ(z), with φ
and Φ denoting the standard normal density and the standard normal
17
distribution function, respectively.
  Given (8), the conditional probabil-
ity of positive trade, Φ Xijt β /σ , is less than 1, which gives rise to at-
tenuation.18 The reason for this is straightforward. Suppose β r > 0, and
r
consider a reduction in Xijt . The marginal effect on the conditional mean
∗
of the uncensored variable Tijt must clearly be larger (in absolute value)
than the effect on the conditional mean of the “cornered” variable Tijt .
However, the economic interpretation of (10) in the gravity context is not
straightforward, as it is an amalgam of the intensive and the extensive
margin.
The intensive margin relates to the expected value of trade, conditional
on the gravity force Xijt (see Greene 2003: 670):
E(ln Tijt |Xijt , ln Tijt > 0) = Xijt β + σλ (z) , with (11)
∂ E(ln Tijt |Xijt , ln Tijt > 0)/∂Xijt
r
= β r − β r λ(z)[z + λ (z)] , (12)
where λ(z) again is the inverse Mills ratio, with z ≡ Xijt β/σ. The extensive
margin relates to the probability of a given country pair already having
taken up a trade relationship, again given the gravity force and time-specific
effects:
Pr(ln Tijt ≥ 0|Xijt ) = Φ(Xijt β/σ) , (13)
where the marginal coefficient, ∂ Pr[ln Tijt ≥ 0|Xijt ]/∂Xijt r
, may be
derived by differentiating the equality E(ln Tijt |Xijt ) = Pr(ln Tijt ≥ 0|Xijt )

17 The first term in (9) is the discrete part of the censored distribution, measuring the
probability mass at zero, while the second term is the conditional mean of the correspond-
ing truncated normal distribution. λ(y) gives the hazard function of the standard-normal
distribution; see Greene (2003: 762–763).  
18 The interpretation of Φ(X β/σ) follows from Pr ln T ∗ − X β /σ < −X β/σ =
 ∗
 ijt ijt ijt ijt
Pr ln Tijt < 0 = Φ(−Xijt β/σ). Due to symmetry, we have Φ(−Xijt β/σ) = 1 − Φ
(Xijt β/σ). Hence, Φ(Xijt β/σ) is the complementary probability for Φ(−Xijt −/σ), and thus
the probability of positive trade.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 663

×E(ln Tijt |Xijt , ln Tijt > 0), and invoking the associated marginal effects;
see Wooldridge (2002: 523). Note that the marginal effects are not con-
stant on either of the two margins. Moreover, it is clear from (13) that
the extensive margin defined in this way does include the term nt,t in equa-
tion (3), but not the pseudo-extensive margin ∆Vt , as repeatedly mentioned
above.
From equation (11), we may write

ln Tijt = Xijt β+ σλ(Xijt β/σ) + εijt

with E(εijt |Xijt , ln Tijt > 0) = 0 . (14)

It is clear that running OLS of ln Tijt on Xijt for ln Tijt > 0 amounts to omit-
ting the variable λ(Xijt β/σ). If the covariance between Xijt and λ(Xijt β/σ)
is nonzero, then the coefficients β are inconsistently estimated. Moreover,
regressing ln Tijt on Xijt using all of the data, including where Tijt = 1, will
not consistently estimate β either, since E(ln Tijt |Xijt ) is nonlinear in Xijt , β
and σ.
The bias involved becomes clear from looking at the simple univariate
case, using y and x to denote the dependent and independent variable.
The model then becomes E(y|x, y > 0) = βx + σλ(βx/σ). Ignoring the
second term and running a regression of y on x for all observations y > 0
introduces the aforementioned omitted-variable bias. The estimated coeffi-
cient β̂ is β̂ = β + σcov[x, λ (βx/σ)] /var(x). Since λ (c) = φ (c) /Φ (c) > 0
and λ (c) = −λ (c) [c + λ (c)] < 0, we know that cov[x, λ (βx/σ)] > 0.
Moreover, we know that in a sample where no corner solutions appear
E(y|x, y > 0) =E(y|x) = βx. Hence, since cov[x, λ(βx/σ)] > 0, we con-
clude that β̂ > β. The omitted-variable bias causes an overestimation of β.
In the long run, the extensive margin should disappear, as trade eventually
does cover the whole world, whence β̂ → β with time. Accordingly, β̂ falls
toward β as time unfolds.

5 Estimation Results

We now proceed toward estimating our corner-solutions model on a subset

of the data set used by Rose (2004). In particular, we focus on the time
span between 1970 and 1990, because in that period very few new countries
came into existence. Hence, we may expect that our results are note overly
contaminated by the pseudo-extensive margin (see above). The summary
664 Review of World Economics 2006, Vol. 142 (4)

statistics of the dependent variables and the covariates used in all of the
models are presented in Table 1. As argued in the introduction, we expect
that, among other things, our estimation sheds new light on the distance
puzzle in that we may separately identify the time pattern of the distance
coefficient at the extensive and the intensive margin of trade. A similar
expectation also pertains to the role of WTO membership for bilateral
trade.

Table 1: Summary Statistics

Model

Restricted sample, Full sample

T > 0 only including T = 0

Ln real value of
(exports+imports) 10.0369 3.1836 5.5360 5.5233
Share of active trade
relationships in sample 1.0000 0.0000 0.5620 0.4961
Ln distance 8.1778 0.8010 8.4887 0.8126
Ln distance × time 95.7212 49.9438 100.3139 51.3477
Dummy: both coun-
tries in WTO 0.4787 0.4995 0.3581 0.4794
Dummy: no country
in WTO 0.4423 0.4967 0.4781 0.4995
#0,1,2: sum island 0.3409 0.5364 0.3963 0.5633
#0,1,2: sum land-
locked 0.2482 0.4668 0.3585 0.5415
#0,1,2: sum Sachs-
Warner open countries 0.7001 0.6463 0.5091 0.6210
Ln product of real
GDPs 48.0002 2.5641 47.4561 2.6695
Dummy: both coun-
tries in same RTA 0.4777 0.4995 0.4154 0.4928
Ln price of oil 3.0924 0.4476 3.1016 0.4476
Ln real world GDP 16.8174 0.2186 16.8210 0.2165
Dummy: common
border 0.0299 0.1704 0.0195 0.1383
Dummy: colonial
ties 0.0240 0.1530 0.0136 0.1157
Dummy: common
language 0.2015 0.4011 0.1734 0.3786

Number of observations 102,823 186,419

Source: We use the data kindly made available by Rose on his website. See [Link]
[Link]/arose/[Link]
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 665

Before turning to the estimation results of our corner-solutions model,

we provide some “naive” exploration of the time dependency of the dis-
tance coefficient, using OLS and Probit techniques. Thus, Figure 4a shows
estimates of the distance coefficient when (7) is applied to the data re-
peatedly, year-by-year. The upper-left panel plots the estimates obtained
from OLS, restricting the sample to country pairs that do in fact re-
port positive trade. This ignores the extensive margin of trade, which
we have emphasized in our accounting framework for trade in Section 3
and which we have argued to be a promising candidate for resolving
the distance puzzle in Section 2. The regressions include a list of stan-
dard covariates (see Table 2) and a comprehensive set of country-fixed
effects. The absolute value of the distance coefficient increases over time
from about 1.10 to approximately 1.50. While the coefficient is at the
higher range of the studies discussed in the meta analysis by Disdier and
Head (2004), the time pattern neatly highlights what the literature has
dubbed the distance puzzle. We shall turn to the question of statistical
significance in this time trend below. For the present moment, the reader
should contrast the upper-left panel in Figure 4a with that in Figure 4b.
The latter shows time-varying estimates of the distance coefficients δ̃t ob-
tained in a panel framework. Time dependence of distance is captured by
substituting
 
(1 − σ) δ ln Dij = δ̃1970 φ1970 + δ̃1971 φ1971 + ... + δ̃1990 φ1990 ln Dij (15)

into (7).
In contrast to the year-by-year estimates, the panel estimation allows for
time dependency only of the distance coefficient, in addition to year-fixed
effects that account for the influence of the business cycle. All other coeffi-
cients, including the large array of country-fixed (importer plus exporter)
effects, are constrained to be constant over time. This procedure leads to pre-
cise estimation of country-fixed effects, thus controlling for country-specific
unobserved variables that are time-invariant. Interestingly, while the abso-
lute value of the distance coefficient still rises over time, the pattern is less
clear-cut than in Figure 4a. Moreover, the series seems to exhibit a structural
break in the late seventies, rather than a linear time trend. Column (1) in
Table A1 in the Appendix shows that the distance coefficients are estimated
with satisfactory precision. The important thing to note here is that the
distance puzzle survives the inclusion of fixed effects, both quantitatively
and qualitatively. It also survives other, less fundamental methodological
666 Review of World Economics 2006, Vol. 142 (4)

Figure 4a: Year-by-Year Estimates of the Distance Coefficient

Figure 4b: Panel Estimates of the Distance Coefficient

 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 667

Table 2: Regression Results: Naive Regressions of the Intensive vs. the Extensive
Margin, Panel Estimations (Dependent Variable: ln of Real Bilateral Trade)

(1) (2) (3)

Intensive mar- Extensive mar- Both margins
gin only OLS gin only PROBIT OLS
E(T|X,T > 0) P(T > 0|X) E(T|X)

Ln distance −1.3114 −0.7305 −3.2627

(0.0389) (0.0340) (0.0902)
Ln distance × time −0.0043 0.0055 0.0231
(0.0018) (0.0011) (0.0023)
Dummy: both countries 0.0934 0.1506 0.5020
in WTO (0.0846) (0.0243) (0.0945)
Dummy: no country 0.0004 0.0507 0.0679
in WTO (0.0705) (0.0180) (0.0682)
#0,1,2: sum island −0.9691 −0.5821 −1.9431
(0.4754) (0.0462) (0.1682)
#0,1,2: sum landlocked −0.3076 5.6736 5.6883
(0.0945) (0.2352) (0.1947)
#0,1,2: sum Sachs-Warner −0.0622 −0.0072 −0.0767
open countries (0.0329) (0.0093) (0.0332)
Ln product of real GDPs 0.5697 −0.0982 −0.2325
(0.0422) (0.0029) (0.0094)
Dummy: both countries −0.2760 0.1421 0.4912
in same RTA (0.0544) (0.0349) (0.0803)
Ln price of oil 0.2970 0.0358 0.3341
(0.0153) (0.0053) (0.0158)
Ln real world GDP 5.0138 1.2527 6.5843
(0.2669) (0.0995) (0.2667)
Dummy: common border 0.3309 −0.6421 −0.9815
(0.1337) (0.0301) (0.2468)
Dummy: colonial ties 1.2220 0.2732 1.2828
(0.1144) (0.0376) (0.1881)
Dummy: common language 0.3605 0.0261 0.3413
(0.0557) (0.0184) (0.0740)

Number of observations 102,823 170,432a 186,419

Number of dyads 7,539 8,154 10,416
Pseudo R-squared 0.7265 0.7116 0.7799
RMSE 1.6673 – 2.5932
Pseudo log likelihood – −33,693 –

Note: Robust standard errors (clustering by country pairs) in parentheses. Exporter and importer
fixed effects, as well as a linear time trend are included in each regression but not reported. Marginal
effects are evaluated at sample means. For dummy variables marginal effects are for discrete change
of dummy variable from 0 to 1.
a Probit estimation is over fewer observation than OLS in column (3) since the presence of exporter
and importer dummies completely predicts some outcomes, so that they drop out of the sample.
668 Review of World Economics 2006, Vol. 142 (4)

modifications, as discussed in Brun et al. (2004) and the meta analysis of

Disdier and Head (2004).19
Column (1) in Table 2 replaces (15) with a parametric specification of the
time dependency of the distance coefficient. In particular, we model a linear
trend in the distance coefficient.20 While the estimates shown in the upper-
left panel of Figure 4b suggest caution with the use of a linear trend, this
procedure allows for a simple way of testing whether there is a statistically
significant trend in the coefficients. Indeed, the time-distance interaction is
statistically significant at the 1 percent level and enters with a negative sign.
But the time effect is small in magnitude: It would take about 300 years for
the distance coefficient to double. However, there is no evidence whatsoever
pointing to a decline in the absolute value of the distance coefficient over
time. The remaining covariates included in the regression have the usual
signs and are mostly estimated with high precision.
Next, we turn to the time dependency of the distance coefficient in Probit
estimations of the extensive margin. We replace the left hand side of (7) by
an indicator variable which is unity if a given trade relationship was active
at time t, and zero otherwise. The resulting specification is estimated using
a Probit model, employing the same covariates as in the OLS regressions
discussed above. The upper-right panels in Figures 4a and 4b show that the
distance coefficient obtained in this model falls in absolute value over time.
Column (2) in Table 2 establishes that the trend in the estimates is statisti-
cally and economically significant. If the trend were to continue, the panel
estimates would imply that by the early years of the next century distance
no longer has any explanatory power for predicting the existence of bilateral
trade relationships between countries. These results are in stark contrast
to the ones discussed above, and they provide some tentative evidence for
a declining role of distance along the extensive margin.
The natural next step is to include observations with zero trade and run
an OLS regression, using the same covariates as above. The results are shown
in the lower-left panels of Figures 4a and 4b. It turns out that the extensive

19 We have run a large number of robustness checks on the persistence of the distance
puzzle in our data, using specifications that vary in parsimony (basic gravity-type regres-
sors versus additional regressors in line with Table 1), in treatment of global year-specific
events (including the price of oil and real world GDP versus unspecified year-specific ef-
fects), and error specification (pooled OLS versus random effects estimator). The distance
puzzle shows up as a robust phenomenon across all estimation procedures. More detailed
results may be found in Felbermayr and Kohler (2004).
20 (1 − σ)δ ln D is substituted by δ̄ (ln D × time) + δ̄ ln D + δ̄time in equation (7).
ij 1 ij 2 ij
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 669

margin dominates the overall assessment of the time dependency of the

distance effect. The results indicate a downward trend in the absolute value
of the distance coefficient which is statistically and economically significant.
However, it is important to remember from our discussion above that OLS
does not yield consistent estimates if we include zero trade observations.
The lower-right panels in Figures 4a and 4b show that there is evidence for
a decrease in the importance of distance in line with the OLS and Probit
estimates presented above. The time trend is statistically significant and
economically meaningful. If this trend were to continue, distance would
lose its bite on bilateral trade volumes around the year 2200.
We now turn to the time pattern of the distance effect as revealed by
a consistent estimation of our corner-solutions model. Table 3 shows re-
gression results for the corner-solutions model. Column (1) reports the
marginal effects of the covariates on the trade potential. As argued above,
these are difficult to interpret economically. Column (2) presents coeffi-
cients obtained by differentiating E ( T| X, T > 0). We interpret this column
as a consistent estimate of the intensive margin. Column (3) presents the
coefficients relating to E ( T| X), which is the unconditional expectation of
bilateral trade. Finally, column (4) contains estimates of the probability of
an interior solution, i.e., P ( T > 0| X), which we interpret as the extensive
margin. All estimates reported are marginal coefficients. Since the model
is nonlinear, they are not constant. The values shown are those for the
respective sample means.21
Importantly, we find a positive time trend in the distance coefficient
which is statistically and economically significant. This implies that the
importance of distance, if estimated in a consistent way, is indeed falling
over time, not only on the extensive margin, but also on the intensive
margin. While the exact time pattern of the coefficients may be nonlinear in
the year-by-year model (Figure 4a, lower-right panel), there is no remaining
evidence whatsoever indicating the presence of a distance puzzle. Moreover,
when looking at the panel model (Figure 4b, lower-right panel), there seems
to be a linear time trend; this also bears out in the results presented in Table 3.
As similar conclusion also holds for the effect of WTO membership on
bilateral trade. Evidence presented by Rose (2004), based on a standard
OLS estimation of the equation restricted to cases where reported trade is
positive, seems to indicate that there is no statistically significant effect of

21 In computing the partial derivatives, sums of binary variables which take integer values
over the interval [0, 1] are treated as continuous variables.
670 Review of World Economics 2006, Vol. 142 (4)

Table 3: Regression Results: The Corner Solutions Model. (Dependent Variable:

ln of Real Bilateral Trade)

(1) (2) (3) (4)

Latent Expectation Expectation Probability of
variable conditional on unconditional on interior solution
interior solution interior solution
E(T*|X) E(T|X,T > 0) E(T|X) P(T > 0|X)

Ln distance −5.0747 −2.7061 −3.7978 −0.3423

(0.1026) (0.0547) (0.0768) (0.0069)
Ln distance × time 0.0215 0.0115 0.0161 0.0015
(0.0080) (0.0043) (0.006) (0.0005)
Dummy: both coun- 2.1130 1.1721 1.6256 0.1348
tries in WTO (0.1815) (0.1007) (0.1396) (0.0116)
Dummy: no country 1.2386 0.6611 0.9266 0.0834
in WTO (0.1557) (0.0831) (0.1165) (0.0105)
#0,1,2: sum −1.6625 −0.8865 −1.2441 −0.1121
island (0.1459) (0.0778) (0.1092) (0.0098)
#0,1,2: sum land- −2.7534 −1.4682 −2.0606 −0.1857
locked (0.1115) (0.0595) (0.0835) (0.0075)
#0,1,2: sum
Sachs-Warner 1.9643 1.0474 1.4700 0.1325
open countries (0.1346) (0.0718) (0.1007) (0.0091)
Ln product of 0.5014 0.2674 0.3753 0.0338
real GDPs (0.0201) (0.0107) (0.015) (0.0014)
Dummy: both
countries 0.8828 0.4758 0.6660 0.0587
in same RTA (0.1254) (0.0676) (0.0946) (0.0083)
Ln price of oil 0.5595 0.2983 0.4187 0.0377
(0.0371) (0.0198) (0.0278) (0.0025)
Ln real world GDP 7.8197 4.1698 5.8520 0.5274
(0.5047) (0.2691) (0.3777) (0.0340)
Dummy: common −2.2660 −1.0664 −1.5137 −0.1723
border (0.4177) (0.1966) (0.2791) (0.0318)
Dummy: colonial 0.8053 0.4486 0.6231 0.0512
ties (0.2660) (0.1482) (0.2058) (0.0169)
Dummy: common 0.4637 0.2515 0.3516 0.0306
language (0.0813) (0.0441) (0.0617) (0.0054)

Number of
observations 186,419
Number of dyads 10,416
Pseudo R-squared 0.7265
RMSE 1.6673
Pseudo log likelihood −345,840

Note: Robust standard errors (clustering by country pairs) in parentheses. Exporter and im-
porter fixed effects, as well as a linear time trend are not reported. Marginal effects are eval-
uated at sample means. For dummy variables marginal effects are for a discrete change of
dummy variable from 0 to 1.

WTO membership. In effect, our estimates are a robustness check of the

Rose results, since the WTO puzzle does not disappear upon introduction of
a comprehensive set of importer and exporter dummies. However, once we
switch attention to the extensive margin of world trade, the effect of WTO
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 671

membership is statistically significant and economically meaningful: Col-

umn (2) in Table 2 shows that compared to country pairs which do not both
belong to the WTO, the probability to find an existing trade relationship be-
tween two member countries is 10 percentage points higher. This carries over
to the corner-solutions model, where the difference between the two cases
is somewhat smaller, but still highly significant. While other explanations of
the WTO puzzles have been proposed, such as the importance of restricting
the sample to countries that really abide by the WTO rules (see Subramanian
and Wei 2004), the effect of WTO membership on the extensive margin of
world trade seems a plausible, yet usually neglected possibility.

6 Conclusions

Our Tobit estimates of a corner-solutions version of the gravity model indi-

cate that the distance puzzle is a reflection of the extensive margin of world
trade expansion. This becomes most strikingly clear in our Probit estimates.
There are several routes to pursue for future research. First, distance-related
trade costs are likely to be nonlinear in distance. It seems questionable that
the oft quoted technological improvements in transport and communica-
tion have been equally relevant for long-distance and short-distance trade.
Similarly, improvements may have been differently pronounced for land and
sea transport. Obviously, an improved explicit treatment of the transport
sector, or—more generally—of trade costs, is required to fully address these
points. Moreover, distance-related trade costs are likely to be of differing
importance across goods or industries.
Meanwhile, our results should reinforce the widespread notion that the
post–World War II era is, indeed, characterized by a long-run decline in the
trade inhibiting force of geographic distance. This is most obvious when
studying the time pattern of the distance coefficient in Probit regressions
for the extensive margin of world trade. However, our view of the distance
puzzle is that, even if resolved in the strict sense of the word, it serves as
a “warning shot” against exaggerated views of a dramatically “shrinking
world geography”. Our results do not warrant the conclusion that distance
and trade costs have become negligibly small. The world has not yet become
a “global village”, and it probably never will. Indeed, a further important con-
clusion of our study is that world trade, even after 5 decades of spectacular
growth, is still far away from covering the whole world. The extensive margin
of trade, where dormant bilateral trade relationships become utilized, still
leaves much ground to be covered, and much gains from trade to be reaped.
672 Review of World Economics 2006, Vol. 142 (4)

Appendix

Table A1: Regression results: Time Pattern of the Distance Coefficient, Panel
Estimates, 1970–1990 (Dependent Variable: ln of Real Bilateral Trade)

(1) (2) (3) (4)

OLS OLS PROBIT TOBIT
dE(T|X,T > 0)/dD dE(T|X)/dD dP(T > 0|X)/dD dE(T|X)/dD

1970 −1.3218 −2.8284 −0.7304 −4.5582

(0.0412) (0.0548) (0.0302) (0.1113)
1971 1.3750 −2.7806 −0.6641 −4.5019
(0.407) (0.0558) (0.0292) (0.1123)
1972 1.3178 −2.7465 −0.6507 −4.3714
(0.3957) (0.0552) (0.0285) (0.1100)
1973 −1.3308 −2.7901 −0.6607 −4.3563
0.3898 (0.0551) (0.0293) (0.1091)
1974 −1.3177 −2.7789 −0.6460 −4.3132)
(0.3926) (0.0564) (0.0297) (0.1084)
1975 −1.3450 −2.8006 −0.6707 −4.4468
(0.3966) (0.0536) (0.0273) (0.1072)
1976 −1.3189 −2.7764 −0.6815 −4.4320
(0.3907) (0.0513) (0.0274) (0.1046)
1977 −1.2874 −2.7006 −0.6714 −4.3519
(0.3944) (0.0520) (0.0273) (0.1038)
1978 −1.3544 −2.7324 −0.6487 −4.3709
(0.3886) (0.0512) (0.0258) (0.1030)
1979 −1.360 −2.6812 −0.6216 −4.2718
(0.3870) (0.0501) (0.0248) (0.1008)
1980 −1.3977 −2.7144 −0.6196 −4.3144
(0.3888) (0.0504) (0.0245) (0.1020)
1981 −1.3770 −2.6390 −0.6224 −4.2467
(0.3813) (0.0485) (0.0240) (0.0997)
1982 −1.3994 −2.6239 −0.6290 −4.3072
(0.3811) (0,0483) (0.0242) (0.1009)
1983 −1.3901 −2.5431 −0.6128 −4.2416
(0.3829) (0.0476) (0.0232) (0.0999)
1984 −1.3579 −2.4728 −0.5975 −4.1811
(0.3746) (0.0481) (0.0229) (0.0998)
1985 −1.3694 −2.44936 −0.6126 −4.2162
(0.0374) (0.0476) (0.0234) (0.1000)
1986 −1.3706 −2.4764 −0.6149 −4.1951
(0.3719) (0.0471) (0.0235) (0.0994)
1987 −1.3586 −2.4831 −0.6175 −4.1993
(0.3688) (0.0467) (0.0237) (0.0988)
1988 −1.3369 −2.4189 −0.5983 −4.0608
(0.0369) (0.0466) (0.0233) (0.0977)
1989 −1.3903 −2.3767 −0.5729 −3.9322
(0.3712) (0.0462) (0.0226) (0.0960)
1990 −1.3850 −2.3744 −0.5747 −3.9244
(0.3733) (0.0460) (0.0210) (0.0954)

Note: Robust standard errors (clustering by country pairs) in parentheses. The regressions contain
the same list of covariates as the regressions presented in Table 2. Exporter and importer fixed effects,
as well as a linear time trend are included but not reported.
 Felbermayr/Kohler: Exploring the Intensive and Extensive Margins of World Trade 673

References

Anderson, J. E. (1979). A Theoretical Foundation for the Gravity Equation. Amer-

ican Economic Review 69 (1): 106–116.
Anderson, J. E., and E. van Wincoop (2003). Gravity with Gravitas: A Solution to
the Border Puzzle. American Economic Review 93 (1): 170–192.
Baier, S., and J. H. Bergstrand (2001). The Growth of World Trade: Tariffs, Trans-
port Costs, and Income Similarity. Journal of International Economics 53 (1):
1–27.
Bernard, A. B., J. B. Jensen, and P. K. Schott (2006). Trade Costs, Firms and Pro-
ductivity. Journal of Monetary Economics 53 (1): 917–937.
Buch, C., J. Kleinert, and F. Toubal (2004). The Distance Puzzle: On the Interpre-
tation of the Distance Coefficient in Gravity Equations. Economics Letters 83
(3): 293–298.
Brun, J.-F., C. Carrère, P. Guillaumont, and J. de Melo (2004). Has Distance Died?
Evidence from a Panel Gravity Model. World Bank Economic Review 19 (1):
99–120.
Coe, D. T., A. Subramanian, N. T. Tamirisa, and R. Bhavnani (2002). The Miss-
ing Globalization Puzzle. IMF Working Paper 02/171. International Monetary
Fund, Washington, D.C.
Deardorff, A. V. (1998). Determinants of Bilateral Trade: Does Gravity Work in
a Neoclassical World? In J. A. Frankel (ed.), The Regionalization of the World
Economy. Chicago: University of Chicago Press.
Deardorff A. V. (2004). Local Comparative Advantage: Trade Costs and the Pat-
tern of Trade. Research Seminar in International Economics Discussion Paper
No. 500. University of Michigan, Ann Arbor.
Disdier, A.-C., and K. Head (2004). The Puzzling Persistence of the Distance Effect
on Bilateral Trade. Mimeo. University of British Columbia.
Eaton, J., and S. Kortum (2002). Technology, Geography, and Trade. Econometrica
70 (5): 1741–1779.
Eaton, J., and A. Tamura (1994). Bilateralism and Regionalism in Japanese and
U.S. Trade and Direct Foreign Investment Patterns. Journal of the Japanese and
International Economies 8 (4): 478–510.
Egger, P., and M. Pfaffermayr (2004). Distance, Trade and FDI: A Hausman–
Taylor SUR Approach. Journal of Applied Econometrics 19 (2): 227–246.
Eichengreen, B. (1995). Trade Blocs, Currency Blocs and the Reorientation of
World Trade in the 1930s. Journal of International Economics 38 (1–2): 1–24.
Eichengreen, B., and D. A. Irwin (1997). The Role of History in Bilateral Trade
Flows. In J. A. Frankel (ed.), The Regionalization of the World Economy.
Chicago: University of Chicago Press.
Evenett, S. J., and A. J. Venables (2002). Export Growth in Developing Countries:
Market Entry and Bilateral Trade Flows. Mimeo. London School of Economics.
674 Review of World Economics 2006, Vol. 142 (4)

Feenstra, R. (2004). Advanced International Trade – Theory and Evidence. Prince-

ton: Princeton University Press.
Feenstra, R., and A. K. Rose (1997). Putting Things in Order: Patterns of Trade
Dynamics and Macroeconomics. NBER Working Paper 5975. National Bureau
of Economic Research, Cambridge, Mass.
Feenstra, R., J. A. Markusen, and A. K. Rose (2001). Using the Gravity Equation
to Differentiate among Alternative Theories of Trade. Canadian Journal of Eco-
nomics 34 (2): 430–447.
Felbermayr, G. J., and W. Kohler (2004). Exploring the Intensive and Extensive
Margins of World Trade. CESifo Working Paper No. 1276. Center for Eco-
nomic Studies, Munich.
Gleditsch, K. S. (2002). Expanded Trade and GDP Data. Journal of Conflict Reso-
lution 46 (5): 712–724.
Greene, W. (2003). Econometric Analysis. New York: Prentice Hall.
Haveman, J., and D. L. Hummels (2004). Alternative Hypotheses and the Volume
of Trade: Evidence on the Extent of Specialization. Canadian Journal of Eco-
nomics 37 (1): 199–218.
Helpman, E. (1987). Imperfect Competition and International Trade: Evidence
from Fourteen Industrial Countries. Journal of the Japanese and International
Economies 1 (1): 62–81.
Helpman, E., M. Melitz, and Y. Rubinstein (2004). Trading Patterns and Trading
Volumes. Mimeo. Harvard University.
Hummels, D., and P. J. Klenow (2005). The Variety and Quality of a Nation’s Ex-
ports. American Economic Review 95 (3): 704–723.
Markusen, J. R. (2002). Multinational Firms and the Theory of International Trade.
Cambridge, Mass.: MIT Press.
Neary, J. P. (2005). Trade Costs and Foreign Direct Investment. CER Working
Paper 05/12. Available at [Link]
[Link] .
Rose, A. (2004). Do We Really Know That the WTO Increases Trade? American
Economic Review 94 (1): 98–114.
Santos Silva, J. M. C., and S. Tenreyro (2003). Gravity-Defying Trade. Working Pa-
per 03, 1. Federal Reserve Bank of Boston, Boston, Mass.
Subramanian, A., and S.-J. Wei. (2003). The WTO Promotes Trade Strongly but
Unevenly. NBER Working Paper 10024. National Bureau of Economic Re-
search, Cambridge, Mass.
Wang, Z. K., and A. Winters (1992). The Trading Potential of Eastern Europe.
Journal of Economic Integration 7 (2): 113–136.
Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data.
Cambridge, Mass.: MIT Press.
Yi, K.-M. (2003). Can Vertical Specialization Explain the Growth of World Trade?
Journal of Political Economy 111 (1): 52–102.