4

The Gravity Model of Bilateral Trade

Our plan is to examine data on bilateral trade between pairs of countries

in order to sort out the influence of geographical proximity versus preferential trading policies in creating regional concentration in trade. The
natural framework with which to attack this question is the gravity model
of bilateral trade.
The gravity model has long been something of an ugly duckling of
international economics: obscure and allegedly lacking respectable theoretical foundations. It has recently enjoyed a swan-like revival, however.
There are at least three reasons for that revival: (1) its empirical success
at predicting bilateral trade flows, (2) its improved theoretical foundations
arising mostly from modern theories of trade in imperfect substitutes,
and (3) a new interest among economists in the subject of geography and
trade, which seeks to treat countries or regions as physically placed at
particular locations rather than as disembodied constructs.
It is not easy to decide whom to anoint as the inventor of the gravity
model. The concept is so natural that it seems always to have been
used to describe economic links between pairs of geographical units,
either with or without the word gravity. Perhaps the most classic and
extensive early application of the model to international trade was by
Linnemann (1966), who continued work first reported in Tinbergen (1962),

who in turn was contemporaneous with Poyhonen (1963). Specialists in
other fields, however, had used versions of the gravity model before
international economists did. Regional economists and urban sociologists
used it by name as far back as 1946 (Isard 1990; Stewart 1948; Zipf 1946).
No doubt there are other early references. It seems safest to cite Isaac
Newton as the original progenitor of the gravity model!
49

Institute for International Economics

http://www.iie.com

Newtons gravitational model says that the attraction between two

heavenly bodies is proportional to the product of their masses and
inversely related to the distance between them. The gravity model of
bilateral trade, in its most basic form, says that trade between country i
and country j is proportional to the product of GDP i and GDP j and
inversely related to the distance between them. Other explanatory variables that are often added are other measures of sizenamely, populations (or per capita GDPs) and land areasand dummy variables representing other measures of geographical or cultural proximitynamely,
landlockedness, common borders, common languages, and common
membership in regional trading arrangements. We elaborate on these
factors below.

The Technique of OLS Regression

One fits an equation such as the gravity model to the data by means of
ordinary least-squares (OLS) regression analysis. The wonderful thing
about OLS is that the technique holds constant for various factors in
order to ascertain the effect of another factor. Those unfamiliar with
econometrics must rest assured that the technique is not invalidated when
the explanatory variables are correlated with each other. To take an example, there is a strong correlation between proximity, as measured by the
distance between a pair of countries, and whether they share a common
boundary. Yet OLS regression can estimate the independent effect of each
factor, so long as we have correctly specified the additive form of the
equation. Trade between France and the United Kingdom will be high
due to their proximity, but trade between France and Germany will be
further boosted by the effect of their common border in addition to
their proximity.
The estimates of the effects will be subject to a margin of error, as
always. But the estimates will be the best they can be, given the data, so
long as the model is correctly specified. Moreover, the standard errors
reported for the coefficient estimates will be the correct ones. Thus we
will be able to judge whether the estimates are reliable, or whether, to
the contrary, the data set is too small to give us the information we want.
(The same point holds with respect to whether the explanatory variables
are too highly correlated to give us the information we wantwhat is
known as the problem of high multicollinearity.) Fortunately, there is
such a huge amount of information in bilateral trade data that we can
in fact obtain relatively reliable estimates of the effects of country size,
proximity, common borders, and the other variables in the gravity model.
When we have finished thinking of all the other variables that should
be expected to determine bilateral trade, we add dummy variables to
represent the bloc effects. The dummy variable is equal to 1 when both
50 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

countries in a given pair belong to the same regional group and 0 otherwise. The estimated coefficient will then tell us how much of the trade
within each region can be attributed to a special regional effect. Again,
the decision to form a free trade area (FTA) is correlated with geographical
proximity. That is what makes regional trading arrangements regional.1
Despite this correlation, the regression analysis can still separate out the
independent effects that each has on trade (so long as we have not omitted
any correlated factors from the list of explanatory variables).
Answers depend on what questions one asks. If one is aware of the
existence of the 15-country European Union and tests for it while also
testing for the existence of an operational FTA among a subset of European
countries, such as the original EEC-6, one will generally get a different
answer than if one tests for the smaller subset alone. This is as it should
be. The EEC-6 variable has a different interpretation when the EU variable
is simultaneously included than when it appears alone. If the regression
technique tells us that some arbitrary subset is an apparently significant
trade bloc, and the conclusion is erroneous in that it spuriously reflects
the effect of the more comprehensive group, then the error is ours in
omitting the proper group dummy, not the regression techniques or the
computers. It will be important to keep this in mind when we interpret
the bloc effects in the next chapter.
Where hypotheses of possible interest include groups that do not yet
formally operate, such as the Asia Pacific Economic Cooperation forum
(APEC) or the continentwide groups, then it is a matter of judgment
whether one is interested in the regression results that include these
groups. In other words, those readers who do not consider interesting
the hypothesis that there is currently an intraregional bias to trade in
APEC, the Free Trade Area of the Americas (FTAA), or East Asia are
welcome to skip the results for those cases. They can turn directly to the
results for bloc effects among the European Union, the North American
Free Trade Agreement (NAFTA), and the other formal FTAs if they wish.
One could argue that dummy variables for all possible groups should
be tested, so that the data can decide what questions are important: if the
true effect of TACBLF (the Trading Area of Countries Beginning with the
Letter F) is zero, then the regression technique can be expected to give
an estimate of zero for the coefficient on that dummy variable.2 There is
the problem, however, that even a data set of 1,953 country pairs has
only so much information to give. If one tires it out by asking a lot

1. Nonregional preferential agreements exist, of course, such as the British Commonwealth

system of preferences or the US-Israel Free Trade Agreement. Indeed, Krugman (1991b)
has given them a name, unnatural trading blocs, for reasons that are explained in chapter 8.
2. More technically, the OLS estimate is an unbiased, efficient, and consistent estimate of the
true parameter, which in this case is zero.
THE GRAVITY MODEL OF BILATERAL TRADE 51

Institute for International Economics

http://www.iie.com

of silly questions, the answers will become increasingly haphazard

including answers to the questions we really care about.3 In the extreme,
if we were to try to estimate more coefficients than there are data points,
we would not get an answer at all. The set of all possible groups among
63 countries is in fact very large (2 to the 63rd power, minus 65). Thus
we must use some discretion in choosing what groups to test for.
Less silly than fabricating nonexistent groups would be to allow each
country to have its own dummy variable or constant term. This would
reflect the possibility that some countries are more open than others to
all partners, regardless of whether they share membership in a regional
trading arrangement. We think that most of the variation in openness
may be captured by the effects of per capita incomes (richer countries are
more open), dummy variables for Hong Kong and Singapore (which

function as entrepots), and a dimension of openness that is shared by

other members of the groups we test (e.g., East Asian countries are on
average more open than would be predicted). Since we have tried including each of these factors, we omit adding separate dummy variables for
openness of each individual country in the central results reported here.4
Having generally discussed the OLS regression technique, we now turn
to examination of the gravity equation itself.

The Gravity Equation

Theoretical Foundations
Although aspersions have been cast on the respectability of the gravity
model of trade in the past, by now its theoretical pedigree has been
proven. Earlier work surveyed by Deardorff (1984, 503-06) provided a
partial foundation for the approach. Leamer and Stern (1970), for example,
3. Technically, estimating many needless parameters uses up degrees of freedom. Then
the parameter estimates, even if unbiased in small samples and accurate in large samples,
will be needlessly wide of the mark in our sample.
4. We have tried also giving each country its own intercept, and we report some of the
results below. But the results appeared unreliable when testing country effects and bloc
effects at the same time. Thus our standard equation does not allow each country to have
its own intercept.
At least one study even allows the coefficients on distance and other variables to vary
from country to country (Dhar and Panagariya 1995). If one believes that distance has a
bigger effect (e.g., shipping costs are higher per kilometer, even after holding constant for
per capita income and the other variables) for some countries trade than others, then this
is the right thing to do. If one is unpersuaded of the importance of allowing for such
variation, however, then one might suspect that the resultant loss of degrees of freedom
for any surprising results that may come out for estimates of some of the parameters about
which we care most.
52 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

noted that bilateral trade is indeterminate in the absence of transport costs

(i.e., there is nothing to determine whether Japan imports apparel from
China or from Morocco) and so assumed that countries essentially draw
their trading partners out of a hat, according to various probabilities. More
formal approaches relied on product differentiation. The assumption that
products are differentiated symmetrically by country of origin has become
associated with Armington (1969). Anderson (1979) adopted a linear
expenditure system in which the preferences for a countrys goods are
assumed to be homothetic and uniform across importing countries. Bergstrand (1985) assumed a more flexible utility function that allowed him to
find evidence that imports were closer substitutes for each other than
for domestic goods. He called his equation a generalized gravity model
because it also included price terms.
The best-known theoretical rationale for the idea that bilateral trade
depends on the product of GDPs comes from work by Helpman (1987)
and Helpman and Krugman (1985, section 1.5). Their approach is the one
we build on when we turn to a theoretical analysis of the effects of
preferential trading arrangements on the volume of trade and on economic
welfare (chapters 7 and 8). In this theory, consumers seek variety in the
products they consume, products are differentiated by firm, not just by
country, and firms are monopolistically competitive. The authors argued
that the classical Heckscher-Ohlin theory of comparative advantage does
not have the property that bilateral trade depends on the product of
incomes, as it does in the gravity model. Deardorff (1984, 500-04) concurred. Since the data do have the property that bilateral trade depends
on the product of incomes, as we shall see, this seemed to be a point in
favor of the differentiated products model of trade and a vote against the
classical models.
More recently, Deardorff (1997) has changed his mind, having discovered how to derive the gravity model from Heckscher-Ohlin theory almost
as easily as from the theory of imperfect substitutes. His main purpose
is to show that the empirical success of the gravity equation does not
necessarily support the imperfect-substitutes model relative to the Heckscher-Ohlin model. For our purposes, the main point is that it seems
possible to derive the gravity model from a variety of leading theories.
The equation has thus apparently gone from an embarrassing poverty of
theoretical foundations to an embarrassment of riches!
To most readers who have not studied enough trade theory to have
lost sight of the obvious, the assumption that trade between countries
depends positively on their size and inversely on distance may seem selfevident. Those trade theorists who previously questioned the foundations
of the gravity model did not have an alternative model of bilateral trade
to offer. It was just that economists had not tried very hard to model
bilateral trade. Deardorff (1997) concludes:
THE GRAVITY MODEL OF BILATERAL TRADE 53

Institute for International Economics

http://www.iie.com

All that the gravity equation says, after all, aside from its particular functional
form, is that bilateral trade should be positively related to the two countries
incomes and negatively related to the distance between them. Transport costs
would surely yield the latter in just about any sensible model. And the dependence
on incomes would also be hard to avoid.

While the derivation of a proportionate relationship between trade

flows and country size is an important foundation, the theories of Helpman (1987) and most of the other authors cited do not include a role for
distance and thus cannot properly be called foundations of the full gravity
model. The few exceptions include Bergstrands (1985) version of the
imperfect-substitutes theory, which incorporated a role for shipping costs,
proxied in practice by distance. Distance is also included in the second
of the two Heckscher-Ohlin-based models developed by Deardorff (1997).
The proportionality between bilateral trade and the product of incomes,
as well as the inverse dependence on distance, are also properties of our
theoretical model, introduced in chapter 7. We assume that transportation
costs raise the price of a good in the importing country and that distance
has a positive effect on transportation costs.
Once one has incorporated a role for distance in raising the cost of
trade, it is a small step to think of similar roles for dummy variables
indicating whether the pair of countries shares a common border or
common language. Each of these links helps reduce the cost of doing
business abroad, just as proximity does. Near the border, consumers can
cross over to shop in the other country and firms can source intermediate
inputs in the other country, much more readily than would be possible
if the countries did not share a common border. Linguistic links and other
historical and cultural links are particularly important at reducing what
in the preceding chapter we called the cost of unfamiliarity in international
trade, what Linnemann (1966) called psychic costs, and Garnaut (1994)
subjective resistance.
The variable that calls for more serious modeling than anyone has
yet done is per capita income. These three variablescommon border,
common language, and per capita incomeare discussed below, as we
go through the estimation results. We speak of these three variables as
constituting part of a standard full gravity model, while income and
distance constitute the basic gravity model. In recent years, there have
been many other extensions of the model, some pursued as part of our
research, some by other authors. These extensions are also discussed
below.

Our Data Sample

The dependent variable in most of our tests is total merchandise trade
(exports plus imports), in log form, between pairs of countries in a given
54 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

year. We have also tried distinguishing between imports and exports, as

do many studies; these results are described as one of the later extensions
in chapter 6.
We have run tests at five-year intervals and more frequently at the end
of the sample. Thus we have estimates for 1965, 1970, 1975, 1980, 1985,
1987, 1990, 1992, and some for 1994. There are enough data in the crosssection dimension that we can estimate each year separately. Our data
source is the UN trade matrix, supplemented in the later years by the
International Monetary Funds Direction of Trade Statistics. The UN data
cover 63 countries in our sample so that there are 1,953 data points (63
62/2) for a given year. This trade constituted 88.7 percent of world trade
reported to the United Nations in 1992. Expanding the set to a larger
number of countries would have allowed testing for more and smaller
regional arrangements, but it would have cut short the time span for
which sufficient data are available to allow useful estimation. The 63
countries are listed in table 4.1 by geographic group.
Many of the early gravity tests concentrated on data for trade among
industrialized countries. There are three possible reasons for this. First,
data were more available for these countries. Second, where the motivation was to learn about the effect of regional trading arrangements, it
was usually the European Community and the European Free Trade
Association (EFTA) that investigators had in mind. Third, when the modern theories of trade in imperfect substitutes were introduced as the
justification for the model, they were thought to apply only to the industrialized countries, not to the developing countries. Theories based on abundant endowments of unskilled labor or natural resources seemed more
applicable to the developing countries.
Each of these reasons for limiting the analysis to industrialized countries
is no longer convincing, even if they once were. First, enough data exist
to include many developing countries. Second, regional arrangements in
Latin America and elsewhere in the developing world certainly deserve
serious attention now. Third, even if the goal is only to assess FTAs in
Europe, one should use as broad a set of data as possible to estimate
normal patterns of trade so that one can ascertain how European links
differ from what is normal. Fourth, some of the countries that have previously been omitted from the list of industrialized countries (Singapore
and Hong Kong) are now richer than most of the ones that are included.
Fifth, we believe that the gravity model is in fact applicable to developing
countries, not just to industrialized countries.
Hummels and Levinsohn (1995) extended the test of Helpman (1987)
to a data set including developing countries, taking it for granted that
the monopolistic competition model would only do a good job of explaining trade among OECD countries. They were surprised to find that it
works equally well for the larger set of countries. Product varieties from
THE GRAVITY MODEL OF BILATERAL TRADE 55

Institute for International Economics

http://www.iie.com

56

Table 4.1

Countries included in the gravity equation by regional group and main city a

Institute for International Economics

Americas
(WH.13)
Canada, Ottawa
United States, Chicago
Argentina, Buenos Aires

Brazil, Sao Paulo

Chile, Santiago
Colombia, Bogota
Ecuador, Quito
Mexico, Mexico City
Peru, Lima
Venezuela, Caracas
Bolivia, La Paz
Paraguay, Asuncion
Uruguay, Montevideo

European Community
(EC.11)
West Germany, Bonn
France, Paris
Italy, Rome
United Kingdom, London
Belgium, Brussels
Denmark, Copenhagen
Netherlands, Amsterdam
Greece, Athens
Ireland, Dublin
Portugal, Lisbon
Spain, Madrid

European Free Trade Area

(EFTA.6)
Austria, Vienna
Finland, Helsinki
Norway, Oslo
Sweden, Stockholm
Switzerland, Geneva
Iceland, Reykjavik

Other countries
(23)

Japan, Tokyo
Indonesia, Jakarta
Taiwan, Taipei
Hong Kong, Hong Kong
South Korea, Seoul
Malaysia, Kuala Lumpur
Philippines, Manila
Singapore, Singapore
Thailand, Bangkok
China, Shanghai

South Africa, Pretoria

Turkey, Ankara
Yugoslavia, Belgrade
Israel, Jerusalem
Algeria, Algiers
Libya, Tripoli
Nigeria, Lagos
Egypt, Cairo
Morocco, Casablanca
Tunisia, Tunis
Sudan, Khartoum
Ghana, Accra
Kenya, Nairobi
Ethiopia, Addis Adaba
Iran, Tehran
Kuwait, Kuwait
Saudi Arabia, Riyadh
India, New Delhi
Pakistan, Karachi
Hungary, Budapest
Poland, Warsaw
Australia, Sydney
New Zealand, Wellington

East Asia
(EAEG.10)b

http://www.iie.com
a. The distance between countries was computed as the great circle distance between the relevant pair of cities.

b. APEC consists of East Asia, plus Australia, New Zealand, Canada, and the United States. Latecomers Mexico and Chile are not counted in APEC.

developing countries are often quite imperfect substitutes, for each other
as well as for goods from rich countries. This applies both to agricultural
products (say, Chilean wine versus Romanian wine versus French wine)
or manufactures (Chinese jeans versus Mexican jeans versus Italian jeans).
Furthermore, as we have already noted, Deardorff has derived the
gravity model for a version of Heckscher-Ohlin trade, which has always
been considered well-suited to North-South trade. This is not to deny
that trade patterns for countries at earlier stages of development are
different from those of countries at later stages. We shall be taking into
account countries incomes per capita. The point is that developing countries should be included in the data set.

Size, Income Per Capita, and Competing Models of Trade

There are two standard ways of measuring the size of countries in the
gravity model: GNP (output) or population.5 When holding constant for
GNP, the coefficient on population is generally negative. This captures
the well-known phenomenon that larger countries tend to be relatively
less open to trade as a percentage of GNP. A Singapore or a Luxembourg
is highly dependent on trade, in part because it lacks many natural endowments and because it lacks room to exploit economies of scale in the
domestic market. A Japan or United States, while engaging in far more
trade in absolute terms, will engage in less trade as a percentage of GNP
because it can find more within its own borders. There is an additional
reason for this pattern. Interstate trade in the United States is considered
domestic; interstate trade within the European Union is considered international trade. The accidents of political history alone will give the result
that the ratio of trade to income falls with the size of the unit.
In some studies these size variables are supplemented by a measure of
land area. This is generally a way of getting at natural resources. A country
with a large land area, holding constant for the other measures of size,
is relatively more self-sufficient and less dependent on trade. In one
gravity equation, we find that, for every 1 percent increase in land area,
trade falls by about 0.2 percent.6 But we shall focus here on population
as the measure of size and self-sufficiency.
Mathematically, it is precisely equivalent, whether we express the
explanatory variables as GNP and per capita GNP, or as GNP and population. To see this, consider the basic gravity equation, with some typical
parameter values substituted in for concreteness:
5. Our standard tests use GNP rather than GDP (gross domestic product) because of greater
data availability. We have also tried the latter, however. The difference does not seem
important in this context (Linnemann 1966).
6. The estimate is highly significant statistically (Frankel and Romer 1996, table 1).
THE GRAVITY MODEL OF BILATERAL TRADE 57

Institute for International Economics

http://www.iie.com

logTij .7 log(GNPi GNPj ) .3 log[(GNP/popi )(GNP/popj )]

.7 log(Dist. ij ),

(4.1)

where Tij is trade between country i and country j. This equation is precisely the same as an equation that expresses the income and population
terms as explanatory variables:
logTij 1.0 log(GNPi GNPj ) .3 log(popi popj ) .7 log(Dist. ij )

(4.2)

In equation (4.1), the tendency of trade to rise less than proportionately

with size is reflected in a GNP coefficient less than one; in equation (4.2),
it is reflected in a population coefficient less than zero. We will choose
to use the first formulation. Although the estimation is precisely equivalent either way, mathematically, the reader of the results may be led to
different interpretations.
Observing equation (4.1), one is usefully led to think about how a
countrys trade depends on its stage of development. Why should trade
depend on development? We do not get much guidance from the standard
gravity model foundationtheories of trade based on imperfect substitutes. In models of the Krugman-Helpman type, there is no role for GNP
per capita. The models are simply expressed in terms of economic size,
without distinguishing between the roles of output and population. It is
clear empirically that the two effects are in fact independent. So this is
an area that bears further research.
One possible explanation for the independent effect of income per capita
is that exotic foreign varieties are superior goods in consumption. Lowincome countries are dominated by subsistence farming. Other possibilities come out of the literature on endogenous growth.7 For example, the
process of development may be led by the innovation or invention of
new products that are then demanded as exports by other countries. It
has also been suggested that the more developed countries have more
advanced transportation infrastructures, including seaports and airports,
which facilitate trade.
Perhaps the most important reason industrialized countries trade more
than less developed countries is that countries tend to liberalize as they
develop. One reason for this pattern is that governments of very poor
countries depend on tariff revenue for a large share of their budget, while
more advanced countries can apply other forms of direct and indirect
taxes to the domestic economy. If we had good direct measures of trade
policy that we could use in the equation, such as a countrys overall level
of tariffs, then we would not expect any of this effect to show up in the
7. Of the many relevant works, some of the more important are Grossman and Helpman
(1989, 1991a) and Rivera-Batiz and Romer (1991). For further references on the connections
between trade and growth, see Frankel, Romer, and Cyrus (1995).
58 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

term for per capita income. But given the absence of good measures of
this kind and the importance of nontariff barriers, the most important
reason to expect low per capita GNP to inhibit trade may be the correlation
with protectionist trade policies.
To take an example, China and Japan have roughly the same aggregate
outputs. Yet China trades less with its partners than does Japan. This is
what the equation would lead us to expect, because Chinas large output
derives primarily from its large population, while Japans derives from
its high level of GNP per capita.
Perhaps the easiest way of keeping in mind all the factors that may be
involved in this relationship is to think in terms of a third way of expressing the same equation:
logTij 1.0 log[(GNP/popi )(GNP/popj )] .7 log(pop)i (pop) j
.7 log(Dist.ij )

(4.3)

Now, to see the effect of growth on trade, we first ask what is the source
of growth. If an increase in GNP (relative to other countries) takes the
form of an increase in GNP per capita, then the effect on trade is proportionate, as called for in the trade theories based on imperfect substitutes.
If an increase in GNP instead is entirely accounted for by an increase in
population, then the effect on trade is somewhat less than proportionate
(0.7 instead of 1.0), as economies of scale make the country proportionately
less dependent on trade. The rapid growth in incomes of most East Asian
countries in recent years has been based primarily on productivity growth
rather than population growthcall it 8 percent a year. Then the equation
predicts that their trade with slow-growing countries will increase by 8
percent a year and their trade with each other will increase by 16 percent
a year (8 8).8
It is also instructive to focus explicitly on the product of per capita
GNPs as a determinant of trade, as in equation (4.1). The prediction that
the product of per capita incomes enters the equation positively contradicts
the prediction of traditional Heckscher-Ohlin theories of trade. If the two
factors of production are capital and labor, then these theories predict
that countries with dissimilar levels of output per capita will trade more
than countries with similar levels. (Or, more precisely, dissimilar capital/
labor ratios.) But the standard gravity model predicts that countries with
similar levels of output per capita will trade more than countries with dissimilar
8. Assume that the country in question is small and gross world product can be held
constant. Then we can jump to the time-series context, even though these estimates are
from a pure cross-section. Otherwise, we need to express growth rates as relative to the
worldwide average: If East Asian countries grow 6 percent per year faster than the world
average, then their trade with other countries also grows 6 percent faster, and their trade
with each other grows 12 percent faster.
THE GRAVITY MODEL OF BILATERAL TRADE 59

Institute for International Economics

http://www.iie.com

levels.9 Seldom do competing theories have such directly contradictory

empirical implications.
Before there was Krugman (and others), there was Linder. The Linder
(1961) hypothesis says that countries with similar levels of per capita
income will have similar preferences and similar but differentiated products, and thus will trade more with each other. It is often viewed as
similar to the Krugman-Helpman hypothesis in its empirical predictions,
if somewhat different in derivation. There is in fact a crucial difference
in the empirical implications, however. The Krugman-Helpman theory
predicts that the sum of the logs of (GNP/popi ) and (GNP/popj ) will have
a positive effect on the log of trade. The Linder hypothesis is usually
described as predicting that the absolute value of the difference of the two
variables will have a negative effect. (There is no theory to predict whether
they enter in log form or simply as ratios.) Heckscher-Ohlin is then seen
as the third case: The absolute value of the difference of the two variables
will have a positive effect. The diametric opposite of Heckscher-Ohlin is
thus the Linder theory, not the Helpman-Krugman theory. The Linder
hypothesis is properly viewed as a distinct third school of thought, not
just a forerunner of Helpman-Krugman.
Deardorff (1997, 15) has argued that it can be viewed as having a certain
kinship to Heckscher-Ohlin. He opinesas Markusen (1986) has already
shownthat if high-income consumers tend to consume larger budget
shares of capital-intensive goods, which Heckscher-Ohlin tells us are produced by capital-rich countries, then it follows that (1) capital-rich countries will trade more with other capital-rich countries than with capitalpoor countries, and (2) capital-poor countries likewise will trade more
with their own kind. These predictions are the same as those of the Linder
hypothesis, derived in a different way. The first of them is borne out by
the common finding in gravity equations that the product of per capita
GNPs has a positive effect on trade. However, the second prediction is
contradicted by the common gravity equation finding, which says instead
that a poor country will trade more with a rich country than with another
poor country. In other words, trade stems from economic development,
not from similarity of the stage of development.
To distinguish among these influencesHeckscher-Ohlin-style factorendowment differences, Linder-style taste differences, and the effect of

9. The Helpman-Krugman sort of theory and its child, the standard gravity equation, also
predict that if the distribution of national incomes across countries becomes more equal
over time, the volume of trade should increase. The United States declined from almost
half of gross world product among market economies in the 1950s to 32 percent in 1970
and then to 26 percent in 1993. Thus a more uniform size distribution among economies is
indeed one explanation for the increase in global trade (Helpman 1987; Krugman 1995, 341).
As noted, Hummels and Levinsohn (1995) find that the same pattern that Helpman found
for OECD countries holds for less developed countries.
60 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

development on tradeone must try to capture the distinctive features

of each. One of the experiments we try below, within the standard gravity
model formulation of bilateral trade, is adding a term for the difference
in per capita GNPs. A negative sign on this term would support the
Linder hypothesis, while a positive sign would support the HeckscherOhlin hypothesis.
One must note that our specification of these other terms remains more
ad hoc than the basic gravity specification that we attribute to HelpmanKrugman, Bergstrand (1985), and our own chapters 7 and 8. This is in
part for the reason that, despite the progress made by Deardorff (1997)
and the earlier authors in deriving theories of bilateral trade, geography
has seldom been introduced into the classical models.
Gruber and Vernon (1970) and Thursby and Thursby (1987) have already
added absolute differences in per capita incomes to the gravity equation
in an attempt to capture the Linder effect. Bergstrand (1989) generalized
his gravity model foundations further to include both a role for factorendowment differences in the spirit of Heckscher-Ohlin and a role for taste
variables in the spirit of Linder. The resulting equation, however, uses only
per capita income variables. Specific commodities are revealed to be capitalintensive in production if the per capita income of the exporting country
is estimated to have a positive coefficient and revealed to be labor-intensive
if it has a negative coefficient.10 Commodities are revealed to be luxuries
in consumption if per capita income of the importing country is estimated
to have a positive coefficient and necessities if it has a negative coefficient.
One can also try to measure factor endowments directly to get at the
Heckscher-Ohlin model more appropriately. Leamer (1974) added factorendowment variables to a gravity-type equation and found that they
performed less well than the standard income and population variables.
We have also tried adding terms for differences in factor endowments as
a more direct test of the Heckscher-Ohlin hypothesis. The results are
described in chapter 6.

Our Results for the Basic Gravity Variables

The equation to be estimated, in its basic form, is:
log(Tij ) 1log(GNPi GNPj ) 2 log(GNP/popi GNP/popj )
3 log(Distij ) 4 (ADJij) 5 (Langij)
1 (WEij ) 2 (WHij ) 3 (EAij ) uij .

(4.4)

10. More precisely, the technical difficulties in generalizing two-commodity two-factor

Heckscher-Ohlin theory to a world with more than two commodities require that we say
that specific commodities tend in an average sense to be capital-intensive or labor-intensive
(Bergstrand 1989; Deardorff 1984).
THE GRAVITY MODEL OF BILATERAL TRADE 61

Institute for International Economics

http://www.iie.com

62

Table 4.2

Gravity model estimations of explicit regional trading arrangements,a 196592

(with trade between country pairs as dependent variable)

Institute for International Economics

1965

1970

1975

7.910**
(0.532)

9.632**
(0.619)

9.157**
(0.591)

10.763**
(0.664)

GNP

0.637**
(0.018)

0.685**
(0.019)

0.646**
(0.019)

0.702**
(0.021)

Per capita GNP

0.235**
(0.026)

0.284**
(0.028)

0.337**
(0.026)

0.483**
(0.044)

0.447**
(0.052)

0.562**
(0.042)

Adjacency

0.433**
(0.161)

0.482**
(0.162)

Language

0.550**
(0.095)

EU15 bloc
NAFTA bloc

1985

1990

1992

12.006**
(0.530)

13.564**
(0.635)

10.956**
(0.492)

12.146**
(0.576)

9.599**
(0.464)

10.523**
(0.509)

12.146**
(0.469)

13.521**
(0.530)

0.744*
(0.018)

0.786**
(0.019)

0.775**
(0.016)

0.804**
(0.017)

0.797**
(0.016)

0.834**
(0.017)

0.796**
(0.016)

0.832**
(0.016)

0.930**
(0.018)

0.963**
(0.018)

0.403**
(0.028)

0.255**
(0.023)

0.294**
(0.025)

0.283**
(0.022)

0.323**
(0.025)

0.247**
(0.022)

0.264**
(0.024)

0.080**
(0.017)

0.128**
(0.018)

0.128**
(0.019)

0.153**
(0.020)

0.594**
(0.049)

0.698**
(0.042)

0.683**
(0.048)

0.588**
(0.039)

0.555**
(0.048)

0.715**
(0.039)

0.707**
(0.047)

0.572**
(0.037)

0.656**
(0.043)

0.770**
(0.038)

0.733**
(0.044)

0.458**
(0.165)

0.394*
(0.170)

0.398*
(0.160)

0.400*
(0.166)

0.571**
(0.174)

0.602**
(0.182)

0.658**
(0.165)

0.626**
(0.171)

0.751**
(0.189)

0.609**
(0.189)

0.445**
(0.157)

0.506**
(0.170)

0.586**
(0.096)

0.348**
(0.094)

0.410**
(0.096)

0.446**
(0.099)

0.675**
(0.093)

0.754**
(0.098)

0.474**
(0.093)

0.571**
(0.097)

0.572**
(0.090)

0.635**
(0.088)

0.768**
(0.090)

0.823**
(0.090)

0.218#
(0.116)

0.143
(0.114)

0.061
(0.110)

0.078
(0.107)

0.140
(0.104)

0.229*
(0.104)

0.021
(0.103)

0.076
(0.103)

0.227*
(0.099)

0.134
(0.100)

0.267**
(0.102)

0.158
(0.103)

0.083
(0.097)

0.135
(0.099)

0.020
(0.311)

0.178
(0.263)

0.227
(0.333)

0.050
(0.275)

0.313
(0.298)

0.028
(0.269)

0.098
(0.274)

0.379
(0.290)

0.264
(0.268)

0.185
(0.289)

0.152
(0.292)

0.367
(0.339)

0.226
(0.294)

0.201
(0.333)

Mercosur bloc

0.343
(0.444)

0.051
(0.444)

0.311
(0.331)

0.451
(0.358)

0.277
(0.326)

0.427
(0.351)

0.561*
(0.236)

0.746**
(0.253)

0.808*
(0.356)

0.686#
(0.379)

1.918**
(0.235)

1.324**
(0.264)

0.690*
(0.340)

0.934*
(0.364)

Andean bloc

1.310**
(0.446)

1.198**
(0.467)

0.307
(0.253)

0.283
(0.275)

0.311
(0.321)

0.351
(0.342)

0.082
(0.248)

0.103
(0.263)

0.103
(0.466)

0.046
(0.479)

0.204
(0.481)

0.965**
(0.238)

1.187**
(0.256)

ASEAN bloc

1.621**
(0.487)

1.274**
(0.503)

1.824**
(0.315)

1.512**
(0.324)

2.272**
(0.393)

1.925**
(0.403)

1.704**
(0.370)

1.487**
(0.378)

1.757**
(0.335)

1.196**
(0.316)

1.766**
(0.281)

1.126**
(0.286)

1.263**
(0.106)

1.448**
(0.159)

1.399**
(0.106)

1.380**
(0.158)

1.732**
(0.097)

1.768**
(0.151)

1.716**
(0.095)

1.688**
(0.140)

Distance

9.326**
(0.544)

1980

10.820**
(0.619)

Intercept

0.368**
(0.094)

|
http://www.iie.com

2.045**
(0.379)

1.570**
(0.393)

AUS-NZ bloc
EU15 openness

0.120
(0.074)

0.159*
(0.072)

0.096
(0.074)

0.036
(0.070)

0.077
(0.067)

0.104
(0.467)

0.186**
(0.063)

0.132*
(0.064)

Institute for International Economics

NAFTA openness

0.600**
(0.114)

0.664**
(0.112)

0.630**
(0.126)

0.491**
(0.115)

Mercosur openness

0.289**
(0.096)

0.091
(0.093)

0.136
(0.097)

0.132
(0.092)

Andean openness

0.147
(0.110)

0.030
(0.092)

0.032
(0.101)

0.058
(0.101)

ASEAN openness

0.451
(0.112)

0.620**
(0.110)

0.392**
(0.103)

0.434**
(0.094)

0.751**
(0.096)

0.252**
(0.098)

0.818**
(0.101)

0.295**
(0.098)

0.087
(0.101)

0.469**
(0.106)

0.106
(0.103)

0.312**
(0.105)

0.331*
(0.145)

AUS-NZ openness

0.616**
(0.112)

0.190#
(0.108)

0.640**
(0.085)

0.072
(0.132)

0.610**
(0.094)

0.272*
(0.136)

0.154
(0.128)

Number of
observations

1,194

1,194

1,274

1,274

1,453

1,453

1,708

1,708

1,647

1,647

1,573

1,573

1,546

1,546

Adjusted R2

0.660

0.674

0.684

0.701

0.703

0.713

0.694

0.703

0.721

0.730

0.750

0.776

0.798

0.816

Standard error
of regression

1.096

1.072

1.126

1.094

1.200

1.180

1.242

1.223

1.204

1.185

1.115

1.057

1.135

1.085

**, *, # denotes significant at the 99%, 95%, and 90% levels, respectively.
a. All variables except dummy variables are in logs.

|
http://www.iie.com
63

The last five explanatory factors are dummy variables. ADJij , short for
Adjacency, is equal to 1 when countries i and j share a common border
and 0 when they do not. Langij is equal to 1 when the countries share a
common language or past colonial links and 0 otherwise. WE, WH, and
EA are three examples of the dummy variables we use when testing the
effects of membership in a common regional groups, standing in this case
for Western Europe, Western Hemisphere, and East Asia.
Table 4.2 reports the results of OLS estimation of the equation, where
the trading blocs tested for are six groups that currently have formal
status as regional trading arrangements: the European Union, NAFTA,
Mercosur, the Andean Community, ASEAN, and the Australia-New
Zealand Closer Economic Relations (CER) pact. Table 4.3 reports results
when, as in equation (4.4), we test for broader, less formal, blocs that
currently exist only as proposals or hypotheses: Western Europe, the
Western Hemisphere, East Asia, APEC, and TAFTA. In both cases, the
tests are run separately at five-year intervals from 1965 to 1990 and are
also run on more recent data available for 1992. In these tables, for comparability across time, we elect to test the effects of country groups defined
to have the same membership in every year, notwithstanding that in the
1970s, for example, the European Economic Community contained nine
countries, rather than the 15 now in the European Union, or that NAFTA
did not exist at all. If we find no significant effects in the early years, we
will know the likely reason why.
We find all five standard gravity variables to be highly significant
statistically (i.e., significant at greater than the 99 percent confidence level).
We discuss them, before turning to the bloc effects in the next chapter.

Coefficients on Size and Per Capita Income

The log of the product of the two countries GNPs is always highly
significant statistically. It is also generally significantly less than 1. Thus
the results indicate that, though trade increases with a countrys size, it
increases less than proportionately (holding GNP per capita constant).
This confirms, as expected, the familiar pattern that small economies tend
to be more open to international trade than larger, more diversified,
economies. In tables 4.2 and 4.3, the coefficient shows an upward trend,
from about .6 in 1965 to more than .9 in 1992. (Earlier research usually
did not show a trend like this in the income coefficient.)
The estimated coefficient on GNP per capita is also highly significant
statistically, indicating that richer countries do indeed trade more than
poor ones. It shows a moderate downward trend, declining from about
.3 early in the sample period to about .1 toward the end of the sample
64 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

period.11 In most cases, a test would fail to reject the constraint that the
sum of the coefficients on GNP and per capita GNP is 1.0. Holding constant
for population, trade between two countries is simply proportionate to
the product of their GNPs.
The reported results measure GNPs and per capita GNPs at current
exchange rates. An alternative is to measure them at purchasing power
parity (PPP) rates (Boisso and Ferrantino 1993, 1996).12 In theory, the PPP
rates are probably to be preferred. Otherwise, large temporary swings in
the nominal exchange rate can create large swings in the real exchange
rate and distort the comparison of incomes. (It is possible that our gravity
estimates for 1985, the year when the dollar had reached its peak real
appreciation, are distorted.) The disadvantage of using the PPP rates is
that they are subject to large measurement errors, as Srinivasan (1995)
has pointed out. Most of our reported results thus measure incomes at
current exchange rates. We have, however, also tried the Summers-Heston
PPP-rate measure. One effect was to restore to a statistically significant
positive value the coefficient on per capita incomes in 1991, when the
conventional exchange rate measure showed a coefficient of the wrong
sign. 13 The results for the other variables were little affected. When
exchange-rate-based incomes and PPP-based incomes are entered into the
equation at the same time, the data seem to prefer the former, though
the multicollinearity is too high for a clear and consistent verdict.

Distance: Measurement and Connection to the Question of

Subnational Provinces
The calculation of the distance variable requires some elaboration. The
proximity measure that we use in most of our tests is the log of distance
between the two major cities of the respective countries. The cities are
usually the capitals of the two countries, but in a few cases we substitute
for the capital a major city that seems closer to the countrys economic
center of gravity (Chicago for the United States rather than Washington,
DC, and Shanghai for China rather than Beijing). Table 4.1 lists the cities
that we use for each of the 63 countries.14
11. Linnemann (1966) obtains similar estimates for this parameter (in the range of .21 to
.27) for the year 1959, as do others. Eichengreen and Irwin (1995), however, obtain much
higher estimates for the interwar period: .59 to .85.
12. Linnemann (1966) tried both and found that it made little difference for the results,
except for the coefficient on income itself.
13. The data are not as complete as for earlier years. Appendix table B5.4a shows the results
for preliminary 1991 data, with the standard exchange-rate-evaluated GNPs, and appendix
table B5.4b shows the same regression with the PPP-evaluated numbers.
14. Boisso and Ferrantino (1996) find very little difference in gravity equation results whether
distance is measured between the most populous cities or the geographical centers.
THE GRAVITY MODEL OF BILATERAL TRADE 65

Institute for International Economics

http://www.iie.com

66

Institute for International Economics

Table 4.3

Gravity model estimations of prospective trade blocs,a 196592

(with trade between country pairs as dependent variable)
1965

1970

1975

9.135**
(0.524)

11.671**
(0.771)

10.040**
(0.544)

11.473**
(0.777)

GNP

0.643**
(0.018)

0.671**
(0.018)

0.647**
(0.018)

0.661**
(0.019)

Per capita GNP

0.262**
(0.028)

0.357**
(0.035)

0.369**
(0.027)

0.401**
(0.049)

0.277**
(0.067)

Adjacency

0.568**
(0.160)

Language
Western Europe
bloc

Intercept

Distance

http://www.iie.com

Western Hemisphere
bloc

1985

1990

1992

10.538**
(0.840)

12.261**
(0.527)

12.377**
(0.753)

10.780**
(0.496)

11.167**
(0.758)

9.558**
(0.461)

9.841**
(0.652)

11.776**
(0.472)

12.043**
(0.643)

0.730*
(0.018)

0.733**
(0.019)

0.752**
(0.016)

0.752**
(0.017)

0.771**
(0.016)

0.774**
(0.017)

0.767**
(0.016)

0.770**
(0.016)

0.901**
(0.018)

0.883**
(0.018)

0.425**
(0.034)

0.284**
(0.024)

0.324**
(0.031)

0.300**
(0.022)

0.306**
(0.028)

0.247**
(0.023)

0.248**
(0.028)

0.087**
(0.018)

0.134**
(0.022)

0.139**
(0.020)

0.163**
(0.022)

0.519**
(0.045)

0.456**
(0.063)

0.695**
(0.048)

0.649**
(0.070)

0.541**
(0.041)

0.567**
(0.059)

0.677**
(0.042)

0.664**
(0.062)

0.525**
(0.040)

0.597**
(0.054)

0.762**
(0.041)

0.732**
(0.050)

0.660**
(0.163)

0.621**
(0.156)

0.640**
(0.161)

0.497**
(0.154)

0.525**
(0.159)

0.668**
(0.158)

0.634**
(0.162)

0.742**
(0.148)

0.740**
(0.154)

0.781**
(0.164)

0.662**
(0.166)

0.561**
(0.144)

0.574**
(0.143)

0.505**
(0.093)

0.561**
(0.092)

0.333**
(0.091)

0.393**
(0.092)

0.326**
(0.094)

0.376**
(0.097)

0.520**
(0.090)

0.611**
(0.091)

0.338**
(0.092)

0.438**
(0.092)

0.359**
(0.084)

0.450**
(0.084)

0.620**
(0.088)

0.654**
(0.085)

0.336**
(0.118)

0.095
(0.128)

0.196#
(0.115)

0.073
(0.125)

0.006
(0.118)

0.132
(0.127)

0.251*
(0.110)

0.151
(0.117)

0.344**
(0.110)

0.270*
(0.118)

0.411**
(0.107)

0.234*
(0.117)

0.095
(0.105)

0.004
(0.160)

0.199#
(0.116)

0.048
(0.134)

0.067
(0.136)

0.357*
(0.156)

0.338*
(0.133)

0.687**
(0.151)

0.241#
(0.141)

0.522**
(0.150)

0.821**
(0.157)

0.943**
(0.173)

0.362**
(0.116)

0.824**
(0.132)

0.321*
(0.138)

9.521**
(0.562)

1980

0.075
(0.110)

East Asia bloc

1.626**
(0.264)

1.791**
(0.273)

1.822**
(0.260)

1.838**
(0.273)

0.949**
(0.229)

0.900**
(0.238)

0.940**
(0.183)

0.784**
(0.189)

0.666**
(0.181)

0.615**
(0.193)

0.762**
(0.173)

0.716**
(0.178)

0.458*
(0.199)

0.266
(0.192)

APEC bloc

0.364#
(0.187)

0.088
(0.206)

0.626**
(0.160)

0.510**
(0.186)

0.828**
(0.167)

0.688**
(0.188)

1.273**
(0.118)

1.152**
(0.134)

1.113**
(0.116)

1.045**
(0.137)

1.183**
(0.108)

0.833**
(0.130)

1.159**
(0.114)

0.882**
(0.124)

0.113
(0.090)

0.133
(0.096)

TAFTA bloc
Western Europe
openness

0.219#
(0.124)

0.236*
(0.094)

0.090
(0.102)
0.006
(0.117)

0.326**
(0.092)

0.165
(0.103)
0.002
(0.130)

0.238**
(0.090)

0.139
(0.100)
0.314**
(0.115)

0.105
(0.096)

0.011
(0.105)
0.253*
(0.119)

0.096
(0.085)

0.050
(0.093)
0.051
(0.114)

0.320**
(0.086)

0.008
(0.084)
0.030
(0.117)

0.465**
(0.101)

Western Hemisphere
openness
East Asia openness

0.223*
(0.100)

Institute for International Economics

0.626**
(0.146)

0.243*
(0.101)

0.430**
(0.133)

0.061
(0.087)

0.465**
(0.145)

0.009
(0.091)

0.672**
(0.114)

0.150#
(0.084)

0.598**
(0.111)

0.375**
(0.091)

0.936**
(0.110)

APEC openness

0.498**
(0.142)

0.300*
(0.130)

0.221
(0.151)

0.179
(0.125)

0.312*
(0.132)

0.546**
(0.126)

TAFTA openness

0.049
(0.090)

0.105
(0.084)

0.036
(0.095)

0.070
(0.083)

0.036
(0.085)

0.054
(0.080)

0.823**
(0.104)
0.275*
(0.113)
0.059
(0.079)

Number of
observations

1,194

1,194

1,274

1,274

1,453

1,453

1,708

1,708

1,647

1,647

1,573

1,573

1,546

1,546

Adjusted R2

0.686

0.698

0.713

0.718

0.717

0.723

0.719

0.730

0.738

0.745

0.773

0.785

0.812

0.829

Standard error of
regression

1.053

1.033

1.073

1.064

1.172

1.160

1.190

1.165

1.167

1.153

1.064

1.035

1.097

1.046

**, *, denotes significant at the 99%, 95%, and 90% levels, respectively.
a. All variables except dummy variables are in logs.
b. TAFTA variables are defined as EU countries plus NAFTA countries.

|
http://www.iie.com
67

Let us note in passing that, in an ideal world, we would have data on

bilateral trade among provinces, or even among smaller geographical
units. Such a data set would have at least three major advantages. The
first is that we could then be more precise about the distances rather than
being forced to assume, in effect, that the entire economic activity of a
large country is concentrated at a single point of mass. This first advantage
turns out apparently to be the least important of the three empirically.
Second, we would have an even larger number of observations with which
to work. The more information in the data set, the more reliably we can
answer the questions in which we are interested.
The third and most important advantage of having data at the provincial
level is that we could ascertain how trade between two geographical
entities is affected by their common membership in a political union. One
might infer the intranational bias to trade in other waysfor example,
by pondering ratios of trade to output that are low in most countries
when judged by the standard of the supposedly borderless world. (This
home-country bias to trade is further analyzed in chapter 6 below.)
When one sees that such links as sharing a common language or common membership in an FTA have big effects on trade between geographical units, extrapolation then suggests that common membership in a
political union should have effects even larger than those. It would be
very instructive to estimate these effects econometrically. It would help
us predict declines in the volume of trade among the former constituents
of the Soviet Union, or between the Czech and Slovak Republics. It would
also help us predict increases in the volume of trade between the old East
and West Germany, or among the members of the European Union, if
and when hopes for full political union are realized in the 21st century.
Unfortunately, data on trade among such subnational units as provinces
or states are seldom available. A rare exception, data on trade undertaken
by Canadian provinces, is discussed in chapter 6.15 The data analyzed in
most of this book, however, are all at the national level.
One can measure the distance between two points on the globe in a
number of ways. Most of our econometric results use distance measured
as the crow flieswhat is technically called the great-circle distance
between the two latitude-longitude combinations.16 Attempts to distinguish between land and sea distances and to measure distances along the
15. One can obtain reasonably recent data on the international trade of the 50 individual
American states but not their trade with each other, which would be crucial. Richardson
and Smith (1995) analyze 1987-89 exports. The only known source of data that includes
interstate trade pertains to 1963 (even in this case, states international trade is not broken
down by country). Greytak, Richardson, and Smith (1995) plan to analyze the 1963 data,
which might shed light on the effect of political federation on trade between states.
16. Similar straight-line measures of distance are used in gravity equations by Linnemann
(1966), Eichengreen and Irwin (1995), and many others.
68 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

Table 4.4 US trade by mode of transportation, 1970-93

Shipping weight
(billions of
kilograms)

Fraction of value
going by air
(percent)

All methodsa

Vessel

Air

Vessel

Air

US imports
1970
1980
1985
1990
1991
1992
1993
1994
1995

8.5
11.6
14.9
18.4
19.4
19.8
20.6
21.6
23.4

40.0
240.8
345.3
495.3
487.1
532.4
580.5
663.4
743.5

24.8
165.1
208.4
283.4
272.6
293.1
310.4
339.4
356.6

3.4
28.0
51.3
90.9
94.3
105.3
119.7
143.0
174.2

271.4
443.1
361.5
496.3
449.0
476.0
531.0
586.9
562.5

0.3
0.6
1.3
1.7
1.6
1.7
1.9
2.2
2.3

US exports
1970
1980
1985
1990
1991
1992
1993
1995
1995

14.1
20.9
24.5
28.1
27.3
27.0
29.1
29.3
31.1

43.2
220.7
213.1
393.0
421.7
447.5
464.9
512.4
583.0

24.6
120.9
91.7
150.8
162.4
169.8
166.6
177.6
215.9

6.1
46.1
52.3
110.5
115.1
121.0
135.1
150.3
181.1

218.0
363.7
317.7
372.4
385.5
379.3
349.5
334.5
401.1

0.4
1.0
0.8
1.5
1.6
1.7
1.7
2.0
2.3

Value (billions of dollars)

a. Includes types other than vessel and air (i.e., land) and revisions that are not distributed
by method of transport.

Source: US Bureau of the Census, Statistical Abstract of the United States (CD-ROM), Highlights of U.S. Export and Import Trade, through 1988; thereafter, U.S. Merchandise Trade,
Selected Highlights.

shipping routes actually followed, described below, do not turn out to

shed a great deal of additional light.
This seems to be a good place in the book to observe that much trade
goes by neither land nor sea these days, but by air: 23 percent of US
imports in 1995, up from 8 percent in 1970, and 31 percent of US exports,
up from 14 percent (table 4.4). This represents a large increase in the use
of air transport at the expense of sea transport. (The estimated share by
land is up a little in the case of exports and down in the case of imports.)
As recently as 1970, the ratio of value of US imports shipped by vessel
to shipments by air was 7.3; by 1995 it had fallen to 2.0. For US exports,
the ratio of value shipped by vessel to shipments by air has fallen from
4.0 to 1.2. Extrapolating logarithmically, more American goods, by value,
will be shipped abroad by air in 1999 than by sea. Air routes, whether
used for shipping goods or human travel, would be the most convenient
justification for using the straight-line or great-circle measure of distance.
THE GRAVITY MODEL OF BILATERAL TRADE 69

Institute for International Economics

http://www.iie.com

The ultimate justification, however, is that it seems to be a reasonable

way of averaging across different modes of transportation and works
well in practice.
We noted when analyzing the c.i.f. measure of transport costs in chapter
3 that it does not seem to rise linearly with distance, but rather less rapidly
than that. This is as one would expect: there is a large cost associated
with loading cargo onto a vessel and then a relatively small marginal cost
per mile of distance traveled. The logarithm has the property that average
cost diminishes with distance (assuming the coefficient comes out less
than 1). We usually specify distance, like the other variables, as entering
the equation in log form. We have also tried other functional forms,
however, as explained below.
It must be admitted that transport costs will not always and everywhere
be monotonically increasing in distance, let alone in a convenient logarithmic form. Lipsey and Weiss (1974) showed that the distance a product is
shipped is positively associated with the level of transport costs, but with
more unexplained variation than one would expect. Other factors found
to be important determinants of shipping costs were bulk (cubic feet per
ton), small consignments (under one ton), the unit value of exports (which
they attribute to price discrimination on the part of oligopolistic shippers),
and the commodity in question (grains can be shipped in more open,
competitive markets, specifically by tanker rather than liner).
One can imagine a variety of ways to measure transport costs more
directly than simply using distance. The c.i.f.-f.o.b. differential has drawbacks, which have already been mentioned in chapter 3. A gravity model
by Geraci and Prewo (1977) supplements distance data (air routes) with
numbers on the ratio of c.i.f. to f.o.b. values (evidently bilaterally), while
acknowledging serious measurement error, for 18 OECD countries. This
approach has the virtue of allowing rough estimates of the effects of
shipping costs per se on trade, expressed as a percentage of value, as
opposed to the indirect effects of distance. Their average estimate for
transport costs within the OECD is 12.8 percenthigher for the nonEuropean countries and lower (averaging 5.2 percent) for the European
countries. The elasticity of transport cost with respect to distance is
approximately 0.53. In the tests described in chapter 3, we found somewhat higher elasticities of c.i.f. margins with respect to distance, though
the estimates are very imprecise.

Estimated Effects of Distance and Adjacency

When the adjacency variable is not included in the gravity equation, the
estimated coefficient on the log of distance is about .75 (Frankel 1993).
This means that when the distance between two countries is increased
by 1.0 percent, trade between them falls by about three-quarters of a
70 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

percent. The adjacency variable should be included, however. One has

only to think of the Mexican maquiladora strip along the US border, or
the large amount of intermediate products and consumer goods that go
back and forth across the Canadian border, to see the relevance of adjacency, beyond distance. The Netherlands is close to France and Korea to
Japan, but without the common border the effect is not the same.
When we hold constant for common borders, the estimated coefficient
on the distance variable is diminished in magnitude, to the range of about
0.5 to 0.7, as seen in tables 4.2 and 4.3. In pooled time-series estimates
(to be discussed later), the distance coefficients cluster around 0.6. Such
an estimate implies that when the distance is increased by 1 percent, trade
falls by about .6 percent.
The coefficient on the dummy variable for a common border (adjacency)
itself is estimated at around .6. Because trade is specified in logarithmic
form, the way to interpret the coefficient on a dummy variable is to take
the exponent. Two countries that share a common border are estimated
to engage in 82 percent more trade than two otherwise-similar countries
[exp(.6)1.82].
We have tested for possible interactive effects of the common-border
variable. For example, one might imagine that small adjacent countries
are more highly integrated than would be predicted by the simple sum
of their size and common-border effects. But we found no significant
interactive effects of adjacency with any of the other variables (reported
in Frankel and Romer 1996).
One can also include an effect for landlockedness, which adds to transportation costs. We have estimated the coefficient on this dummy variable
at 0.36. This means that the lack of ocean ports reduces trade by about
one-third, holding constant distance, population, and land area.17
We checked for possible nonlinearity in the log-distance term, as it
could conceivably be the cause of any apparent bias toward intraregional
trade that is left after controlling linearly for distance. The log of distance
appears to be sufficient; the level and square of distance add little.18 We
have also tried a more sophisticated measure of shipping distances. Wang
(1992) enters measures of such sea distances and land distances separately
in a gravity model. She finds a small, though statistically significant,
difference in coefficients between the two. Her measure of sea distance
17. The estimate is quite significant statistically. Both its magnitude and significance fall,
however, when the regular GNP measure of size, in addition to population and land, is
included in the equation, and estimated by instrumental variables to account for the possible
endogeneity of income with respect to trade (Frankel and Romer 1996, table 1).
18. A significant positive coefficient on the quadratic term confirms the property that trade
resistance increases less than linearly with distance, but the log is able to capture this
property. Quadratic and cubic terms in the log were not at all significant when tried alongside
the log (Frankel 1993).
THE GRAVITY MODEL OF BILATERAL TRADE 71

Institute for International Economics

http://www.iie.com

Table 4.5 Estimates of coefficients on log

distance
Source

Estimate

Without controlling for adjacency

Brada and Mendez (1983)
Bikker (1987)
Boisso and Farrantino (1995)
Linnemann (1966)
Oguledo and MacPhee (1994)
Mansfield and Bronson (1997)

0.76
0.9 to 1.1
1.0 to 1.5
.77
.76
.51 to .69a

Controlling for adjacency

Aitken (1973)
Biessen (1991)
Bergstrand (1985)
Tinbergen (1962)

.35 to .51
.74 to .86
.75 to .78
.56

a. For 1950-90.

takes into account the lengthier trips involved in sea voyages around
obstacles such as the Cape of Good Hope and Cape Horn, and she also
adds the land distance from the center of the country to the major port.
We tried these data, generously supplied by Winters and Wang (1994),
in place of own simpler distance measure. The change had relatively little
effect on most of the results (Frankel, Wei, and Stein 1995).19 In short, the
precise method of measuring distance appears to be less of an issue than
one might have thought.
Our estimate for the effect of distance on bilateral trade is similar to
that estimated by many others (table 4.5). Those controlling for adjacency
tend to get lower coefficients on the log of distanceas we do and as
one would expect.20
Bikker (1987), who measures distance by sea routes, tries a clever way
of isolating the role of physical shipping costs from the other costs of
doing business at a distance. For those years when the Suez Canal was
closed by blockade, 1967-75, he adds a variable for the additional sea
distance that had to be covered between the country pair in question,
divided by the normal distance. The Suez variable is statistically significant. Its estimated coefficient is low, however (0.2)less than one-fifth
of the effect of the regular distance variable. This leads him to conclude
that physical shipping costs are less important than conventionally
assumed and that other sorts of costs to doing business at a distance are
19. The coefficient of distance varies a bit over the course of the observations but with no
clear trend.
20. Leamer (1993) obtains a similar elasticity, .68, for West German trade. He is struck by
the importance of distance and concludes that, under NAFTA, Southern California will
experience the greatest increase in trade with Mexico.
72 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

correspondingly more important. While air transport may explain a bit

of the difference, the result suggests that shipping costs do not constitute
the majority of costs to doing business at a distance.
We have tried disaggregating trade into three categories. The results
show higher distance effects for manufactures than for agricultural products or other raw materials, even though the latter are bulkier. In 1980,
for example, the coefficient on distance (0.58 if estimated on trade
in aggregate form) is only 0.30 on agricultural goods, but 0.53 on
manufactures (and 0.57 on other raw materials).21 In other years, the
coefficient on manufactures is the highest in absolute value of all three
categories: 1965, 1970, 1975, and 1985 (the last year for which we have
this disaggregated data).22 These findings confirm that physical transport
costs are not necessarily the most important component of costs associated
with distance. The costs associated with transport time and cultural unfamiliarity may be greater, and these costs are more important for manufactured goods than for agriculture.
There is no observable tendency for the effect of distance to fall over
time during the course of our sample: 1965-92. If anything, the trend
seems to be upward. The same is true for the distance coefficients in
the gravity model estimated by others, most extensively by Boisso and
Ferrantino (1996) in their year-by-year study between 1950 and 1988.23
Indeed, our distance coefficient seems to be quite similar to what other
gravity-equation researchers estimated in the 1920s, 1930s, and 1950s. De
Menil and Maurel (1994) found a coefficient of .70 during 1924-26,
despite what must have been much higher shipping costs at that time.
In analyzing 34 countries (561 bilateral flows) later in the interwar period,
Eichengreen and Irwin (1995, 14) also get coefficients similar to ours: .71
in 1928, .55 in 1935, and .51 in 1938. Related estimates in Eichengreen
and Frankel (1995) show a distance coefficient of .48 in 1928, .37 in

21. The standard errors are all 0.06 to 0.07. The agriculture estimates are significantly less
than the others.
22. This was reported in Frankel, Stein, and Wei (1994, table 2) but omitted in the published
version to save space. See also Frankel, Wei, and Stein (1995, tables 4a and 4b). Smeets
(1994) gets similar results when studying Germanys bilateral trade in 1978: the negative
effect of distance is least evident when estimated for crude materials, is more evident for
food and live animals (followed by beverages and tobacco), and is consistently very highly
significant for mineral fuels, manufactures (the bulk of Germanys trade), and other categories.
23. Earlier, when Boisso and Ferrantino (1993) estimated the gravity model without allowing
a role for FTAs, they found that the apparent coefficient on distance did decline steadily
over the period. But the coefficient must have been appropriating some of the effect of the
regional trading arrangements. This illustrates the need to allow for both bloc effects and
proximity effects in explaining existing regional concentration, neither omitting the former
(as in the Krugman-Summers school) nor the latter (as in the Bhagwati-Panagariya school).
THE GRAVITY MODEL OF BILATERAL TRADE 73

Institute for International Economics

http://www.iie.com

1935, and .34 in 1938. The coefficient thus tends to be lower in magnitude
than estimates from the postwar period.
This trend goes against what one might expect from declining transportation costs. The authors attribute it to missing variables. But we suspect
that the coefficient of distance is not reliably linked to physical shipping
costs. Eichengreen and Irwin (1997) confirm those estimates for the interwar period and find similar effects for distance in 1949, 1954, and 1964,
as well. Compounding the apparent failure of the distance effect to decline
over time, Flandreau (1995) finds coefficients for the 19th century that are
even lower in absolute value: .48 for 1860, .49 for 1870, and .27 for
1880.24 In short, we have more than a century of gravity estimates, and
nowhere is there evidence of a decline over time in the distance coefficient.
This trend in the distance coefficient, or lack thereof, is at first thought
surprising. At second thought, it looks like a consequence of the proposition that physical shipping is only one of several sorts of costs associated
with doing business at a distance. Neither reaction is correct. The other
costs to doing business at a distance, such as unfamiliarity with foreign
cultures, should be declining over time right along with physical shipping
costs, as a consequence of increased ease of travel and communication.
The most likely of the possible explanations is rather that, even though
the average cost of shipping per kilometer has undoubtedly declined over
time, there is no reason to think that the marginal cost per percentage increase
in distance has declined over time. But this is what the coefficient in
the gravity equation measures. The reverse could even be the case, if
technology has for some reason reduced transportation costs at relatively
short distances more rapidly than it has reduced costs at long distances.
Trucking, for example, may have reduced shipping costs at short distances
by more than innovations in air and sea shipping have reduced costs at
long distances. Imagine that technological progress reduced shipping
costs at all distances by some fixed percentage of their previous level. Then
there would be no reason for the coefficient on log distance to fall.25

Linguistic Links
Next, we added a dummy variable to represent when both countries of
a pair speak a common language or had colonial links earlier in the
century. We allowed for English, Spanish, Chinese, Arabic, French, German, Japanese, Dutch, and Portuguese.26 We allow countries to speak
24. Eichengreen-Frankel, Eichengreen-Irwin, and Flandreau all hold constant for adjacency.
25. I am most indebted to J. David Richardson for this point.
26. Havrylyshyn and Pritchett (1991) find that three languages are significant in the gravity
modelPortuguese, Spanish, and English, in decreasing order of magnitude. In a study of
poor countries, Foroutan and Pritchett (1993) find that French, Spanish, and English are
statistically significant.
74 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com

more than one language. Switzerland, for example, is counted as having

linguistic links with both France and Germany.
The results show a highly significant effect when all languages are
constrained to have the same coefficient. The estimate fluctuates over
time between .33 and .77. Pooled time-series estimates of the coefficient
cluster around 0.44. The implication is that two countries sharing linguistic/colonial links tend to trade roughly 55 percent more than they would
otherwise [exp(.44)1.55]. In other words, the effect of sharing a common
language, even for far-removed countries, is very similar in magnitude
to the effect of sharing a common border.
We tested whether some of the major languages were more important
than others. When we tried supplementing the general language term
and allowing each of the five major languages to have an independent
extra coefficient, the language effects lost statistical significance for half
the years, due to multicollinearity (Frankel and Wei 1995c, table 2).27 Nonetheless, two languages, English and Chinese, appear to qualify as especially important. Two Chinese-speaking countries appear to trade four
times as much as other countries. There is a problem here, however. Direct
Taiwan-China trade is not reported in the official statistics because it was
officially nonexistent during our sample period. Much of it goes through
Hong Kong and is thus counted twice. An attempt to correct for this
factor turns out to eliminate the extra effect of the Chinese language term.
The difference between the Chinese effect and that for other languages
is not, after all, highly significant (Frankel and Wei 1993a, table 3).
Somewhat surprisingly, the inclusion of the linguistic/colonial terms
has little effect on the other coefficients. The regional bloc effects, which
we will turn to in the next chapter, are not much lower when linguistic
links are included than when they are not.28
Boisso and Ferrantino (1996) construct a new measure of linguistic
distance that is a continuous scalar rather than a discrete dummy variable,
thereby taking into account linguistic diversity within countries. They
generate the percentages of the population in a given country speaking
each language, and then construct an index of linguistic dissimilarity for
each pair of countries. The coefficient on their index exhibits a trend
whereby trade in the postwar period has taken place among countries
that are increasingly similar linguistically. They interpret this as possible
evidence of increased cultural barriers to trade. That interpretation is
consistent with the upward trend in our estimated language coefficients.
Their result, however, does not hold constant for the other variables in

27. When manufactured goods are considered alone and the five individual major languages
are estimated independently, the language coefficients lose all statistical significance (Frankel, Stein, and Wei 1994, table 5).
28. Estimates of the equation without the language effect were reported in Frankel (1993).
THE GRAVITY MODEL OF BILATERAL TRADE 75

Institute for International Economics

http://www.iie.com

the gravity model. It could just be a reflection of the effect of regional

trading arrangements that themselves tend to be drawn along linguistic
lines. When the gravity model takes into account the increasing importance of FTAs, the significance and trend in the coefficient on the BoissoFerrantino linguistic dissimilarity variable are less clear.

The Panel ApproachPooling Time Series

and Cross-Section
To bring the most information possible to bear at once, one can pool the
data across the cross-section and time-series dimensions. Estimates will
be reported in the next chapter. We flag one immediate benefit of pooling.
We can now conduct meaningful tests for such FTAs as NAFTA and the
Australia-New Zealand Closer Economic Relationship that could not be
reliably tested in the cross-section dimension because of the very small
number of observations (three for NAFTA and one for Australia-New
Zealand).
To pool observations from different years without either allowing each
to have a separate constant term or else converting the dollar figures to
suitable real terms would let inflation distort the estimates.29 Furthermore,
our theoretical rationale for the gravity model (chapter 7) says that, in a
time-series dimension, we should normalize for growth in real gross
world product. We have adopted the approach of computing year-specific
intercept terms, which absorb the effects of global inflation and growth.30

29. Bikker (1987, note 3) points out that this is a problem with Aitken (1973).
30. Appendix C discusses some of the details of the interpretation of the intercept. Brada
and Mendez (1988) also allow yearly constant terms.
76 REGIONAL TRADING BLOCS IN THE WORLD ECONOMIC SYSTEM

Institute for International Economics

http://www.iie.com