The Gravity Model

Author(s): James E. Anderson

Source: Annual Review of Economics , 2011, Vol. 3 (2011), pp. 133-160
Published by: Annual Reviews

Stable URL: https://www.jstor.org/stable/42940183

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Annual Reviews is collaborating with JSTOR to digitize, preserve and extend access to Annual
Review of Economics

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 The Gravity Model
James E. Anderson
Department of Economics, Boston College, Chestnut Hill, Massachusetts 02467, and
NBER, Cambridge, Massachusetts 02138; email: james.anderson.l@bc.edu

Annu. Rev. Econ. 2011. 3:133-60

Keywords
First published online as a Review in Advance on incidence, multilateral resistance, trade costs, migration
February 25, 201 1

The Annual Review of Economics is online at Abstract

economics.annualreviews.org
Gravity has long been one of the most successful empirical models
This article's doi:
in economics. Incorporating deeper theoretical foundations of
1 0. 1 1 46/annurev-economics-l 1 1 809-1 25 114
gravity into recent practice has led to a richer and more accurate
Copyright © 201 1 by Annual Reviews.
estimation and interpretation of the spatial relations described by
All rights reserved
gravity. Wider acceptance has followed. Recent developments are
JEL codes: FIO, F22, J61, FIO, R40
reviewed here, and suggestions are made for promising future
1 94 1 - 1 3 83/1 1/0904-01 33 $20.00 research.

1 33

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 1. INTRODUCTION

The gravity model in economics was until relatively recently an intellectual orp
unconnected to the rich family of economic theory. This review is a tale of the orp
reunion with its heritage and the benefits that continue to flow from connections to
distant relatives.
Gravity has long been one of the most successful empirical models in economi
ordering remarkably well the enormous observed variation in economic interact
across space in both trade and factor movements. The good fit and relatively tig
clustering of coefficient estimates in the vast empirical literature suggested that
underlying economic law must be at work, but in the absence of an accepted conne
to economic theory, most economists ignored gravity. The authoritative survey
Learner & Levinsohn (1995) captures the state of professional thinking in the m
1990s: "These estimates of gravity have been both singularly successful and singul
unsuccessful. They have produced some of the clearest and most robust empirical
ings in economics. But, paradoxically, they have had virtually no effect on the subj
international economics. Textbooks continue to be written and courses designed wit
any explicit references to distance, but with the very strange implicit assumption
countries are both infinitely far apart and infinitely close, the former referring to fa
and the latter to commodities." Subsequently, gravity first appeared in textbook
2004 (Feenstra 2004), following success in connecting gravity to economic theory,
subject of Section 4.
Reviews are not intended to be surveys. My take on the gravity model, thus license
be idiosyncratic, scants or omits some topics that others have found important wh
emphasizes some topics that others have scanted. My emphases and omissions are int
to guide the orphan to maturity. An adoptive parent's biases may have contaminate
judgment, caveat emptor.
My focus is on theory. Incorporating the theoretical foundations of gravity into re
practice has led to richer and more accurate estimation and interpretation of the s
relations described by gravity, so where appropriate I point out this benefit. The ha
reaped from empirical work applying the gravity model is recently surveyed elsew
(Anderson van Wincoop 2004, Bergstrand Sc Egger 2011).
From a modeling standpoint, gravity is distinguished by its parsimonious and tracta
representation of economic interaction in a many-country world. Most internationa
nomic theory is concentrated on two-country cases, occasionally extended to three-cou
cases with special features. The tractability of gravity in the many-country case re
from its modularity: The distribution of goods or factors across space is determine
gravity forces conditional on the size of economic activities at each location. Modul
readily allows for disaggregation by goods or regions at any scale and permits infe
about trade costs that is not dependent on any particular model of production and ma
structure in full general equilibrium. The modularity theme recurs often below b
missing from some other prominent treatments of gravity in the literature.

2. TRADITIONAL GRAVITY

The story begins by setting out the traditional gravity model and noting clues to
union with economic theory. The traditional gravity model drew on analogy with

134 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
Newton's law of gravitation. A mass of goods or labor or other factors of production
supplied at origin /, Y„ is attracted to a mass of demand for goods or labor at destina-
tion /, Ej , but the potential flow is reduced by the distance between them, dir Strictly
applying the analogy,

X,i = Y¡Ej/dfj

gives the predicted movement of goods or labor between i and /, Xz/. Ravenstein (1889)
pioneered the use of gravity for migration patterns in the nineteenth-century United
Kingdom. Tinbergen (1962) was the first to use gravity to explain trade flows. Departing
from strict analogy, traditional gravity allowed the exponents of 1 applied to the mass
variables and of 2 applied to bilateral distance to be generated by data to fit a statisti-
cally inferred loglinear relationship between data on flows and the mass variables and
distance.

Generally, across many applications, the estimated coefficients on the mass variables
cluster close to 1 and the distance coefficients cluster close to -1, while the estimated
equation fits the data well: Most data points cluster close to the fitted line in the sense that
80%-90% of the variation in the flows is captured by the fitted relationship. The fit of
traditional gravity improved when supplemented with other proxies for trade frictions,
such as the effect of political borders and common language.
Notice that bilateral frictions alone would appear to be inadequate to fully explain the
effects of trade frictions on bilateral trade because the sale from i to j is influenced by the
resistance to movement on z' s other alternative destinations and by the resistance on
movement to / from /' s alternative sources of supply. Prodded by this intuition, the tradi-
tional gravity literature recently developed remoteness indexes of each country's average
effective distance to or from its partners (£]• d¡j/Yi was commonly defined as the remote-
ness of country /) and used them as further explanatory variables in the traditional gravity
model, with some statistical success.
The general problem posed by the intuition behind remoteness indexes is analogous
to the N-body problem in Newtonian gravitation. An economic theory of gravity is
required for an adequate solution. Because there are many origins and many destina-
tions in any application, a theory of the bilateral flows must account for the relative
attractiveness of origin-destination pairs. Each sale has multiple possible destinations,
and each purchase has multiple possible origins: Any bilateral sale interacts with all
others and involves all other bilateral frictions. This general equilibrium problem is
neatly solved with structural gravity models. For expositional ease, the discussion
focuses below on goods movements, except when migration or investment is specifically
treated.

3. FRICTIONLESS GRAVITY LESSONS

Taking a step toward structure, an intuitively appealing starting point is the description
of a completely smooth homogeneous world in which all frictions disappear. Developing
the implications of this structure yields a number of useful insights about the pattern of
world trade.

A frictionless world implies that each good has the same price everywhere. In a homo-
geneous world, economic agents everywhere might be predicted to purchase goods in the
same proportions when faced with the same prices. In the next section, the assumptions on

www.annualreviews.org • The Gravity Model 135

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 preferences and/or technology that justify this plausible prediction are the focus, but here
the focus is on the implications for trade patterns. In a completely frictionless and homo-
geneous world, the natural benchmark prediction is that X/y/Ey = Y,-/Y, where the propor-
tion of spending by / on goods from i is equal to the global proportion of spending on
goods from /, and Y denotes world spending.
Any theory must impose adding up constraints, which for goods requires that the sum
of sales to all destinations must equal Y¿, the total sales by origin /, and the sum of
purchases from all origins must equal Ey, the total expenditure for each destination /. Total
sales and expenditures must be equal: i.e., ^ Y, = Ey = Y.
One immediate payoff is an implication for inferring trade frictions. Multiplying
both sides of the frictionless benchmark prediction X/y/Ey = Y,/Y by Ey yields predicted
frictionless trade Y,Ey/Y. The ratio of observed trade X/y to predicted frictionless trade
YiEj/Y represents the effect of frictions along with random influences. (Bilateral trade
data are notoriously rife with measurement error.) Fitting the statistical relationship
between the ratio of observed to frictionless trade and various proxies for trade costs is
justified by this simple theoretical structure as a proper focus of empirical gravity
models.
Thus far, the treatment of trade flows has been of a generic good that most of the
literature has implemented as an aggregate: the value of aggregate bilateral trade in
goods, for example. But the model applies more naturally to disaggregated goods
(and factors) because the frictions to be analyzed below are likely to differ markedly
by product characteristics. The extension to disaggregated goods, indexed by k , is
straightforward:

ykpk

x»=-W = sWYk. (1)

Here sf = Y^/Yk is country i's share of the world's sales of goods class k , and
is country /' s share of the world's spending on k , equal to Y*, the world's sales
The notation and logic also readily apply to the disaggregation of count
regions, and indeed a prominent portion of the empirical literature has exam
eral flows between city pairs or regions, motivated by the observation tha
nomic interaction is concentrated at very short distances. The model can be
to reflect individual decisions aggregated with a probability model (see Se
below).
In aggregate gravity applications (i.e., most applications), it has been common to use
origin and destination mass variables equal to gross domestic product (GDP). This is
conceptually inappropriate and leads to inaccurate modeling unless the ratio of gross
shipments to GDP is constant (in which case the ratio goes into a constant term). A possible
direction for aggregate modeling is to convert trade to the same value-added basis as GDP,
but this seems more problematic than using disaggregated gravity to explain the pattern of
gross shipments and then uniting estimated gravity models within a superstructure to
connect to GDP. That is the strategy of the structural gravity model research program
reviewed here.
Equation 1 generates a number of useful implications, (a) Big producers have big
market shares everywhere. ( b ) Small sellers are more open in the sense of trading more
with the rest of the world, (c) The world is more open the more similar in size, and the
more specialized, the countries are. (d) The world is more open the greater the number of

136 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 countries, (e) World openness rises with convergence under the simplifying assumption of
balanced trade.

Implication (a), that big producers have big market shares everywhere, follows because,
reverting to the generic notation and omitting the k superscript, the frictionless gravity
prediction is that

Xij/E¡ = Sļ.

Implication (¿?), that small sellers are more open in the sense of trading more with the rest
of the world, follows from

^X,;/£, = l-y;/Y=l-s;
¥i

using Ysj Ej - Si Y» which is balanced trade for the world.

Implication ( c ) is that the world is more open the more similar in size, and the more
specialized, the countries are. It is convenient to define world openness as the ratio of
international shipments to total shipments, X¡j/Y. Dividing Equation 1 through
by Yk and suppressing the goods index k , world openness is given by

E E x"/y = E M1 - */> = 1 - E v,-

i ¥i i i

Using standard st

^2 bjSj = Nrb
/

where N is the number of countries or regions, Var denotes variance, r¿,s is the correla-
tion coefficient between b and s, and 1 /N = S/S//N = the average share. This
equation follows from the shares summing to one and using standard properties of
covariance. Here, Var(s) and Var(b) measure size dissimilarity, and the correlation of
s and b, r^s, is an inverse measure of specialization. Substituting into the expression for
world openness,

EE Xjj/Y = 1 - 1/N - Nrhsy/Var(s)Var(b). (2)

/ ¥i

Implication (c) follows from Equation 2 because on the right-hand side the simila
country size shrinks the variances while specialization shrinks the correlation r^s.
The country-size similarity property has been prominently stressed in the mon
tic competition and trade literature. (It is sometimes taken as evidence for mono
competition in a sector rather than as a consequence of gravity no matter what e
the pattern of the b's and s's.) The specialization property has also been noted
literature as reflecting forces that make for greater net international trade, the
value of Sj-bj. Making comparisons across goods classes, variation in the right-
side of Equation 2 results from variation in specialization and in the dispersion
shipment and expenditure shares. Notice again that the cross-commodity varia
world openness arises here in a frictionless world, a reminder that measures of
home bias in a world with frictions must be evaluated relative to the frictionless
benchmark.

www.annualreviews.org • The Gravity Model 137

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 Country-size similarity also tends to increase bilateral trade between any pair of coun-
tries, all else equal. This point (Bergstrand & Egger 2007) is seen most clearly with
aggregate trade that is also balanced, hence s; - bj. Equation 1 can be rewritten as

X
X A'/
A'/ -¿¿Willi ~ s/
~ s/ sj y >

where s'j = Y,/(Y;- + Y/), the share of / in the joint GDP of i and /. The product sļsļ
is maximized at s¡ = s- - 1/2, so for a given joint GDP size, bilateral trade is
increasing in country similarity. (With unbalanced trade or specialization, an analo-
gous similarity property holds for the bilateral similarity of income and expenditure
shares. Let yj = Ej/Yj. Then the same equation as before holds with the right-hand
side multiplied by y;.)
A more novel implication of Equation 2 is implication (d), that world openness is
ordinarily increasing in the number of countries. Increasing world openness due to a rise
in the number of countries reflects the property that smaller countries are more naturally
open and division makes for more and smaller countries.
This effect is seen by differentiating the left-hand side of ^ Yli^j XV/Y=1
yielding -J2j(bjdsj + s¡dbj). Increasing the number of countries tends to imply reducing
the share of each existing country while increasing the share (from zero) of the new
country. The preceding differential expression should thus ordinarily be positive.
The qualification with the term ordinarily is needed because the pattern of share
changes will depend on the underlying structure as revealed by the left-hand side of
Equation 2. On the one hand, the average share 1 /N decreases as N rises, raising world
openness. On the other hand, the change in the number of countries will usually change
rbs'/vär(b)Var(s) in ways that depend on the type of country division (or confederation)
as well as on indirect effects on shares as prices change. [The apparent direct effect of N in
the first term on the right-hand side of Equation 2 vanishes because 1/N scales
y/Var(b)Var(s).]
A practical implication of this discussion is that intertemporal comparisons of ratios of
world international trade to world income, to be economically meaningful, should be
controlled for changes in the size distribution and the number of countries, a correction of
large practical importance in the past 50-100 years. Alternatively, measures of openness
meant to reflect the effects of trade frictions should be constructed in relation to the
frictionless benchmark.
Applied to aggregate trade data, gravity yields implication (č), that world openness rises
with convergence under the simplifying assumption of balanced trade for each country,
bj = Sj , V/. The right-hand side of Equation 2 becomes NVar(s) + 1/N under balanced
trade, and per-capita income convergence lowers Var(s) toward the variance of population.
Baier Sc Bergstrand (2001) use the convergence property to partially explain postwar
growth in world trade/income, finding relatively little action, although presumably more
recent data influenced by the rise of China and India might give more action.
Pointing toward a connection with economic theory, the shares s¡ and bj and the
plausible hypothesis of the frictionless model must originate from an underlying structure
of preferences and technology. Also, the deviation of observed Xř/ from the frictionless
prediction reflects frictions as they act on the pattern of purchase decisions of buyers and
the sales decisions of sellers, which originate from an underlying structure of preferences
and technology.

13 S Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
4. STRUCTURAL GRAVITY

Modeling economies with trade costs works best if it moves backward from the end user.
Start by evaluating all goods at user prices, applying demand-side structure to determine
the allocation of demand at those prices. Treat all costs incurred between production and
end use as being incurred by the supply side of the market, even though there are often
significant costs directly paid by the user. What matters economically in the end is the full
cost between production and end use, and the incidence of that cost on the producer and
end user. Many of these costs are not directly observable, and the empirical gravity litera-
ture indicates the total is well in excess of the transportation and insurance costs that are
observable (see Anderson &c van Wincoop 2004 for a survey of trade costs).
The supply side of the market under this approach both produces and distributes the
delivered goods, incurring resource costs that are paid by end users. The factor markets for
those resources must clear at equilibrium factor prices, determining costs that link to end-
user prices. Budget constraints require national factor incomes to pay for national expen-
ditures plus net lending or transfers including remittances. Below the national accounts,
individual economic agents also meet budget constraints. Goods markets clear when prices
are found such that demand is equal to supply for each good. The full general equilibrium
requires a set of bilateral factor prices and bilateral goods prices such that all markets clear
and all budget constraints are met.
This standard description of general economic equilibrium is too complex to yield
something like gravity. A hugely useful simplification is modularity, subordinating the
economic determination of equilibrium distribution of goods within a class under the
superstructure determination of the distribution of production and expenditure between
classes of goods. Anderson &c van Wincoop (2004) call this property trade separability.
Observing that goods are typically supplied from multiple locations, even within fine
census commodity classes, it is natural to look for a theoretical structure that justifies
grouping in this way. The structural gravity model literature has uncovered two struc-
tures that work, one on the demand side and one on the supply side, detailed in
Sections 4.1 and 4.2.

Modularity (trade separability) permits the analyst to focus exclusively on inference

about distribution costs from the pattern of distribution of goods (or factors) without
having to explain at the same time what determines the total supplies of goods to all
destinations or the total demand for goods from all origins. This is a great advantage for
two reasons. First, it simplifies the inference task enormously. Second, the inferences about
the distribution of goods or factors are consistent with a great many plausible general
equilibrium models of national (or regional) production and consumption.
Modularity also requires a restriction on trade costs, so that only the national aggregate
burden of trade costs within a goods class matters for allocation between classes. The most
popular way to meet this requirement is to restrict the trade costs so that the distribution of
goods uses resources in the same proportion as the production of those same goods.
Samuelson (1952) invented iceberg melting trade costs in which the trade costs were
proportional to the volume shipped, as the amount melted from the iceberg is proportional
to its volume. The iceberg metaphor still applies when allowing for a fixed cost, as if a
chunk of the iceberg breaks off as it parts from the mother glacier. Mathematically, the
generalized iceberg trade cost is linear in the volume shipped. Economically, distribution
continues to require resources to be used in the same proportion as in production. Fixed

www.annualreviews.org • The Gravity Model 139

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 costs are realistic and potentially play an important role in explaining why many potential
bilateral flows are equal to zero.
More general nonlinear trade cost functions continue to satisfy the production propor-
tionality restriction and thus meet the requirements of modularity, but depart from the
iceberg metaphor. Bergstrand (1985) derives a joint cost function that is homogeneous of
degree one with constant elasticity of transformation. This setup allows for substitution
effects in costs between destinations rather than the cost independence because of fixed
coefficients in the iceberg model. Bilateral costs have a natural aggregator that is an iceberg
cost facing monopolistically competitive firms. A nice feature of the joint cost model is its
econometric tractability under the hypothesis of the profit-maximizing choice of destina-
tions. Although potentially more realistic, the joint cost refinement turns out to make
relatively little difference empirically.
Arkolakis (2008) develops a nonlinear (in volume) trade cost function in which hetero-
geneous customers are obtained by firms with a marketing technology featuring a
fixed-cost component (running a national advertisement) and a variable-cost component
(leafletting or telemarketing) subject to diminishing returns as the less likely customers are
encountered. Because of the Ricardian production and distribution technology, resource
requirements in distribution remain proportional to production resource requirements.
Arkolakis shows that the marketing technology model can rationalize features of the firm-
level bilateral shipments data that cannot be explained with the linear fixed-costs model.
His setup is not econometrically tractable but is readily applicable as a simulation model.
In all applications based on the preceding cost functions, proxies for costs are entered in
some convenient functional form, usually loglinear in variables such as bilateral distance,
contiguity, membership of a country, continent or regional trade agreement, common
language, and common legal traditions (see Anderson &c van Wincoop 2004 for more
discussion). More generality in trade costs that violates the production proportionality
restriction comes at the price of losing modularity (see Matsuyama 2007 for a recent
exploration of the implications of noniceberg trade costs in a two-country Ricardian
model; see Deardorff 1980 for a general treatment of the resource requirements of trade
costs as a setting for his demonstration that the law of comparative advantage holds quite
generally).

4.1. Demand-Side Structure

The second requirement for modularity can be met by restricting the preferences and/or
technology such that the cross effects in demand between classes of goods (either interme-
diate or final) flow only through aggregate price indexes. This demand property is satisfied
when preferences or technologies are homothetic and weakly separable with respect to a
partition into classes with members defined by location, a partition structure called the
Armington assumption. Thus, for example, steel products from all countries are members
of the steel class. Notice that the assumption implies that goods are purchased from
multiple sources because they are evaluated differently by end users, and goods are differ-
entiated by place of origin.
It is usual to impose identical preferences across countries. Differences in demand across
countries, such as a home bias in favor of locally produced goods, can be accommodated,
understanding that "trade costs" now include the effect of a demand-side home bias. In
practice it is difficult to distinguish demand-side home bias from the effect of trade costs, as

140 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
the proxies used in the literature (e.g., common language, former colonial ties, or internal
trade dummies) plausibly pick up both demand and cost differences. Henceforth, the term
trade cost is used without qualification but is understood to potentially reflect demand-side
home bias. Declines in trade costs can be understood as reflecting homogenization
of tastes.
Separability implies that each goods class has a natural quantity aggregate and a natural
price aggregate, with substitution between goods classes occurring as if the quantity aggre-
gates were goods in the standard treatment. The separability assumption implies that
national origin expenditure shares within the steel class are not altered by changes in the
prices of nonsteel products, although, of course, the aggregate purchase of steel is affected
by the aggregate cross effect. Homotheticity ensures that relative demands are functions
only of relative aggregate prices.
The first economic foundation for the gravity model was based on specifying the
expenditure function to be a constant elasticity of substitution (CES) function (Anderson
1979). Expenditure shares in the CES case are given by

ww-
where P; is the CES price index, o is the elasticity of substitution parameter, ßt is
the distribution parameter for varieties shipped from /, p¿ is their factory gate price,
and tļj > 1 is the trade cost factor between origin / and destination /. The CES price
index is given by

/ y/O-*)
P / = ÍÇ (ftM,)'""] •
Notice that the same parameters characterize expenditure behavior in
ences are common across the world by assumption. Notice also that t
ant to income; preferences are homothetic. With frictionless trade
therefore all the buyers' shares of good / must equal the sellers' sha
destination prices), Y;/Y. Thus the frictionless benchmark is justified
homothetic preferences. For intermediate goods, the same logic works
diture shares with cost shares.
The distribution parameters ß, bear several interpretations. They
taste parameters. Alternatively, in applications to monopolistically c
ß i is proportional to the number of firms from i offering distinct
1989). Countries with more active firms get bigger weights. In lon
competition, the number of firms is endogenous. Owing to fixed entr
tries have more active firms in equilibrium, all else equal. The num
contributes to determining the Y,'s that are given in the gravity modul
The other building block in the structural gravity model is market
ered prices Y i = Multiplying both sides of Equation 3 by E¡ a
/ yields a solution for ßip]~a:

B.p'- =

Define the denominator as II- _fT.

www.annualreviews.

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 Substituting into Equations 3 and 4 yields the structural gravity model:

(î)

= (6)

(7)

The second ratio on the right-hand side of Equation 5 is a decreasing function (under the
empirically valid restriction a > 1 ) of direct bilateral trade costs relative to the product of
two indexes of all bilateral trade costs in the system.
Anderson &c van Wincoop (2003) call the terms P¡ and II,- inward and outward multi-
lateral resistance, respectively. Note that {Pj -<T, IIř1-ťT} can be solved from Equations 6
and 7 for given Ef s, and Y¡ s combined with a normalization.1 Under the assumption
of bilateral trade cost symmetry t¡j = ř//,V/,/ and balanced trade £; = Y/,V/, the natural
normalization is II, = P¿. Anderson & van Wincoop estimated their gravity equation
for Canada's provinces and U.S. states with a full information estimator that utilized
Equation 7 with ü; = P¿. Subsequent research has focused mostly on estimating Equation
5 with directional country fixed effects to control for Ej/Pj ~a and Yi/Ylj~a.
Multilateral resistance is, on the face of it, an index of inward and outward bilateral
trade costs, but because of the simultaneity of the system (Equations 6 and 7), all
bilateral trade costs in the world contribute to the solution values. This somewhat
mysterious structure has a simple and intuitive interpretation: Inward and outward
multilateral resistance measure average buyers' and sellers' incidence of trade costs,
respectively.
The incidence interpretation follows because the uniform preferences assumption in
demand implies that the seller in effect makes a single shipment at a uniform markup
factor n, to a world market with a share determined by

ī lirj • <8)
The right-hand side of Equation 8, referring
interpreted as the global expenditure share o
world market, where the world price index
summing Equation 8. = 1 is a convenient
i <7^/(i
price. Then with given ßjpfs, the= normaliz
1 is a u
normalization in solving for multilateral re
II/ is straightforwardly interpreted as the s
(Anderson &; Yotov 2010a).

^or any solution to the system {Pj3, n®}, {AP;°, nf/A} is a

& Yotov (2010a) find that the system given in Equations 6
quadratic in the 1 - a power transforms of the P* s and ITs

142 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 Pj is now interpreted as buyers' incidence. Solving Equation 7 for P;, it is a CES index of
bilateral buyers' incidences fy/ri/, V/, equivalent to buyers paying a uniform markup factor
Pj on their entire bundle of shipments (from all i)ū Sellers' incidence n, similarly is then
interpretable as a CES index of the bilateral sellers' incidences řz//P,-, from Equation 6.
The interpretation of n and P as buyers' and sellers' incidence generalizes the elemen-
tary economics idea of incidence in the one-good case. If the actual set of trade costs were
to be replaced with hypothetical trade costs = FI ,P;, market clearance, and budget
constraints, Equations 6 and 7 would still hold with the initial equilibrium shares; hence
the sellers' factory gate prices would remain the same, and the aggregate buyers' prices
would remain the same.2 In this sense, the set of bilateral is equivalent to the set of ttf s
that decompose into the product of buyers' and sellers' incidence factors. (Unlike the one-
good case, it is the aggregate sales and purchases that is constant; bilateral flows would
change in the hypothetical equilibrium.)
The model given in Equations 5-7 applies to any goods class. For disaggregated gravity,
all variables and parameters given in Equations 5-7 should be understood as having
superscript k's to denote the goods class in question. When accounting for substitution
between goods classes, aggregate expenditure (or the cost of intermediate inputs) is given
by the expenditure (or cost) function C(Pj , . . .,Pf)uj, where C( ) is the aggregate cost of
living index for /', and w7 is the utility of the representative agent (or quantity of aggregate
output). Then, by Shephard's lemma, E * = P* dC(-) / dP1- . Each class of goods has expendi-
ture shares described by Equations 3 and 4 but is amended with the addition of a super-
script k to every variable and parameter. Anderson van Wincoop (2004) argue
theoretically for estimating disaggregated gravity, and Anderson &c Yotov (2010a, 2010b)
demonstrate that aggregation bias is large in practice.
The buyers' and sellers' incidence measures are usefully interpreted as the incidence of
total factor productivity (TFP) frictions in distribution. They contrast with standard TFP-
type measures of productivity in distribution. The sectoral TFP friction in distribution
is defined by the uniform friction that preserves the value of sectoral shipments at destination
prices: t) = ^ÎjJÎjlYhj where y* denotes the number of units of product class k received
from / at destination /. řf is a Laspeyres index of outward trade frictions facing seller i in
good k.
The TFP measure if is useful for analyzing the distribution productivity of the world
economy as a whole, but it is misleading for purposes of understanding comparative
economic performance and the national patterns of production and trade, if gives the
sellers' incidence only under the partial equilibrium and inconsistent assumption that all
incidence falls on the seller i. Anderson &c Yotov (20 10a, b) show that in practice these
differences are significant: Laspeyres TFP measures and the incidence of TFP in distribu-
tion differ in magnitude, and in the case of inward measures, the correlation between them
is low.3

2This property of Equations 6 and 7 was noted by Anderson &c van Wincoop (2004), foreshadowing the interpreta-
tion of multilateral resistance as incidence.

3An alternative measure proposed by Redding &c Venables (2004) resembles multilateral resistance but does not
measure incidence. Their measure of market access uses essentially the same formula as Equation 6, whereas their

measure of supplier access uses the CES price index formula P* = 1^,- (ßfpf í,y)1-<7* j • These variables are
constructed without taking account of the simultaneous determination of the two variables, so they do not measure
incidence.

www.annualreviews.org • The Gravity Model 141

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 For consistency of the gravity modules with full general equilibrium, involving alloca-
tion across the sectors k in each country, the ITs are normalized in each sector k for given
parameters and factory gate price pf by

£ (/Jfpfnf)1-" = l. (9)
/

In practice, when analyzing a gravity module, it is often convenient to normalize one of the
P's to 1. The choice of normalization is irrelevant to the distribution of the goods because
only relative incidence matters.
Now let us return to the interpretation of the gravity equation (Equation 5), reproduced
here for convenience:

" y vn,p;j '

The right-hand side is the product of two ratios. The first ratio is the predicted frictionless
trade flow given the E's and Y's, Y¿E¡/Y. The second ratio is thus interpreted as the ratio of
predicted trade (given the ťs) to predicted frictionless trade.
The useful measure of constructed home bias (CHB) (Anderson &c Yotov 2010a) is
interpreted as the predicted value of internal trade of / with itself to the predicted value of
internal trade in the frictionless equilibrium. CHB is thus given by

cms(m)"'- "0>
CHB varies substantially by country, product, and time because of changing expend
ture and supply shares, even when gravity coefficients are constant (Anderson & Yot
2010a, b).
Policy makers often focus on overall import penetration ratios such as an<^
the analogous ratio for exports. These concerns are acute for certain goods
classes. The import and export penetration ratios are a linear function of CHB for any
goods class k :

£ Xl/Ef = 1 - (tg/řfn*)1-" Y^/Y*, (11)

¥i

¿2 X$/Y' = 1 - (12)
¥i

Anderson Yotov show that CHBs vary a lot across goo

policy concerns, they exhibit a lot of intertemporal move
shipment shares at constant i*' s, implying a lot of expla
export ratios.
The interpretation of the second ratio in Equation 5 a
bilateral flow: It is equal to the ratio of predicted bilate
trade; hence (ří//níP/ )1_<J is the constructed trade bias on
buyer's bilateral incidence from i relative to the averag
Alternatively, the same statistic viewed from the expo
bilateral seller's incidence relative to the average seller

144 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
shift about because of changes in production and expenditure shares of world ship-
ments, as implied by the frictionless gravity model, but also because of the general equilib-
rium force of share changes that alters incidence even when trade costs {ří;} are constant
(Anderson &; Yotov 2010a).
The gravity model also readily disaggregates within countries, allowing useful investi-
gations of interregional versus international trade costs. Indeed, the development of the
structural gravity model (Anderson van Wincoop 2003) was provoked to solve a puzzle
posed by one of the most provocative and useful empirical findings of the traditional
gravity literature. McCallum (1995) found that crossing the Canadian border had an
enormous trade-destroying effect on the trade flows of Canada's provinces. He found that
Canada's provinces traded 22 times more with each other than with U.S. states, all else
equal. This was too large to be a sensible component of bilateral trade costs tiv
Structural gravity solved the puzzle by showing that the border dummy variable in
McCallum's traditional model reflected the effect of multilateral resistance. The border
dummy in the McCallum regression shifts the ratio of interprovincial trade to province-
state trade. Because it is a traditional gravity regression, it does not control for multilateral
resistance. Using Equation 5 to form this ratio for a pair of such flows in the structural
gravity model and rearranging terms, one obtains, for British Columbia's exports to adja-
cent Alberta and across the U.S. border to adjacent Washington,

XBC,AB _ Í tßc,wA Pab Y 1

XßC,WA 'tßC,AB PwaJ

The expression on the right-hand side of the equation reflects not only the direct trade
cost increase at the U.S. border that raises tBC,wA/tBC,AB> but also the effect of the ratio
of multilateral resistances for a province and a state, in this case Alberta and Washing-
ton, PAB/PWA. Because Canada's provinces must do far more of their trade with the
outside world than do U.S. states (Canada is about one-tenth the size of the United
States in GDP), the provinces naturally have higher multilateral resistance than the
states, thereby greatly increasing interprovincial trade. In McCallum's traditional gravity
regression, the border dummy variable has a regression coefficient that is an average
of such terms, although a biased estimate of it because of the omission of the multilat-
eral resistance controls from his regression. Estimating the structural gravity model,
Anderson &c van Wincoop (2003) find a more plausible border-cost component of tih
in the range of 20%-50%.
Interregional versus international trade cost implications of structural gravity were
further developed by Anderson Yotov (2010a). They offer a decomposition of incidence
into domestic and international components and calculate sellers' incidence for Canada's
provinces on trade within Canada as compared with trade with the rest of the world. They
find that, although incidence overall declined substantially from 1990 to 2002, it was
entirely on the external trade; sellers' incidence on domestic trade remained constant.
Similar investigations are likely to provide a useful context for regional integration policy
in many countries and economic areas around the world where separatism and economic
integration are important concerns.
Notice that the trade flows in Equation 5 are invariant to a uniform rise in trade costs
(including costs of internal shipment). This follows because Equations 6 and 7 imply that
raising all ří;'s by the factor / > 1 will raise each n and P by the factor À ^2. This formal
homogeneity property has useful empirical content: If the world really were getting smaller

www.annualreviews.org • The Gravity Model 14 5

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 uniformly, the gravity model would be unable to reveal it. The empirical literature tends to
indicate little change in gravity coefficients (see, especially, Anderson 6c Yotov 2010a, b),
contrary to intuition about globalization driven by falling communications costs and
improving quality of transport but consistent with uniform shrinkage of resistance to
trade.
Anderson (1979) was the first to derive gravity from the Armington/CES preference
structure, noting that Armington preferences implied a bilateral trade flow gravity equa-
tion of the form of Equation 5 that would require controlling for the importer and exporter
trade cost indexes. By using a units choice to set all equilibrium factory gate prices equal
to 1, Anderson's (1979) derivation concealed how Equations 5-7 formed a conditional
general equilibrium module that would be the foundation for the useful comparative statics
to come a generation later. The comparative statics of inward and outward multilateral
resistance were first used by Anderson &; van Wincoop (2003). Recognition that multilat-
eral resistance is interpreted as incidence is in Anderson Sc Yotov (2010a).

4.2. Supply-Side Structure

An alternative derivation of a mathematically equivalent structural gravity model was
proposed by Eaton tk, Kortum (2002), based on homogeneous goods on the demand side,
iceberg trade costs, and Ricardian technology with heterogeneous productivity for each
country and good due to random productivity draws from a Fréchet distribution. Despite
CES structure for the intermediate goods demand, in equilibrium the share of goods
demanded from i by country / is determined only on the supply side; the influence of a
disappears into a constant term. In equilibrium each country will be assigned a subset of the
goods, and except for knife-edge cases, it is the only supplier of these goods. The bilateral
trade flows obey the same formulas as Equations 5-7. 1 - a is interpreted as -0, where 0 is
the dispersion parameter of the Fréchet distribution. In contrast to the Armington/CES
model, all action is on the extensive margin of trade. Eaton &c Kortum derive their model
for one sector only, a specification generalized by Costinot et al. (2010), so that 6 ¿ is the
dispersion parameter for the distribution describing productivity draws in sector k.
The Ricardian structure of supply leads to a simple general equilibrium superstructure,
an appealing feature that has led to a growing literature combining estimation and simula-
tion. The general equilibrium superstructure is discussed below in Section 7.
Chaney (2008) derives a similar supply-side gravity structure based on Ricardian pro-
ductivity draws from a Pareto distribution in which the dispersion parameter of the Pareto
distribution plays essentially the same role as 6 in the Eaton-Kortum model. For each firm,
changes in variable trade costs act on the intensive margin, but for the total sectoral
bilateral trade flow, these effects disappear and the aggregate effect is effectively on the
extensive margin of trade. Chaney's model includes a fixed cost of export for monopolisti-
cally competitive firms, and in equilibrium the elasticity of substitution affects the pattern
of trade by being part of the elasticity of equilibrium trade volume with respect to the
fixed cost.

4.3. Zeroes

In practice, many potential bilateral trade flows are not active. The data presented to the
analyst may record a zero that is a true zero or it may reflect shipments that fall below a

146 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
threshold above zero. In addition there may be missing observations that may or may not
reflect true zeroes. The prevalence of zeroes rises with disaggregation, so that in finely
grained data, a large majority of bilateral flows appear to be inactive. Finally, over time,
the small bilateral flows in finely disaggregated data appear to wink on and off. The zeroes
present two distinct issues for the analyst: appropriate specification of the economic model
and appropriate specification of the error term on which to base econometric inference.
Discussion of the specification of the error term is deferred to Section 5.
In specifying the economic model, zero trade flows present a problem for the CES/
Armington model of demand and the Eaton-Kortum supply-side structure. With elasticities
of substitution greater than one (or the equivalent dispersion/comparative advantage pa-
rameter restriction for the Eaton-Kortum model), the empirically relevant case, some
volume will be purchased no matter how high the price. One way to rationalize zeroes is
to modify the demand specification so as to allow choke prices above which all demand is
choked off. A start is made by Novy (2010), who derives gravity in a highly restricted one-
slope parameter translog expenditure function case that allows for zeroes in demand.4
More general translog treatments are feasible and desirable. Anderson & Neary (2005)
present a general homothetic preferences structure, showing that multilateral resistance is
defined and solved from a similar equation system once the functional form and its param-
eters are specified, along with data on shipment and expenditure shares.
An alternative economic specification explanation retains CES/Armington preferences
and rationalizes zeroes as resulting from fixed costs of export facing monopolistic compet-
itive firms. If no firm in / is productive enough to make incurring the fixed cost of exporting
to / profitable (given the cost of production in /, variable trade cost ř/;, and willingness to
pay in /), then zero trade results. Helpman, Melitz, and Rubinstein (HMR; Helpman et al.
2008) develop this idea. The selection effect determines which markets are active and also
determines a volume effect Vz/ due to productivity heterogeneity among firms whereby
markets that are active have a greater or lesser number of firms active depending on the
same selection mechanism. The gravity model becomes

EkYk ( th V "k
Xk - 1 ' Vk 11
yk i i I pknk J

( tk V °k VkEk
= E ( tk V' VkEk

( tk V Gk Vk Yk

if I'- = E (fji ( tk J -yr- Vk Yk

HMR report results suggesting that this mechanism is indeed potent and that inference
without accounting for it biases estimates of the variable trade costs downward.
The key mechanism is a Pareto productivity distribution of potential trading firms. The
Pareto distribution is capable of capturing the empirical observation that the largest and
most productive firms export the most and to the most destinations. The Pareto distribu-
tion allows a tractable estimation procedure that requires only aggregate bilateral trade

4Novy's aggregate bilateral OECD trade flow data contain no zeroes, so this feature is not exploited yet.

www.annualreviews.org • The Gravity Model 147

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 data, an important advantage because firm-level trade data are not widely available. In
practice, identification of the parameters in estimating the HMR model requires a plausible
exclusion restriction - a proxy for the fixed cost of export that is not also a proxy for the
variable cost of trade. HMR use common religion, a specification that many find dubious.
An important challenge for the future is combining the HMR mechanism with the
translog expenditure system. This combination might be able to distinguish between fixed
export costs and choke prices as an explanation of zeroes.

4.4. Discrete Choice Structure

The third alternative model of structural gravity is based on modeling individual discrete
choice in a setting in which the individual trader faces costs or receives benefits not
observable to the econometrician. Of all possible bilateral pairs, the trader chooses one
because it yields the greatest gain. A population of such traders has observable character-
istics such as bilateral distance that condition the probability of each choice; the econome-
trician observes the resulting masses allocated and uses a probability model to structure
statistical inference. An early attempt along these lines was made by Savage & Deutsch
(1960) and followed by Learner 8c Stern (1970). Several problems with the model limited
its appeal. It did not offer a rationale for the linear homogeneity of the mass variables in
gravity, and its characterization of cross effects did not have a sound rationale.
Discrete choice modeling was greatly advanced by McFadden (1974), who proved that
under plausible restrictions in this setting (the random variable, to the econometrician,
results in the observed choices following the type 1 extreme value distribution), the
resulting probability model is the multinomial logit. Building on the multinomial logit, it
is easy to generate a structural gravity model. This reasoning has rationalized recent work
on models of migration (e.g., Grogger Sc Hanson 2008, Beine et al. 2009).
It is straightforward to combine the discrete choice setup with the market clearance
conditions to derive the buyers' and sellers' incidence of trade costs exactly as in the
preceding models. The development is postponed to the next section, but is noted here
because exactly the same reasoning applies to goods traders making discrete choices on
where to sell or buy their goods. Thus the discrete choice probability model rationalizes
structural gravity equally well. It may be fruitful to explore the applicability of two-sided
matching models in the trade context as well as the job-market context.

5. ESTIMATION

As an empirical model, gravity is fundamentally about inferring trade costs in a settin

which much of what impedes trade is not observable to the econometrician. Wha
observable are the trade flows and a set of proxies for various types of trade costs,
with direct measures of some components of trade costs. Most issues with modeling tr
costs are discussed in Anderson & van Wincoop (2004). Since that time, there have b
several notable advances in modeling and inferring trade costs.
Two of the advances deal with the implications of zeroes in the bilateral trade flow d
One view of zeroes is that they stand for flows too small to report, an interpretation
indeed represents reporting practices of government trade ministries. Interpreting zero
this way, it is legitimate to drop the zero observations from estimation because there i
economic significance to the zeroes relative to the nonzero observations.

148 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 In the presence of heteroskedastic errors, Santos Silva ÔC Tenreyro (2006) point out that
inconsistent estimation arises from the usual econometric gravity practice using logarith-
mic transforms of Equation 5 augmented with a normal disturbance term and estimated
with ordinary least squares (OLS). Because the data have a lot of zeroes, the disturbance
term must have a substantial mass at very small values, violating the normal distribution
assumption. They propose instead to model the disturbance term as generated from a
Poisson distribution, leading to estimation with a Poisson pseudo-maximum likelihood
(PPML) technique. Their results show that PPML leads to smaller estimates of trade costs
compared to OLS.
The heteroskedastic error problem identified by Santos Silva and Tenreyro is important,
but their solution has not convinced all researchers. Martin ÔC Pham (2008) argue, based
on Monte Carlo simulations, that when heteroskedasticity is properly controlled, Tobit
estimators outperform PPML when zeroes are common. Heteroskedasticity is likely to be
attenuated using size-adjusted trade Xi;/Y/E; as the dependent variable, as advocated by
Anderson &c van Wincoop (2003, 2004).
An alternative view of zeroes, encountered above, is that economically meaningful
selection generates the zeroes. All firms in origin / face fixed costs of entering exporting to
any particular destination /, and only the sufficiently productive ones can afford to pay the
fixed cost. When a destination ; is so expensive to reach that no firm in i can afford the
fixed cost, zeroes are generated in the data. In this case, OLS estimation without account-
ing for selection is biased for two reasons: the standard left censored selection reason and
because, for bilateral pairs with positive flows, of a volume effect due to selection of firms
along with the bilateral trade cost t¿¡ that is the object of investigation in OLS or PPML
estimation. HMR find that their technique also results in lower cost estimates than does
OLS. (They report that estimation with Poisson error terms as opposed to normal ones
does not alter their findings.)
In principle, an economic model of zeroes is attractive, but many researchers are
suspicious of the exclusion restriction used by HMR to identify their volume effect. They
assume that common religion affects the fixed cost of export but not the variable cost t¡¡.
Moreover, the tractability of the HMR model depends on a restrictive distributional
assumption on the productivity draws distribution of firms, which in turn is a specializa-
tion of a particular model of monopolistic competition that is not applicable to all sectors.
Anderson Yotov (2010b) report that estimation with PPML, HMR, or OLS leads to
essentially identical results for buyers' and sellers' resistance and CHB because it leads to
gravity coefficients that are almost perfectly correlated. The homogeneity property of Equa-
tions 5-7 implies that only relative trade costs can be inferred by gravity; hence the differ-
ences in techniques effectively amount to different implicit normalizations. Anderson &;
Yotov report this near-perfect correlation finding based on estimation with the three tech-
niques over 18 three-digit manufacturing sectors, 76 countries, and 13 years of data.

5.1. Traditional

Some researchers continue to use a traditional form of the gravity model, presumably in the
belief that the structural model featured above is not sufficiently well established. It seems
useful to review a generic traditional model along with my objections.
A typical traditional gravity model regresses the log of bilateral trade on log trade costs
proxied by a vector of bilateral variables that are not at issue here, log GDP for origin and

www.annualreviews.org • The Gravity Model 149

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 destination, and log population for origin and destination. In addition, a number of
authors include remoteness indexes of each country's distance from its partners, atheoretic
measures that are inadequate attempts to control for multilateral resistance. [Anderson &
van Wincoop (2003) report significant differences between gravity estimated with remote-
ness and with multilateral resistance.]
The first objection to the traditional model is its aggregation, which causes two prob-
lems. There is aggregation bias because of sectorally varying trade costs and sectorally
varying elasticities of trade with respect to costs (see Anderson van Wincoop 2004 for
analysis and Anderson & Yotov 2010a, b for evidence on downward bias). The second
aggregation problem is specification bias because GDP is a value-added concept with a
variable relationship to gross trade flows. Much recent attention to the vertical disintegra-
tion of production and its international aspect emphasizes the variable intertemporal
relationship of gross trade to GDP, and its variation across countries is also significant.
Disaggregation and use of the appropriate sectoral output and expenditure variables fix
both problems.
The second objection is omitted variable bias from the perspective of the structural
gravity model - the traditional model leaves out multilateral resistance. Multilateral
resistance has only low correlation with remoteness indexes, and the omitted variable
will be correlated with the other right-hand-side variables and thus bias estimation.
The traditional model's inclusion of mass variables such as GDP and population
presumably picks up a part of the missing explanatory power of multilateral resistance,
as Anderson and Yotov's work shows that multilateral resistance is associated with
country size. Estimation with country fixed effects controls appropriately for all these
issues.

5.2. Structural

Anderson &c van Wincoop (2003) combine Equations 6 and 7 with the stochastic version
of Equation 5 to form a full information estimator of the coefficients of the proxies for
trade costs such as distance and international borders. Utilizing the unitary elasticities on
the E's and Y's, their dependent variable is X;,/ Y¡Eh size-adjusted trade.
An alternative fixed effects estimator controls for the unobservable multilateral resis-
tances and activity variables:

X¡¡ = (13)

where £¡¡ is the random err

is the fixed effect for cou
information estimator bu
(2004) argues for the fixe
however, another and per
there may be other countr
full information estimatio
A major drawback to fixe
metrician blows up the bui
trade costs. Fortunately, in
building like an archeologi
Thus Anderson &; Yotov (2

i jo Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
form, but then estimate the multilateral resistances by calculating the fitted ř;/s and plug-
ging them into Equations 6 and 7.
This technique is used to test the structural gravity model by comparing the estimates of
fixed effects (ximj) with the structural gravity term ( YjE/Il^Pp1 ). The results are remark-
ably close in an economic sense (the fitted regression line has an estimated elasticity around
0.96, compared with the theoretical value of 1.0) across 76 countries and 18 manufactur-
ing sectors over 13 years. Although this result suggests that the constraints that legitimize
full information methods are very close to being valid, fixed effects estimation still seems
the better, more cautious practice to follow.
Baier Sc Bergstrand (2009) propose an alternative direct estimator of multilateral resis-
tance based on a Taylor's series approximation of Equation 5. They report reasonably good
results, but I suspect that many researchers will be wary of the approximation error. In
contrast, the method of Anderson and Yotov avoids approximation error. As Baier Sc
Bergstrand emphasize, the advantage of their method relative to panel estimation with
fixed effects is that it avoids the upper bound on the number of fixed effects imposed by
typical econometric packages at the time of this writing. (STATA currently imposes a limit
of 11,000 independent variables, whereas 100 countries over 10 years require approxi-
mately 200,000 fixed effects, and even yearly estimation requires 20,000.) In principle, the
Baier Sc Bergstrand estimator could be used to construct trf s and then could be combined
with data on the Ts and E's using Equations 6 and 7 to obtain the incidence measures and
perform comparative statics with them. The constructed multilateral resistances in Baier Sc
Bergstrand's method can be compared to the point estimates, differences being attributed
to random error and approximation error.

5.3. Foreign Affiliate Sales

A large share of international trade comprises sales by foreign affiliates of multinational
firms. Standard trade gravity models include this trade along with that of domestically
owned firms. If the trade costs are the same for both types of firms, this treatment is
entirely appropriate.
There is reason to believe, however, that the trade cost structure facing foreign affiliate
sales differs from that facing domestic firms. For trade in intermediate inputs, information
and other transactions costs are reduced for intrafirm trade, but even for horizontal trade
there are likely to be transactions cost advantages when a foreign affiliate sells into its
home country. This reasoning suggests that one should split the home and foreign firms
into separate sectors for more accurate and informative inference about trade costs.
This approach to gravity with multinationals follows the conditional general equilib-
rium strategy, treating total sales as exogenous. It avoids taking a stand on determinants of
the location of production. A significant literature that is at least loosely related to gravity
attempts to explain this location decision along with the volume of foreign affiliate sales. It
is treated below in the discussion of foreign direct investment (FDI).

6. GRAVITY AND FACTOR FLOWS

Gravity has long been applied to empirically model factor movements. As with trade flows,
the model always fits well. But, in contrast to the recent development of an economic
structural gravity model of trade, there has been little progress in building a theoretical

www.annualreviews.org • The Gravity Model 151

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 foundation. This section sets out a structural model of migration, reviews promising steps
toward a structural model of FDI, and closes by pointing to the unsolved puzzle of
modeling international portfolio capital movements.

6.1. Migration
The decision to migrate is a discrete choice from a menu of locations. Each worker that
migrates faces a flow cost common to all workers who migrate in a particular bilateral
link, but each worker also has an idiosyncratic component of cost or utility from the move.
We may think of an idiosyncratic cost component as plausibly associated with a fixed cost,
but in the migration decision, the distinction between fixed and variable cost plays no
important role because the decision to migrate has no volume decision accompanying it.
This stands in contrast to the export selection model of Helpman et al. (2008) in which the
decision to export and the decision of how much to export are distinct.
Let wl denote the wage at location /,Vi. The worker h who migrates from origin / to
destination / faces a cost of migration modeled with iceberg cost factor ò]i > 1, receiving net
wage (wl/&1). Worker h's idiosyncratic utility from migration is represented by e!ih, private
information to him. He chooses to migrate if (wl /$l)ďh > for at least some /. Among
alternative destinations, he chooses the one with the largest surplus. Suppose that the worker
has logarithmic utility. Then his observable component of utility of migration from j to i is
= In it/ - In 3}t - In wL In this sort of setting, McFadden (1974) shows that if In s had the
type 1 extreme value distribution, the probability that a randomly drawn individual would
pick any particular migration destination is given by the multinomial logit form.
Building on this insight, migration models subsequently used the multinomial logit to
model bilateral migration flows (for two recent examples, see Beine et al. 2009, Grogger 8c
Hanson 2008). This section develops a novel gravity model representation of the migration
model by making use of the market-clearing conditions to derive the appropriate multilat-
eral resistance variables.
At the aggregate level, the probability is equal to the proportion of migrants from j
(assumed to be identical except for their values of s) that pick destination i. Let N; denote
the population of natives of /. The predicted migration flow from / to / that results from the
setup is

M" = G(uji)N' (14)

where

}2k exp (ut*)

With logarithmic utility, the migration equation is

M" - N'. (15)

ZkuA/Vk

Equation 15 is a structure analogous to the CES demand (in the Armington model) or
Ricardian supply (in the Eaton-Kortum model) shares that underpin the trade gravity
equation. The connection of the share equation (Equation 15) to the structural gravity
form of the model is completed by using the labor market balance equations to solve for
and substitute out the equilibrium m/s.

152 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 Define W = wk/ô'k and define the labor force supplied to i from all origins:

V = M"- (16)
/

Also, N = ^2 J N; = ^ i L ', the world labor supply N. The labor market clearance equa-
tion is

L' = w<Y, ((! /^)/W')Ni.

i

Then

w> = -, (17)
Si'N

where

= l/^Ņ/
I
¿.s W N y '

Using Equation 17 to substitute for the wage in W7,

(19)
Y N
Substituting for the wage in Equation 15 using Equation 17
equation of migration:

(20)
N Q'W/

The first ratio represents the migration pattern of a frictionless world. The implica-
tion is that in a frictionless world, populations originating in / would be found in equal
proportions to their share of world population in all destinations: Mn /V = N' /N.
The second term represents the effect of migration frictions. The bilateral migration
friction ô;í reduces migration. It is divided by the product of the weighted averages of
the inverse of migration frictions, one for inward migration to i from all origins and one
for outward migration from / to all destinations. The system given in Equations 18 and
19 can be solved for the Q's and Ws (subject to a normalization). Their interpretation
and their connection to multilateral resistance in the more familiar trade gravity model
are easier to see in the case in which utility is generalized to the log of a constant
relative risk aversion function.5
Let the coefficient of relative risk aversion be 0. In this case Equation 20 becomes

N 'q-WJ '

5A tractable gravity equation results from Equation 14 by using a restriction on utility to convert exp«7' into a
tractable form. When utility is given by the log of any power function of the wage net of migration costs, the CES-
type form of gravity results, with consequent ease of estimation and resemblance to the trade flow structural gravity
model.

www.annualreviews.org • The Gravity Model 153

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 where

Q'= [v- (y)1-8^]17'1""'

y w' N
and

. r tx/i'l-0 ( ' r <1 1/(l"fl)

W'= . T tx/i'l-0 ( ' - r
ļr ñ' - N_
Here, fi' and W' are CES price indexes of migration frictions, one for inward (fi') and
one for outward (WJ) migration frictions. These equations are exactly analogous to the
Anderson and van Wincoop model's inward and outward multilateral resistance equations
for trade, but applied to migration. As with the trade gravity model, outward multilateral
resistance gives the sellers' incidence of the migration costs on average while the inward
multilateral resistance gives the buyers' incidence of migration costs. Equations 18-20
result from the special case 6 = 2.
Q and W are general equilibrium concepts as is clear because their solution in the
simultaneous systems above involves every bilateral migration cost in the world. They are
conditional general equilibrium concepts because the L's are endogenous in a full general
equilibrium. It is possible in a Ricardian production setting to combine the migration
system with the trade gravity model to derive equilibrium labor supplies that are functions
of the incidence of both migration frictions and trade frictions.
As with trade gravity models, Q's and W's can be computed once the <5's are economet-
rically constructed and the labor supplies V and population stocks N7 are observed.
A normalization is needed (see Anderson &; Yotov 2010a for details).
A similar model has been applied to services trade by Head et al. (2009). Instead of
actually changing locations, the foreign worker does the job in her home location. The cost
of migration becomes the cost of monitoring the distant worker. Worker productivities in
each location have the Fréchet distribution, as in the Eaton-Kortum model. The firm selects
workers so as to minimize the log of the delivered unit labor cost. Then the distribution of
log productivities takes the Gumbel form. The fraction of service jobs in origin i going to
workers in location / has the multinomial logit form. The total number of workers and of
jobs in each location enters the model in the same way as in the migration model. I suspect
that the choice between off shoring the service job and migrating the worker can be fruit-
fully addressed with some combination of the two models.
The preceding treatment applies to a stationary equilibrium in which the L's are the
result of M's fully adjusting labor supplied at each location to its equilibrium value given
the initial stocks of labor {N7} and the set of migration frictions, the <5's. In adapting the
model to fit actual data, the N's, L's, and M's are observed at points in time, and with panel
data, the observations are linked over time.
If the sequence of observations is regarded as reaching the static equilibrium each
period, the observed migration is just that amount needed to reach the equilibrium in each
period. This model would be consistent with naive expectations about future wages, or
with a pure guest worker model in which migration is determined by contemporaneous
variables only. So, in principle, under this interpretation, the preceding model could be
applied at each date, all variables now having a time subscript.

IJ4 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 The alternative is a dynamic model in which the migrants form expectations about the
sequence of future wages based on underlying expectations about the future evolution of
the distribution of trade frictions, the populations, and, as shown below following the
development of the integrated trade and migration model, variables that predict the
demand for labor at all locations. This sophistication requires a big increase in complexity,
with dubious applicability of rational expectations to unskilled workers.
The other issue raised by thinking of dynamics is the issue of partial adjustment -
migration in any one year may not suffice to reach the static equilibrium of the preceding
section. In this case, the standard ad hoc approach of partial adjustment due to quadratic
adjustment costs might be applied without too large an increase in complexity.

6.2. Foreign Direct Investment

FDI has been successfully explained by gravity structures without a theoretical foundation.
More recent work has made progress on foundations. Satisfactory foundations are more
difficult to find for at least two reasons. First, the question of the location of production
must be answered in an upper-level general equilibrium model, which requires taking a
stand on one of many possible production, preference, and market structures restricted so
as to produce tractable results. Second, the determinants of location depend on whether the
good in question is vertically or horizontally linked to other sources of firm profits.
A key element in explaining the location of horizontally linked production is the
proximity-concentration tradeoff: A firm with fixed cost reduces per-unit production cost
by concentrating production at one location but can save distribution costs by allocating
production in proximity to markets. Even under strong restrictions, the models obtained so
far are nonlinear and require approximation to be taken to data.
Helpman et al. (2004) model interaction between horizontally linked exports and
foreign affiliate sales, in which the firm chooses between exporting from home or investing
abroad and selling from a foreign plant. They are able to draw inferences from aggregate
data by modeling heterogeneous productivity of firms with a Pareto distribution. Fixed
costs of export and of investing abroad serve to select firms into nontraders, exporters, and
multinationals with ratios that vary market by market because of trade costs modeled as
transport costs and tariffs only, omitting the usual gravity variables. Their empirical appli-
cation with U.S. data obtains fairly good results in explaining the ratio of exports to
foreign affiliate sales with a linear approximation to their underlying nonlinear model.
The model fits much less well than standard export gravity equations, which is not surpris-
ing because the dependent variable is different and the question addressed is more difficult
to answer.

Kleinert &c Toubal (2010) extend Helpman et al. (2004) to allow for fixed setup co
that rise with distance, a wrinkle that can explain why foreign affiliate sales can fall r
than rise with distance, as the earlier proximity-concentration tradeoff suggested.
also derive a gravity-type relationship from two other structures, a vertical integr
model and a two-country factor proportions model of fragmentation.
Bergstrand &c Egger (2007) offer a gravity model of FDI derived from the knowledg
based capital theory of horizontal multinational enterprises. Their objective is a
general equilibrium model that can explain trade, foreign affiliate sales, and FDI. T
simulate a theoretical model that generates nonlinear relationships between exports, af
iate sales, and their exogenous determinants. Then they fit an approximate empi

www.annualreviews.org • The Gravity Model ijj

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 relationship to the generated data and take the same relationship to actual data, with some
success. A limitation of their model is that, although the factor proportions model with
three factors is used to explain simultaneous exports and affiliate sales, the countries in
their simulation setup have identical endowment proportions and differ only in size.
Keller Sc Yeaple (2009) develop a gravity model of vertically integrated intrafirm trade
featuring trade costs with two elements, a standard iceberg trade cost and a communica-
tion cost that rises with the complexity of the firm's technology. Input complexity raises
technology transfer costs, whereas the costs of embodied technology transfer are indepen-
dent of complexity and are increasing in trade costs. An increase in trade costs reduces
foreign affiliate sales, and this effect is strongest in the most complex sectors. In contrast,
an increase in trade costs reduces the imports of foreign affiliates, and this effect is weakest
in the most complex sectors. Like the standard trade gravity model, Keller Sc Yeaple's
model of foreign affiliate sales permits inference about trade costs from observable trade
flows.
Keller Sc Yeaple report fairly good results estimating the model using confidential data
on U.S. multinational firm activity from the Bureau of Economic Analysis. The role played
by communication cost interacting with technological complexity thus appears likely to be
helpful in explaining the rising share (in total trade) of intrafirm trade and also the rising
share of trade in intermediates.
An alternative strategy along the lines of the conditional general equilibrium approach
outlined above for migration appears useful. The migration decision model of Section 2
could apply to FDI because the location decision for a plant is similar to the location
decision of a migrant. (Unlike migration but like trade, FDI involves a volume decision
along with a participation decision.) The rate of return on investment could be taken as
exogenous in a conditional general equilibrium approach just as wages are taken as exog-
enous in the migration gravity model, while market-clearing conditions apply just as in the
migration model. Idiosyncratic cost factors would apply to the various investment projects,
just as they do to the individual migrants. The Keller- Yeaple model of vertically integrated
intrafirm trade offers a structure for identifying one type of cost. A weakness in the
extension by analogy is in risk diversification. Migrants cannot diversify their risks, but
firms can, although with limited possibilities that may be very limited for FDI. The poten-
tial risk diversification would modify the utility derived from each location choice. The
discrete choice approach faces truly formidable modeling challenges in endogenizing the
investment rates of return, unlike the wage equation suggested by the migration model.
A promising start on these lines is by Head Sc Ries (2008). Potential acquisitions go to
the highest bidder, who bids based on his anticipated return net of monitoring costs that
rise with distance and other standard gravity variables. The probability of the winning bid
going to source country i takes the multinomial logit form. The mass variables are the
stocks of projects in each host country and each source country's share of world bidders.

6.3. Portfolio Investment

Martin Sc Rey (2004) offer the first gravity-type model of international portfolio invest-
ment. The coefficient of relative risk aversion plays the role, in equilibrium, of the elasticity
of substitution in the CES demand specification. Although appealing as a rationale for the
gravity application of Portes Sc Rey (2005), the Martin Sc Rey model does not provide a
fully satisfactory foundation for gravity models of investment flows because (a) trade is

ij6 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
assumed to be frictionless, (b) investment costs are uniform, and, most importantly, (c) the
analysis is restricted to two countries. The third-party effects that play a big role in the
gravity model of trade (and of migration) cannot be treated.

7. INTEGRATED SUPERSTRUCTURE

The gravity model nests inside a general equilibrium superstructure. As pointed out by
Anderson & van Wincoop (2004), modularity implies that the problem of resource
and expenditure allocation across sectors in the general equilibrium superstructure can
be treated separably from the gravity module problem of distribution within sectors to
destinations or from origins. Consistency between the two levels of the problem requires
fixed-point calculations in general, but the economy of thought and computation due to
separability is extremely useful and, in particular, makes it possible to integrate gravity
with a wide class of general equilibrium production models. So far, only simple production
models have been used for full general equilibrium comparative statics, but I anticipate
that this situation will change.
The simplest production structure is an endowments economy. Anderson & van
Wincoop (2003) use the endowments model to calculate the effect of eradicating the U.S.-
Canada border on their estimated gravity model of trade between U.S. states, Canadian
provinces, and the aggregated rest of the world.
Another attractive candidate is the Ricardian production model. Eaton & Kortum
(2002) nest gravity inside a Ricardian model of production, a choice followed by a host of
subsequent researchers such as Arkolakis (2008). An important feature of these models is
the action on the extensive margin, as industries arise or disappear. In the 2002 Eaton-
Kortum model, the extensive margin is the only margin. Arkolakis and others have variants
in which both extensive and intensive margins are active. This is an important feature
because disaggregated trade data and especially firm-level data indicate that both margins
are active.

Between the two extremes of zero and infinite elasticity of transformation of the
endowments and Ricardian models lie a host of more complex production structures in
which action occurs on the intensive margin of production when relative prices change,
leading to another channel of interaction between the gravity modules in each sector
(and resulting buyers' and sellers' incidences) and the pattern of production. Consistency
between the modules is achieved by using Equation 9 to normalize the ITs in each
sector. I think the future will see work with these more complex general equilibrium
features.
Migration of labor and capital in the form of FDI has been given a complete gravity
representation in this review. In the integrated superstructure, it can be treated simulta-
neously with the trade modules. In this setting, multilateral resistance in trade has signifi-
cant effects on migration and vice versa. I anticipate that development of this link will be
useful.

A number of authors have constructed integrated models that motivate econometric

work aimed at discriminating between one or another specification of the upper-level
production and market structure. A summary of work along these lines is presented
in Feenstra (2004, chapter 5), where the main focus is on the link between gravity
and increasing returns to scale. Research has continued on these lines, but I do not
review it here.

www.annualreviews.org • The Gravity Model 157

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 I think the gravity model is a poor vehicle for inferences about returns to scale, market
structure, and the global general equilibrium links between economies. This review and my
previous work argue that gravity is about the distribution of given amounts of goods in
each origin drawn by given amounts of expenditure in each destination, enabling inference
about trade costs from the deviation of observed distribution from the frictionless equilib-
rium. The determinants of total shipments and total expenditures are irrelevant to this
inference because country fixed effects are a consistent control that does not require taking
a stand on any particular production or market structure model. Conversely, the cross-
section variation of bilateral trade does not seem likely to have much useful information
about the determination of national total shipments or expenditure. Interdependence is so
deeply wound between these variables in the full general equilibrium model that inference
about structure seems implausible. In contrast, simulation models look reasonably prom-
ising as a source of insight.

8. CONCLUSION

This idiosyncratic review of work on the gravity model suggests that the story is not ov
so a conclusion can only point to potential future chapters. Distribution broadly def
consumes a very large share of the world's resources, and gravity has proven to be the m
generally useful empirical model for understanding the distribution of goods and factor
production. It appears to work well at almost any scale.
The progress in structural modeling of gravity has yielded three distinct rationales f
the same observationally equivalent model of the distribution of economic flows betw
origins and destinations, one based on the demand side (the CES/Armington model),
based on the supply side (the Eaton-Kortum model), and one based on a discrete cho
model of the individual actor transferring the goods or factors. Further work may sug
ways to discriminate between these.
The structural modeling of gravity imposes trade separability, permitting gravity m
ules to be nested inside a wide range of general equilibrium superstructures. Future w
with simulation models may suggest which of many candidate general equilibrium prod
tion models do better.
The problem of zeroes in the trade and factor flows data has been addressed with som
success, particularly by HMR. But I expect future work to do better. The CES framew
(with elasticity of substitution greater than one) is unsuitable for describing small amo
of trade. The translog cost function, in particular, seems likely to yield better descript
and better understanding of why so many potential flows are equal to zero. This is so e
if, as in HMR, fixed export costs play an important role in selecting firms to export.
Featured in this review are incidence measures produced by Anderson and Yotov. If t
profession agrees that they are as interesting and useful as they appear to me, more wor
needed to see how believable the measures are. As it stands, they are completely reliant
CES structure. How well does the CES do in representing the world economy? This is
especially important question in light of the zeroes question in the preceding paragr
I look forward to the development of the translog case to help answer this question.

DISCLOSURE STATEMENT

The author is not aware of any affiliations, memberships, funding, or financial holdi
that might be perceived as affecting the objectivity of this review.

15 8 Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
ACKNOWLEDGMENTS

I thank Jeffrey Bergstrand, Keith Head, and Yoto Yotov for helpful comments on earlier
versions of this review.

LITERATURE CITED

Anderson JE. 1979. A theoretical foundation for the gravity equation. Am. Econ. Rev. 69:106-16
Anderson JE, Neary JP. 2005. Measuring the Restrictiveness of International Trade Policy. Cambridge:
MIT Press

Anderson JE, van Wincoop E. 2003. Gravity with gravitas. Am. Econ. Rev. 93:170-92
Anderson JE, van Wincoop E. 2004. Trade costs./. Econ. Lit. 42:691-751
Anderson JE, Yotov YV. 2010a. The changing incidence of geography. Am. Econ. Rev. 100:2157-86
Anderson JE, Yotov YV. 2010b. Specialization: pro- and anti-glob alizing, 1990-2002. NBER Work.
Pap. 16301
Arkolakis C. 2008. Market penetration costs and the new consumers margin in international trade.
NBER Work. Pap. 14214
Baier SL, Bergstrand JH. 2001. The growth of world trade: tariffs, transport costs, and income
similarity./. Int. Econ. 53(1): 1-27
Baier SL, Bergstrand JH. 2009. Bonus vetus OLS: a simple method for approximating international
trade-cost effects using the gravity equation./. Int. Econ. 77(l):77:77-85
Beine M, Docquier F, Ozden C. 2009. Diasporas. Policy Res. Work. Pap. 4984, World Bank,
Washington, DC
Bergstrand JH. 1985. The gravity equation in international trade: some microeconomic foundations
and empirical evidence. Rev. Econ. Stat. 67(3):474-81
Bergstrand JH. 1989. The generalized gravity equation, monopolistic competition, and the factor-
proportions theory in international trade. Rev. Econ. Stat. 71(1): 143-53
Bergstrand JH, Egger P. 2007. A knowledge- and physical-capital model of international trade flows,
foreign direct investment and multinational enterprises./. Int. Econ. 73:278-308
Bergstrand JH, Egger P. 2011. Gravity equations and economic frictions in the world economy. In
Palgrave Handbook of International Trade , ed. D Bernhofen, R Falvey, D Greenaway,
U Krieckemeier. New York: Palgrave-Macmillan. In press
Chaney T. 2008. Distorted gravity: the intensive and extensive margins of international trade.
Am. Econ. Rev. 98:1707-21

Costinot A, Donaldson D, Komunjer I. 2010. What goods do countries trade ? A quantitative explora-
tion of Ricar do's ideas. Work. Pap., MIT
Deardorff A. 1980. The general validity of the law of comparative advantage. J. Polit. Econ.
88:941-57

Eaton J, Kortum S. 2002. Technology, geography, and trade. Econometrica 70:1741-79

Feenstra R. 2004. Advanced International Trade: Theory and Evidence. Princeton, NJ: Princeton Univ.
Press

Grogger J, Hanson GH. 2008. Income maximization and the selection and sorting of international
migrants. NBER Work. Pap. 13821
Head K, Mayer T, Ries J. 2009. How remote is the off-shoring threat? Eur. Econ. Rev. 53:429-44
Head K, Ries J. 2008. FDI as an outcome of the market for corporate control: theory and evidence.
/. Int. Econ. 74:2-20
Helpman E, Melitz MJ, Rubinstein Y. 2008. Estimating trade flows: trading partners and trading
volumes. Q.J. Econ. 123:441-87
Helpman E, Melitz MJ, Yeaple SR. 2004. Export versus FDI with heterogeneous firms. Am. Econ.
Rev. 94:300-1 6

Keller W, Yeaple SR. 2009. Gravity in the weightless economy. NBER Work. Pap. 15509

www.annualreviews.org • The Gravity Model 159

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms
 Kleinert J, Toubal F. 2010. Gravity for FDI. Rev. Int. Econ. 1 8( 1 ): 1- 1 3
Learner EE, Levinsohn J. 1995. International trade theory: the evidence. In Handbook of International
Economics , Vol. 3, ed. GM Grossman, K Rogoff, pp. 1339-94. Amsterdam: Elsevier Sci. B.V.
Learner EE, Stern RM. 1970. Quantitative International Economics. Chicago: Aldine
Martin W, Pham CS. 2008. Estimating the gravity equation when zero trade flows are frequent.
MPRA Pap. 9453, Univ. Libr. Munich, Ger.
Martin P, Rey H. 2004. Financial super-markets: Size matters for asset trade. /. Int. Econ. 64:335-61
Matsuyama K. 2007. Beyond icebergs: towards a theory or biased globalization. Rev. tcon. Stud.
74:237 -53

McCallum J. 1995. National borders matter: Canada-U.S. regional trade patterns. Am. Econ. Rev.
85(3):615-23
McFadden D. 1974. Conditional logit analysis of qualitative choice behavior. In Frontiers of Econo-
metrics, ed. P Zarembka, pp. 105-42. New York: Academic
Novy D. 2010. International trade and monopolistic competition without CES: estimating translog
gravity. Unpublished manuscript, Warwick Univ.
Portes R, Rey H. 2005. The determinants of cross-border equity flows. J. Int. Econ. 65:269-96
Ravenstein EG. 1889. The laws of migration./. R. Stat. Soc. 52(2):241-305
Redding S, Venables AJ. 2004. Economic geography and international inequality. J. Int. Econ.
62( 1 ):53- 82
Samuelson PA. 1952. The transfer problem and transport costs: the terms of trade when impediments
are absent. Econ. ]. 62:278-304
Santos Silva J, Tenreyro S. 2006. The log of gravity. Rev. Econ. Stat. 88(4):641-58
Savage IR, Deutsch KW. 1960. A statistical model of the gross analysis of transactions flows.
Econometrica 28:551-72

Tinbergen J. 1962. Shaping the World Economy: Suggestions for an International Economic Policy.
New York: Twentieth-Century Fund

i6o Anderson

This content downloaded from

130.56.64.101 on Fri, 12 Apr 2024 22:26:23 +00:00
All use subject to https://about.jstor.org/terms