See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228303962

The Gravity Model

Article in Annual Review of Economics · December 2010

CITATIONS READS
13 9,786

1 author:

James E. Anderson
Boston College
135 PUBLICATIONS 17,826 CITATIONS

SEE PROFILE

All content following this page was uploaded by James E. Anderson on 16 April 2014.

The user has requested enhancement of the downloaded file.

 The Gravity Model∗
James E. Anderson
Boston College and NBER
January 18, 2011

Abstract
Gravity has long been one of the most successful empirical models in economics. In-
corporating deeper theoretical foundations of gravity into recent practice has led to
a richer and more accurate estimation and interpretation of the spatial relations de-
scribed by gravity. Wider acceptance has followed. Recent developments are reviewed
here and suggestions are made for promising future research.
JEL Classification: F10, R1.
Contact information: James E. Anderson, Department of Economics, Boston Col-
lege, Chestnut Hill, MA 02467, USA.

Keywords: Incidence, multilateral resistance, trade costs, migration.

∗
This review was prepared for Annual Review of Economics, vol. 3. I thank Jeffrey H. Bergstrand, Keith
Head, J. Peter Neary and Yoto V. Yotov for helpful comments.
 The gravity model in economics was until relatively recently an intellectual orphan, un-
connected to the rich family of economic theory. This review is a tale of the orphan’s reunion
with its heritage and the benefits that continue to flow from connections to more distant
relatives.
Gravity has long been one of the most successful empirical models in economics, order-
ing remarkably well the enormous observed variation in economic interaction across space
in both trade and factor movements. The good fit and relatively tight clustering of coeffi-
cient estimates in the vast empirical literature suggested that some underlying economic law
must be at work, but in the absence of an accepted connection to economic theory, most
economists ignored gravity. The authoritative survey of Leamer and Levinsohn (1995) cap-
tures the mid-90’s state of professional thinking: “These estimates of gravity have been both
singularly successful and singularly unsuccessful. They have produced some of the clearest
and most robust empirical findings in economics. But, paradoxically, they have had virtually
no effect on the subject of international economics. Textbooks continue to be written and
courses designed without any explicit references to distance, but with the very strange im-
plicit assumption that countries are both infinitely far apart and infinitely close, the former
referring to factors and the latter to commodities.” Subsequently, gravity first appeared in
textbooks in 2004 (Feenstra, 2004), following on success in connecting gravity to economic
theory, the subject of Section 3.
Reviews are not intended to be surveys. My take on the gravity model, thus licensed
to be idiosyncratic, scants or omits some topics that others have found important while it
emphasizes some topics that others have scanted. My emphases and omissions are intended
to guide the orphan to maturity. An adoptive parent’s biases may have contaminated my
judgment, caveat emptor.
My focus is on theory. Incorporating the theoretical foundations of gravity into recent
practice has led to richer and more accurate estimation and interpretation of the spatial
relations described by gravity, so where appropriate I will point out this benefit. The har-
vest reaped from empirical work applying the gravity model is recently surveyed elsewhere
(Anderson and van Wincoop, 2004; Bergstrand and Egger, 2011).
From a modeling standpoint, gravity is distinguished by its parsimonious and tractable
representation of economic interaction in a many country world. Most international economic
theory is concentrated on two country cases, occasionally extended to three country cases
with special features. The tractability of gravity in the many country case is due to its
modularity: the distribution of goods or factors across space is determined by gravity forces
conditional on the size of economic activities at each location. Modularity readily allows
for disaggregation by goods or regions at any scale and permits inference about trade costs
not dependent on any particular model of production and market structure in full general
equilibrium. The modularity theme recurs often below, but is missing from some other
prominent treatments of gravity in the literature.

1 Traditional Gravity

The story begins by setting out the traditional gravity model and noting clues to its union
with economic theory. The traditional gravity model drew on analogy with Newton’s Law of
Gravitation. A mass of goods or labor or other factors of production supplied at origin i, Yi ,
is attracted to a mass of demand for goods or labor at destination j, Ej , but the potential
flow is reduced by the distance between them, dij . Strictly applying the analogy,

Xij = Yi Ej /d2ij

gives the predicted movement of goods or labor between i and j, Xij . Ravenstein (1889)
pioneered the use of gravity for migration patterns in the 19th century UK. 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 statistically inferred log-linear relationship between

2
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 i’s other alternative destinations and by the resistance on move-
ment to j from j’s alternative sources of supply. Prodded by this intuition the traditional
gravity literature recently developed remoteness indexes of each country’s ‘average’ effective
P
distance to or from its partners ( i dij /Yi was commonly defined as the remoteness of coun-
try j) 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 destinations 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
2 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 homoge-
neous 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 pref-
erences 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 homogeneous
world, the natural benchmark prediction is that Xij /Ej = Yi /Y , the proportion of spending
by j on goods from i is equal to the global proportion of spending on goods from i, where
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 Yi , the total sales by origin i, and the sum of purchases
from all origins must equal Ej , the total expenditure for each destination j. Total sales and
P P
expenditures must be equal: i.e., i Yi = j Ej = Y .
One immediate payoff is an implication for inferring trade frictions. Multiplying both
sides of the frictionless benchmark prediction Xij /Ej = Yi /Y by Ej yields predicted friction-
less trade Yi Ej /Y . The ratio of observed trade Xij to predicted frictionless trade Yi Ej /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

4
the frictions to be analyzed below are likely to differ markedly by product characteristics.
The extension to disaggregated goods, indexed by k, is straightforward.

Yik Ejk
Xijk = = ski bkj Y k . (1)
Yk

Here ski = Yik /Y k is country i’s share of the world’s sales of goods class k and bkj = Ejk /Y k
is country j’s share of the world spending on k, equal to the world’s sales of k, Y k .
The notation and logic also readily apply to the disaggregation of countries into regions,
and indeed a prominent portion of the empirical literature has examined bilateral flows
between city pairs or regions, motivated by the observation that much economic interaction
is concentrated at very short distances. The model can interpreted to reflect individual
decisions aggregated with a probability model; see section 5.1 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 con-
ceptually 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.

1. Big producers have big market shares everywhere,

2. small sellers are more open in the sense of trading more with the rest of the world,

3. the world is more open the more similar in size and the more specialized the countries
are,

4. the world is more open the greater the number of countries, and

5
 5. world openness rises with convergence under the simplifying assumption of balanced
trade.

Implication 1, 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 /Ej = si .

Implication 2, that small sellers are more open in the sense of trading more with the rest of
the world follows from
X
Xij /Ej = 1 − Yj /Y = 1 − sj
i6=j
P P
using j Ej = i Yi , which implies balanced trade for the world.
Implication 3 is that the world is more open the more similar in size and the more special-
ized the countries are. It is convenient to define world openness as the ratio of international
shipments to total shipments, j i6=j Xij /Y . Dividing (1) through by Y k and suppressing
P P

the goods index k, world openness is given by

XX X X
Xij /Y = bj (1 − sj ) = 1 − b j sj .
j i6=j j j

Using standard statistical properties

X p
bj sj = N rbs V ar(s)V ar(b) + 1/N,
j

where N is the number of countries or regions, V ar denotes variance, rbs is the correlation
P P
coefficient between b and s and 1/N = i si /N = j bj /N , the average share.This equation
follows from the shares summing to one and using standard properties of covariance. Here,
V ar(s) and V ar(b) measures size dis-similarity and the correlation of s and b, rbs , is an

6
inverse measure of specialization. Substituting into the expression for world openness:

XX p
Xij /Y = 1 − 1/N − N rbs V ar(s)V ar(b) (2)
j i6=j

Implication 3 follows from equation (2) because on the right-hand side the similarity of
country size shrinks the variances while specialization shrinks the correlation rbs .
The country-size similarity property has been prominently stressed in the monopolistic
competition and trade literature. (It is sometimes taken as evidence for monopolistic com-
petition in a sector rather than as a consequence of gravity no matter what explains the
pattern of the b’s and s’s.) The specialization property has also been noted in that liter-
ature as reflecting forces that make for greater net international trade, the absolute value
of sj − bj . Making comparisons across goods classes, variation in the right-hand side of (2)
results from variation in specialization and in the dispersion of the shipment and expenditure
shares. Notice again that the cross-commodity variation in world openness arises here in a
frictionless world, a reminder that measures of world home bias in a world with frictions
must be evaluated relative to the frictionless world benchmark.
Country-size similarity also tends to increase bilateral trade between any pair of countries,
all else equal. This point (Bergstrand and Egger, 2007) is seen most clearly with aggregate
trade that is also balanced, hence sj = bj . Equation (1) can be rewritten as

(Yi + Yj )2
Xij = sij ij
i sj ,
Y

where sij ij ij
i ≡ Yi /(Yi + Yj ), the share of i in the joint GDP of i and j. The product si sj

is maximized at sij ij
i = sj = 1/2, so for given joint GDP size, bilateral trade is increasing

in country similarity. (With unbalanced trade or specialization, an analogous similarity

property holds for the bilateral similarity of income and expenditure shares. Let γj = Ej /Yj .
Then the same equation as before holds with the right-hand side multiplied by γj .)
A more novel implication of equation (2) is implication 4, that world openness is ordinarily

7
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.
P P P
This effect is seen by differentiating the left-hand side of j i6=j Xij /Y = 1 − j bj sj ,
P
yielding − j (bj dsj + sj 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 ‘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
p
hand, the change in the number of countries will usually change rbs V ar(b)V ar(s) in ways
that depend on the type of country division (or confederation) as well as indirect effects on
shares as prices change. (The apparent direct effect of N in the first term on the right-hand
p
side of equation (2) vanishes because 1/N scales V ar(b)V ar(s).)
A practical implication of this discussion is that inter-temporal comparisons of ratios of
world international trade to world income, to be economically meaningful, should be con-
trolled for changes in the size distribution and the number of countries, a correction of large
practical importance in the past 50 to 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 5, that world openness rises
with convergence under the simplifying assumption of balanced trade for each country, bj =
sj , ∀j. The right-hand side of equation (2) becomes N V ar(s) + 1/N under balanced trade,
and per-capita income convergence lowers V ar(s) toward the variance of population. Baier
and 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.

8
 Pointing toward a connection with economic theory, the shares si and bj and the plau-
sible hypothesis of the frictionless model must originate from an underlying structure of
preferences and technology. Also, the deviation of observed Xij from the frictionless pre-
diction 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.

3 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
the end user. Many of these costs are not directly observable, and the empirical gravity
literature indicates the total is well in excess of the transportation and insurance costs that
are observable (see Anderson and 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 expenditures
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.

9
 This standard description of general economic equilibrium is too complex to yield some-
thing 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 and 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 structures that work, one on the
demand side and one on the supply side, detailed in sections 3.1 and 3.2.
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 is 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. Samuel-
son (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 ice-
berg 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 costs are realistic and
potentially play an important role in explaining why many potential bilateral flows are equal

10
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) derived a joint cost function that is homogeneous of
degree one with Constant Elasticity of Transformation (CET). This setup allows for substi-
tution effects in costs between destinations rather than the cost independence due to 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 profit maximizing choice of destinations.
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 economet-
rically 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, con-
tiguity, membership of a country, continent or regional trade agreement, common language
and common legal traditions. See Anderson and 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 recent exploration of the
implications of non-iceberg trade costs in a 2 country Ricardian model. See Deardorff (1980)

11
for a very general treatment of the resource requirements of trade costs as a setting for his
demonstration that the law of comparative advantage holds quite generally.

3.1 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 technology are homothetic and weakly separable with respect to a par-
tition into classes whose members are 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 differentiated 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 very difficult to distinguish demand-side home bias from the effect of trade costs,
since the proxies used in the literature (common language, former colonial ties, or internal
trade dummies, etc.) plausibly pick up both demand and cost differences. Henceforth 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 nat-
ural price aggregate, with substitution between goods classes occurring as if the quantity
aggregates 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 non-steel products, though of course the aggregate purchase of steel is affected by
the aggregate cross effect. Homotheticity ensures that relative demands are functions only

12
of relative aggregate prices.
The first economic foundation for the gravity model was based on specifying the expendi-
ture function to be a Constant Elasticity of Substitution (CES) function (Anderson, 1979).
Expenditure shares in the CES case are given by

 1−σ
Xij βi pi tij
= (3)
Ej Pj

where Pj is the CES price index, σ is the elasticity of substitution parameter, βi is the
‘distribution parameter’ for varieties shipped from i, pi is their factory gate price and tij > 1
is the trade cost factor between origin i and destination j. The CES price index is given by

!1/(1−σ)
X
Pj = (βi pi tij )1−σ . (4)
i

Notice that the same parameters characterize expenditure behavior in all locations; prefer-
ences are common across the world by assumption. Notice also that the shares are invariant
to income, preferences are homothetic. With frictionless trade, tij = 1, ∀(i, j) and therefore
all the buyers’ shares of good i must equal the sellers share of world sales (at destination
prices), Yi /Y . Thus the frictionless benchmark is justified by assuming identical homothetic
preferences. For intermediate goods, the same logic works replacing expenditure shares with
cost shares.
The ‘distribution parameters’ βi bear several interpretations. They could be exogenous
taste parameters. Alternatively, in applications to monopolistically competitive products, βi
is proportional to the number of firms from i offering distinct varieties (Bergstrand, 1989).
Countries with more active firms get bigger weights. In long run monopolistic competition
the number of firms is endogenous. Due to fixed entry costs, bigger countries have more active
firms in equilibrium, all else equal. The number of active firms contributes to determining
the Yi ’s that are given in the gravity module.
The other building block in the structural gravity model is market clearance: at delivered

13
 P
prices Yi = j Xij . Multiplying both sides of (3) by Ej and summing over j yields a solution
for βi p1−σ
i ,
Yi
βi p1−σ
i =P 1−σ E
.
j (tij /Pj ) j

Define the denominator as Π1−σ

i .
Substituting into (3) and (4) yields the structural gravity model:

Ej Yi tij 1−σ
 
Xij = (5)
Y Pj Πi
X tij 1−σ Ej

(Πi )1−σ = (6)
j
Pj Y
X  tij 1−σ Y
i
(Pj )1−σ = . . (7)
i
Π i Y

The second ratio on the right-hand side of (5) is a decreasing function (under the empirically
valid restriction σ > 1) of direct bilateral trade costs relative to the product of two indexes
of all bilateral trade costs in the system.
Anderson and van Wincoop (2003) called the terms Pj and Πi inward and outward
multilateral resistance respectively. Note that {Pj1−σ , Πi1−σ } can be solved from (6)-(7) for
given t1−σ 1
ij ’s, Ej ’s and Yi ’s combined with a normalization. Under the assumption of bilateral

trade cost symmetry tij = tji , ∀i, j and balanced trade Ej = Yj , ∀j, the natural normalization
is Πi = Pi . Anderson and van Wincoop estimated their gravity equation for Canada’s
provinces and US states with a full information estimator that utilized (7) with Πi = Pi .
Subsequent research has focused mostly on estimating (5) with directional country fixed
effects to control for Ej /Pj1−σ and Yi /Π1−σ
i .
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 (6)-(7), all bilateral trade costs in the
world contribute to the solution values. This somewhat mysterious structure has a simple
1
For any solution to the system {Pj0 , Π0i }, {λPj0 , Π0i /λ} is also a solution. Thus a normalization is needed.
Anderson and Yotov (2010a) find that the system (6)-(7) solves quite quickly, not surprisingly because it is
quadratic in the 1 − σ power transforms of the P ’s and Π’s.

14
and intuitive interpretation: inward and outward multilateral resistance measure average
buyer’s and sellers incidence of trade costs respectively.
The incidence interpretation follows because the uniform preferences assumption in de-
mand implies that the seller in effect makes a single shipment at a uniform markup factor
Πi to a world market with a share determined by

 1−σ
Yi βi pi Πi
= . (8)
Y PW

The right-hand side of (8), referring to the general form (3), is interpreted as the global
expenditure share on the good from i in a hypothetical unified world market, where the world
1/(1−σ)
price index PW = ( i (βi pi Πi )1−σ )
P
is solved from summing (8). PW = 1 is a convenient
normalization of this hypothetical world price. Then with given βi pi ’s the normalization
1/(1−σ)
( i (βi pi Πi )1−σ )
P
= 1 is a useful normalization in solving for multilateral resistances
with (6)-(7). The factor Πi is straightforwardly interpreted as the sellers’ incidence of trade
costs from origin i (Anderson and Yotov, 2010a).
Pj is now interpreted as buyers’ incidence. Solving (7) for Pj , it is a CES index of bilateral
buyers’ incidences tij /Πi , ∀i, equivalent to buyers paying a uniform markup factor Pj on their
entire bundle of shipments (from all i). Sellers incidence Πi similarly is then interpretable
as a CES index of the bilateral sellers’ incidences tij /Pj , from (6).
The interpretation of Π and P as buyers’ and sellers’ incidence generalizes the elementary
economics idea of incidence in the one good case. If the actual set of trade costs were to be
replaced with hypothetical trade costs e
tij = Πi Pj , market clearance and budget constraints
(6)-(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 tij are equivalent to the set of e
tij ’s that decompose into the
product of buyers’ and sellers’ incidence factors. (Unlike the one good case, it is the aggre-
2
This property of (6)-(7) was noted by Anderson and van Wincoop (2004), foreshadowing the interpre-
tation of multilateral resistance as incidence.

15
gate 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 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(Pj1 , ..., PjK )uj , where C(·) is the aggregate cost of living
index for j and uj is the utility of the representative agent (or quantity of aggregate output).
Then, by Shephard’s Lemma, Ejk = Pjk ∂C(·)/∂Pjk . Each class of goods has expenditure
shares described by equations (3)-(4) but amended to add superscript k to every variable
and parameter. Anderson and van Wincoop (2004) argue theoretically for estimating dis-
aggregated gravity and Anderson and 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: t̄ki =
P k k P k k
j tij yij / j yij where yij denotes the number of units of product class k

received from i at destination j. t̄ki is a Laspeyres index of outward trade frictions facing
seller i in good k.
The TFP measure t̄ki is useful for analyzing distribution productivity of the world econ-
omy as a whole, but it is misleading for purposes of understanding comparative economic
performance and the national patterns of production and trade. t̄ki gives the sellers’ inci-
dence only under the partial equilibrium and inconsistent assumption that all incidence falls
on the seller i. Anderson and Yotov (2010a, 2010b) show that in practice these differences
are significant: Laspeyres TFP measures and the incidence of TFP in distribution differ in

16
magnitude and in the case of inward measures the correlation between them is low.3
For consistency of the gravity modules with full general equilibrium, involving allocation
across the sectors k in each country, the Π’s are normalized in each sector k for given
parameters and ‘factory gate’ price pki by

X
(βik pki Πki )1−σk = 1. (9)
j

In practice, when analyzing a gravity module, it is often convenient to normalize one of the
P ’s to one. The choice of normalization is irrelevant to distribution of the goods because
only relative incidence matters.
Now return to the interpretation of the gravity equation (5), reproduced below for con-
venience.
 1−σ
Yi Ej tij
Xij = .
Y Π i Pj

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, Yi Ej /Y . The second ratio is thus interpreted as the ratio
of predicted trade (given the t’s) to predicted frictionless trade.
The useful measure of Constructed Home Bias (Anderson and Yotov, 2010a) is interpreted
as the predicted value of internal trade of i with itself to the predicted value of internal trade
in the frictionless equilibrium. Constructed Home Bias (CHB) is thus given by

 1−σ
tii
CHBi ≡ . (10)
Πi Pi

CHB varies substantially by country, product and time because of changing expenditure
and supply shares, even when gravity coefficients are constant (Anderson and Yotov; 2010a,
3
An alternative measure proposed by Redding and Venables (2004) resembles multilateral resistance but
does not measure incidence. Their measure of ‘market access’ uses essentially
P the same formula as (6) while
their measure of ‘supplier access’ uses the CES price index formula Pjk = [ i (βik pki tkij )1−σk ]1/(1−σk ) . These
variables are constructed without taking account of the simultaneous determination of the two variables, so
they do not measure incidence.

17
2010b).
P
Policy makers are often focused on overall import penetration ratios such as i6=j Xij /Ej
P
and the analogous ratio i6=j Xji /Yj 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:

X
Xijk /Ejk = 1 − (tkjj /Pjk Πkj )1−σk Yjk /Y k . (11)
i6=j
X
Xjik /Yjk = 1 − (tkjj /Pjk Πkj )1−σk Ejk /Y k . (12)
i6=j

Anderson and Yotov show that CHB’s vary a lot across goods and more importantly for
policy concerns, they exhibit a lot of intertemporal movement because of changing world
shipment shares at constant tkij ’s, implying a lot of explanatory power over the import and
export ratios.
The interpretation of the second ratio in equation (5) applies straightforwardly to any
bilateral flow: it is equal to the ratio of predicted bilateral trade to predicted frictionless
trade, hence (tij /Πi Pj )1−σ is the ‘constructed trade bias’ on the link from i to j because of
the buyer’s bilateral incidence from i relative to the average buyer’s incidence for country
j. Alternatively, the same statistic viewed from the exporter’s viewpoint results from the
bilateral seller’s incidence relative to the average seller’s incidence. Bilateral trade flows shift
about due to changes in production and expenditure shares of world shipments, as implied
by the frictionless gravity model, but also because of the general equilibrium force of share
changes that alters incidence even when trade costs {tij } are constant (Anderson and Yotov,
2010a).
The gravity model also readily disaggregates within countries, allowing useful investiga-
tions of inter-regional vs. international trade costs. Indeed, the development of the structural
gravity model (Anderson and 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.

18
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 US states, all else equal. This was too large to be
a sensible component of bilateral trade costs tij .
Structural gravity solved the puzzle by showing that the border dummy variable in Mc-
Callum’s traditional model reflected the effect of multilateral resistance. The border dummy
in the McCallum regression shifts the ratio of inter-provincial 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 yields, for British Columbia’s exports to adjacent Alberta and across
the US border to adjacent Washington

 σ−1
XBC,AB tBC,W A PAB
= .
XBC,W A tBC,AB PW A

The expression on the right-hand side of the equation reflects not only the direct trade
cost increase at the US border that raises tBC,W A /tBC,AB , but also the effect of the ratio
of multilateral resistances for a province and a state, in this case Alberta and Washington,
PAB /PW A . Because Canada’s provinces must do far more of their trade with the outside
world than do US states (Canada is about one tenth the size of the US in GDP), the provinces
naturally have higher multilateral resistance than the states, thereby greatly increasing inter-
provincial 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 due to the omission of the multilateral resistance controls from his regression. Estimating
the structural gravity model, Anderson and van Wincoop (2003) find a more plausible border
cost component of tij , in the range of 20% to 50%.
Inter-regional vs. international trade cost implications of structural gravity were further
developed by Anderson and Yotov (2010a). They offer a decomposition of incidence into do-

19
mestic and international components and calculate sellers’ incidence for Canada’s provinces
on trade within Canada as compared to trade with the rest of the world. They find that while
incidence overall declined substantially from 1990-2002, 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)-(7) imply that
raising all tij ’s by the factor λ > 1 will raise each Π and P by the factor λ1/2 . This formal
homogeneity property has useful empirical content: if the world really were getting smaller
uniformly, the gravity model would be unable to reveal it. The empirical literature tends to
indicate little change in gravity coefficients (see especially Anderson and Yotov, 2010a and
2010b), 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 equation
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 very useful comparative statics to
come a generation later. The comparative statics of inward and outward multilateral resis-
tance were first used by Anderson and van Wincoop (2003). Recognition that multilateral
resistance is interpreted as incidence is in Anderson and Yotov (2010a).

3.2 Supply Side Structure

An alternative derivation of a mathematically equivalent structural gravity model was pro-

posed by Eaton and Kortum (2002), based on homogeneous goods on the demand side, ice-

20
berg trade costs, and Ricardian technology with heterogeneous productivity for each country
and good due to random productivity draws from a Frechet distribution. Despite CES struc-
ture for the intermediate goods demand, in equilibrium the share of goods demanded from
i by country j is determined only on the supply side; the influence of σ disappears into a
constant term. In equilibrium each country will be assigned a subset of the goods, and ex-
cept 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 − σ is interpreted as −θ where θ is the dispersion
parameter of the Frechet distribution. In contrast to the Armington/CES model, all action
is on the extensive margin of trade. Eaton and Kortum derive their model for one ‘sector’
only, a specification generalized by Costinot, Donaldson and Komunjer (2010), so that θk 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 simulation.
General equilibrium superstructure is discussed below in section 6.
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 θ 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 mar-
gin of trade. Chaney’s model includes a fixed cost of export for monopolistically 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.

3.3 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 threshold above zero. In addition there may be missing observations that may or may

21
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 4 on estimation.
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 sub-
stitution greater than one (or the equivalent dispersion/comparative advantage parameter
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 and Neary (2005) present a general homo-
thetic preferences structure, showing that multilateral resistance is defined and solved from
a similar equation system once the functional form and its parameters are specified, along
with data on shipment and expenditure shares.
An alternative economic specification explanation retains CES/Armington preferences
and rationalizes zeroes as due to fixed costs of export facing monopolistic competitive firms.
If no firm in i is productive enough to make incurring the fixed cost of exporting to j
profitable (given the cost of production in i, variable trade cost tij and willingness to pay in
j), then zero trade results. Helpman, Melitz and Rubinstein (HMR, 2008) develop this idea.
The selection effect determines which markets are active and also determines a volume effect
Vij due to productivity heterogeneity among firms whereby markets that are active have a
greater or lesser numbers of firms active depending on the same selection mechanism. The
4
Novy’s aggregate bilateral OECD trade flow data contain no zeroes, so this feature is not exploited yet.

22
gravity model becomes

!1−σk
Ejk Yik k tkij
Xijk = V
Y k ij Pjk Πki
!1−σk
X tkij Vijk Ejk
(Πki )1−σk =
j
Pjk Yk
!1−σk
X tkij Vijk Yik
(Pjk )1−σk = . .
i
Πki 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 distribution
allows a tractable estimation procedure that requires only aggregate bilateral trade data, an
important advantage because firm level trade data is 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. The combination might be able to distinguish between fixed export costs
and choke prices as an explanation of zeroes.

3.4 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 characteristics such as
bilateral distance that condition the probability of each choice, the econometrician observes

23
the resulting masses allocated and uses a probability model to structure statistical inference
An early attempt along these lines was made by Savage and Deutsch (1960) and followed
by Leamer and 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 and Hanson, 2008, and Beine, Docquier and Ozden, 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.

4 Estimation

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

in which much of what impedes trade is not observable to the econometrician. What are
observable are the trade flows and a set of proxies for various types of trade costs, along
with direct measures of some components of trade costs. Most issues with modeling trade
costs are discussed in Anderson and van Wincoop (2004). Since that time there have been
several notable advances in modeling and inferring trade costs.

24
 Two of the advances deal with the implications of zeroes in the bilateral trade flow data.
One view of zeroes is that they stand for flows too small to report, an interpretation that
indeed represents reporting practices of government trade ministries. Interpreting zeroes in
this way, it is legitimate to drop the zero observations from estimation because there is no
economic significance to the zeroes relative to the non-zero observations.
In the presence of heteroskedastic errors, Santos-Silva and Tenreyro (2006) point out
that inconsistent estimation arises from the usual econometric gravity practice using loga-
rithmic 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 Pois-
son 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 and Pham (2008) argue based
on Monte Carlo simulations that when heteroskedasticity is properly controlled, Tobit es-
timators outperform PPML when zeroes are common. Heteroskedasticity is likely to be
attenuated using size-adjusted trade Xij /Yi Ej as the dependent variable, as advocated by
Anderson and van Wincoop (2003, 2004).
An alternative view of zeroes, encountered above, is that economically meaningful selec-
tion generates the zeroes. All firms in origin i face fixed costs of entering exporting to any
particular destination j, and only the sufficiently productive ones can afford to pay the fixed
cost. When a destination j 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 accounting 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

25
the bilateral trade cost tij 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 suspi-
cious 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 tij . 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 specialization of a particular
model of monopolistic competition that is not applicable to all sectors.
Anderson and 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
equations (5)-(7) implies that only relative trade costs can be inferred by gravity, hence the
differences in techniques effectively amount to different implicit normalizations. Anderson
and Yotov report this near-perfect correlation finding based on estimation with the three
techniques over 18 three-digit manufacturing sectors, 76 countries and 13 years of data.

4.1 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 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 and

26
van Wincoop, 2003, report significant differences between gravity estimated with remoteness
and with multilateral resistance.)
The first objection to the traditional model is its aggregation, which causes two problems.
There is aggregation bias because of sectorally varying trade costs and sectorally varying
elasticities of trade with respect to costs (see Anderson and van Wincoop, 2004, for analysis
and Anderson and Yotov, 2010a and 2010b 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 disintegration
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 fixes both problems.
The second objection is omitted variable bias from the perspective of the structural grav-
ity 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, since 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.

4.2 Structural

Anderson and van Wincoop (2003) combine equations (6)-(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 Xij /Yi Ej , size-adjusted trade.
An alternative fixed effects estimator controls for the unobservable multilateral resistances

27
and activity variables
Xij = xi mj t1−σ
ij ij , (13)

where ij is the random error term, xi is the fixed effect for country i as an exporter and
mj is the fixed effect for country j as an importer. Equation (13) is less efficient than a
full information estimator but seems preferable to most subsequent investigators. Feenstra
(2004) argues for the fixed effects estimator because it does not require custom coding, but
another and perhaps better reason is that researchers should be suspicious that there may be
other country-specific unobservables that the fixed effects pick up, but that full information
estimation would drive toward spurious results.
A major drawback to fixed effects estimation is its demolition of structure: the econo-
metrician blows up the building to get at the safe inside containing the inferred bilateral
trade costs. Fortunately, in the case of structural gravity, it is feasible to reconstruct the
building like an archeologist, using structural principles in the form of equations (6)-(7).
Thus Anderson and Yotov (2010a, 2010b) use fixed effects to estimate equation (5) in its
stochastic form, but then calculate the multilateral resistances by calculating the fitted tij ’s
and plugging them into equations (6)-(7).
This technique is used to ‘test’ the structural gravity model by comparing the estimates
of fixed effects (xi mj ) with the structural gravity term (Yi Ej Πσ−1
i Pjσ−1 ). The results are
remarkably close in an economic sense (the fitted regression line has an estimated elasticity
around 0.96, compared to 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 and Bergstrand (2009) propose an alternative direct estimator of multilateral re-
sistance 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 and

28
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 typ-
ical 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 approximately
200,000 fixed effects and even yearly estimation requires 20,000.) In principle the Baier
and Bergstrand estimator could be used to construct tij ’s and then combined with data on
the Y ’s and E’s using equations (6)-(7) in order to obtain the incidence measures and per-
form comparative statics with them. The constructed multilateral resistances in Baier and
Bergstrand’s method can be compared to the point estimates, differences being attributed
to random error and approximation error.

4.3 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 intra-firm 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 the investigator 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 equilibrium
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).

29
5 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
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.

5.1 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, Melitz and Rubinstein (2008) in which
the decision to export and the decision how much to export are distinct.
Let wi denote the wage at location i, ∀i. The worker h who migrates from origin j to
destination i faces a cost of migration modeled with iceberg cost factor δ ji > 1, receiving net
wage (wi /δ ji ). Worker h’s idiosyncratic utility from migration is represented by jih , private
information to him. He chooses to migrate if (wi /δ ji )jih ≥ wj for at least some i. 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 uji = ln wi − ln δ ji − ln wj . In this sort of setting, McFadden (1974) showed that if ln 
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.

30
 Building on this insight, migration models subsequently used the multinomial logit to
model bilateral migration flows. For two recent examples, see Beine, Docquier and Ozden
(2009) and Grogger and Hanson (2008). This section develops a novel gravity model repre-
sentation of the migration model by making use of the market clearing conditions to derive
the appropriate multilateral 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 , that pick destination i. Let N j denote
the population of natives of j. The predicted migration flow from j to i that results from
the setup is

M ij = G(uji )N j . (14)

where
exp(uji )
G(uji ) = P jk
.
k exp(u )

With logarithmic utility, the migration equation is

wi /δ ji
M ji = P k jk N j . (15)
k w /δ

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 equa-
tion. The connection of the share 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 w’s.
Define W j ≡ wk /δ jk and define the labor force supplied to i from all origins
P
k

X
Li ≡ M ji . (16)
j

Nj = Li , the world labor supply N . The labor market clearance equation

P P
Also, N ≡ j i

31
is
X
Li = wi ((1/δ ji )/W j )N j .
j

Then
Li
i
w = i (17)
ΩN

where
X 1/δ ji N j
Ωi = . (18)
j
Wj N

Using equation (17) to substitute for the wage in W j ,

X 1/δ jk Lk
Wj = . (19)
k
Ωk N

Substituting for the wage in equation (15) using equation (17) yields the structural gravity
equation of migration:
Li N j 1/δ ji
M ji = . (20)
N Ωi W j

The first ratio represents the migration pattern of a frictionless world. The implication
is that in a frictionless world, populations originating in j would be found in equal propor-
tions to their share of world population in all destinations: M ji /Li = N j /N . The second
term represents the effect of migration frictions. The bilateral migration friction δ ji reduces
migration. It is divided by the the product of weighted averages of the inverse of migration
frictions, one for inward migration to i from all origins and one for outward migration from
j to all destinations. The system given in equations (18)-(19) can be solved for the Ω’s and
W ’s (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
5
A tractable gravity equation results from (14) by using a restriction on utility to convert exp uji 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.

32
 Let the coefficient of relative risk aversion be θ. In this case equation (20) becomes

1−θ
Li N j δ ji

ji
M =
N Ω̄i W̄ j

where " #1/(1−θ)

X (δ ji )1−θ N j
Ω̄i ≡
j
W̄ j N

and " #1/(1−θ)

X (δ ji )1−θ Li
W̄ j ≡ .
i
Ω̄i N

Here, Ω̄i and W̄ j are CES price indexes of migration frictions, one for inward (Ω̄i ) and one for
outward (W̄ j ) 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 θ = 2.
Ω and W are general equilibrium concepts as is clear because their solution in the si-
multaneous 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 combined the migration sys-
tem 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, Ω̄’s and W̄ ’s can be computed once the δ’s are econo-
metrically constructed and the labor supplies Li and population stocks N j are observed. A
normalization is needed. (See Anderson and Yotov, 2009, for details.)
A similar model has been applied to services trade by Head, Mayer and Ries (2009).
Instead of actually changing locations, the foreign worker does the job in his home location.

33
The cost of migration becomes the cost of monitoring the distant worker. Worker produc-
tivities in each location have the Frechet 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 dis-
tribution of log productivities takes the Gumbel form. The fraction of service jobs in origin i
going to workers in location j has the multinomial logit form. The total numbers of workers
and of jobs in each location enter 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
fruitfully 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 {N j } and the set of migration frictions, the δ’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.
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 develop-
ment 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

34
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.

5.2 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 concen-
trating 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, Melitz and Yeaple (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 non-traders, 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
application with US 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 surprising
because the dependent variable is different and the question addressed is more difficult to

35
answer.
Kleinert and Toubal (2010) extend Helpman, Melitz and Yeaple to allow for fixed setup
costs that rise with distance, a wrinkle that can explain why foreign affiliate sales can fall
rather than rise with distance as the earlier proximity-concentration tradeoff suggested. They
also derive a gravity-type relationship from two other structures, a vertical integration model
and a two country factor proportions model of fragmentation.
Bergstrand and Egger (2007) offer a gravity model of FDI derived from the knowledge-
based capital theory of horizontal multi-national enterprises. Their objective is a full general
equilibrium model that can explain trade, foreign affiliate sales and FDI. They simulate
a theoretical model that generates nonlinear relationships between exports, affiliate sales
and their exogenous determinants. Then they fit an approximate ‘empirical’ relationship to
the generated data and take the same relationship to actual data, with some success. A
limitation of their model is that, though the factor proportions model with 3 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 and Yeaple (2009) develop a gravity model of vertically integrated intra-firm trade
featuring trade costs with two elements, a standard iceberg trade cost and a communication
cost that rises with the complexity of the firm’s technology. Input complexity raises tech-
nology transfer costs, whereas the costs of embodied technology transfer are independent
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 and Yeaple’s model of foreign
affiliate sales permits inference about trade costs from observable trade flows.
Keller and 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

36
helpful in explaining the rising share (in total trade) of intra-firm 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 5.1
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 exogenous 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 intra-
firm 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, though
with limited possibilities that may be very limited for FDI. The potential 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 and 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.

5.3 Portfolio Investment

Martin and Rey (2004) offer the first gravity type model of international portfolio investment.
The coefficient of relative risk aversion plays the role, in equilibrium, of the elasticity of
substitution in the CES demand specification. While appealing as a rationale for the gravity

37
application of Portes and Rey (2005), the Martin and Rey model does not provide a fully
satisfactory foundation for gravity models of investment flows because (i) trade is assumed
to be frictionless, (ii) investment costs are uniform, and (iii) most important, 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.

6 Integrated Superstructure

The gravity model nests inside a general equilibrium superstructure. As pointed out in
Anderson and 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 and van Win-
coop (2003) use the endowments model to calculate the effect of eradicating the US-Canada
border on their estimated gravity model of trade between US states, Canadian provinces and
the aggregated rest of the world.
Another attractive candidate is the Ricardian production model. Eaton and 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 Eaton-Kortum
model of 2002, the extensive margin is the only margin. Arkolakis and others have variants in

38
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 endow-
ments 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 Π’s 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 rep-
resentation in this review. In the integrated superstructure it can be treated simultaneously
with the trade modules. In this setting, multilateral resistance in trade has significant 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 alogn these lines is 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.
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 equilibrium. The
determinants of total shipments and total expenditures are irrelevant to this inference be-
cause 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 deter-

39
mination 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 promising as a source of
insight.

7 Conclusion

This idiosyncratic review of work on the gravity model suggests that the story is not over,
so a conclusion can only point to potential future chapters. Distribution broadly defined
consumes a very large share of the world’s resources and gravity has proven to be the most
generally useful empirical model for understanding the distribution of goods and factors of
production. It appears to work well at almost any scale.
The progress in structural modeling of gravity has yielded three distinct rationales for the
same observationally equivalent model of the distribution of economic flows between origins
and destinations, one based on the demand side (the CES/Armington model), one based
on the supply side (the Eaton-Kortum model), and one based on a discrete choice model of
the individual actor transferring the goods or factors. Further work may suggest ways to
discriminate between these.
The structural modeling of gravity imposes trade separability, permitting gravity modules
to be nested inside a wide range of general equilibrium superstructures. Future work with
simulation models may suggest which of many candidate general equilibrium production
models do better.
The problem of zeroes in the trade and factor flows data has been addressed with some
success, particularly by Helpman, Melitz and Rubinstein. But I expect future work to do
better. The CES framework (with elasticity of substitution greater than one) is unsuitable
for describing small amounts of trade. The translog cost function, in particular, seems likely
to yield better descriptions and better understanding of why so many potential flows are

40
equal to zero. This is so even 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 the
profession agrees that they are as interesting and useful as they appear to me, more work is
needed to see how believable the measures are. As it stands, they are completely reliant on
CES structure. How well does the CES do in representing the world economy? This is an
especially important question in light of the zeroes question in the preceding paragraph. I
look forward to the development of the translog case to help answer this question.

41
References

[1] Anderson, James E. 1979. “A theoretical foundation for the gravity equation”, American
Economic Review 69, 106-116.

[2] Anderson, James E. and J. Peter Neary. 2005. Measuring the Restrictiveness of Inter-
national Trade Policy, Cambridge: MIT Press.

[3] Anderson, James E. and Eric van Wincoop. 2004.“Trade Costs”, Journal of Economic
Literature, 42, 691-751.

[4] Anderson, James E. and Eric van Wincoop. 2003. “Gravity with Gravitas”, American
Economic Review, 93, 170-92.

[5] Anderson, James E. and Yoto V. Yotov. 2010a. “The Changing Incidence of Geography,”
American Economic Review, 100, 2157-86.

[6] Anderson, James E. and Yoto V. Yotov. 2010b. “Specialization: Pro- and Anti-
globalizing, 1990-2002”, NBER Working Paper No. 16301.

[7] Arkolakis, Costas. 2008. “Market Penetration Costs and the New Consumers Margin in
International Trade”, NBER Working Paper No. 14214.

[8] Baier, Scott L., and Jeffrey H. Bergstrand, 2001. “The Growth of World Trade: Tariffs,
Transport Costs, and Income Similarity”, Journal of International Economics, vol. 53,
no. 1, 1-27.

[9] Baier, Scott L. and Jeffrey H. Bergstrand, 2009. “Bonus Vetus OLS: A Simple Method
for Approximating International Trade-Cost Effects using the Gravity Equation”, Jour-
nal of International Economics, vol. 77, 77-85.

[10] Beine, Michel, Frederic Docquier and Caglar Ozden. 2009. “Diasporas”, Policy Research
Working Paper No. 4984. World Bank.

42
[11] Bergstrand, Jeffrey H., 1985. “The Gravity Equation in International Trade: Some Mi-
croeconomic Foundations and Empirical Evidence”, Review of Economics and Statistics,
Vol. 67, No. 3, August 1985, 474-481.

[12] Bergstrand, Jeffrey H., 1989, “The Generalized Gravity Equation, Monopolistic Compe-
tition, and the Factor-Proportions Theory in International Trade,” Review of Economics
and Statistics, Vol. 71, No. 1, February 1989, 143-153.

[13] Bergrstrand, Jeffrey H. and Peter Egger. 2007. “A Knowledge- and Physical-capital
Model of International Trade Flows, Foreign Direct Investment and Multinational En-
terprises”, Journal of International Economics, 73, 278-308.

[14] Bergrstrand, Jeffrey H. and Peter Egger. 2011. “Gravity Equations and Economic Fric-
tions in the World Economy”, in Daniel Bernhofen, Rodney Falvey, David Green-
away and Udo Krieckemeier, eds., Palgrave Handbook of International Trade, Palgrave-
Macmillan Press, in press.

[15] Chaney, Thomas,. 2008. “Distorted Gravity: The Intensive and Extensive Margins of
International Trade”, American Economic Review, 98, 1707-21.

Costinot, Arnaud, David Donaldson and Ivana Komunjer. 2010. “What Goods Do Coun-
tries Trade? A Quantitative Exploration of Ricardo’s Ideas”, unpublished manuscript,
MIT.

[16] Deardorff, Alan, 1980. “The General Validity of the Law of Comparative Advantage”,
Journal of Political Economy, 88, 941-57.

Eaton, Jonathan, and Samuel Kortum. 2002. “Technology, Geography, and Trade”,
Econometrica, 70: 1741-1779.

[17] Feenstra, Robert. 2004. Advanced International Trade: Theory and Evidence, Princeton,
NJ: Princeton University Press.

43
[18] Grogger, Jeffrey and Gordon H. Hanson. 2008. “Income Maximization and the Selection
and Sorting of International Migrants”, NBER Working Paper No. 13821.

[19] Head, Keith, Thierry Mayer and John Ries. 2009. “How Remote is the Off-shoring
Threat?”, European Economic Review, 53, 429-44.

[20] Head, Keith and John Ries. 2008. “FDI as an Outcome of the Market for Corporate
Control: Theory and Evidence”, Journal of International Economics, 74, 2-20.

[21] Helpman, Elhanan, Marc J. Melitz and Yona Rubinstein. 2008. “Estimating Trade
Flows: Trading Partners and Trading Volumes”, Harvard University, Quarterly Journal
of Economics, 123: 441-487.

[22] Helpman, Elhanan, Marc J. Melitz and Stephen R. Yeaple. 2004. “Export Versus FDI
with Heterogeneous Firms”, American Economic Review, 94, 300-316.

[23] Keller, Wolfgang and Stephen R. Yeaple. 2009. “Gravity in the Weightless Economy”,
NBER Working Paper No. 15509.

[24] Kleinert, Jörn and Farid Toubal, 2010. “Gravity for FDI”, Review of International
Economics, 18 (1), 1-13.

[25] Leamer, Edward E. and Robert M. Stern, 1970. Quantitative International Economics,
Chicago: Aldine Press.

LeamerLevinsohn, Leamer, Edward E. and James Levinsohn, 1995. “International Trade

Theory: the Evidence” in Gene M. Grossman and Kenneth Rogoff, eds., Handbook of
International Economics, vol. 3, Amsterdam: Elsevier Science B.V., 1339-94.

[26] Martin, Philippe and Helene Rey, (2004), “Financial Super-markets: Size Matters for
Asset Trade”, Journal of International Economics, 64, 335-61.

[27] Martin, Will and Pham S. Cong (2008), “Estimating the Gravity Equation when Zero
Trade Flows Are Frequent”, MPRA Paper No. 9453, University Library of Munich.

44
 [28] Matsuyama, Kiminori, 2007, “Beyond Icebergs: Towards a Theory of Biased Globaliza-
tion”, Review of Economic Studies, 74, 237-53.

[29] McCallum, John (1995) “National Borders Matter: Canada-U.S. Regional Trade Pat-
terns,” American Economic Review, 1995, 85(3), pp. 615-623.

[30] McFadden, Daniel, 1974. “Conditional Logit Analysis of Qualitative Choice Behavior”.
In Paul Zarembka, ed. Frontiers of Econometrics, Academic Press.

Novy10, Novy, Dennis. 2010. “International Trade and Monopolistic Competition with-
out CES: Estimating Translog Gravity”, unpublished manuscript, Warwick University.

[31] Redding, Stephen and Anthony J. Venables. 2004. “Economic Geography and Interna-
tional Inequality”, Journal of International Economics, 62(1), 53-82.

[32] Portes, Richard and Helene Rey, (2005), “The Determinants of Cross-Border Equity
Flows”, Journal of International Economics, 65, 269-96.

[33] Ravenstein, Edward George, 1889, “The Laws of Migration.” Journal of the Royal Sta-
tistical Society, Vol. 52, No. 2. (June, 1889), pp. 241305.

[34] Samuelson, Paul A. (1952), “The Transfer Problem and Transport Costs: The Terms
of Trade When Impediments are Absent”, The Economic Journal, 62, 278-304.

[35] Santos Silva, Jorge and Sylvana Tenreyro. 2006. “The Log of Gravity”, Review of Eco-
nomics and Statistics, Vol. 88, No. 4: 641-658.

[36] Savage, I. Richard, and Karl W. Deutsch, 1960. “A Statistical Model of the Gross
Analysis of Transactions Flows”, Econometrica, 28, 551-72.

[37] Tinbergen, Jan. 1962. Shaping the World Economy: Suggestions for an International
Economic Policy. New York: The Twentieth Century Fund.

45

View publication stats