Rest 88 4 641

THE LOG OF GRAVITY
J. M. C. Santos Silva and Silvana Tenreyro*
Abstract—Although economists have long been aware of Jensen’s in- models are severely biased, distorting the interpretation of
equality, many econometric applications have neglected an important
implication of it: under heteroskedasticity, the parameters of log-
the model. These biases might be critical for the compara-
linearized models estimated by OLS lead to biased estimates of the true tive assessment of competing economic theories, as well as
elasticities. We explain why this problem arises and propose an appropri- for the evaluation of the effects of different policies. In
ate estimator. Our criticism of conventional practices and the proposed
solution extend to a broad range of applications where log-linearized contrast, our method is robust to the different patterns of
equations are estimated. We develop the argument using one particular heteroskedasticity considered in the simulations.
illustration, the gravity equation for trade. We find significant differences We next use the proposed method to provide new esti-
between estimates obtained with the proposed estimator and those ob-
tained with the traditional method. mates of the gravity equation in cross-sectional data. Using
standard tests, we show that heteroskedasticity is indeed a
severe problem, both in the traditional gravity equation
I. Introduction introduced by Tinbergen (1962), and in a gravity equation
that takes into account multilateral resistance terms or fixed
E CONOMISTS have long been aware that Jensen’s in-
equality implies that E(ln y) ⫽ ln E(y), that is, the effects, as suggested by Anderson and van Wincoop (2003).
Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

expected value of the logarithm of a random variable is We then compare the estimates obtained with the proposed
different from the logarithm of its expected value. This PML estimator with those generated by OLS in the log
basic fact, however, has been neglected in many economet- linear specification, using both the traditional and the fixed-
ric applications. Indeed, one important implication of Jen- effects gravity equations.
sen’s inequality is that the standard practice of interpreting Our estimation method paints a very different picture of
the parameters of log-linearized models estimated by ordi- the determinants of international trade. In the traditional
nary least squares (OLS) as elasticities can be highly mis- gravity equation, the coefficients on GDP are not, as gen-
leading in the presence of heteroskedasticity. erally estimated, close to 1. Instead, they are significantly
Although many authors have addressed the problem of smaller, which might help reconcile the gravity equation
obtaining consistent estimates of the conditional mean of with the observation that the trade-to-GDP ratio decreases
the dependent variable when the model is estimated in the with increasing total GDP (or, in other words, that smaller
log linear form (see, for example, Goldberger, 1968; Man- countries tend to be more open to international trade). In
ning & Mullahy, 2001), we were unable to find any refer- addition, OLS greatly exaggerates the roles of colonial ties
ence in the literature to the potential bias of the elasticities and geographical proximity.
estimated using the log linear model. Using the Anderson–van Wincoop (2003) gravity equa-
In this paper we use the gravity equation for trade as a tion, we find that OLS yields significantly larger effects for
particular illustration of how the bias arises and propose an geographical distance. The estimated elasticity obtained
appropriate estimator. We argue that the gravity equation, from the log-linearized equation is almost twice as large as
and, more generally, constant-elasticity models, should be that predicted by PML. OLS also predicts a large role for
estimated in their multiplicative form and propose a simple common colonial ties, implying that sharing a common
pseudo-maximum-likelihood (PML) estimation technique. colonial history practically doubles bilateral trade. In con-
Besides being consistent in the presence of heteroskedas- trast, the proposed PML estimator leads to a statistically and
ticity, this method also provides a natural way to deal with economically insignificant effect.
zero values of the dependent variable. The general message is that, even controlling for fixed
Using Monte Carlo simulations, we compare the perfor- effects, the presence of heteroskedasticity can generate
mance of our estimator with that of OLS (in the log linear strikingly different estimates when the gravity equation is
specification). The results are striking. In the presence of log-linearized, rather than estimated in levels. In other
heteroskedasticity, estimates obtained using log-linearized words, Jensen’s inequality is quantitatively and qualitatively
important in the estimation of gravity equations. This sug-
Received for publication March 29, 2004. Revision accepted for publi- gests that inferences drawn on log-linearized regressions
cation September 13, 2005. can produce misleading conclusions.
* ISEG/Universidade Técnica de Lisboa and CEMAPRE; and London
School of Economics, CEP, and CEPR, respectively. Despite the focus on the gravity equation, our criticism of
We are grateful to two anonymous referees for their constructive the conventional practice and the solution we propose ex-
comments and suggestions. We also thank Francesco Caselli, Kevin tend to a broad range of economic applications where the
Denny, Juan Carlos Hallak, Daniel Mota, John Mullahy, Paulo Parente,
Manuela Simarro, and Kim Underhill for helpful advice on previous equations under study are log-linearized, or, more generally,
versions of this paper. The usual disclaimer applies. Jiaying Huang transformed by a nonlinear function. A short list of exam-
provided excellent research assistance. Santos Silva gratefully acknowl- ples includes the estimation of Mincerian equations for
edges the partial financial support from Fundação para a Ciência e
Tecnologia, program POCTI, partially funded by FEDER. A previous wages, production functions, and Euler equations, which are
version of this paper circulated as “Gravity-Defying Trade.” typically estimated in logarithms.
The Review of Economics and Statistics, November 2006, 88(4): 641–658

© 2006 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology
642 THE REVIEW OF ECONOMICS AND STATISTICS
The remainder of the paper is organized as follows. where ␩ij is an error factor with E(␩ij兩Yi, Yj, Dij) ⫽ 1,
Section II studies the econometric problems raised by the assumed to be statistically independent of the regressors,
estimation of gravity equations. Section III considers leading to
constant-elasticity models in general; it introduces the PML
estimator and specification tests to check the adequacy of E共T ij 兩Y i ,Y j ,D ij 兲 ⫽ ␣ 0 Y i␣ 1Y j␣ 2D ij␣ 3.
the proposed estimator. Section IV presents the Monte Carlo
simulations. Section V provides new estimates of both the There is a long tradition in the trade literature of log-
traditional and the Anderson–van Wincoop gravity equa- linearizing equation (2) and estimating the parameters of
tion. The results are compared with those generated by interest by least squares, using the equation
OLS, nonlinear least squares, and tobit estimations. Section ln Tij ⫽ ln ␣0 ⫹ ␣1 ln Yi ⫹ ␣2 ln Yj
VI contains concluding remarks. (3)
⫹ ␣3 ln Dij ⫹ ln ␩ij .
II. The Econometrics of the Gravity Equation The validity of this procedure depends critically on the
assumption that ␩ij, and therefore ln ␩ij, are statistically
independent of the regressors. To see why this is so, notice

A. The Traditional Gravity Equation
that the expected value of the logarithm of a random
The pioneering work of Jan Tinbergen (1962) initiated a variable depends both on its mean and on the higher-order
vast theoretical and empirical literature on the gravity equa- moments of the distribution. Hence, for example, if the
tion for trade. Theories based on different foundations for variance of the error factor ␩ij in equation (2) depends on Yi,
trade, including endowment and technological differences, Yj, or Dij, the expected value of ln ␩ij will also depend on the
increasing returns to scale, and Armington demands, all regressors, violating the condition for consistency of OLS.2
predict a gravity relationship for trade flows analogous to In the cases studied in section V we find overwhelming
Newton’s law of universal gravitation.1 In its simplest form, evidence that the error terms in the usual log linear speci-
the gravity equation for trade states that the trade flow from fication of the gravity equation are heteroskedastic, which
country i to country j, denoted by Tij, is proportional to the violates the assumption that ln ␩ij is statistically indepen-
product of the two countries’ GDPs, denoted by Yi and Yj, dent of the regressors and suggests that this estimation
and inversely proportional to their distance, Dij, broadly method leads to inconsistent estimates of the elasticities of
construed to include all factors that might create trade interest.
resistance. More generally, A related problem with the analogy between Newtonian
gravity and trade is that gravitational force can be very
T ij ⫽ ␣ 0 Y i␣ 1Y j␣ 2D ij␣ 3, (1) small, but never zero, whereas trade between several pairs
of countries is literally zero. In many cases, these zeros
where ␣0, ␣1, ␣2, and ␣3 are unknown parameters. occur simply because some pairs of countries did not trade
The analogy between trade and the physical force of in a given period. For example, it would not be surprising to
gravity, however, clashes with the observation that there is find that Tajikistan and Togo did not trade in a certain year.3
no set of parameters for which equation (1) will hold exactly These zero observations pose no problem at all for the
for an arbitrary set of observations. To account for devia- estimation of gravity equations in their multiplicative form.
tions from the theory, stochastic versions of the equation are In contrast, the existence of observations for which the
used in empirical studies. Typically, the stochastic version dependent variable is zero creates an additional problem for
of the gravity equation has the form
2 As an illustration, consider the case in which ␩ follows a log normal
ij
distribution, with E(␩ij兩Yi, Yj, Dij) ⫽ 1 and variance ␴ij2 ⫽ f(Yi, Yj, Dij). The
T ij ⫽ ␣ 0 Y i␣ 1Y j␣ 2D ij␣ 3␩ ij , (2) error term in the log-linearized representation will then follow a normal
distribution, with E [ln ␩ij兩Yi, Yj, Dij] ⫽ ⫺ 21 ln(1 ⫹ ␴ij2), which is also a
function of the covariates.
1 See, for example, Anderson (1979), Helpman and Krugman (1985), 3 The absence of trade between small and distant countries might be
Bergstrand (1985), Davis (1995), Deardoff (1998), and Anderson and van explained, among other factors, by large variable costs (for example,
Wincoop (2003). A feature common to these models is that they all assume bricks are too costly to transport) or large fixed costs (for example,
complete specialization: each good is produced in only one country. information on foreign markets). At the aggregate level, these costs can be
However, Haveman and Hummels (2001), Feenstra, Markusen, and Rose best proxied by the various measures of distance and size entering the
(2000), and Eaton and Kortum (2001) derive the gravity equation without gravity equation. The existence of zero trade between many pairs of
relying on complete specialization. Examples of empirical studies framed countries is directly addressed by Hallak (2006) and Helpman, Melitz, and
on the gravity equation include the evaluation of trade protection (for Rubinstein (2004). These authors propose a promising avenue of research
example, Harrigan, 1993), regional trade agreements (for example, using a two-part estimation procedure, with a fixed-cost equation deter-
Frankel, Stein, & Wei, 1998; Frankel, 1997), exchange rate variability (for mining the cutoff point above which a country exports, and a standard
example, Frankel & Wei, 1993; Eichengreen & Irwin, 1995), and currency gravity equation. Their results, however, rely heavily on both normality
unions (for example, Rose, 2000; Frankel & Rose, 2002; and Tenreyro & and homoskedasticity assumptions, the latter being the particular concern
Barro, 2002). See also the various studies on border effects influencing the of this paper. A natural topic for further research is to develop and
patterns of intranational and international trade, including McCallum implement an estimator of the two-part model that, like the PML estimator
(1995), and Anderson and van Wincoop (2003), among others. proposed here, is robust to distributional assumptions.
THE LOG OF GRAVITY 643
the use of the log linear form of the gravity equation. As before, log-linearization of equation (5) raises the prob-
Several methods have been developed to deal with this lem of how to treat zero-value observations. Moreover,
problem [see Frankel (1997) for a description of the various given that equation (5) is a multiplicative model, it is also
procedures]. The approach followed by the large majority of subject to the biases caused by log-linearization in the
empirical studies is simply to drop the pairs with zero trade presence of heteroskedasticity. Naturally, the presence of
from the data set and estimate the log linear form by OLS. the individual effects may reduce the severity of this prob-
Rather than throwing away the observations with Tij ⫽ 0, lem, but whether or not that happens is an empirical issue.
some authors estimate the model using Tij ⫹ 1 as the In our empirical analysis we provide estimates for both
dependent variable or use a tobit estimator. However, these the traditional and the Anderson–van Wincoop gravity equa-
procedures will generally lead to inconsistent estimators of tions, using alternative estimation methods. We show that,
the parameters of interest. The severity of these inconsis- in practice, heteroskedasticity is quantitatively and qualita-
tencies will depend on the particular characteristics of the tively important in the gravity equation, even when control-
sample and model used, but there is no reason to believe that ling for fixed effects. Hence, we recommend estimating the
they will be negligible. augmented gravity equation in levels, using the proposed
Zeros may also be the result of rounding errors.4 If trade PML estimator, which also adequately deals with the zero-

is measured in thousands of dollars, it is possible that for value observations.
pairs of countries for which bilateral trade did not reach a
minimum value, say $500, the value of trade is registered as III. Constant-Elasticity Models
0. If these rounded-down observations were partially com- Despite their immense popularity, empirical studies in-
pensated by rounded-up ones, the overall effect of these volving gravity equations still have important econometric
errors would be relatively minor. However, the rounding flaws. These flaws are not exclusive to this literature, but
down is more likely to occur for small or distant countries, extend to many areas where constant-elasticity models are
and therefore the probability of rounding down will depend used. This section examines how the deterministic multipli-
on the value of the covariates, leading to the inconsistency cative models suggested by economic theory can be used in
of the estimators. Finally, the zeros can just be missing empirical studies.
observations that are wrongly recorded as 0. This problem is In their nonstochastic form, the relationship between the
more likely to occur when small countries are considered, multiplicative constant-elasticity model and its log linear
and again the measurement error will depend on the covari- additive formulation is trivial. The problem, of course, is
ates, leading to inconsistency. that economic relations do not hold with the accuracy of
physical laws. All that can be expected is that they hold on
B. The Anderson–van Wincoop Gravity Equation average. Indeed, here we interpret economic models like the
gravity equation as yielding the expected value of the
Anderson and van Wincoop (2003) argue that the tradi- variable of interest, y ⱖ 0, for a given value of the explan-
tional gravity equation is not correctly specified, as it does atory variables, x (see Goldberger, 1991, p. 5). That is, if
not take into account multilateral resistance terms. One of economic theory suggests that y and x are linked by a
the solutions for this problem that is suggested by those constant-elasticity model of the form yi ⫽ exp (xi␤), the
authors is to augment the traditional gravity equation with function exp(xi␤) is interpreted as the conditional expecta-
exporter and importer fixed effects, leading to tion of yi given x, denoted E[yi兩x].6 For example, using the
notation in the previous section, the multiplicative gravity
T ij ⫽ ␣ 0 Y i␣ 1Y j␣ 2D ij␣ 3e ␪ id i⫹␪ jd j, (4) relationship can be written as the exponential function exp
[ln ␣0 ⫹ ␣1 ln Yi ⫹ ␣2 ln Yj ⫹ ␣3 ln Dij], which is interpreted
where ␣0, ␣1, ␣2, ␣3, ␪i, and ␪j are the parameters to be as the conditional expectation E(Tij兩Yi, Yj, Dij).
estimated and di and dj are dummies identifying the exporter Because the relation yi ⫽ exp(xi␤) holds on average but
and importer.5 not for each i, an error term is associated with each obser-
Their model also yields the prediction that ␣1 ⫽ ␣2 ⫽ 1, vation, which is defined as εi ⫽ yi ⫺ E[yi兩x].7 Therefore, the
which leads to the unit-income-elasticity model stochastic model can be formulated as
T ij ⫽ ␣ 0 Y i Y j D ij␣ 3e ␪ id i⫹␪ jd j, 6 Notice that if exp(x ␤) is interpreted as describing the conditional
i
median of yi (or some other conditional quantile) rather than the condi-
whose stochastic version has the form tional expectation, estimates of the elasticities of interest can be obtained
estimating the log linear model using the appropriate quantile regression
estimator (Koenker & Bassett, 1978). However, interpreting exp(xi␤) as a
E共T ij 兩Y i ,Y j ,D ij ,d i ,d j 兲 ⫽ ␣ 0 Y i Y j D ij␣ 3e ␪ id i⫹␪ jd j. (5) conditional median is problematic when yi has a large mass of zero
observations, as in trade data. Indeed, in this case the conditional median
of yi will be a discontinuous function of the regressors, which is generally
4 Trade data can suffer from many other forms of errors, as described in not compatible with standard economic theory.
Feenstra, Lipsey, and Bowen (1997). 7 Whether the error enters additively or multiplicatively is irrelevant for
5 Note that, throughout the paper, T denotes exports from i to j. our purposes, as explained below.
ij
y i ⫽ exp共xi ␤兲 ⫹ εi , (6) It may be surprising that the pattern of heteroskedasticity

and, indeed, the form of all higher-order moments of the
with yi ⱖ 0 and E[εi兩x] ⫽ 0. conditional distribution of the error term can affect the
As we mentioned before, the standard practice of log- consistency of an estimator, rather than just its efficiency.
linearizing equation (6) and estimating ␤ by OLS is inap- The reason is that the nonlinear transformation of the
propriate for a number of reasons. First of all, yi can be 0, in dependent variable in equation (7) changes the properties of
which case log-linearization is infeasible. Second, even if the error term in a nontrivial way because the conditional
all observations of yi are strictly positive, the expected value expectation of ln ␩i depends on the shape of the conditional
of the log-linearized error will in general depend on the distribution of ␩i. Hence, unless very strong restrictions are
covariates, and hence OLS will be inconsistent. To see the imposed on the form of this distribution, it is not possible to
point more clearly, notice that equation (6) can be expressed recover information about the conditional expectation of yi
as from the conditional mean of ln yi, simply because ln ␩i is
correlated with the regressors. Nevertheless, estimating
y i ⫽ exp共xi ␤兲 ␩i , equation (7) by OLS will produce consistent estimates of
the parameters of E[ln yi兩x] as long as E[ln (yi)兩x] is a linear

with ␩i ⫽ 1 ⫹ εi/exp(xi␤) and E[␩i兩x] ⫽ 1. Assuming for the function of the regressors.10 The problem is that these
moment that yi is positive, the model can be made linear in parameters may not permit identification of the parameters
the parameters by taking logarithms of both sides of the of E[yi兩x].
equation, leading to In short, even assuming that all observations on yi are
positive, it is not advisable to estimate ␤ from the log linear
ln yi ⫽ xi ␤ ⫹ ln ␩i . (7) model. Instead, the nonlinear model has to be estimated.
To obtain a consistent estimator of the slope parameters A. Estimation

in equation (6) estimating equation (7) by OLS, it is neces- Although most empirical studies use the log linear form
sary that E[ln ␩i兩x] does not depend on xi.8 Because ␩i ⫽ of the constant-elasticity model, some authors [see Frankel
1 ⫹ εi/exp(xi␤), this condition is met only if εi can be and Wei (1993) for an example in the international trade
written as εi ⫽ exp(xi␤) vi, where vi is a random variable literature] have estimated multiplicative models using non-
statistically independent of xi. In this case, ␩i ⫽ 1 ⫹ vi and linear least squares (NLS), which is an asymptotically valid
therefore is statistically independent of xi, implying that estimator for equation (6). However, the NLS estimator can
E[ln ␩i兩x] is constant. Thus, only under very specific con- be very inefficient in this context, as it ignores the het-
ditions on the error term is the log linear representation of eroskedasticity that, as discussed before, is characteristic of
the constant-elasticity model useful as a device to estimate this type of data.
the parameters of interest. The NLS estimator of ␤ is defined by
When ␩i is statistically independent of xi, the conditional
variance of yi (and εi) is proportional to exp(2xi␤). Although
冘关 y ⫺ exp共x b兲兴 ,
n
ˆ ⫽ arg min
␤ 2
economic theory generally does not provide any informa- i⫽1
i i
tion on the variance of εi, we can infer some of its properties
b
from the characteristics of the data. Because yi is nonnega- which implies the following set of first-order conditions:
tive, when E[yi兩x] approaches 0, the probability of yi being
positive must also approach 0. This implies that V[yi兩x], the
冘关 y ⫺ exp共x ␤兲兴
n
conditional variance of yi, tends to vanish as E[yi兩x] passes i

ˆ exp共x ␤兲
ˆ x ⫽ 0.
i i i (8)
i⫽1
to 0.9 On the other hand, when the expected value of y is far
away from its lower bound, it is possible to observe large These equations give more weight to observations where
deviations from the conditional mean in either direction, exp(xi␤ˆ ) is large, because that is where the curvature of the
leading to greater dispersion. Thus, in practice, εi will conditional expectation is more pronounced. However,
generally be heteroskedastic and its variance will depend on these are generally also the observations with larger vari-
exp(xi␤), but there is no reason to assume that V[yi兩x] is ance, which implies that NLS gives more weight to noisier
proportional to exp(2xi␤). Therefore, in general, regressing observations. Thus, this estimator may be very inefficient,
ln yi on xi by OLS will lead to inconsistent estimates of ␤. depending heavily on a small number of observations.
If the form of V[yi兩x] were known, this problem could be
8 Consistent estimation of the intercept would also require E[ln ␩i 兩x] ⫽ 0. eliminated using a weighted NLS estimator. However, in
9 In the case of trade data, when E [ yi兩x] is close to its lower bound (that
is, for pairs of small and distant countries), it is unlikely that large values
of trade are observed, for they cannot be offset by equally large deviations 10 When E [ln y 兩 x] is not a linear function of the regressors, estimating
i
in the opposite direction, simply because trade cannot be negative. equation (7) by OLS will produce consistent estimates of the parameters
Therefore, for these observations, dispersion around the mean tends to be of the best linear approximation to E [ln yi 兩 x] (see Goldberger, 1991, p.
small. 53).
practice, all we know about V[yi兩x] is that, in general, it goes natural to give the same weight to all observations.13 Even
to 0 as E[yi兩x] passes to 0. Therefore, an optimal weighted if E[yi兩x] is not proportional to V[yi兩x], the PML estimator
NLS estimator cannot be used without further information based on equation (9) is likely to be more efficient than the
on the distribution of the errors. In principle, this problem NLS estimator when the heteroskedasticity increases with
can be tackled by estimating the multiplicative model using the conditional mean.
a consistent estimator, and then obtaining the appropriate The estimator defined by equation (9) is numerically
weights estimating the skedastic function nonparametri- equal to the Poisson pseudo-maximum-likelihood (PPML)
cally, as suggested by Delgado (1992) and Delgado and estimator, which is often used for count data.14 The form of
Kniesner (1997). However, this nonparametric generalized equation (9) makes clear that all that is needed for this
least squares estimator is rather cumbersome to implement, estimator to be consistent is the correct specification of the
especially if the model has a large number of regressors. conditional mean, that is, E[yi兩x] ⫽ exp(xi␤). Therefore, the
Moreover, the choice of the first-round estimator is an open data do not have to be Poisson at all—and, what is more
question, as the NLS estimator may be a poor starting point important, yi does not even have to be an integer—for the
due to its considerable inefficiency. Therefore, the nonpara- estimator based on the Poisson likelihood function to be
metric generalized least squares estimator is not appropriate consistent. This is the well-known PML result first noted by

to use as a workhorse for routine estimation of multiplica- Gourieroux, Monfort, and Trognon (1984).
tive models.11 Indeed, what is needed is an estimator that is The implementation of the PPML estimator is straight-
consistent and reasonably efficient under a wide range of forward: there are standard econometric programs with
heteroskedasticity patterns and is also simple to implement. commands that permit the estimation of Poisson regression,
even when the dependent variables are not integers. Because
A possible way of obtaining an estimator that is more
the assumption V[yi兩x] ⬀ E[yi兩x] is unlikely to hold, this
efficient than the standard NLS without the need to use
estimator does not take full account of the heteroskedastic-
nonparametric regression is to follow McCullagh and
ity in the model, and all inference has to be based on an
Nelder (1989) and estimate the parameters of interest using
Eicker-White (Eicker, 1963; White, 1980) robust covariance
a PML estimator based on some assumption on the func-
matrix estimator. In particular, within Stata (StataCorp.,
tional form of V[yi兩x].12 Among the many possible specifi- 2003), the PPML estimation can be executed using the
cations, the hypothesis that the conditional variance is following command:
proportional to the conditional mean is particularly appeal-
ing. Indeed, under this assumption E[yi兩x] ⫽ exp(xi␤) ⬀ poisson exporti, j ln共distij 兲
V[yi兩x], and ␤ can be estimated by solving the following set
of first-order conditions: ln Yi ln Yj 共other variables兲ij , robust
where export (or import) is measured in levels.

冘
n
关 y i ⫺ exp共xi ␤˜ 兲兴xi ⫽ 0. (9) Of course, if it were known that V[yi兩x] is a function of

i⫽1 higher powers of E[yi兩x], a more efficient estimator could be
obtained by downweighting even more the observations
Comparing equations (8) and (9), it is clear that, unlike with large conditional mean. An example of such an esti-
the NLS estimator, which is a PML estimator obtained mator is the gamma PML estimator studied by Manning and
assuming that V[yi兩x] is constant, the PML estimator based Mullahy (2001), which, like the log-linearized model, as-
on equation (9) gives the same weight to all observations, sumes that V[yi兩x] is proportional to E[yi兩x]2. The first-order
rather than emphasizing those for which exp(xi␤) is large. conditions for the gamma PML estimator are given by
This is because, under the assumption that E[yi兩x] ⬀ V[yi兩x],
冘关 y ⫺ exp共x ˇ␤兲兴 exp共 ⫺ x ˇ␤兲 x ⫽ 0.
n
all observations have the same information on the parame-
i i i i
ters of interest as the additional information on the curvature i⫽1
of the conditional mean coming from observations with
large exp(xi␤) is offset by their larger variance. Of course, In the case of trade data, however, this estimator may
this estimator may not be optimal, but without further have an important drawback. Trade data for larger countries
information on the pattern of heteroskedasticity, it seems (as gauged by GDP per capita) tend to be of higher quality
(see Frankel & Wei, 1993; Frankel, 1997); hence, models
11 A nonparametric generalized least squares estimator can also be used assuming that V[yi兩x] is a function of higher powers of
to estimate linear models in the presence of heteroskedasticity of unknown E[yi兩x] might give excessive weight to the observations that
form (Robinson, 1987). However, despite having been proposed more
than 15 years ago, this estimator has never been adopted as a standard tool
by researchers doing empirical work, who generally prefer the simplicity 13 The same strategy is implicitly used by Papke and Wooldridge (1996)
of the inefficient OLS, with an appropriate covariance matrix. in their pseudo-maximum-likelihood estimator for fractional data models.
12 See also Manning and Mullahy (2001). A related estimator is proposed 14 See Cameron and Trivedi (1998) and Winkelmann (2003) for more
by Papke and Wooldridge (1996) for the estimation of models for details on the Poisson regression and on more general models for count
fractional data. data.
are more prone to measurement errors.15 Therefore, the 共 y i ⫺ y̆ i 兲 2 ⫽ ␭ 0 共 y̆ i 兲 ␭ 1 ⫹ ␰ i , (12)

Poisson regression emerges as a reasonable compromise,
giving less weight to the observations with larger variance using an appropriate PML estimator. The approach based on
than the standard NLS estimator, without giving too much equation (12) is asymptotically valid, and inference about ␭1
weight to observations more prone to contamination by can be based on the usual Eicker-White robust covariance
measurement error and less informative about the curvature matrix estimator. For example, the hypothesis that V[yi兩x] is
of E[yi兩x].16 proportional to E[yi兩x] is accepted if the appropriate confi-
dence interval for ␭1 contains 1. However, if the purpose is
B. Testing to test the adequacy of a particular value of ␭1, a slightly
simpler method based on the Gauss-Newton regression (see
In this subsection we consider tests for the particular Davidson & MacKinnon, 1993) is available.
pattern of heteroskedasticity assumed by PML estimators, Specifically, to check the adequacy of the PPML for
focusing on the PPML estimator. Although PML estimators which ␭1 ⫽ 1 and y̆i ⫽ exp(xi␤˜ ), equation (12) can be
are consistent even when the variance function is misspeci- expanded in a Taylor series around ␭1 ⫽ 1, leading to
fied, the researcher can use these tests to check if a different
共 y i ⫺ y̆ i 兲 2 ⫽ ␭ 0 y̆ i ⫹ ␭ 0 共␭ 1 ⫺ 1兲共ln yi 兲y̆i ⫹ ␰i .

PML estimator would be more appropriate and to decide
whether or not the use of a nonparametric estimator of the
variance is warranted. Now, the hypothesis that V[yi兩x] ⬀ E[yi兩x] can be tested
Manning and Mullahy (2001) suggested that if against equation (10) simply by checking the significance of
the parameter ␭0(␭1 ⫺ 1). Because the error term ␰i is
V关 y i 兩x兴 ⫽ ␭ 0 E关 y i 兩x兴 ␭ 1, (10) unlikely to be homoskedastic, the estimation of the Gauss-
Newton regression should be performed using weighted
the choice of the appropriate PML estimator can be based on least squares. Assuming that in equation (12) the variance is
a Park-type regression (Park, 1966). Their approach is based also proportional to the mean, the appropriate weights are
on the idea that if equation (10) holds and an initial consis- given by exp(⫺xi␤˜ ), and therefore the test can be performed
tent estimate of E[yi兩x] is available, then ␭1 can be consis- by estimating
tently estimated using an appropriate auxiliary regression.
Specifically, following Park (1966), Manning and Mullahy 共 y i ⫺ y̆ i 兲 2 / 冑y̆ i ⫽ ␭ 0 冑y̆ i
(2001) suggest that ␭1 can be estimated using the auxiliary (13)
model ⫹ ␭0 共␭1 ⫺ 1兲共ln y̆i 兲冑y̆i ⫹ ␰*i
ln 共 yi ⫺ y̆i 兲2 ⫽ ln ␭0 ⫹ ␭1 ln y̆i ⫹ vi , (11) by OLS and testing the statistical significance of ␭0(␭1 ⫺ 1)
using a Eicker-White robust covariance matrix estimator.17
where y̆i denotes the estimated value of E[yi兩x]. Unfortu- In the next section, a small simulation is used to study the
nately, as the discussion in the previous sections should Gauss-Newton regression test for the hypothesis that V[yi兩x]
have made clear, this approach based on the log- ⬀ E[yi兩x], as well as the Park-type test for the hypothesis
linearization of equation (10) is valid only under very that the constant-elasticity model can be consistently esti-
restrictive conditions on the conditional distribution of yi. mated in the log linear form.
However, it is easy to see that this procedure is valid when
the constant-elasticity model can be consistently estimated IV. A Simulation Study
in the log linear form. Therefore, using equation (11) a test
for H0 : ␭1 ⫽ 2 based on a nonrobust covariance estimator This section reports the results of a small simulation
provides a check on the adequacy of the estimator based on study designed to assess the performance of different meth-
the log linear model. ods to estimate constant-elasticity models in the presence of
A more robust alternative, which is mentioned by Man- heteroskedasticity and rounding errors. As a by-product, we
ning and Mullahy (2001) in a footnote, is to estimate ␭1 also obtain some evidence on the finite-sample performance
from of the specification tests presented above. These experi-
ments are centered around the following multiplicative
15 Frankel and Wei (1993) and Frankel (1997) suggest that larger model:
countries should be given more weight in the estimation of gravity
equations. This would be appropriate if the errors in the model were just 17 Notice that to test V [ y 兩x] ⬀ E [ y 兩x] against alternatives of the form
i i
the result of measurement errors in the dependent variable. However, if it V [ yi兩x] ⫽ ␭0 exp [xi (␤ ⫹ ␭)], the appropriate auxiliary regression would
is accepted that the gravity equation does not hold exactly, measurement be
errors account for only part of the dispersion of trade data around the
gravity equation.
16 It is worth noting that the PPML estimator can be easily adapted to 共 y i ⫺ y̆ i 兲 2 / 冑y̆ i ⫽ ␭ 0 冑y̆ i ⫹ ␭ 0 ␭x i 冑y̆ i ⫹ ␰ i* ,
deal with endogenous regressors (Windmeijer & Santos Silva, 1997) and and the test could be performed by checking the joint significance of the
panel data (Wooldridge, 1999). These extensions, however, are not pur- elements of ␭0␭. If the model includes a constant, one of the regressors in
sued here. the auxiliary regression is redundant and should be dropped.
E关 y i 兩x兴 ⫽ ␮共 x i ␤兲 ⫽ exp共␤0 ⫹ ␤1 x1i ⫹ ␤2 x2i 兲, not only corrects the heteroskedasticity in the data, but,
because ␩i is log normal, it is also the maximum likelihood
i ⫽ 1, . . . , 1000. (14)
estimator. The GPML is the optimal PML estimator in this
case, but it should be outperformed by the true maximum
Because, in practice, regression models often include a
likelihood estimator. Finally, case 4 is the only one in which
mixture of continuous and dummy variables, we replicate
this feature in our experiments: x1i is drawn from a standard the conditional variance does not depend exclusively on the
normal, and x2 is a binary dummy variable that equals 1 with mean. The variance is a quadratic function of the mean, as
a probability of 0.4.18 The two covariates are independent, in case 3, but it is not proportional to the square of the mean.
and a new set of observations of all variables is generated in We carried out two sets of experiments. The first set was
each replication using ␤0 ⫽ 0, ␤1 ⫽ ␤2 ⫽ 1. Data on y are aimed at studying the performance of the estimators of the
generated as multiplicative and the log linear models under different
patterns of heteroskedasticity. In order to study the effect of
y i ⫽ ␮共 x i ␤兲␩ i , (15) the truncation on the performance of the OLS, and given
that this data-generating mechanism does not produce ob-
where ␩i is a log normal random variable with mean 1 and servations with yi ⫽ 0, the log linear model was also

variance ␴2i . As noted before, the slope parameters in equa- estimated using only the observations for which yi ⬎ 0.5
tion (14) can be estimated using the log linear form of the [OLS(y ⬎ 0.5)]. This reduces the sample size by approxi-
model only when ␴2i is constant, that is, when V[yi兩x] is mately 25% to 35%, depending on the pattern of heteroske-
proportional to ␮(xi␤)2. dasticity. The estimation of the threshold tobit was also
In these experiments we analyzed PML estimators of the performed using this dependent variable. Notice that, al-
multiplicative model and different estimators of the log- though the dependent variable has to cross a threshold to be
linearized model. The consistent PML estimators studied observable, the truncation mechanism used here is not equal
were: NLS, gamma pseudo-maximum-likelihood (GPML), to the one assumed by Eaton and Tamura (1994). Therefore,
and PPML. Besides these estimators, we also considered the in all these experiments the ET-tobit will be slightly mis-
standard OLS estimator of the log linear model (here called specified and the results presented here should be viewed as
simply OLS); the OLS estimator for the model where the a check of its robustness to this problem.
dependent variable is yi ⫹ 1 [OLS(y ⫹ 1)]; a truncated OLS The second set of experiments studied the estimators’
estimator to be discussed below; and the threshold tobit of performance in the presence of rounding errors in the
Eaton and Tamura (1994) (ET-tobit).19 dependent variable. For that purpose, a new random vari-
To assess the performance of the estimators under differ- able was generated by rounding to the nearest integer the
ent patterns of heteroskedasticity, we considered the four values of yi obtained in the first set of simulations. This
following specifications of ␴2i : procedure mimics the rounding errors in official statistics
and generates a large number of zeros, a typical feature of
Case 1: ␴2i ⫽ ␮(xi␤)⫺2; V[yi兩x] ⫽ 1. trade data. Because the model considered here generates a
Case 2: ␴2i ⫽ ␮(xi␤)⫺1; V[yi兩x] ⫽ ␮ (xi␤). large proportion of observations close to zero, rounding
Case 3: ␴2i ⫽ 1; V[yi兩x] ⫽ ␮(xi␤)2. down is much more frequent than rounding up. As the
Case 4: ␴2i ⫽ ␮(xi␤)⫺1 ⫹ exp(x2i); V[yi兩x] ⫽ ␮(xi␤) ⫹ probability of rounding up or down depends on the covari-
exp(x2i) ␮(xi␤)2. ates, this procedure will necessarily bias the estimates, as
discussed before. The purpose of the study is to gauge the
In case 1 the variance of εi is constant, implying that the magnitude of these biases. Naturally, the log linear model
NLS estimator is optimal. Although, as argued before, this cannot be estimated in these conditions, because the depen-
case is unrealistic for models of bilateral trade, it is included dent variable equals 0 for some observations. Following
in the simulations for completeness. In case 2, the condi- what is the usual practice in these circumstances, the trun-
tional variance of yi equals its conditional mean, as in the cated OLS estimation of the log-linear model was per-
Poisson distribution. The pseudo-maximum-likelihood esti- formed dropping the observations for which the dependent
mator based on the Poisson distribution is optimal in this variable equals 0. Notice that the observations discarded
situation. Case 3 is the special case in which OLS estimation with this procedure are exactly the same that are discarded
of the log linear model is consistent for the slope parameters by OLS(y ⬎ 0.5) in the first set of experiments. Therefore,
of equation (14). Moreover, in this case the log linear model
this estimator is also denoted OLS(y ⬎ 0.5).
18 For example, in gravity equations, continuous variables (which are all
The results of the two sets of experiments are summa-
strictly positive) include income and geographical distance. In equation rized in table 1, which displays the biases and standard
(14), x1 can be interpreted as (the logarithm of) one of these variables. errors of the different estimators of ␤ obtained with 10,000
Examples of binary variables include dummies for free-trade agreements, replicas of the simulation procedure described above. Only
common language, colonial ties, contiguity, and access to water.
19 We also studied the performance of other variants of the tobit model, results for ␤1 and ␤2 are presented, as these are generally the
finding very poor results. parameters of interest.
TABLE 1.—SIMULATION RESULTS UNDER DIFFERENT FORMS OF HETEROSKEDASTICITY

Results Without Rounding Error Results with Rounding Error
␤1 ␤2 ␤1 ␤2
Estimator: Bias S.E. Bias S.E. Bias S.E. Bias S.E.
Case 1: V [ yi兩x] ⫽ 1
PPML ⫺0.00004 0.016 0.00009 0.027 0.01886 0.017 0.02032 0.029

NLS ⫺0.00006 0.008 ⫺0.00003 0.017 0.00195 0.008 0.00274 0.018
GPML 0.01276 0.068 0.00754 0.082 0.10946 0.096 0.09338 0.108
OLS 0.39008 0.039 0.35568 0.054 — — — —
ET-tobit ⫺0.47855 0.030 ⫺0.47786 0.032 ⫺0.49981 0.030 ⫺0.49968 0.032
OLS( y ⬎ 0.5) ⫺0.16402 0.027 ⫺0.15487 0.038 ⫺0.22121 0.026 ⫺0.21339 0.036
OLS( y ⫹ 1) ⫺0.40237 0.014 ⫺0.37683 0.022 ⫺0.37752 0.015 ⫺0.34997 0.024
Case 2: V [ yi兩x] ⫽ ␮ (xi␤)
PPML ⫺0.00011 0.019 0.00009 0.039 0.02190 0.020 0.02334 0.041

NLS 0.00046 0.033 0.00066 0.057 0.00262 0.033 0.00360 0.057

GPML 0.00376 0.043 0.00211 0.062 0.13243 0.073 0.11331 0.087
OLS 0.21076 0.030 0.19960 0.049 — — — —
ET-tobit ⫺0.42394 0.028 ⫺0.42316 0.033 ⫺0.45518 0.028 ⫺0.45513 0.033
OLS( y ⬎ 0.5) ⫺0.17868 0.026 ⫺0.17220 0.043 ⫺0.24405 0.026 ⫺0.23889 0.040
OLS( y ⫹ 1) ⫺0.42371 0.015 ⫺0.39931 0.025 ⫺0.39401 0.016 ⫺0.36806 0.028
Case 3: V [ yi兩x] ⫽ ␮ (xi␤)2
PPML ⫺0.00526 0.091 ⫺0.00228 0.130 0.02332 0.091 0.02812 0.133

NLS 0.23539 3.066 0.07323 1.521 0.23959 3.082 0.07852 1.521
GPML ⫺0.00047 0.041 ⫺0.00029 0.083 0.17134 0.068 0.14442 0.104
OLS 0.00015 0.032 ⫺0.00003 0.064 — — — —
ET-tobit ⫺0.31908 0.044 ⫺0.32161 0.058 ⫺0.36480 0.043 ⫺0.36789 0.056
OLS( y ⬎ 0.5) ⫺0.34480 0.039 ⫺0.34614 0.064 ⫺0.41006 0.037 ⫺0.41200 0.060
OLS( y ⫹ 1) ⫺0.51804 0.021 ⫺0.50000 0.038 ⫺0.48564 0.022 ⫺0.46597 0.040
Case 4: V [ yi兩x] ⫽ ␮ (xi␤) ⫹ exp(x2i) ␮ (xi␤)2
PPML ⫺0.00696 0.103 ⫺0.00647 0.144 0.02027 0.104 0.01856 0.146

NLS 0.35139 7.516 0.08801 1.827 0.35672 7.521 0.09239 1.829
GPML 0.00322 0.057 ⫺0.00137 0.083 0.12831 0.085 0.10245 0.129
OLS 0.13270 0.039 ⫺0.12542 0.075 — — — —
ET-tobit ⫺0.29908 0.049 ⫺0.42731 0.063 ⫺0.34351 0.047 ⫺0.46225 0.060
OLS( y ⬎ 0.5) ⫺0.39217 0.042 ⫺0.41391 0.070 ⫺0.45188 0.040 ⫺0.46173 0.066
OLS( y ⫹ 1) ⫺0.51440 0.021 ⫺0.58087 0.041 ⫺0.48627 0.022 ⫺0.56039 0.044
As expected, OLS only performs well in case 3. In all this estimator, despite being consistent, leads to very poor
other cases this estimator is clearly inadequate because, results because of its erratic behavior.20 Therefore, it is clear
despite its low dispersion, it is often badly biased. More- that the loss of efficiency caused by some of the forms of
over, the sign and magnitude of the bias vary considerably. heteroskedasticity considered in these experiments is strong
Therefore, even when the dependent variable is strictly enough to render this estimator useless in practice.
positive, estimation of constant-elasticity models using the In the first set of experiments, the results of the gamma
log-linearized model cannot generally be recommended. As PML estimator are very good. Indeed, when no measure-
for the modifications of the log-linearized model designed ment error is present, the biases and standard errors of the
to deal with the zeros of the dependent variable—ET-tobit, GPML estimator are always among the lowest. However,
OLS(y ⫹ 1), and OLS(y ⬎ 0.5)—their performance is also this estimator is very sensitive to the form of measurement
very disappointing. These results clearly emphasize the error considered in the second set of experiments, consis-
need to use adequate methods to deal with the zeros in the
tently leading to sizable biases. These results, like those of
data and raise serious doubts about the validity of the results
the NLS, clearly illustrate the danger of using a PML
obtained using the traditional estimators based on the log
estimator that gives extra weight to the noisier observations.
linear model. Overall, except under very special circum-
stances, estimation based on the log-linear model cannot be As for the performance of the Poisson PML estimator, the
recommended. results are very encouraging. In fact, when no rounding
One remarkable result of this set of experiments is the error is present, its performance is reasonably good in all
extremely poor performance of the NLS estimator. Indeed,
when the heteroskedasticity is more severe (cases 3 and 4), 20 Manning and Mullahy (2001) report similar results.
cases. Moreover, although some loss of efficiency is notice- TABLE 2.—REJECTION FREQUENCIES AT THE 5% LEVEL
FOR THE TWO SPECIFICATION TESTS
able as one moves away from case 2, in which it is an
optimal estimator, the biases of the PPML are always Frequency
small.21 Moreover, the results obtained with rounded data Case GNR Test Park Test
suggest that the Poisson-based PML estimator is relatively Without Measurement Error
robust to this form of measurement error of the dependent 1 0.91980 1.00000
variable. Indeed, the bias introduced by the rounding-off 2 0.05430 1.00000
errors in the dependent variable is relatively small, and in 3 0.58110 0.06680
4 0.49100 0.40810
some cases it even compensates the bias found in the first
set of experiments. Therefore, because it is simple to im- With Measurement Error
plement and reliable in a wide variety of situations, the 1 0.91740 1.00000
2 0.14980 1.00000
Poisson PML estimator has the essential characteristics 3 0.57170 1.00000
needed to make it the new workhorse for the estimation of 4 0.47580 1.00000
constant-elasticity models.
Obviously, the sign and magnitude of the bias of the

estimators studied here depend on the particular specifica-
tion considered. Therefore, the results of these experiments logarithmic tradition. We perform the comparison of the two
cannot serve as an indicator of what can be expected in techniques using both the traditional and the Anderson–van
other situations. However, it is clear that, apart from the Wincoop (2003) specifications of the gravity equation.
Poisson PML method, all estimators will often be very For the sake of completeness, we also compare the PPML
misleading. estimates with those obtained from alternative ways re-
These experiments were also used to study the finite- searchers have used to deal with zero values for trade. In
sample performance of the Gauss-Newton regression particular, we present the results obtained with the tobit
(GNR) test for the adequacy of the Poisson PML based on estimator used in Eaton and Tamura (1994), OLS estimator
equation (13) and of the Park test advocated by Manning applied to ln(1 ⫹ Tij), and a standard nonlinear least squares
and Mullahy (2001), which, as explained above, is valid estimator. The results obtained with these estimators are
only to check for the adequacy of the estimator based on the presented for both the traditional and the Anderson–van
log linear model.22 Given that the Poisson PML estimator is Wincoop specifications.
the only estimator with a reasonable behavior under all the
cases considered, these tests were performed using residuals
A. The Data
and estimates of ␮ (xi␤) from the Poisson regression. Table
2 contains the rejection frequencies of the null hypothesis at The analysis covers a cross section of 136 countries in
the 5% nominal level for both tests in the four cases 1990. Hence, our data set consists of 18,360 observations of
considered in the two sets of experiments. In this table the bilateral export flows (136 ⫻ 135 country pairs). The list of
rejection frequencies under the null hypothesis are given in countries is reported in table A1 in the appendix. Informa-
bold. tion on bilateral exports comes from Feenstra et al. (1997).
In as much as both tests have adequate behavior under the Data on real GDP per capita and population come from the
null and reveal reasonable power against a wide range of World Bank’s (2002) World Development Indicators. Data
alternatives, the results suggest that these tests are important on location and dummies indicating contiguity, common
tools to assess the adequacy of the standard OLS estimator language (official and second languages), colonial ties (di-
of the log linear model and of the proposed Poisson PML rect and indirect links), and access to water are constructed
estimator. from the CIA’s (2002) World Factbook. The data on lan-
guage and colonial links are presented on tables A2 and A3
in the appendix.23 Bilateral distance is computed using the
V. The Gravity Equation great circle distance algorithm provided by Andrew Gray
In this section, we use the PPML estimator to quantita- (2001). Remoteness—or relative distance—is calculated as
tively assess the determinants of bilateral trade flows, un- the (log of) GDP-weighted average distance to all other
covering significant differences in the roles of various countries (see Wei, 1996). Finally, information on preferen-
measures of size and distance from those predicted by the tial trade agreements comes from Frankel (1997), comple-
mented with data from the World Trade Organization. The
21 These results are in line with those reported by Manning and Mullahy list of preferential trade agreements (and stronger forms of
(2001). trade agreements) considered in the analysis is displayed in
22 To illustrate the pitfalls of the procedure suggested by Manning and
table A4 in the appendix. Table A5 in the appendix provides
Mullahy (2001), we note that the means of the estimates of ␭1 obtained
using equation (11) in cases 1, 2, and 3 (without measurement error) were
0.58955, 1.29821, and 1.98705, whereas the true values of ␭1 in these 23 Alternative estimates based on Boisso and Ferrantino’s (1997) index
cases are, respectively, 0, 1, and 2. of language similarity are available, at request, from the authors.
TABLE 3.—THE TRADITIONAL GRAVITY EQUATION

Estimator: OLS OLS Tobit NLS PPML PPML
Dependent Variable: ln(Tij) ln(1 ⫹ Tij) ln(a ⫹ Tij) Tij Tij ⬎ 0 Tij
Log exporter’s GDP 0.938** 1.128** 1.058** 0.738** 0.721** 0.733**
(0.012) (0.011) (0.012) (0.038) (0.027) (0.027)
Log importer’s GDP 0.798** 0.866** 0.847** 0.862** 0.732** 0.741**
(0.012) (0.012) (0.011) (0.041) (0.028) (0.027)
Log exporter’s GDP per capita 0.207** 0.277** 0.227** 0.396** 0.154** 0.157**
(0.017) (0.018) (0.015) (0.116) (0.053) (0.053)
Log importer’s GDP per capita 0.106** 0.217** 0.178** ⫺0.033 0.133** 0.135**
(0.018) (0.018) (0.015) (0.062) (0.044) (0.045)
Log distance ⫺1.166** ⫺1.151** ⫺1.160** ⫺0.924** ⫺0.776** ⫺0.784**
(0.034) (0.040) (0.034) (0.072) (0.055) (0.055)
Contiguity dummy 0.314* ⫺0.241 ⫺0.225 ⫺0.081 0.202 0.193
(0.127) (0.201) (0.152) (0.100) (0.105) (0.104)
Common-language dummy 0.678** 0.742** 0.759** 0.689** 0.752** 0.746**
(0.067) (0.067) (0.060) (0.085) (0.134) (0.135)
Colonial-tie dummy 0.397** 0.392** 0.416** 0.036 0.019 0.024
(0.070) (0.070) (0.063) (0.125) (0.150) (0.150)

Landlocked-exporter dummy ⫺0.062 0.106* ⫺0.038 ⫺1.367** ⫺0.873** ⫺0.864**
(0.062) (0.054) (0.052) (0.202) (0.157) (0.157)
Landlocked-importer dummy ⫺0.665** ⫺0.278** ⫺0.479** ⫺0.471** ⫺0.704** ⫺0.697**
(0.060) (0.055) (0.051) (0.184) (0.141) (0.141)
Exporter’s remoteness 0.467** 0.526** 0.563** 1.188** 0.647** 0.660**
(0.079) (0.087) (0.068) (0.182) (0.135) (0.134)
Importer’s remoteness ⫺0.205* ⫺0.109 ⫺0.032 1.010** 0.549** 0.561**
(0.085) (0.091) (0.073) (0.154) (0.120) (0.118)
Free-trade agreement dummy 0.491** 1.289** 0.729** 0.443** 0.179* 0.181*
(0.097) (0.124) (0.103) (0.109) (0.090) (0.088)
Openness ⫺0.170** 0.739** 0.310** 0.928** ⫺0.139 ⫺0.107
(0.053) (0.050) (0.045) (0.191) (0.133) (0.131)
Observations 9613 18360 18360 18360 9613 18360
RESET test p-values 0.000 0.000 0.204 0.000 0.941 0.331
a description of the variables and displays the summary ficients differ—oftentimes significantly—from those ob-
statistics. tained using OLS. This suggests that in this case, heteroske-
dasticity (rather than truncation) is responsible for the
B. Results differences between PPML results and those of OLS using
only the observations with positive exports. Further evi-
The Traditional Gravity Equation: Table 3 presents the
dence on the importance of the heteroskedasticity is pro-
estimation outcomes resulting from the various techniques
for the traditional gravity equation. The first column reports vided by the two-degrees-of-freedom special case of
OLS estimates using the logarithm of exports as the depen- White’s test for heteroskedasticity (see Wooldridge, 2002, p.
dent variable; as noted before, this regression leaves out 127), which leads to a test statistic of 476.6 and to a p-value
pairs of countries with zero bilateral trade (only 9,613 of 0. That is, the null hypothesis of homoskedastic errors is
country pairs, or 52% of the sample, exhibit positive export unequivocally rejected.
flows). Poisson estimates reveal that the coefficients on import-
The second column reports the OLS estimates using er’s and exporter’s GDPs in the traditional equation are not,
ln(1 ⫹ Tij) as dependent variable, as a way of dealing with as generally believed, close to 1. The estimated GDP elas-
zeros. The third column presents tobit estimates based on ticities are just above 0.7 (s.e. ⫽ 0.03). OLS generates
Eaton and Tamura (1994). The fourth column shows the significantly larger estimates, especially on exporter’s GDP
results of standard NLS. The fifth column reports Poisson (0.94, s.e. ⫽ 0.01). Although all these results are conditional
estimates using only the subsample of positive-trade pairs. on the particular specification used,25 it is worth pointing out
Finally, the sixth column shows the Poisson results for the that unit income elasticities in the simple gravity framework
whole sample (including zero-trade pairs). are at odds with the observation that the trade-to-GDP ratio
The first point to notice is that PPML-estimated coeffi- decreases with increasing total GDP, or, in other words, that
cients are remarkably similar using the whole sample and smaller countries tend to be more open to international
using the positive-trade subsample.24 However, most coef- trade.26
24 The reason why truncation has little effect in this case is that 25 This result holds when one looks at the subsample of OECD countries.
observations with zero trade correspond to pairs for which the estimated It is also robust to the exclusion of GDP per capita from the regressions.
value of trade is close to zero. Therefore, the corresponding residuals are 26 Note also that PPML predicts almost equal coefficients for the GDPs
also close to zero, and their elimination from the sample has little effect. of exporters and importers.
TABLE 4.—RESULTS OF THE TESTS FOR TYPE OF HETEROSKEDASTICITY agreement dummy will not reflect the net effect of trade
(p-VALUES)
agreements. To account for the possibility of diversion, we
Test (Null Hypothesis) Exports ⬎ 0 Full Sample include an additional dummy, openness, similar to that used
GNR (V [ yi兩x] ⬀ ␮ (xi␤)) 0.144 0.115 by Frankel (1997). This dummy takes the value 1 whenever
Park (OLS is valid) 0.000 0.000 one (or both) of the countries in the pair is part of a
preferential trade agreement, and thus it captures the extent
of trade between members and nonmembers of a preferen-
The role of geographical distance as trade deterrent is
tial trade agreement. The sum of the coefficients on the trade
significantly larger under OLS; the estimated elasticity is
agreement and the openness dummies gives the net creation
⫺1.17 (s.e. ⫽ 0.03), whereas the Poisson estimate is ⫺0.78
effect of trade agreements. OLS suggests that trade destruc-
(s.e. ⫽ 0.06). This lower estimate suggests a smaller role for
tion comes from trade agreements. Still, the net creation
transport costs in the determination of trade patterns. Fur-
thermore, Poisson estimates indicate that, after controlling effect is around 40%. In contrast, Poisson regressions pro-
for bilateral distance, sharing a border does not influence vide no significant evidence of trade diversion, although the
trade flows, whereas OLS, instead, generates a substantial point estimates are of the same order of magnitude under
effect: It predicts that trade between two contiguous coun- both methods.

tries is 37% larger than trade between countries that do not Hence, even when allowing for trade diversion effects, on
share a border.27 average, the Poisson method estimates a smaller effect of
We control for remoteness to allow for the hypothesis that preferential trade agreements on trade, approximately half
larger distances to all other countries might increase bilat- of that indicated by OLS. The contrast in estimates suggests
eral trade between two countries.28 Poisson regressions that the biases generated by standard regressions can be
support this hypothesis, whereas OLS estimates suggest that substantial, leading to misleading inferences and, perhaps,
only exporter’s remoteness increases bilateral flows be- erroneous policy decisions.
tween two given countries. Access to water appears to be We now turn briefly to the results of the other estimation
important for trade flows, according to Poisson regressions; methods. OLS on ln(1 ⫹ Tij) and tobit give very close
the negative coefficients on the land-locked dummies can be estimates for most coefficients. Like OLS, they yield large
interpreted as an indication that ocean transportation is estimates for the elasticity of bilateral trade with respect to
significantly cheaper. In contrast, OLS results suggest that distance. Unlike OLS, however, they produce insignificant
whether or not the exporter is landlocked does not influence coefficients for the contiguity dummy. They both generate
trade flows, whereas a landlocked importer experiences extremely large and statistically significant coefficients for
lower trade. (These asymmetries in the effects of remote- the trade-agreement dummy. The first method predicts that
ness and access to water for importers and exporters are trade between two countries that have signed a trade agree-
hard to interpret.) We also explore the role of colonial ment is on average 266% larger than that between countries
heritage, obtaining, as before, significant discrepancies: without an agreement. The second predicts that trade be-
Poisson regressions indicate that colonial ties play no role in tween countries in such agreements is on average 100%
determining trade flows, once a dummy variable for com- larger. NLS tends to generate somewhat different estimates.
mon language is introduced. OLS regressions, instead, gen- The elasticity of trade with respect to the exporter’s GDP is
erate a sizeable effect (countries with a common colonial significantly smaller than with OLS, but the corresponding
past trade almost 45% more than other pairs). Language is elasticity with respect to importer’s GDP is significantly
statistically and economically significant under both estima- larger than with OLS. The estimated distance elasticity is
tion procedures. smaller than with OLS and bigger than with Poisson. Like
Strikingly, in the traditional gravity equation, preferential- the other methods, NLS predicts a significant and large
trade agreements play a much smaller—although still sub- effect for free-trade agreements.
stantial—role according to Poisson regressions. OLS esti- It is noteworthy that all methods, except the PPML, lead
mates suggest that preferential trade agreements raise to puzzling asymmetries in the elasticities with respect to
expected bilateral trade by 63%, whereas Poisson estimates importer and exporter characteristics (especially remoteness
indicate an average enhancement below 20%. and access to water).
Preferential trade agreements might also cause trade di- To check the adequacy of the estimated models, we
version; if this is the case, the coefficient on the trade- performed a heteroskedasticity-robust RESET test (Ramsey,
1969). This is essentially a test for the correct specification
27 The formula to compute this effect is (ebi ⫺ 1) ⫻ 100%, where b is
i of the conditional expectation, which is performed by
the estimated coefficient.
28 To illustrate the role of remoteness, consider two pairs of countries, (i, checking the significance of an additional regressor con-
j) and (k, l), and assume that the distance between the countries in each structed as (x⬘b)2, where b denotes the vector of estimated
pair is the same (Dij ⫽ Dkl), but i and j are closer to other countries. In this parameters. The corresponding p-values are reported at the
case, the most remote countries, k and l, will tend to trade more between
each other because they do not have alternative trading partners. See bottom of table 3. In the OLS regression, the test rejects the
Deardoff (1998). hypothesis that the coefficient on the test variable is 0. This
TABLE 5.—THE ANDERSON–VAN WINCOOP GRAVITY EQUATION

Estimator: OLS OLS Tobit NLS PPML PPML
Dependent variable: ln (Tij) ln (1 ⫹ Tij) ln (a ⫹ Tij) Tij Tij ⬎ 0 Tij
Log distance ⫺1.347** ⫺1.332** ⫺1.272** ⫺0.582** ⫺0.770** ⫺0.750**
(0.031) (0.036) (0.029) (0.088) (0.042) (0.041)
Contiguity dummy 0.174 ⫺0.399* ⫺0.253 0.458** 0.352** 0.370**
(0.130) (0.189) (0.135) (0.121) (0.090) (0.091)
Common-language dummy 0.406** 0.550** 0.485** 0.926** 0.418** 0.383**
(0.068) (0.066) (0.057) (0.116) (0.094) (0.093)
Colonial-tie dummy 0.666** 0.693** 0.650** ⫺0.736** 0.038 0.079
(0.070) (0.067) (0.059) (0.178) (0.134) (0.134)
Free-trade agreement dummy 0.310** 0.174 0.137** 1.017** 0.374** 0.376**
(0.098) (0.138) (0.098) (0.170) (0.076) (0.077)
Fixed effects Yes Yes Yes Yes Yes Yes
Observations 9613 18360 18360 18360 9613 18360
RESET test p-values 0.000 0.000 0.000 0.000 0.564 0.112

means that the model estimated using the logarithmic spec- As with the standard specification of the gravity equation,
ification is inappropriate. A similar result is found for the we find that, using the Anderson–van Wincoop (2003)
OLS estimated using ln(1 ⫹ Tij) as the dependent variable specification, estimates obtained with the Poisson method
and for the NLS. In contrast, the models estimated using the vary little when only the positive-trade subsample is used.
Poisson regressions pass the RESET test, that is, the RESET Moreover, we find again strong evidence that the errors of
test provides no evidence of misspecification of the gravity the log linear model estimated using the sample with posi-
equations estimated using the PPML. With this particular tive trade are heteroskedastic. With this specification, the
specification, the model estimated using tobit also passes two-degree-of-freedom special case of White’s test for het-
the test for the traditional gravity equation. eroskedasticity leads to a test statistic of 469.2 and a p-value
Finally, we also check whether the particular pattern of of 0.
heteroskedasticity assumed by the models is appropriate. As Because we are now conditioning on a much larger set of
explained in section III B, the adequacy of the log linear controls, it is not surprising to find that most coefficients are
model was checked using the Park-type test, whereas the sensitive to the introduction of fixed effects. For example, in
hypothesis V[yi兩x] ⬀ ␮(xi␤) was tested by evaluating the the Poisson method, although the distance elasticity remains
significance of the coefficient of (ln y̆i)公y̆i in the Gauss- about the same and the coefficient on common colonial ties
Newton regression indicated in equation (13). The p-values is still insignificant, the effect on common language is now
of the tests are reported in table 4. Again, the log linear smaller and the coefficient on free-trade agreements is
specification is unequivocally rejected. On the other hand, larger. The results of the other estimation methods are
these results indicate that the estimated coefficient on (ln y̆i) generally much more sensitive to the inclusion of the fixed
公y̆i is insignificantly different from 0 at the usual 5% level. effects.
This implies that the Poisson PML assumption V[yi兩x] ⫽ Comparing the results of PPML and OLS for the positive-
␭0E[yi兩x] cannot be rejected at this significance level. trade subsample, the following observations are in order.
The distance elasticity is substantially larger under OLS
The Anderson–van Wincoop Gravity Equation: Table 5 (⫺1.35 versus ⫺0.75). Sharing a border has a positive effect
presents the estimated coefficients for the Anderson–van on trade under Poisson, but no significant effect under OLS.
Wincoop (2003) gravity equation, which controls more Sharing a common language has similar effects under the
properly for multilateral resistance terms by introducing two techniques. Common colonial ties have strong effects
exporter- and importer-specific effects. As before, the col- under OLS (with an average enhancement effect of 100%),
umns show, respectively, the estimated coefficients obtained whereas Poisson predicts no significant effect. Finally, the
using OLS on the log of exports, OLS on ln(1 ⫹ Tij), tobit,
NLS, PPML on the positive-trade sample, and PPML. Note values of the product of the GDPs, implicitly assuming that the variance
that, because this exercise uses cross-sectional data, we can of the error term is proportional to the square of this product. This is
only identify bilateral variables.29 contrary to what is advocated by Frankel and Wei (1993) and Frankel
(1997), who suggest that larger countries should be given more weight in
the estimation of gravity equations because they generally have better
29 Anderson and van Wincoop (2003) impose unit income elasticities by data. In any case, whether this should be done or not is an empirical
using as the dependent variable the log of exports divided by the product question, and the right course of action depends on the pattern of
of the countries’ GDPs. Because we are working with cross-sectional data heteroskedasticity. With our data, using the ratio of exports to the product
and the model specification includes importer and exporter fixed effects, of GDPs as the dependent variable leads to models that are rejected by the
income elasticities cannot be identified, and there is no need to impose specification tests. Therefore, the implied assumptions about the pattern of
restrictions on them. Still, the estimation of the PML models could be heteroskedasticity are not supported by our data. Hence, we use exports as
performed using as the dependent variable the ratio of exports to the the dependent variable of the gravity equation, and not the ratio of exports
product of the GDPs. This would downweight the observations with large to the product of GDPs.
TABLE 6.—RESULTS OF THE TESTS FOR TYPE OF HETEROSKEDASTICITY The basic problem is that log-linearization (or, indeed, any
(p-VALUES)
nonlinear transformation) of the empirical model in the
Test (Null Hypothesis) Exports ⬎ 0 Full Sample presence of heteroskedasticity leads to inconsistent esti-
GNR (V [ yi兩x] ⬀ ␮ (xi␤)) 0.100 0.070 mates. This is because the expected value of the logarithm
Park (OLS is valid) 0.000 0.000 of a random variable depends on higher-order moments of
its distribution. Therefore, if the errors are heteroskedastic,
two techniques produce reasonably similar estimates for the the transformed errors will be generally correlated with the
coefficient on the trade-agreement dummy, implying a trade- covariates. An additional problem of log-linearization is that
enhancement effect of the order of 40%. it is incompatible with the existence of zeros in trade data,
As before, the other estimation methods lead to some which led to several unsatisfactory solutions, including
puzzling results. For example, OLS on ln(1 ⫹ Tij) now truncation of the sample (that is, elimination of zero-trade
yields a significantly negative effect of contiguity, and under pairs) and further nonlinear transformations of the depen-
NLS, the coefficient on common colonial ties becomes dent variable.
significantly negative. We argue that the biases are present both in the traditional
To complete the study, we performed the same set of specification of the gravity equation and in the Anderson–

specification tests used before. The p-values of the het- van Wincoop (2003) specification, which includes country-
eroskedasticity-robust RESET test at the bottom of table 5 specific fixed effects.
suggest that with the Anderson–van Wincoop (2003) spec- To address the various estimation problems, we propose
ification of the gravity equation, only the models estimated a simple Poisson pseudo-maximum-likelihood method and
by the PPML method are adequate. The p-values of the tests assess its performance using Monte Carlo simulations. We
to check whether the particular pattern of heteroskedasticity find that in the presence of heteroskedasticity the standard
assumed by the models is appropriate are reported in table methods can severely bias the estimated coefficients, cast-
6. As in the traditional gravity equation, the log linear ing doubt on previous empirical findings. Our method,
specification is unequivocally rejected. On the other hand, instead, is robust to different patterns of heteroskedasticity
these results indicate that the estimated coefficient on (ln y̆i) and, in addition, provides a natural way to deal with zeros in
公y̆i is insignificantly different from 0 at the usual 5% level. trade data.
This implies that the Poisson PML assumption V[yi兩x] ⫽ We use our method to reestimate the gravity equation and
␭0E[yi兩x] cannot be rejected at this significance level. document significant differences from the results obtained
To sum up, whether or not fixed effects are used in the using the log linear method. For example, income elastici-
specification of the model, we find strong evidence that ties in the traditional gravity equation are systematically
estimation methods based on the log-linearization of the smaller than those obtained with log-linearized OLS regres-
gravity equation suffer from severe misspecification, which sions. In addition, in both the traditional and Anderson–van
hinders the interpretation of the results. NLS is also gener- Wincoop specifications of the gravity equation, OLS esti-
ally unreliable. In contrast, the models estimated by PPML mation exaggerates the role of geographical proximity and
show no signs of misspecification and, in general, do not colonial ties. RESET tests systematically favor the Poisson
produce the puzzling results generated by the other meth- PML technique. The results suggest that heteroskedasticity
ods.30 (rather than truncation of the data) is responsible for the
main differences.
VI. Conclusions Log-linearized equations estimated by OLS are of course
In this paper, we argue that the standard empirical meth- used in many other areas of empirical economics and econo-
ods used to estimate gravity equations are inappropriate. metrics. Our Monte Carlo simulations and the regression out-
comes indicate that in the presence of heteroskedasticity this
30 It is worth noting that the large differences in estimates among the practice can lead to significant biases. These results suggest
various methods persist when we look at a smaller subsample of countries that, at least when there is evidence of heteroskedasticity, the
that account for most of world trade and, quite likely, have better data. Poisson pseudo-maximum-likelihood estimator should be used
More specifically, we run similar regressions for the subsample of 63
countries included in Frankel’s (1997) study. These countries accounted as a substitute for the standard log linear model.
for almost 90% of the world trade reported to the United Nations in 1992.
One advantage of this subsample is that the number of zeros is signifi-
cantly reduced. Heteroskedasticity, however, is still a problem: The null REFERENCES
hypothesis of homoskedasticity is rejected in both the traditional and the
fixed-effects gravity equations. As with the full sample, PPML generates Anderson, J., “A Theoretical Foundation for the Gravity Equation,”
a smaller role for distance and common language than OLS, and, unlike American Economic Review 69 (1979), 106–116.
OLS, PPML predicts no role for colonial ties. In line with the findings Anderson, J., and E. van Wincoop, “Gravity with Gravitas: A Solution to
documented in Frankel (1997), the OLS estimated coefficient on the the Border Puzzle,” American Economic Review 93 (2003), 170–
free-trade-agreement dummy is negative in both specifications of the 192.
gravity equation, whereas PPML predicts a positive and significant effect Bergstrand, J., “The Gravity Equation in International Trade: Some
(slightly bigger than that found for the whole sample). These results are Microeconomic Foundations and Empirical Evidence,” this RE-
available—on request—from the authors. VIEW, 69 (1985), 474–481.
Boisso, D., and M. Ferrantino, “Economic and Cultural Distance in Hallak, J. C., “Product Quality and the Direction of Trade,” Journal of
International Trade: Empirical Puzzles,” Journal of Economic International Economics 68 (2006), 238–265.
Integration 12 (1997), 456–484. Harrigan, J., “OECD Imports and Trade Barriers in 1983,” Journal of
Cameron, A. C., and P. K. Trivedi, Regression Analysis of Count Data International Economics 35 (1993), 95–111.
(Cambridge: Cambridge University Press, 1998). Haveman, J., and D. Hummels, “Alternative Hypotheses and the Volume
Central Intelligence Agency, World Factbook, http://www.cia.gov/cia/ of Trade: The Gravity Equation and the Extent of Specialization,”
publications/factbook/ (2002). Purdue University mimeograph, (2001).
Davidson, R., and J. G. MacKinnon, Estimation and Inference in Econo- Helpman, E., and P. Krugman, Market Structure and Foreign Trade
metrics (Oxford: Oxford University Press, 1993). (Cambridge, MA: MIT Press, 1985).
Davis, D., “Intra-industry Trade: A Hecksher-Ohlin-Ricardo Approach,” Helpman, E., M. Melitz, and Y. Rubinstein, “Trading Patterns and Trading
Journal of International Economics 39 (1995), 201–226. Volumes,” Harvard University mimeograph (2004).
Deardoff, A., “Determinants of Bilateral Trade: Does Gravity Work in a Koenker, R., and G. S. Bassett, Jr., “Regression Quantiles,” Econometrica
Neoclassical World?” in Jeffrey Frankel (Ed.), The Regionalization 46 (1978), 33–50.
of the World Economy (Chicago: University of Chicago Press, Manning, W. G., and J. Mullahy, “Estimating Log Models: To Transform
1998). or Not to Transform?” Journal of Health Economics 20 (2001),
Delgado, M., “Semiparametric Generalized Least Squares Estimation in 461–494.
the Multivariate Nonlinear Regression Model,” Econometric The- McCallum, J., “National Borders Matter: Canada-US Regional Trade
ory 8 (1992), 203–222. Patterns,” American Economic Review 85 (1995), 615–623.
Delgado, M., and T. J. Kniesner, “Count Data Models with Variance of McCullagh, P., and J. A. Nelder, Generalized Linear Models, 2nd ed.
Unknown Form: An Application to a Hedonic Model of Worker (London: Chapman and Hall, 1989).

Absenteeism,” this REVIEW, 79 (1997), 41–49. Papke, L. E., and J. M. Wooldridge, “Econometric Methods for Fractional
Eaton, J., and S. Kortum, “Technology, Geography and Trade,” NBER Response Variables with an Application to 401(k) Plan Participa-
working paper no. 6253 (2001). tion Rates,” Journal of Applied Econometrics 11 (1996), 619–632.
Eaton, J., and A. Tamura, “Bilateralism and Regionalism in Japanese and Park, R., “Estimation with Heteroskedastic Error Terms,” Econometrica
US Trade and Direct Foreign Investment Patterns,” Journal of the 34 (1966), 888.
Japanese and International Economics 8 (1994), 478–510. Ramsey, J. B., “Tests for Specification Errors in Classical Linear Least
Eichengreen, B., and D. Irwin, “Trade Blocs, Currency Blocs, and the
Squares Regression Analysis,” Journal of the Royal Statistical
Reorientation of World Trade in the 1930’s,” Journal of Interna-
Society B 31 (1969), 350–371.
tional Economics, 38 (1995), 1–24.
Robinson, P. M., “Asymptotically Efficient Estimation in the Presence of
Eicker, F., “Asymptotic Normality and Consistency of the Least Squares
Estimators for Families of Linear Regressions,” The Annals of Heteroskedasticity of Unknown Form,” Econometrica 55 (1987),
Mathematical Statistics 34 (1963), 447–456. 875–891.
Feenstra, R. C., R. E. Lipsey, and H. P. Bowen, “World Trade Flows, Rose, A., “One Money One Market: Estimating the Effect of Common
1970–1992, with Production and Tariff Data,” NBER working Currencies on Trade,” Economic Policy 15 (2000), 7–46.
paper no. 5910 (1997). StataCorp., Stata Statistical Software: Release 8 (College Station, TX:
Feenstra, R., J. Markusen, and A. Rose, “Using the Gravity Equation to StataCorp LP, 2003).
Differentiate among Alternative Theories of Trade,” Canadian Tenreyro, S., and R. Barro, “Economic Effects of Currency Unions,” FRB
Journal of Economics 34 (2001), 430–447. Boston series working paper no. 02–4 (2002).
Frankel, J., Regional Trading Blocs in the World Economic System Tinbergen, J., The World Economy. Suggestions for an International
(Washington, DC: Institute for International Economics, 1997). Economic Policy (New York: Twentieth Century Fund, 1962).
Frankel, J., and A. Rose, “An Estimate of the Effect of Common Curren- Wei, S., “Intra-national versus International Trade: How Stubborn Are
cies on Trade and Income,” Quarterly Journal of Economics 117 Nation States in Globalization?” NBER working paper no. 5331
(2002), 409–466. (1996).
Frankel, J., E. Stein, and S. Wei, “Continental Trading Blocs: Are They White, H., “A Heteroskedasticity-Consistent Covariance Matrix Estimator
Natural, or Super-Natural?” in J. Frankel (Ed.), The Regionaliza- and a Direct Test for Heteroskedasticity,” Econometrica 48 (1980),
tion of the World Economy (Chicago: University of Chicago Press, 817–838.
1998). Winkelmann, R., Econometric Analysis of Count Data, 4th ed. (Berlin:
Frankel, J., and S. Wei, “Trade Blocs and Currency Blocs,” NBER Springer-Verlag, 2003).
working paper no. 4335 (1993). Windmeijer, F., and J. M. C. Santos Silva, “Endogeneity in Count Data
Goldberger, A. “The Interpretation and Estimation of Cobb-Douglas Models: An Application to Demand for Health Care,” Journal of
Functions,” Econometrica 36 (1968), 464–472. Applied Econometrics 12 (1997), 281–294.
Goldberger, A., A Course in Econometrics (Cambridge, MA: Harvard Wooldridge, J. M., “Distribution-Free Estimation of Some Nonlinear
University Press, 1991). Panel Data Models,” Journal of Econometrics 90 (1999), 77–97.
Gourieroux, C., A. Monfort, and A. Trognon, “Pseudo Maximum Likeli- Econometric Analysis of Cross Section and Panel Data (Cam-
hood Methods: Applications to Poisson Models,” Econometrica 52 bridge, MA: MIT Press, 2002).
(1984), 701–720. World Bank, World Development Indicators CD-ROM (The World Bank,
Gray, A., http://argray.fateback.com/dist/formula.html (2001). 2002).
APPENDIX
TABLE A1.—LIST OF COUNTRIES

Albania Denmark Kenya Romania
Algeria Djibouti Kiribati Russian Federation
Angola Dominican Rep. Korea, Rep. Rwanda
Argentina Ecuador Laos P. Dem. Rep. Saudi Arabia
Australia Egypt Lebanon Senegal
Austria El Salvador Madagascar Seychelles
Bahamas Eq. Guinea Malawi Sierra Leone
Bahrain Ethiopia Malaysia Singapore
Bangladesh Fiji Maldives Solomon Islands
Barbados Finland Mali South Africa
Belgium-Lux. France Malta Spain
Belize Gabon Mauritania Sri Lanka
Benin Gambia Mauritius St. Kitts and Nevis
Bhutan Germany Mexico Sudan
Bolivia Ghana Mongolia Suriname
Brazil Greece Morocco Sweden

Brunei Guatemala Mozambique Switzerland
Bulgaria Guinea Nepal Syrian Arab Rep.
Burkina Faso Guinea-Bissau Netherlands Tanzania
Burundi Guyana New Caledonia Thailand
Cambodia Haiti New Zealand Togo
Cameroon Honduras Nicaragua Trinidad and Tobago
Canada Hong Kong Niger Tunisia
Central African Rep. Hungary Nigeria Turkey
Chad Iceland Norway Uganda
Chile India Oman United Arab Em.
China Indonesia Pakistan United Kingdom
Colombia Iran Panama United States
Comoros Ireland Papua New Guinea Uruguay
Congo Dem. Rep. Israel Paraguay Venezuela
Congo Rep. Italy Peru Vietnam
Costa Rica Jamaica Philippines Yemen
Côte D’Ivoire Japan Poland Zambia
Cyprus Jordan Portugal Zimbabwe
TABLE A2.—COMMON OFFICIAL AND SECOND LANGUAGES

English French Spanish Dutch
Australia Belgium-Lux. Argentina Belgium-Lux.

Bahamas Benin Belize Netherlands
Barbados Burkina Faso Bolivia Suriname
Belize Burundi Chile
Brunei Cameroon Colombia German
Cameroon Canada Costa Rica
Canada Central African Rep. Dominican Rep. Austria
Fiji Chad Ecuador Germany
Gambia Comoros El Salvador Switzerland
Ghana Congo Dem. Rep. Eq. Guinea
Guyana Congo Rep. Guatemala Greek
Hong Kong Côte D’Ivoire Honduras
India Djibouti Mexico Cyprus
Indonesia Eq. Guinea Nicaragua Greece
Ireland France Panama
Israel Gabon Paraguay Hungarian
Jamaica Guinea Peru

Jordan Haiti Spain Hungary
Kenya Lebanon Uruguay Romania
Kiribati Madagascar Venezuela
Malawi Mali Italian
Malaysia Mauritania Arabic
Maldives Mauritius Italy
Malta Morocco Algeria Switzerland
Mauritius New Caledonia Bahrain
New Zealand Niger Chad Lingala
Nigeria Rwanda Comoros
Oman Senegal Djibouti Congo Dem. Rep.
Pakistan Seychelles Egypt Congo Rep.
Panama Switzerland Israel
Papua New Guinea Togo Jordan Russian
Philippines Tunisia Lebanon
Rwanda Mauritania Mongolia
Seychelles Malay Morocco Russian Federation
Sierra Leone Oman
Singapore Brunei Saudi Arabia Swahili
South Africa Indonesia Sudan
Sri Lanka Malaysia Syria Kenya
St. Helena Singapore Tanzania Tanzania
St. Kitts and Nevis Tunisia
Suriname Portuguese United Arab Em. Chinese
Tanzania Yemen
Trinidad and Tobago Angola China
Uganda Brazil Turkish Hong Kong
United Kingdom Guinea-Bissau Malaysia
United States Mozambique Cyprus Singapore
Zambia Portugal Turkey
Zimbabwe
TABLE A3.—COLONIAL TIES

United Kingdom France Spain
Australia Mauritius Algeria Argentina

Bahamas New Zealand Benin Bolivia
Bahrain Nigeria Burkina Faso Chile
Barbados Pakistan Cambodia Colombia
Belize Saint Kitts and Nevis Cameroon Costa Rica
Cameroon Seychelles Central African Rep. Cuba
Canada Sierra Leone Chad Ecuador
Cyprus South Africa Comoros El Salvador
Egypt Sri Lanka Congo Eq. Guinea
Fiji Sudan Djibouti Guatemala
Gambia Tanzania Gabon Honduras
Ghana Trinidad and Tobago Guinea Mexico
Guyana Uganda Haiti Netherlands
India United States Laos Nicaragua
Ireland Zambia Lebanon Panama
Israel Zimbabwe Madagascar Paraguay
Jamaica Mali Peru

Jordan Mauritania Venezuela
Kenya Morocco Portugal
Kuwait Niger
Malawi Senegal Angola
Malaysia Syria Brazil
Maldives Togo Guinea-Bissau
Malta Tunisia Mozambique
Vietnam Oman
TABLE A4.—PREFERENTIAL TRADE AGREEMENTS IN 1990

EEC/EC CARICOM CACM
Belgium Bahamas Costa Rica

Denmark Barbados El Salvador
France Belize Guatemala
Germany Dominican Rep. Honduras
Greece Guyana Nicaragua
Ireland Haiti
Italy Jamaica Bilateral Agreements
Luxembourg Trinidad and Tobago
Netherlands St. Kitts and Nevis EC-Cyprus
Portugal Suriname EC-Malta
Spain EC-Egypt
United Kingdom SPARTECA EC-Syria
EC-Algeria
EFTA Australia EC-Norway
New Zealand EC-Iceland
Iceland Fiji EC-Switzerland
Norway Kiribati Canada–United States
Switzerland Papua New Guinea Israel–United States
Liechtenstein Solomon Islands
CER PATCRA
Australia Australia
New Zealand Papua New Guinea
TABLE A5.—SUMMARY STATISTICS

Full Sample Exports ⬎ 0
Variable Mean Std. Dev. Mean Std. Dev.
Trade 172132.2 1828720 328757.7 2517139
Log of trade — — 8.43383 3.26819
Log of exporter’s GDP 23.24975 2.39727 24.42503 2.29748
Log of importer’s GDP 23.24975 2.39727 24.13243 2.43148
Log of exporter’s per capita GDP 7.50538 1.63986 8.09600 1.65986
Log of importer’s per capita GDP 7.50538 1.63986 7.98602 1.68649
Log of distance 8.78551 0.74168 8.69497 0.77283
Contiguity dummy 0.01961 0.13865 0.02361 0.15185
Common-language dummy 0.20970 0.40710 0.21284 0.40933
Colonial-tie dummy 0.17048 0.37606 0.16894 0.37472
Landlocked exporter dummy 0.15441 0.36135 0.10767 0.30998
Landlocked importer dummy 0.15441 0.36135 0.11401 0.31784
Exporter’s remoteness 8.94654 0.26389 8.90383 0.29313
Importer’s remoteness 8.94654 0.26389 8.90787 0.28412
Preferential–trade-agreement dummy 0.02505 0.15629 0.04452 0.20626

Openness dummy 0.56373 0.49594 0.65796 0.47442

Rest 88 4 641

Uploaded by

Copyright:

Available Formats

Rest 88 4 641

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Rest 88 4 641

Uploaded by

Copyright:

Available Formats

THE LOG OF GRAVITY

J. M. C. Santos Silva and Silvana Tenreyro*

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

The Review of Economics and Statistics, November 2006, 88(4): 641–658

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

y i ⫽ exp共xi ␤兲 ⫹ εi , (6) It may be surprising that the pattern of heteroskedasticity

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

To obtain a consistent estimator of the slope parameters A. Estimation

conditional variance of yi, tends to vanish as E[yi兩x] passes i

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

where export (or import) is measured in levels.

关 y i ⫺ exp共xi ␤˜ 兲兴xi ⫽ 0. (9) Of course, if it were known that V[yi兩x] is a function of

are more prone to measurement errors.15 Therefore, the 共 y i ⫺ y̆ i 兲 2 ⫽ ␭ 0 共 y̆ i 兲 ␭ 1 ⫹ ␰ i , (12)

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

TABLE 1.—SIMULATION RESULTS UNDER DIFFERENT FORMS OF HETEROSKEDASTICITY

PPML ⫺0.00004 0.016 0.00009 0.027 0.01886 0.017 0.02032 0.029

Case 2: V [ yi兩x] ⫽ ␮ (xi␤)

PPML ⫺0.00011 0.019 0.00009 0.039 0.02190 0.020 0.02334 0.041

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

Case 3: V [ yi兩x] ⫽ ␮ (xi␤)2

PPML ⫺0.00526 0.091 ⫺0.00228 0.130 0.02332 0.091 0.02812 0.133

Case 4: V [ yi兩x] ⫽ ␮ (xi␤) ⫹ exp(x2i) ␮ (xi␤)2

PPML ⫺0.00696 0.103 ⫺0.00647 0.144 0.02027 0.104 0.01856 0.146

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

TABLE 3.—THE TRADITIONAL GRAVITY EQUATION

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

TABLE 5.—THE ANDERSON–VAN WINCOOP GRAVITY EQUATION

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

TABLE A1.—LIST OF COUNTRIES

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

TABLE A2.—COMMON OFFICIAL AND SECOND LANGUAGES

Australia Belgium-Lux. Argentina Belgium-Lux.

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

TABLE A3.—COLONIAL TIES

Australia Mauritius Algeria Argentina

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

TABLE A4.—PREFERENTIAL TRADE AGREEMENTS IN 1990

Belgium Bahamas Costa Rica

TABLE A5.—SUMMARY STATISTICS

Downloaded from http://direct.mit.edu/rest/article-pdf/88/4/641/1982805/rest.88.4.641.pdf by guest on 16 July 2024

You might also like