Mage man, Glenn; S udnicka, Zuzanna; Van Ho e, Jan
Wo king Pape
Dis ance and bo de e ec s in in e na ional ade: A
compa ison o es ima ion me hods
Economics Discussion Pape s, No. 2015-69
P o ided in Coope a ion wi h:
Kiel Ins i u e o he Wo ld Economy – Leibniz Cen e o Resea ch on Global Economic Challenges
Sugges ed Ci a ion: Mage man, Glenn; S udnicka, Zuzanna; Van Ho e, Jan (2015) : Dis ance and
bo de e ec s in in e na ional ade: A compa ison o es ima ion me hods, Economics Discussion
Pape s, No. 2015-69, Kiel Ins i u e o he Wo ld Economy (I W), Kiel
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/125169
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h p://c ea i ecommons.o g/licenses/by/3.0/
Recei ed Decembe 4, 2015 Accep ed as Economics Discussion Pape Decembe 7, 2015
Published Decembe 17, 2015
© Au ho (s) 2015. Licensed unde he C ea i e Commons License - A ibu ion 3.0
Discussion Pape
No. 2015-69 | Decembe 17, 2015 | h p://www.economics-ejou nal.o g/economics/discussionpape s/2015-69
Dis ance and Bo de E ec s in In e na ional T ade:
A Compa ison o Es ima ion Me hods
Glenn Mage man, Zuzanna S udnicka, and Jan Van Ho e
Abs ac
This pape compa es a ious es ima ion echniques used o de e mine he impac o dis ance and
bo de s on in e na ional ade. The esul s consis en ly con i m he signi ican ly nega i e dis ance
e ec , while he bo de e ec , measu ed by e alua ing whe he in a-con inen al ade exceeds
in e -con inen al ade, appea s o be ambiguous and dependen on he es ima ion me hod. In
addi ion, also he size o bo h e ec s a ies subs an ially ac oss es ima ion me hods. Finally, he
au ho s gene ally ind ha he es ima ions a e in line wi h he espec i e weigh ing schemes o
each es ima ion me hod.
(Published in Special Issue Dis ance and Bo de E ec s in Economics)
JEL F10 F14
Keywo ds Dis ance e ec ; bo de e ec
Au ho s
Glenn Mage man, KU Leu en – Cen e o Economic S udies (CES) & Leu en Cen e
o I ish S udies Add ess: KU Leu en – Uni e si y o Leu en, Facul y o Business and
Economics, Cen e o Economic S udies, Naamses aa 69, 3000 Leu en, Belgium,
[email p o ec ed]
Zuzanna S udnicka, KU Leu en – Cen e o Economic S udies (CES) & Leu en Cen e o I ish
S udies, Belgium
Jan Van Ho e, KU Leu en – Cen e o Economic S udies (CES) & Leu en Cen e o I ish
S udies, Belgium, & INFER
Glenn Mage man g a e ully acknowledges inancial suppo om he USA Fulb igh Founda ion Belgium-Luxembou g,
he O le -La Fon aine schola ship g an ed by Y es Mo eau a he Depa men o Enginee ing a he Uni e si y o Leu en,
he NBB 2014 Sponso ship p og am a he Na ional Bank o Belgium and he Junio Mobili y g an by he Uni e si y
o Leu en. Zuzanna S udnicka g a e ully acknowledges inancial suppo om FWO Flande s (Resea ch G an
G.0365.10N) as well as om he Flemish Go e nmen . The au ho s acknowledge use ul eedback on p e ious e sions by
Jo is Tielens, Janez K en, Jan-Pie e Laleman, Sophie Soe e, Gee Dhaene, Huw Edwa ds and wo anonymous e e ees.
Ci a ion Glenn Mage man, Zuzanna S udnicka, and Jan Van Ho e (2015). Dis ance and Bo de E ec s in
In e na ional T ade: A Compa ison o Es ima ion Me hods. Economics Discussion Pape s, No 2015-69, Kiel Ins i u e o
he Wo ld Economy. h p://www.economics-ejou nal.o g/economics/discussionpape s/2015-69
1 In oduc ion
“The e is e y li le ha economis s ully unde s and abou global ade bu
he e is one hing ha we do know – comme ce declines d ama ically wi h he
dis ance” (Leame , 2007).The nega i e impac o dis ance on ade is indeed
one o he mos obus indings in in e na ional economics (see e.g., Leame
1993; F ankel 1997a; Disdie and Head, 2008). T ade is no only educed by
dis ance, bu also by in e na ional bo de s (see e.g., McCallum, 1995; Wei,
1996; Ande son and an Wincoop, 2003; Obs eld and Rogo , 2001; Coughlin
and No y, 2012). Adjacen coun ies ade mo e han non-adjacen ones (see
e.g., Leame 1993; Helliwell, 1997), leading o he so-called adjacency o
con ingency e ec . Consequen ly in a-na ional/con inen al ade exceeds
in e -na ional/con inen al ade.
These well-es ablished empi ical esul s appea in he es ima es o he
so-called g a i y equa ion based on a ious kinds o ade models, ei he
assuming pe ec compe i ion (e.g., Ande son, 1979; Dea do , 1995; Ea on
and Ko um, 2002), monopolis ic compe i ion (e.g., Be gs and, 1989, 1990)
o a demand sys em wi h anslog p e e ences (No y, 2013).
The impo ance o dis ance and bo de s as de e minan s o ade lows
may be explained by he a ie y o ba ie s o ade hey e lec . Dis ance
and bo de e ec s accoun no only o he geog aphical ba ie s be ween wo
ading pa ne s, bu also o a ious cos s ade s may incu when anspo -
ing a good o i s inal consume (see e.g., Ande son and an Wincoop, 2004,
o su eys on he ela ionship be ween physical dis ance and ade cos s).
Mo eo e , as a gued by Blum and Gol a b (2006), dis ance may also cap u e
consume s’ as es, since i educes ade e en in online p oduc s whe e ade
cos s should be ze o.
Theo e ical wo k a gues ha he magni ude o he dis ance e ec should
be equal o one. Empi ical e idence by and la ge con i ms his heo e ical
p edic ion. In gene al, acco ding o he me a-analysis by Disdie and Head
(2008), based on 1,467 es ima es om 103 pape s, he size o he dis ance
e ec is close o 0.9. Howe e , he es ima ed magni ude o he dis ance e ec
a ies depending on he coun ies o pe iods s udied. In addi ion, because o
inc easing globaliza ion and ad ances in anspo echnology, he wo ld is
sh inking. The e o e, one may expec ha he dis ance coe icien dec eases
o e ime. Empi ical s udies measu ing he e olu ion o ade elas ici y wi h
espec o dis ance a e, howe e , no conclusi e. Some au ho s ind li le
change in he ade elas ici y o dis ance (see e.g., Leame 1993). Also Disdie
and Head (2008) a gue ha he dis ance e ec is a he cons an a e a ise
a ound mid wen ie h cen u y. F ankel (1997a), Soloaga and Win e s (2001),
Be helon and F eund (2008), among o he s, ob ain e idence o an inc easing
2
dis ance e ec , whe eas Boisso and Fe an ino (1993), Eicheng een and I win
(1998), B un e al. (2005), Felbe may and Kohle (2006), Coe e al. (2007),
amongs o he s, obse e a nega i e e olu ion in he dis ance e ec o e ime.
The e a e se e al possible explana ions o hese con adic o y esul s. Fo
example, B un e al. (2005) a gue ha in as uc u e is esponsible o he
decline o he dis ance e ec . Acco ding o Felbe may and Kohle (2006),
he non-dec easing dis ance e ec ound in p e ious s udies, can be explained
by he ac ha hese s udies do no ake in o accoun he ex ensi e ma gin
o ade. Finally, Be helon and F eund (2008) show ha he inc ease o
he o e all dis ance coe icien is due o he changes o dis ance coe icien s
ac oss indus ies. They explo e wo possible easons o hese changes. Fi s ,
in some indus ies, goods ha e become mo e subs i uable. Second, ade cos s
ha e changed oo. The au ho a gues ha he i s phenomenon is he mos
impo an one.
The empi ical li e a u e on bo de e ec s was inspi ed by he seminal
wo k o McCallum (1995), who shows ha Canadian p o inces ade up o 22
imes mo e wi h each o he han wi h US s a es ( he so-called “home bias”).
This inding was con i med, o a longe ime pe iod by Helliwell (1996) and
Helliwell and McCallum (1995). Simila ly, Wei (1996) inds ha OECD
coun ies buy abou 2.5 imes mo e om hemsel es han om iden ical
o eign coun ies1Helliwell (1997) poin s o an e en la ge bo de e ec , bu
i is app oxima ely hal ed by o coun ies sha ing a common bo de and
common language. Following a simila app oach as Wei (1996) and Helliwell
(1997), Ni sch (2000) inds ha domes ic ade wi hin a Eu opean Union
coun y is se en o en imes la ge han ade wi h ano he Eu opean Union
coun y.2Finally, no e ha mos o hese s udies obse e a ade inc easing
e ec o adjacen coun ies.
These indings, and in pa icula he inding by McCallum (1995), we e
e isi ed by Ande son and an Wincoop (2003) who show ha he spec-
acula ly high bo de e ec s come om omi ing he mul ila e al esis ance
e m in McCallum’s speci ica ion and om he small size o he Canadian
economy. Mo eo e , al hough mos s udies ollowing McCallum (1995) in-
clude a mul ila e al esis ance e m in he o m o a emo eness a iable, hey
s ill do no accoun o na ional bo de ba ie s. Thus, Ande son and an
Wincoop (2003) show ha he inclusion o he mul ila e al esis ance e m
1Wei (1996) assumes ha a coun y’s pu chases om i sel equal he di e ence be-
ween i s p oduc ion and expo s. In his analysis, he assumes ha he in e nal dis ance
equals hal o he dis ance om he coun y’s economic cen e o he bo de o he nea es
neighbou .
2No e ha Helliwell (1996, 1997), Wei (1996) and Ni sch (2000) employ he me hod o
seemingly un ela ed eg essions (SUR).
3
conside ably educes McCallum’s a io o in e -p o incial ade o p o ince-
s a e ade ( om 16.4 o 10.7). Mo eo e , when using US da a ins ead o
Canadian da a, hey ind ha ade be ween s a es exceeds ade be ween
s a es and p o inces only by a ac o 15. Finally, hey also ind ha bo de s
educe ade be ween he US and Canada by 44 pe cen and among o he
indus ialized coun ies by 29 pe cen .
In e es ingly, as demons a ed by Wol (2000), he home bias exis s no
only a he in e na ional le el, bu also a he in ana ional le el. Acco ding
o his au ho , ade be ween US s a es is abou h ee imes lowe han
ade wi hin s a es.3Adjacen s a es ade 2.6 imes mo e wi h each o he .
In addi ion, he dis ance coe icien is simila o he coe icien s ound o
in e na ional ade. Hillbe y and Hummels (2003) explain he inding o
Wol (2000) by he impo ance o wholesale ac i i ies o in a-s a e ade.
Mo e ecen ly, Coughlin and No y (2012) compa e in e na ional bo de s
wi h domes ic bo de s. Mo e p ecisely, hey compa e ade be ween and
wi hin indi idual US s a es wi h ade be ween s a es and o eign coun ies.
They ind ha a s a e’s bo de is a la ge ade ba ie han an in e na ional
US bo de . One o possible explana ion is ela ed o Hillbe y and Hummels
(2008) who ind ha ade wi hin he US is hea ily concen a ed a he local
le el. T ade wi hin a single ZIP code is on a e age h ee imes highe han
ade wi h pa ne s ou side he ZIP code. Hillbe y and Hummels (2008)
explain hei inding by co-loca ion o p oduce s in supply chains o exploi
in o ma ional spillo e s, minimize anspo a ion cos s and acili a e jus -in-
ime p oduc ion. Acco ding o Coughlin and No y (2012), p oduce s also
concen a e in o de o bene i om ex e nal economies o scale in he p es-
ence o in e media e goods and associa ed agglome a ion e ec s (see e.g.,
Rossi-Hansbe g, 2005), as well as om he hub-and-spoke dis ibu ion sys-
ems and wholesale shipmen s (see, Hillbe y and Hummels, 2003). I means
ha he domes ic bo de e ec e lec s he local concen a ion o economic
ac i i y a he han ade ba ie s associa ed wi h c ossing a s a e bo de .
F om his li e a u e e iew i appea s ha he sensi i i y o he dis ance
and bo de e ec s in ade ha e been es ed o a ious coun ies, egions
and pe iods. So a , he sensi i i y o hese e ec s o he applied es ima ion
me hods has no been es ed ye in a consis en manne . This pape aims
o ill his gap. The emainde o he pape is o ganized as ollows. In he
nex sec ion we discuss he main econome ic app oaches mainly o ecen ly
ollowed in he g a i y li e a u e. In he hi d sec ion we p esen he da a and
ou empi iical app oach. Sec ion 4 discusses he esul s om applying a ious
econome ic echniques measu ing dis ance and bo de e ec s. Sec ion 5
3No e ha he does no ake in o accoun he mul ila e al esis ance e m.
4

p esen s some obus ness checks. Sec ion 6 concludes.
2 The Econome ics o G a i y
While he ea lies implemen a ion o he g a i y model in in e na ional ade
was jus an in ui i e copy o i s coun e pa in physics, mos models o in-
e na ional ade now de i e an agg ega e bila e al demand sys em ha can
be w i en as a o m o he o iginal g a i y equa ion. Following he no a ion
o Head and Maye (2014), we w i e he gene al g a i y model as:
Xij =GSiMjφij (1)
whe e Xij deno es nominal expo s om coun y i o j,Gis a g a i y
cons an , Siand Mja e he capabili ies o expo e and impo e espec i ely,
and φij is a unc ion o he impac o ade ba ie s o bila e al ade lows,
wi h 0 ≤φij ≤1. Using homo he ic budge sha es and gene al equilib ium
ma ke clea ing condi ions o he expo e , one can de i e a s uc u al basis
o eq.1, so ha :
Xij =Yi
Pi
Xj
Πj
φij (2)
whe e Yiis g oss ou pu o expo e i,Xjis he o al consump ion alue
o goods in j,Piand Πja e mul ila e al ade esis ance e ms (MTR).4Sub-
sequen ly in mos empi ic applica ions, Yiand Xja e p oxied by expo e ’s
GDP and impo e ’s GDP espec i ely.
The bulk o heo y in he g a i y li e a u e is ela ed o s a ic and c oss-
sec ional models. A he same ime mos empi ics a e pe o med in a panel
se ing, and his o wo main easons: i) he e is plen y o panel da a a ail-
able a he coun y le el and e en a he sec o o p oduc le el; and ii) using
ime-in a ian eg esso s (such as dis ance and bo de s) can in e causa ion
o he model wi h espec o p edic ed ade lows.5Howe e , e en in panel
se ings almos all he es ima ed models a e s ill s a ic, no dynamic.6
4In Ande son and an Wincoop (2003), he au ho s en o ce Xi=Yi(balanced ade)
and φij =φji (symme ic ade cos s), which leads o Pi= Πjas a unique solu ion o
hei sys em o ma ke clea ing condi ions.
5A he same ime, obse a ions o indi iduals (coun ies in ou case he e) a e no
independen o e ime. This in oduces spu ious co ela ion and gene a es s anda d e o s
ha a e oo small. Tha ’s why (a leas wi h la ge Nand small T) we should clus e
obse a ions a he highes le el o agg ega ion, i.e. he coun y-pai le el since we obse e
bila e al lows.
6The e is some wo k on dynamic panel models in in e na ional ade, o ins ance Ha is
5
2.1 F om OLS o NLS...
The gene al unc ional o m o he empi ical g a i y model is gi en by
Y=exp(Xβ)η(3)
whe e Xis a ec o o eg esso s wi h elemen s xij,βis a ec o o
coe icien s o be es ima ed, and ηis a ec o o idiosync a ic e o e ms
wi h andom noise so ha E(ηij|X) = 1.7Clea ly, eq.3 can accomoda e bo h
eq.1 and eq.2. T adi ional es ima ion o he g a i y model log-linea izes he
model and uses OLS o es ima e he pa ame e s o in e es , β:
y=Xβ+ε(4)
whe e y=ln(Y) and ε=ηexp(Xβ). This linea ans o ma ion is o en
applied in empi ical ade esea ch, bu i causes h ee issues. The i s wo
issues a e poin ed a by San os Sil a and Ten ey o (2006), he hi d one is
new, and cons i u es he ocus o his pape . i) The alidi y o he model
depends on he o hogonali y o ηwi h espec o he eg esso s - which is
iola ed wi h he e oscedas ic e o s; ii) he es ima ion uns on only posi i e
alues, as ln(0) is unde ined, which leads o he exclusion o ze o- ade lows
in es ima ing bila e al ade; iii) he speci ied loss unc ion ha minimizes he
objec i e implies how obse a ions a e weigh ed in es ima ing he pa ame e s
o in e es . Le ’s elabo a e a bi on each o hese.
1. He e oscedas ici y One cause o he e ogenei y is omi ed a iable bias.
I he model is misspeci ied due o omi ed a iables o he exclusion
o a (non)-linea combina ion o eg esso s which a e co ela ed wi h
he e o e m, his leads o a non-homogeneous pa e n o he esid-
uals o he model (see also obus ness es s). When es ima ing eq.4,
one assumes ha he e is no in o ma ion in he noise, o equi alen ly
Y|X∼N(·), whe e N(µ,σ) is he No mal dis ibu ion wi h a gi en
mean µ=Xβand s anda d de ia ion σand lnη ∼N(0, σ). Howe e ,
when he e o e m is he e oscedas ic, he a iance o he e o e m
is no cons an (σi6=σ, ∀i). He e oscedas ici y does no a ec he un-
biasedness o he OLS es ima o , bu i a ec s he e iciency, since i
does no minimize he a iance. I also a ec s he es ima ed p- alues
and Ma yas (2004), Ha is, Kos enko, Ma yas and Timol (2009) and Bal agi e al. (2014).
7He e exp(·) is he exponen ial unc ion, E(·) s ands o he expec a ions ope a o , and
Xis a ec o o a iables wi h app op ia e leng h clea om he con ex .
6
and o a lesse ex en con idence in e als and p edic ion in e als. The
es ima ed s anda d e o s a e biased and he bias can go ei he way. I
he e oscedas ici y is mode a e, we can ans o m he es ima ion equa-
ion o use obus me hods o co ec o he s anda d e o s such as
Whi e’s (1980) s anda d e o s i we conside he es ima ion equa ion
co ec ly speci ied. Also, Weigh ed Leas Squa es can be used o o -
se he he e oscedas ici y p oblem and p oduce an e icien es ima o .
Howe e , de i ing he co ec weigh ing ma ix h ough i e a ion can
be a edious ask (see below).
2. Posi i e alues
Running he es ima ion p ocedu e only on posi i e alues can bias he
es ima ed coe icien s, as ze o ade lows can con ain aluable in o ma-
ion. San os Sil a and Ten ey o (2006) ad oca e a Poisson Pseudo Max-
imum Likelihood (PPML) es ima o o deal wi h bo h he e oscedas ic-
i y and ze o- ade lows simul aneously, and we will desc ibe he PPML
below. Howe e , PPML does no di ec ly accoun o s uc u al ze os,
as de i ed om models wi h ixed cos s o expo ing (see Meli z, 2003;
Helpman e al., 2008) o models o Be and compe i ion as in Ea on
and Ko um (2002). Some pa ches ha e been p oposed such as selec-
ion models wi h a 2-s age es ima ion p ocedu e, whe e he i s s age
es ima es he amoun o ze os in he sys em, and he second s age sub-
sequen ly es ima es he bila e el ade alues. While Helpman e al.
(2008) use a selec ion model ha is de i ed om heo y and accoun s
o i m he e ogenei y, al e na i es such as Ze o In la ed models deli e
biased esul s as he g a i y model does no ela e o coun models,
only he i s -o de condi ions o he PPML coincide wi h hose o he
Poisson model.8
3. Loss unc ion
The speci ied loss unc ion o any es ima ion p ocedu e o be minimized
a ec s how es ima es o βa e ob ained. The loss unc ion used in
OLS is he leas squa ed e o s unc ion, which pu s la ge weigh on
la ge obse ed e o s. The objec i e o minimize is ha o he Sum o
8Since he Helpman e al. (2008) p ocedu e can only be pe o med on a small subse o
coun ies (in o de o be compu a ionally able o use ixed e ec s), we do no p esen he
esul s o hose es ima ions he e. In addi ion, since he coun da a al e na i es o Nega i e
Binomial and Ze o In la ed models a e biased, we do no go in o u he de ails on hei
es ima ion in his pape .
7
Squa ed Residuals (SSR)
ˆ
β= a g min
β
SSR(β) = a g min
βX(y−Xβ)2(5)
whe e ˆ
βis he es ima e o β ha minimizes he objec i e unc ion.
The i s -o de condi ions a e ∂SSR(β)
∂β =−2X0y+ 2X0Xβ= 0, o
β= (X0X)−1Xy, whe e X0is he anspose o X.9The e is a unique
minimum i Xhas ull ank. In he linea model and unde no mali y
o he e o e ms, he i s -o de condi ions wi h espec o βo he ob-
jec i e o be op imized unde Leas Squa es and Maximum Likelihood
(ML) coincide. In he linea model wi h no mally dis ibu ed e o s,
he log-likelihood unc ion `(β|X) = −n
2ln(2π)−n
2ln(σ2)−1
2σ2(y−
Xβ)0(y−Xβ) is he objec i e unc ion o be maximized. The i s -
o de condi ions w i e ∂`
∂β = (X0X)−1Xy =ˆ
βOLS.
Ins ead o log-linea izing equa ion 3, we can es ima e he coe icien s om
he model in he o iginal exponen ial unc ion. Using non-linea leas squa es
(NLS) and op imizing SSR, he objec i e o es ima e pa ame e s o he model
becomes:
ˆ
β= a g min
β
SSR(β) = a g min
βX[Y−exp(Xβ)]2(6)
wi h a sys em o i s -o de condi ions:
∂ˆ
β
∂β =X[Y−exp(Xβ)]exp(Xβ)X= 0 (7)
The i s ac o (Y−exp(Xβ)) is he model o be es ima ed, minimiz-
ing he e o s, and he ac o exp(Xβ)X ep esen s weigh s o each obse -
a ion in minimizing hose e o s. Some au ho s (F ankel and Wei, 1993;
F ankel, 1997b; Ande son and an Wincoop, 2003) ha e p oposed using he
NLS me hod in es ima ing he g a i y equa ion: he unc ion gi es mo e
weigh o obse a ions whe e exp(Xβ) is la ge, so ha coun ies wi h la ge
Siand Mj o ins ance, ge mo e weigh .10 The e is economic in ui ion o
his weigh ing scheme, as coun ies wi h highe GDP end o epo mo e
accu a ely and he e o e ge mo e weigh in es ima ing he model. How-
e e , San os Sil a and Ten ey o (2006) s a e ha i) his does no add ess
9F om he second-o de condi ions, his is a minimum: ∂2SSR(β)
∂β2= 2X0X≥0.
10This is no only GDP, bu also o he a iables ha migh be used in his dimension
such as he MTR, and also he dis ance unc ion.
8
Figu e 1: Es ima ion o he dis ance coe icien c oss me hods
-1.6 -1.4 -1.2 -1 -.8 -.6 -.4
Coe icien
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Yea
OLS LSDV
BB PPML
GPML
No e: Mul iple c oss-sec ion es ima es ac oss es ima ion me hods. The es ima ion equa ion is gi en
by eq.2, con olling o he obse able a iables in he dis ance unc ion: ln(GDP) expo e , ln(GDP
impo e , ln(dis ance), adjacency, common o icial language, colonial ies and RTAs. We also con ol o
WTO membe ship s a us o he expo e and impo e . Expo e and impo e ixed e ec s a e used in
he LSDV, PPML and GPML models. The coe icien s o dis ance a e all signi ican a he 0,1% le el
ac oss all yea s and es ima ion me hods. S anda d e o s a e obus and clus e ed a he coun y-pai
le el.
Table 2 shows he es ima ed pa ame e s o in e - e sus in a-con inen al
ade lows o e he pooled pe iod 1998-2011. No e ha we obse e ha he
dis ance coe icien is in line wi h he c oss-sec ional es ima es as be o e.
Mo eo e , we see ha he adjacency e ec is posi i e, ough insigni ican
in he BB se ing. Fu he mo e, all he con ol a iables ha e he expec ed
signs and sizes.
We ob ain in e es ing insigh s in o he bo de e ec by con inen s. We
i s ocus on he LSDV speci ica ion. A e co ec ing o mul ila e al ade
esis ance, GDP and bila e al obse ables, we see ha some con inen s ade
ela i ely mo e globally han in a-con inen ally. This is he case o Eu ope
and Asia, wi h Eu ope being he mos open con inen . In o he wo ds, hese
con inen s a e globally mo e connec ed, a inding we see in eali y. No e ha
his is no a odds wi h he well-known ac ha e.g. in a-Eu opean ade
exceeds ex a-Eu opean ade since i is he e ec a e con olling o he
egula g a i y explana ions. The Paci ic is he mos “closed” con inen , in
he sense ha i ades ela i ely mo e inside he Paci ic han ac oss he
globe. When we u n o o he es ima ion me hods, esul s change d ama -
ically depending on he p ocedu e used. Hence con a y o ou indings o
he dis ance e ec , he bo de e ec appea s o be much mo e sensi i e o
15

he selec ed es ima ion me hody. In he biased OLS se ing (no co ec ing
o MTR), mos esul s a e insigni ican , and Asia is ading ela i ely mo e
inside Asia. Simila o ou indings o he dis ance e ec , also now LSDV
and BB line up nicely. Based on bo h me hods, he Ame icas a e mo e open,
while Eu ope is no . Fo he non-linea me hods o PPML and GPML, we
ind con adic o y coe icien s o Eu ope. I is also in e es ing o no e ha
his con inen al bo de e ec is in con as wi h he egional lows as in An-
de son and an Wincoop (2003). In hei indings, he bo de e ec is always
nega i e. Appa en ly hings change in he global con ex .
Finally, Figu e 2 shows he e olu ion o he es ima ed coe icien s o he
in a-con inen al p edic ed ade o e he yea s 1998 o 2011, whe e we ha e
used he LSDV me hod as ep esen a ion.
16
Table 2: Bo de s by con inen s
(1) (2) (3) (4) (5)
OLS LSDV BB PPML GPML
lnGDP(expo e ) 1.092*** 1.099***
(0.0198) (0.0195)
lnGDP(impo e ) 0.857*** 0.864***
(0.0299) (0.0298)
ln(dis ance) -1.023*** -1.298*** -1.399*** -0.494*** -1.385***
(0.0992) (0.0848) (0.109) (0.0888) (0.0227)
Adjacency 0.912** 0.777* 0.508 0.415*** 1.029***
(0.221) (0.293) (0.241) (0.110) (0.0646)
Common Language 0.700*** 0.702*** 0.568*** 0.0751 0.654***
(0.0508) (0.0625) (0.0478) (0.0619) (0.0350)
Colonial Ties 1.015*** 0.926*** 1.026*** 0.403* 1.408***
(0.128) (0.0573) (0.135) (0.193) (0.0621)
RTA 0.851* 0.711* 0.833* 0.477*** 0.486***
(0.243) (0.206) (0.233) (0.0587) (0.0273)
WTO expo e 0.561** 0.528***
(0.110) (0.0756)
WTO impo e 0.289* 0.271**
(0.0774) (0.0420)
Eu ope 0.006 -0.506* -0.0539 0.440*** -0.922***
(0.197) (0.130) (0.131) (0.119) (0.0478)
Ame icas 0.343 0.410 -0.494* 0.675*** 0.300***
(0.198) (0.194) (0.174) (0.122) (0.0780)
Asia 0.240** -0.350** -0.237* -0.100 -0.121**
(0.0498) (0.0629) (0.0904) (0.151) (0.0396)
A ica -0.0309 0.295* -0.0793 0.669*** 0.430***
(0.213) (0.0904) (0.147) (0.0545) (0.0569)
Paci ic 2.495*** 1.333** 0.923* 1.110*** 1.915***
(0.237) (0.202) (0.270) (0.173) (0.191)
Cons an -31.28*** 5.953** -40.30*** 11.85*** 10.73***
(1.951) (0.894) (1.096) (1.017) (0.332)
adj. R20.671 0.735 0.674
Coun y FE No Yes No Yes Yes
BIC 1232841.3 1324651.3 1230787.3 6.40034e+10 8388126.5
N 283586 319276 283586 532029 532029
No es: Model speci ica ions a e (1) OLS, (2) Leas Squa es Dummy Va iable, (3) Baie and Be gs and
(2009) Taylo app oxima ion me hod, (4) Poisson Pseudo-Maximum Likelihood and (5) Gamma Pseudo-
Maximum Likelihood. In model (3), all he bila e al a iables a e i s -o de Taylo app oxima ed. Coun-
y FE depic expo e and impo e ixed e ec s. Adjus ed R2and/o he Bayesian In o ma ion C i e ion
(BIC) a e gi en whe e possible. Robus and clus e ed s anda d e o s a e be ween pa en hesis, clus e ed
a he con inen le el. Signi icance le els: 5% (*), 1% (**) and 0,1%(***).
17
Figu e 2: E olu ion o he dummy es ima es o con inen s o e ime
-1 0 1 2
Coe icien
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Yea
Eu ope Ame icas
Asia A ica
Paci ic
No e: Mul iple c oss-sec ion es ima es ac oss es ima ion me hods. The es ima ion equa ion is gi en
by eq.2, con olling o he obse able a iables in he dis ance unc ion: ln(GDP) expo e , ln(GDP
impo e , ln(dis ance), adjacency, common o icial language, colonial ies and RTAs. We also con ol o
WTO membe ship s a us o he expo e and impo e . Expo e and impo e ixed e ec s a e used in
he LSDV, PPML and GPML models. The coe icien s con inen s dummies a e all signi ican a he 0,1%
le el ac oss all yea s and es ima ion me hods. S anda d e o s a e obus and clus e ed a he coun y-pai
le el.
5 Robus ness Checks
We check o se e al po en ial sou ces o misspeci ica ion. Fi s , ollowing he
commen s in Head and Maye (2014), we check con e gence o he di e en
es ima o s unde di e en sample sizes. Since he es ima ed coe icien s based
on OLS and GPML a e close, while he PPML coe icien s a e lowe , Head
and Maye (2014) ague ha his migh be due o misspeci ica ion o he model
i he sample size is big enough. We d aw andom subsamples om ou da a,
o 75, 50 and 25% espec i ely. Figu e 3 shows he es ima ed coe icien s o
dis ance o e ime by es ima ion me hod. Ac oss all sample sizes, PPML
consis en ly deli e s lowe es ima es o dis ance, and he anking o he
o he es ima es emains he same. In his se ing, we canno ec ea e he
con e gence o all es ima es as p oposed by Head and Maye (2014) in hei
simula ion se ing, who s a e ha , i sample size is la ge enough, and absen
o misspeci ica ion, he es ima es o OLS, PPML and GPML should coincide.
Secondly, we d aw he esidual e sus i ed alues plo o he LSDV
me hod in Figu e 4.21 Since we deal wi h so many obse a ions, adi ional
sca e plo s a e no e y e icien . Ins ead, we p opose o use a local polyno-
21We also un he esidual plo s o he o he es ima ion me hods. Pa e ns a e simila .
18
mial smoo he plo o ep esen he unde lying sca e s. The local polynomial
has wo added ad an ages o esidual analysis: i) he polynomial smoo he
is an indica o o po en ial non-linea i ies o o he pa e ns in he esidu-
als, and ii) he 95% con idence in e als o he smoo he a e a nice way o
depic po en ial he e oscedas ici y: i he e a e i egula i ies in he wid h
o he con idence in e als, his indica es non-cons an a iance o he e o
e ms. We e un he o iginal speci ica ion o Sec ion 4.1, bu now o he yea
2005 only, o e ade po en ial au o-co ela ion.22 The e is i) a clea s uc u e
in he esiduals ha esembles a hi d-deg ee polynomial and ii) po en ial
he e oscedas ici y could show up in he le ail o he dis ibu ion o he
i ed alues. We a e con iden ha he e oscedas ici y is no a ec ing ou e-
sul s in any majo way (we also check he pa e n o he e oscedas ici y using
PPML, gi ing almos iden ical esul s), and ocus on he non-linea pa e n
o he esiduals. To see whe e he s uc u e comes om, we plo he esiduals
agains each eg esso . The esidual plo s o all eg esso s look ine, excep
he esidual plo agains dis ance unco e s he same s uc u e as he esidu-
als e sus i ed plo . We should he e o e e un he model wi h a polynomial
app oxima ion o dis ance: i migh be he case ha he linea speci ica ion
o he dis ance unc ion is jus no co ec , and we jus assumed i ollowing
he bulk o he g a i y li e a u e. Also, gi en ou global sample, he e ec o
dis ance migh no be well app oxima ed by a linea ela ionship, whe e his
migh be mo e app op ia e o local sub samples such as Eu ope. We e un
he model wi h a second and hi d o de polynomial o dis ance, which in
e ec lowe s he pa e n in he esidual plo s, bu comple ely dis up s all o
he g a i y es ima es. We also check i he e is a non-linea pa e n inside
con inen s, so o see ha he non-linea i ies do no come om he global
sample. We see he same esidual pa e n ecu ing o each isola ed sub-
sample. We he e o e p e e o keep wi h he main s eam o he li e a u e
and use he log-linea dis ance unc ion. Howe e , he co ec speci ica ion
o he dis ance unc ion is an in e es ing opic in i s own.
Finally, we check o po en ial mul icollinea i y be ween dis ance and he
bo de e ec , since his migh also d i e misspeci ica ion. We ind VIF es
esul s o all a iables in he model (excluding he ixed e ec s) be ween 1
and 1.5, whe e VIF alues o abo e 5 o 6 migh indica e po en ial mul i-
collinea i y p oblems. We he e o e also ejec his po en ial p oblem.
22We also an he model on he pooled e sion and he panel e sion, all gi ing e y
simila esul s.
19
Figu e 3: Subsample es ima es o dis ance
-1.6 -1.4 -1.2 -1 -.8 -.6 -.4
Coe icien
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Yea
OLS LSDV
BB PPML
GPML
(a) Subsample 75%
-1.6 -1.4 -1.2 -1 -.8 -.6 -.4
Coe icien
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Yea
OLS LSDV
BB PPML
GPML
(b) Subsample 50%
-1.6 -1.4 -1.2 -1 -.8 -.6 -.4
Coe icien
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Yea
OLS LSDV
BB PPML
GPML
(c) Subsample 25%
No e: Mul iple c oss-sec ion es ima es ac oss es ima ion me hods. Expo e and impo e ixed e ec s a e
used in all models excep OLS and BB. The coe icien s o dis ance a e all signi ican a he 0,1% le el
ac oss all yea s and es ima ion me hods. S anda d e o s a e obus and clus e ed a he coun y-pai
le el.
Figu e 4: Residual e sus i ed plo s
-5 0 5 10
Residuals
-5 0 5 10 15 20
Linea p edic ion
95% CI Polynomial Smoo he
ke nel = epanechniko , deg ee = 0, bandwid h = .26, pwid h = .4
Local polynomial smoo h
No e:A local polynomial smoo he ep esen s he unde lying sca e plo . The g ay band indica es he
95% con idence in e al, he y= 0 line is indica ed in ed.
6 Conclusion
This pape compa ed he dis ance and bo de e ec s on global bila e al ade
lows using a ious econome ic echniques. We clea ly con i m he nega i e
dis ance e ec , bu i s magni ude appea s o a y ac oss es ima ion me hods.
The dis ance e ec is also cons an o e ime in all me hods. The obse ed
a ie y is in line wi h he heo e ical expec a ions. Ou e idence o he
bo de e ec by con inen s is mo e ambiguous. Gene ally speaking, we can
con i m ha in acon inen al ade exceeds in e con inen al ade, con olling
o a ious o he ade explana ions. Howe e , his gene al inding b eaks
down using some es ima ion me hods. These esul s call o cau ion when
including dis ance and bo de e ec s in u u e empi ical ade s udies. Ou
esul s do no a ou pa icula es ima ion me hods, as hey all ha e hei
20

me i s and sho comings. Ra he , esea che s should be awa e o he impac
o he selec ed me hod on he magni u e o he dis ance e ec and on he
magni ude, di ec ion and signi icance o he bo de e ec .
21
Re e ences
[1] Ande son, J. E. (1979). A heo e ical ounda ion o he g a i y equa-
ion. Ame ican Economic Re iew, 69(1), 106–16.
[2] Ande son, J. E. (2010). The incidence o g a i y. Bos on College Wo k-
ing Pape s in Economics 732, Bos on College Depa men o Economics.
[3] Ande son, J. E. and an E. Wincoop (2003). G a i y wi h G a i as:
A Solu ion o he Bo de Puzzle. Ame ican Economic Re iew, 93(1),
170–192.
[4] Ande son, J. E. and E. an Wincoop (2004). T ade cos s. Jou nal o
Economic Li e a u e, 42(3), 691–751.
[5] Bal agi, B., Egge , P. and M. P a e may (2014). Panel Da a G a i y
Models o In e na ional T ade. CESIFO Wo king Pape no. 4616.
[6] Be gs and, J. H. (1989). The gene alized g a i y equa ion, monopolis ic
compe i ion, and he ac o -p opo ions heo y in in e na ional ade.
The Re iew o Economics and S a is ics, 71(1), 143–153.
[7] Be gs and, J. H. (1990). The Hecksche -Ohlin-Samuelson Model, he
Linde Hypo hesis and he De e minan s o Bila e al In a-indus y
T ade. Economic Jou nal, 100(403), 1216–1229.
[8] Be helon, M. and C. F eund (2008). On he conse a ion o dis ance in
in e na ional ade. Jou nal o In e na ional Economics, 75(2), 310-320.
[9] Blum, B. S. and A. Gold a b (2006). Does he in e ne de y he law o
g a i y? Jou nal o In e na ional Economics, 70(2), 384–405.
[10] Boisso, D. and M. Fe an ino (1993). Is he Wo ld Ge ing Smalle ? A
G a i y Expo Model o he Whole Plane Wo ld? Wo king Pape s
9225, Richa d Johnson Cen e Wo king Pape , Sou he n Me hodis Uni-
e si y.
[11] Bosque , C. and H. Boulhol (2010). Scale-dependence o he Nega i e
Binomial Pseudo-Maximum Likelihood Es ima o . Documen s de a ail
du Cen e d’Economie de la So bonne 10092.
[12] B un, J. F., Ca `e e, C., Guillaumon , P. and J. de Melo (2005). Has
dis ance died? E idence om a panel g a i y model. The Wo ld Bank
Economic Re iew, 19(1), 99–120.
22
[13] Bu ge , M., an Oo , F. and G.-J. Linde s (2009). On he Speci ica ion
o he G a i y Model o T ade: ze os, Excess ze os and Ze o-in la ed
Es ima ion. Spa ial Economic Analysis, 4(2), 167–190.
[14] Coe, D. T., Sub amanian A., and N.T. Tami isa (2007). The Missing
Globaliza ion Puzzle: E idence o he Declining Impo ance o Dis ance.
IMF S a Pape s, 54(1), 34–58.
[15] Coughlin, C. C. and D. No y (2012). Is he In e na ional Bo de E ec
La ge han he Domes ic Bo de E ec ? E idence om U.S. T ade.
Discussion Pape s 09/29, Uni e si y o No ingham, GEP.
[16] Dea do , A. (1995). De e minan s o bila e al ade: Does g a i y wo k
in a neoclassical wo ld? Wo king Pape s 382, Resea ch Semina in In-
e na ional Economics, Uni e si y o Michigan.
[17] de Sousa, J. (2012). The Cu ency Union E ec on T ade is Dec easing
O e Time. Economics Le e s, 117(3), 917–920.
[18] Disdie , A. and K. Head (2008). The puzzling pe sis ence o he dis ance
e ec on bila e al ade. The Re iew o Economics and S a is ics, 90(1),
37–48.
[19] Ea on, J. and S. Ko um (2002). Technology, Geog aphy, and T ade.
Econome ica, 70, 1741–1779.
[20] Eicheng een, B., and D. A. I win (1998). The Role o His o y in Bila e al
T ade Flows. In: F ankel, J. (1998). The Regionaliza ion o he Wo ld
Economy , F ankel.
[21] Egge , P. and K. S aub (2015). GLM es ima ion o ade g a i y models
wi h ixed e ec s (unpublished manusc ip ).
[22] Feens a, R. C., Ma kusen, J. R. and A. K. Rose (2001). Using he
G a i y Equa ion o Di e en ia e Among Al e na i e Theo ies o T ade.
Canadian Jou nal o Economics, 34(2), 430–447.
[23] Felbe may , G. J., and W. Kohle (2006). Explo ing he In ensi e and
Ex ensi e Ma gins o Wo ld T ade. Re iew o Wo ld Economics, 142(4),
642–674.
[24] F ankel, J. (1997a). Regional T ading Blocs in he Wo ld o Economic
Sys em. Washing on, D.C.: Ins i u e o In e na ional Economics.
23
[25] F ankel, J. (1997b). S e iliza ion o Money In lows: Di icul (Cal o) o
Easy (Reisen)? Es udios de Economia, 24(2), 263–285.
[26] F ankel, J. and S.-J. Wei (1993). A Paci ic Economic Bloc: Is The e
Such an Animal? FRBSF Economic Le e , Fede al Rese e Bank o
San F ancisco.
[27] Gaulie , G. and S. Zignago (2010). BACI: In e na ional T ade Da abase
a he P oduc -le el The 1994-2007 Ve sion. CEPII, WP No 2010-23.
[28] Gou ie oux, C., Mon o , A. and A. T ognon (1984). Pseudo Maximum
Likelihood Me hods: Applica ions o Poisson Models. Econome ica,
52(3), 701-720.
[29] G ossman, G. (1998). Commen on Dea do In: F ankel, J. (1998).
The Regionaliza ion o he Wo ld Economy, 33–57.
[30] Ha is, M., Kos enko, W., Ma yas, M. and I. Timol (2009):. The Robus -
ness o Es ima o s o Dynamic Panel Da a Models o Misspeci ica ion.
Sin. Economic Re iew, 51, 399–426.
[31] Ha is, M. and L. Ma yas (2004): A Compa a i e Analysis o Di e en
IV and GMM Es ima o s o Dynamic Panel Da a Models. In e na ional
S a is ical Re iew, 72, 397–408.
[32] Head, K., and T. Maye (2014). G a i y Equa ions: Wo kho se, Toolki ,
and Cookbook, chap e 3 in Gopina h, G, E. Helpman and K. Ro-
go (eds), ol. 4 o he Handbook o In e na ional Economics, Else ie :
131–195.
[33] Heckman, J.J. (1979). Sample Selec ion Bias as a Speci ica ion E o .
Econome ica, 47, 153–161.
[34] Helliwell, J. F. (1996). Do Na ional Bo de s Ma e o Quebec’s T ade?
Canadian Jou nal o Economics, 29(3), 507–522.
[35] Helliwell, J. F. (1997). Na ional Bo de s, T ade and Mig a ion. NBER
Wo king Pape s 6027, Na ional Bu eau o Economic Resea ch, Inc.
[36] Hillbe y, R. and D. Hummels (2003). In ana ional Home Bias: Some
Explana ions. The Re iew o Economics and S a is ics 85(4), 1089–1092.
[37] Hillbe y, R. and D. Hummels (2008). T ade esponses o geog aphic
ic ions: A decomposi ion using mic o-da a. Eu opean Economic Re-
iew, 52 (3), 527–550.
24