Coimb a Viei a, Ca olina; Lohmann, Sophie; Zagheni, Emilio
A icle — Published Ve sion
The Value o Cul u al Simila i y o P edic ing Mig a ion:
E idence om Food and D ink In e es s in Digi al T ace
Da a
Popula ion and De elopmen Re iew
P o ided in Coope a ion wi h:
John Wiley & Sons
Sugges ed Ci a ion: Coimb a Viei a, Ca olina; Lohmann, Sophie; Zagheni, Emilio (2024) : The Value o
Cul u al Simila i y o P edic ing Mig a ion: E idence om Food and D ink In e es s in Digi al T ace
Da a, Popula ion and De elopmen Re iew, ISSN 1728-4457, Wiley, Hoboken, NJ, Vol. 50, Iss. 1, pp.
149-176,
h ps://doi.o g/10.1111/pad .12607
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/294022
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-nc/4.0/
The Value o Cul u al Simila i y o
P edic ing Mig a ion: E idence om Food
and D ink In e es s in Digi al T ace Da a
CAROLINA COIMBRA VIEIRA ,SOPHIE LOHMANN
AND EMILIO ZAGHENI
One o he s onges empi ical egula i ies in spa ial demog aphy is ha lows o
mig an s a e posi i ely associa ed wi h popula ion s ocks a o igin and des ina-
ion and a e in e sely ela ed o dis ance. This pa e n was o malized in o wha
a e known as g a i y models o mig a ion. T adi ionally, dis ance is measu ed
geog aphically, bu o he measu es o dis ance, such as cul u al dis ance, a e also
ele an in explaining mig a ion lows. Howe e , measu es o cul u al dis ance a e
no widely adop ed in he li e a u e on modeling mig a ion lows, pa ially because
o he di icul ies associa ed wi h ope a ionalizing and p oducing hese measu es
ac oss space and ime. In his pape , we use a scalable app oach o ob ain p oxies
o measu ing cul u al simila i y be ween coun ies by using Facebook da a and
illus a e he impac o inco po a ing hese measu es, based on ood and d ink
in e es s, in o g a i y models o p edic ing mig a ion. Ou esul s show ha , despi e
hei limi a ions, he new measu es o cul u al simila i y de i ed om Facebook da a
imp o e he p edic ion powe o adi ional g a i y models and ha e a p edic i e
capaci y compa able o ha o classic a iables used in he li e a u e, such as sha ed
language and his o y. The esul s open up new oppo uni ies o unde s anding he
de e minan s o mig a ion and o p edic ing mig a ion when conside ing b oade
and complemen a y pe spec i es on he meaning and measu emen o dis ance.
In oduc ion
One o he s onges empi ical egula i ies in spa ial demog aphy is ha
lows o mig an s a e posi i ely associa ed wi h popula ion size a o igin and
des ina ion and a e in e sely ela ed o dis ance. This pa e n was obse ed
in he 19 h cen u y by Ra ens ein (1889) and was la e o malized by
Ca olina Coimb a Viei a, Sophie Lohmann and Emilio Zagheni, Max Planck Ins i-
u e o Demog aphic Resea ch, 18057, Ros ock, Ge many. E-mail: ca olcoimb a.dcc@
gmail.com.
POPULATION AND DEVELOPMENT REVIEW 50(1): 149–176 (MARCH 2024) 149
© 2024 The Au ho s. Popula ion and De elopmen Re iew published by Wiley Pe iodicals LLC on behal o Popula ion Council.
This is an open access a icle unde he e ms o he C ea i e Commons A ibu ion-NonComme cial License, which pe mi s
use, dis ibu ion and ep oduc ion in any medium, p o ided he o iginal wo k is p ope ly ci ed and is no used o comme cial
pu poses.
150 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
Zip (1946) in o wha a e known as g a i y models o mig a ion. T adi-
ionally, dis ance is measu ed geog aphically. Howe e , o he measu es,
including hose based on economic and cul u al ac o s, ha e also been
ound o be ele an o explaining mig a ion lows (Ande son 2011; Esses
2018; Ca agliu e al. 2013; Böhme, G öge , and S öh 2020; Lewe and Van
den Be g 2008).
The cul u al dis ance be ween wo coun ies could, he e o e, be a
aluable p edic o o mig a ion lows gi en he bidi ec ional ela ionship
be ween cul u e and mig a ion. Fo ins ance, he cul u al i in e ms o
language, no ms, and alues is an impo an ac o ha people conside be-
o e mo ing be ween coun ies (Ca agliu e al. 2013; Pede sen, Py liko a,
and Smi h 2004). A e mo ing, mig an s hen ansmi cul u al elemen s,
such as ood habi s (Opa e-Obisaw e al. 2000), om hei o igin coun y
o hei des ina ion coun y and back again (Mesoudi 2018).
Measu es o cul u al dis ance a e di icul o es ima e and hus ha e
no ye been widely adop ed in g a i y models o assessing and p edic ing
mig a ion. The ew s udies ha ha e examined he impac o cul u al di-
mensions o cul u al dis ance on mig a ion lows ha e ypically elied on
su ey esponses ega ding no ms, alues, and belie s, such as hose om
he Wo ld Values Su ey (WVS; Ingleha 1997). Immig a ion da a om
Denma k, Ge many, and he Ne he lands illus a e ha g ea e cul u al dis-
ance, as de i ed om he WVS, is associa ed wi h less long- e m mobili y
(Whi e 2013). O e all, cul u al dis ance, as de i ed om he WVS, seems
o play an impo an ole in p edic ing mig a ion lows be ween Eu opean
coun ies (Ca agliu e al. 2013). Howe e , su ey app oaches ha ocus on
mig a ion su e om signi ican limi a ions, such as he di icul y o each-
ing mig an s, and he complexi y and high cos s associa ed wi h unning a
c oss-na ional su ey wi h a mig a ion ocus.
In his pape , we use complemen a y measu es o cul u al simila -
i y based on cul u al no ms, alues, and belie s de i ed om su eys
(i.e., he WVS) o he s udy o mig a ion lows. We expand he analysis
o he impac o cul u e on mig a ion lows by adding measu es o cul u al
simila i y based on cul u al a ibu es ega ding ood and d ink in e es s
de i ed om social media da a (i.e., Fou squa e and Facebook). This a icle
canno dis inguish be ween all he mechanisms unde lying he complex
ela ionship be ween cul u e and mig a ion, and he aim o his s udy is no
o es ablish a causal link be ween hem. We do, howe e , demons a e ha
cul u e is an impo an aspec o conside when s udying mig a ion and
ha he inclusion o measu es o cul u al simila i ies imp o es p edic ions
o mig a ion lows, e en a e accoun ing o classic p edic o s o mig a ion.
We expand he li e a u e by showing he impac o adding measu es o
cul u al simila i y de i ed om social media da a based on ood and d ink
in e es s (i.e., ood and d ink simila i y) o he analysis o mig a ion lows.
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 151
Food and d ink a e wo o he mos basic needs o human beings.
The manne in which people in e ac wi h ood, om he p ocu emen and
selec ion o ood o i s p epa a ion and consump ion, e lec s complex in-
e ela ionships and in e ac ions among indi iduals, he socie y in which
hey li e, and hei cul u e (Axelson 1986; Fe guson, I u bide, and Ra aelli
2020). Food s udies ha e become an impo an in e disciplina y ield o
s udy ha ocuses on he ela ionship be ween ood and human expe ience
and he ela ionships be ween ood, cul u e, and socie y (Alme ico 2014).
Food p oduc ion, dis ibu ion, and consump ion a e all shaped by cul u al
codes (Counihan and Van Es e ik 2012) and ep esen cul u al ac s (Mon-
ana i 2006). Food p epa a ion1is one o he opics included in he Lis s o
In angible Cul u al He i age p o ided by UNESCO,2which co e s cul u al
p ac ices and exp essions o in angible he i age. Mo e gene ally, ood can be
seen as an impo an ma ke o cul u al iden i y (Ki le , Suche , and Nelms
2016). Ou o all ca ego ies o in e es s on Facebook (e.g., ood and d ink,
news and en e ainmen , hobbies and ac i i ies, spo s and ou doo s), ood
and d ink a e he only in e es s ha belong o a uni e sal ca ego y, gi en
ha ood and d ink a e wo o he mos basic needs o human beings. Some
in e es s a e speci ic o ce ain demog aphic g oups; o ins ance, no e e y-
one is in e es ed in spo s o celeb i ies. By con as , ood is popula ac oss a
wide demog aphic spec um. In his con ex , gi en he impo ance o ood
o cul u e (Ashley e al. 2004; De Solie and Du uz 2013; Recchi and Fa ell
2019), we conside measu es o cul u al simila i y based on ood and d ink
in e es s ha a e de i ed om social media da a.
We p opose he use o measu es o ood and d ink simila i y de eloped
by Viei a e al. (2022) and e alua e hei po en ial in p edic ing mig a ion.
Unlike su ey da a ha need o ely on di e en ounds o su ey o be
collec ed, hese measu es a e imely, cos -e ec i e, and scalable as hey a e
based on agg ega e da a om Facebook ha a e eely and publicly a ail-
able h ough he Facebook Ad e ising Pla o m (we will e e o hese da a
as Facebook Ads da a). We illus a e he applicabili y o he p oposed ap-
p oach by showing how hese new measu es o ood and d ink simila i y
can be used o p edic mig a ion and explain mig a ion lows. The measu es
o ood and d ink simila i y de i ed om Facebook Ads ha e, despi e hei
limi a ions, a capaci y o p edic mig a ion lows ha is compa able o ha
o classic a iables used in he li e a u e o ep esen he cul u al dimension,
such as sha ed language and sha ed his o y. Addi ionally, he measu es o
ood and d ink simila i y de i ed om Facebook Ads a e able o cap u e
changes quickly, especially when mig a ion pa e ns change apidly due o
c ises. In his con ex , we expec ha hese measu es o ood and d ink sim-
ila i y de i ed om Facebook Ads could ep esen , almos in eal- ime, he
cul u al changes ha occu du ing big and unexpec ed mig a ion e en s
(e.g., he mig a ion o Uk ainians a e Russia’s in asion). Fu he mo e, he
Facebook Ads measu es o ood and d ink simila i y in oduce a mo e nu-
anced iew o symme ic and nonsymme ic measu es o simila i y, opening
152 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
up new oppo uni ies o p edic ing and unde s anding he de e minan s o
mig a ions.
Backg ound
Cul u al dis ance measu es ope a ional pa ame e s ha can be used as p ox-
ies o cul u al dimensions. They allow esea che s o es ima e he ex en o
which coun ies di e cul u ally (Tung and Ve beke 2010). The cul u al di-
mensions used o measu e cul u e can a y depending on he ocus o he
esea ch (Moh e al., 2019). Fo ins ance, he s udy o cul u e can ocus
on aspec s o daily li e by conside ing cul u al objec s, such as he clo hes
people wea , he music hey lis en o, and he ood hey ea (Recchi and
Fa ell 2019; Kwan es and Glaze 2017). Food s udies is an impo an in e -
disciplina y ield ha ecognizes ood as a cen al aspec o accul u a ion,
cul u al p ac ices, and cul u al iden i y (Fe guson, I u bide, and Ra aelli
2020; Mon ana i 2006; Ashley e al. 2004; De Solie and Du uz 2013; Ki -
le , Suche , and Nelms 2016).
Ope a ionally, cul u e has been adi ionally measu ed in e ms o
no ms, alues, and belie s ia sampling su eys (Kwan es and Glaze 2017)
in which he su ey esponses a e used o cha ac e ize cul u al aspec s
o a coun y (e.g., Schwa z’s alue su ey (Schwa z 1994), he WVS
(Ingleha 1997), and Ho s ede’s3cul u al cha ac e is ics (Ho s ede 1983))
and o e alua e he ela i e dis ance be ween coun ies (Gup a, Hanges, and
Do man 2002; De San is, Mal aglia i, and Sal ini 2016; Muccia di and De
San is 2017; Mu huk ishna e al. 2020).
Such s udies based on su eys a e highly aluable bu also ha e im-
po an limi a ions. In addi ion o measu emen e o (G o es and Lybe g
2010), he esul s may su e om a ious biases (Suchman 1962), like so-
cial desi abili y bias, ques ion o de bias, and acquiescence bias. Fu he -
mo e, su eys a e cos ly and equi e a long ime o un. Fo example, mos
go e nmen s a is ics a e upda ed only once a yea (o en wi h a delay),
and majo su eys such as he Eu opean Values S udy (EVS) and he WVS
a e o en spaced e en u he apa ( he EVS is conduc ed once e e y nine
yea s, and he WVS is ca ied ou once e e y i e yea s). This lack o imely
in o ma ion makes i di icul o decision-make s o espond dynamically o
shi ing ci cums ances. While all demog aphic s udies in ol e unce ain y
and complexi y, mig a ion p edic ions a e pa icula ly unce ain (Bijak and
Bijak 2022). Fo ins ance, mig a ion lows ela ed o e ugee mo emen s4
a e among he mos ola ile o ms o mig a ion and a e, he e o e, he mos
di icul o p edic . These dynamic shi s in mig a ion lows illus a e he
need o mo e equen da a o ack mig a ion lows and imp o e p edic-
ions. To o e come some o hese limi a ions, we p opose an app oach ha
elies on passi ely collec ed da a om social media, which can be used o
complemen da a om exis ing sou ces.
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 153
Social media ad e ising pla o ms p o ide complemen a y ools ha
can be used o measu e cul u al p e e ences and ha allow o compa -
isons ac oss egions ia passi ely collec ed da a (You e al. 2017). As one
o he i s s udies o add ess his ques ion using online da a sou ces, Sil a
e al. (2014) iden i ied cul u al bounda ies and simila i ies ac oss popula-
ions by clus e ing hem based on he analysis o ood and d ink habi s.
Howe e , hei analyses o culina y habi s a ound he wo ld we e limi ed
o Fou squa e check-ins, which conside ed only 101 ca ego ies and hus
unde es ima ed he a ie y o use s’ in e es s.
In one o he i s s udies on his opic using Facebook Ads da a, Viei a
e al. (2020) examined he simila i ies be ween selec ed coun ies and B azil
based on hei popula ion’s in e es s in ypical B azilian dishes. Howe e ,
he esul s we e limi ed o dishes lis ed on Wikipedia, which es ic ed he
po en ial lis o dishes. Mo eo e , because some coun ies do no ha e a
Wikipedia page dedica ed o lis ing hei ypical dishes, he me hodology
was no scalable. Ob ado ich e al. (2022) also used da a om Facebook
Ads o examine c oss-na ional cul u al di e ences ac oss nea ly 60,000 in-
e es s. They alida ed hei wo k by compa ing he cul u al dis ances cal-
cula ed using hei measu emen s wi h hose o adi ional su ey-based
measu es. Howe e , as he au ho s included a wide ange o cul u al ea-
u es, om poli ics o na ional pa ks, and used a “black box” model o ep-
esen coun ies’ cul u es, i is ha d o assess exac ly wha hei index was
measu ing. Mo e ecen ly, Viei a e al. (2022) p esen ed a scalable, da a-
d i en me hodology o measu ing he cul u al simila i ies be ween coun-
ies based on he mos popula ood and d ink in each coun y om a lis
con aining mo e han 200,000 in e es s on Facebook Ads (Speiche e al.
2018). Rela i e o Ob ado ich e al. (2022), he me hodology p oposed by
Viei a e al. (2022) compa ed coun ies using ewe , bu explici ly known
a ibu es selec ed om a e y la ge da a se . In o he wo ds, in e es s ha
we e no ele an o any o he coun ies we e dis ega ded o educe ea u e
spa si y. They p esen ed wo measu es o cul u al simila i y, including he
i s asymme ic measu e o cul u al simila i y de i ed om social media
da a.
The li e a u e ha we jus summa ized sugges ed me hodologies based
on social media da a o measu e cul u al simila i y be ween coun ies and
hen co ela ed hese measu es wi h su ey-based measu es. To he bes o
ou knowledge, ou s is he i s pape o e alua e how sui able he use o an
asymme ic measu e o simila i y is o p edic ing mig a ion. In his wo k,
we decided o e alua e he impac o adding hese measu es o cul u al sim-
ila i y o a g a i y model o p edic mig a ion. In o de o es he Facebook
measu es agains he mos s ingen baseline possible, we compa ed i s p e-
dic i e capaci y wi h ha o measu es o cul u al simila i y de i ed om he
WVS (Ingleha 1997) and Fou squa e da a (Sil a e al. 2014).
154 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
Facebook Ads da a ha e become an impo an ool in demog aphic
esea ch, especially o s udying mig a ion pa e ns (Leasu e e al. 2023;
Zagheni, Webe , and Gummadi 2017; Dubois e al. 2018; Spy a os e al.
2019; Alexande , Polimis, and Zagheni 2019; Palo i e al. 2020). Howe e ,
he s udy o in e na ional mig a ion and he de elopmen o models o ex-
plain and p edic lows o people be ween coun ies a e no new (Massey
e al. 1993). One o he mos adi ional p edic ion app oaches is based on
g a i y- ype models (Tinbe gen 1962; Lewe and Van den Be g 2008; Co-
hen e al. 2008; Ramos 2016). Fo example, Cohen e al. (2008) de eloped
an algo i hm o p ojec u u e numbe s o in e na ional mig an s om any
coun y o egion o any o he coun y o egion. The model conside s he
popula ion and he geog aphical a ea o he o igin and he des ina ion coun-
y and he geog aphic dis ance be ween he o igin and des ina ion. Subse-
quen ly, esea che s ha e added o his model by iden i ying o he a iables,
such as social a iables (e.g., mo ali y a e) (Kim and Cohen 2010); his o -
ical a iables, such as sha ed his o y and sha ed language (Lewe and Van
den Be g 2008; Beine, Be oli, and Mo aga 2016; Kim and Cohen 2010;
Ca agliu e al. 2013; Abel, Rayme , and Guan 2019); and online sea ch key-
wo ds (Böhme, G öge , and S öh 2020). Fo ins ance, Böhme, G öge , and
S öh (2020) showed how geo- e e enced online sea ch da a can be used o
measu e mig a ion in en ions in o igin coun ies and o p edic bila e al mi-
g a ion lows. Mo eo e , dis ance measu es ha go beyond es ima ing geo-
g aphic dis ance o, o example, assess adminis a i e, poli ical, economic,
o cul u al dis ance (Ghemawa 2001) a e impo an a iables ha should
be conside ed by mig a ion p edic ion models.
Mos o he exis ing s udies ha analyzed cul u al changes in ela-
ion o mig a ion we e es ic ed o one o a ew coun ies, o , i hey ook
a b oade in e na ional pe spec i e, used cul u al dis ance measu es ha
we e symme ic by cons uc ion (Rapopo , Sa doschau, and Sil e 2020).
Since mig a ion is nei he homogeneous ac oss coun ies no symme ic,
we apply an asymme ic measu e o cul u al simila i y ac oss many coun-
ies in o de o mo e accu a ely ep esen p ocesses o in e na ional cul-
u al exchange.
Da a
In his sec ion, we desc ibe he main da a sou ces we used o collec in e na-
ional da a o ou p edic ion models. We p esen he da a sou ces we used
o measu e he cul u al and ood and d ink simila i y be ween coun ies:
he WVS (Ingleha 1997), Fou squa e (Sil a e al. 2014), and Facebook
Ads (Viei a e al. 2022). We also p o ide a desc ip ion o he da a sou ces
o g a i y model a iables such as popula ion, a ea, and geog aphic dis-
ance, as well as o mig a ion lows as he ou come a iable. To ensu e he
compa abili y o ou esul s wi h p e iously sugges ed indices o simila i y
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 155
based on Fou squa e da a, ou analysis ocuses on a subse o 16 o he mos
popula coun ies by numbe o Fou squa e check-ins (Sil a e al. 2014).
The coun ies selec ed o he analysis we e chosen based on a com-
p omise ac oss h ee di e en c i e ia. Fi s , we wan ed o ma ch he lis
o coun ies selec ed by Sil a e al. (2014) o allow o compa isons be-
ween he p e ious li e a u e using Fou squa e da a and ou esul s using
Facebook da a. Second, he coun ies we chose co e a la ge po ion o geo-
g aphic a eas and popula ions ac oss he wo ld. Finally, and impo an ly, we
selec ed coun ies wi h high Facebook pene a ion a es: A gen ina, Aus-
alia, B azil, Chile, G ea B i ain, F ance, Indonesia, Japan, Sou h Ko ea,
Malaysia, Mexico, Russia, Singapo e, Spain, Tu key, and he Uni ed S a es.
In o he wo ds, we a o ed a choice o coun ies wi h compa a i ely low
and consis en biases o e a b oade selec ion o coun ies ha would be
mo e he e ogeneous in e ms o biases and o which he in e p e a ion o
esul s would be mo e complex.
Facebook ads da a
Viei a e al. (2022) collec ed da a ega ding Facebook use s’ in e es s in ood
and d ink and p oposed measu es o cul u al simila i y be ween coun ies.
The da a collec ed om Facebook Ads e e o he numbe o Facebook
mon hly ac i e use s (i.e., ac i e o e he pas 30 days) who ma ched he
demog aphic a ibu es a ge ed a he ime o da a collec ion. The Facebook
Ma ke ing API enables ma ke e s and esea che s o es ima e he mon hly
ac i e Facebook use coun o a p oposed ad e isemen , aligning wi h
speci ied inpu c i e ia (Kosinski e al. 2015). The pla o m p o ides a se o
cus omizable demog aphic a ibu es, such as age, gende , home loca ion,
and in e es s, allowing ad e ise s o ailo hei inpu que ies. A ibu es
like age, gende , and loca ion a e explici ly decla ed by he use s in hei
p o iles, whe eas in e es s can be ei he decla ed by he use o in e ed by
Facebook based on use ac i i ies such as pos ing o in e ac ing wi h con-
en (e.g., liking con en , sha ing con en , o upda ing one’s s a us). The
me hodology p oposed by Viei a e al. (2022) consis ed o selec ing a subse
o popula oods and d inks o each coun y and hen c ea ing a ec o ep-
esen a ion acco ding o Facebook use s’ in e es s in hose oods and d inks.
Finally, hey measu ed he simila i y be ween hose coun y-le el ec o s.
Me hodological de ails a e a ailable in Viei a e al. (2022).5Measu es de-
i ed om he Facebook Ads da a, including he code used o analyze he
da a se s and gene a e he igu es, a e a ailable in a public web eposi o y.6
Two measu es o simila i y we e p oposed by Viei a e al. (2022)—
Facebook asymme ic simila i y and Facebook symme ic simila i y—
depending on he subse o ood and d ink used o c ea e he ec o ep-
esen a ions. The asymme ic simila i y be ween wo coun ies, c1and c2,is
measu ed in e ms o he mos popula ood and d ink in c1, whe eas he
156 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
simila i y be ween c2and c1is measu ed in e ms o he mos popula ood
and d ink in c2. In his case, since he simila i y be ween c1and c2is di e en
om he simila i y be ween c2and c1, his measu e is no symme ic.
Howe e , we could also measu e he simila i y be ween wo coun ies
by conside ing a ixed se o in e es s o bo h coun ies. In his case, we can
e e o he measu e as symme ic simila i y, co esponding o he measu e o
simila i y be ween wo coun ies conside ing he union o he mos popu-
la ood and d ink in hese coun ies. Since he subse o in e es s is ixed,
he simila i y be ween c1and c2is equal o he simila i y be ween c2and
c1. In ou models, we e e o he Facebook asymme ic measu e o simi-
la i y as Facebook asymme ic simila i y— ood o igin o ood des ina ion—
depending on which subse o op ood and d ink, om he coun y o o igin
o he coun y o des ina ion, we conside ed in he measu e o simila i y.
We e e o he symme ic measu e o simila i y as Facebook symme ic
simila i y.
Fo he Facebook measu es o ood and d ink simila i y (Viei a e al.
2022), selec ed he op 507 ypes o ood and d ink in each coun y. In his
case, he asymme ic measu e o simila i y be ween wo coun ies, c1and
c2, co esponds o he cosine simila i y be ween he 50-dimensional ec o
ep esen a ion o each coun y in e ms o he 50 op oods and d inks in
coun y c1. The symme ic measu e o simila i y be ween wo coun ies, on
he o he hand, is gi en by he cosine simila i y be ween he ec o ep e-
sen a ion o each coun y in e ms o he 394 oods and d inks. The se o
394 in e es s co esponds o he union o he op 50 in e es s in each o he
16 coun ies.
The main aim o his pape is o assess he ex en o which he consid-
e ed Facebook measu es o cul u al simila i y—using only ood and d ink
as cul u al ma ke s—a e meaning ul p edic o s o mig a ion lows. In his
sense, i is impo an o alida e ou esul s ob ained wi h he Facebook Ads
da a and o ensu e hei compa abili y wi h he esul s o p io esea ch.
We selec ed he wo mos ele an da a se s o compa ing measu es o cul-
u al simila i y: he WVS and Fou squa e. The WVS is an es ablished and
adi ional da a se based on la ge-scale ep esen a i e su ey da a along
se e al cul u al dimensions e lec ing cul u al no ms, alues, and belie s.
The Fou squa e da a se (Sil a e al. 2014) is based on da a on use s’ ood
and d ink habi s collec ed om Fou squa e check-ins. Al hough he mea-
su es o cul u al simila i y om he WVS and he Fou squa e da a ocus
on di e en aspec s o cul u e, bo h measu es a e used as baselines o he
Facebook measu es o ood and d ink simila i y. Howe e , as men ioned
be o e, bo h da a se s ha e signi ican limi a ions. The main disad an ages
o he WVS a e he cos s and he ope a ional ime needed o elease new
su ey wa es, which ha e ypically been conduc ed o abou i e yea s.
The Fou squa e pla o m, in con as , is no as widely used as Facebook. In
addi ion, he pla o m is hea ily biased om a demog aphic poin o iew
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 163
is no eliable, since mo e a iables always inc ease he me ic, e en i new
a iables a e only ma ginally p edic i e. To add ess his issue, we included
o he measu es ha penalize he numbe o a iables in hei calcula ion.
The las column in Table 1 shows he Wa anabe–Akaike o widely appli-
cable in o ma ion c i e ion (WAIC) (Gelman e al. 1995) o each o he
models es ed. We used he ull inpu da a se (wi hou c oss- alida ion),
which consis s o 240 pai s o coun ies. The lowe he Wa anabe–Akaike,
he mo e closely a model can p edic he ac ual obse a ions. Table T1 also
shows he adjus ed R-squa ed and he esul ing coe icien s and s a is ics
o each o hese models. In his able, he columns ep esen he models
and he ows display each o he a iables in he p edic ion model. In he
nex sec ion, he esul ing coe icien s and s a is ics om Table 1 and Table
T1 a e desc ibed in mo e de ail.
Resul s
Table T1 shows in de ail all he coe icien s o each o he a iables in-
cluded in he g a i y models ha we es ed using he ull inpu da a se
co esponding o 240 pai s o coun ies. Table 1 shows he esul s a e aged
ac oss c oss- alida ions, excep o he WAIC, which was calcula ed om
he model using he ull inpu da a se . To e alua e he impac o adding
measu es o cul u al simila i y o he mig a ion model, we i s assess he
co ela ion be ween he a iables conside ed.
Figu e 1 shows he co ela ion be ween each o he measu es o sim-
ila i y and mig a ion lows (in he loga i hm scale) be ween each pai
o coun ies wi hin he 16 coun ies we analyzed. We obse ed ha he
symme ic and asymme ic measu es o ood and d ink simila i y de-
i ed om Facebook Ads da a showed a posi i e co ela ion (0.38). Sim-
ila ly, he measu es o ood and d ink simila i y based on Fou squa e da a
and Facebook Ads da a we e also posi i ely co ela ed (0.37 be ween he
Fou squa e measu e and he Facebook symme ic measu e and 0.32 be-
ween he Fou squa e measu e and he Facebook asymme ic measu e).
The Fou squa e and Facebook measu es o ood and d ink simila i y we e
highly co ela ed wi h each o he and cap u ed simila pa e ns o ood and
d ink in e es s ac oss coun ies. Nex , we compa ed he cul u al simila i-
ies de i ed om social media wi h hose de i ed om he WVS. Al hough
cul u al simila i ies based on he WVS da a and he Facebook symme ic
measu e did no cap u e he same cul u al a ibu es and we e no subs an-
ially associa ed (0.07), he WVS da a we e posi i ely co ela ed wi h bo h
Facebook asymme ic (0.14) and Fou squa e measu es o ood and d ink
simila i y (0.28).
The measu es o ood and d ink simila i y de i ed om social media
da a did no exhibi a s ong co ela ion wi h he me ics ob ained om he
su ey da a. Whe eas he me ics de i ed om he WVS da a encompassed
164 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
cul u e in e ms o no ms, alues, and belie s, he me ics de i ed om he
Fou squa e and Facebook da a p ima ily emphasized he in e es in ood
and d ink as cul u al ma ke s. The di e ences in he na u e o he da a
sugges ha he WVS cul u al simila i y measu e cap u ed di e en aspec s
o cul u e ha we e no e lec ed by he in e es s in ood and d ink d awn
om he social media da a. The measu es de i ed om he Fou squa e and
Facebook da a, on he o he hand, we e highly co ela ed o each o he and
cap u ed a simila pa e n o ood and d ink in e es s ac oss coun ies.
We obse ed ha cul u al simila i y based on he WVS da a was no
posi i ely co ela ed wi h mig a ion lows (−0.01). This esul means ha
coun ies ha we e close o each o he in he WVS cul u al map had sligh ly
smalle mig a ion lows be ween hem. Table T1 shows a signi ican nega-
i e e ec o he WVS cul u al simila i y on mig a ion lows, meaning ha
a high WVS cul u al simila i y was associa ed wi h smalle mig a ion lows.
Despi e he unexpec ed nega i e e ec o he WVS cul u al simila i y on
mig a ion lows, he adjus ed R-squa ed imp o ed (0.83) when he WVS
cul u al simila i y was added o he g a i y model. This esul con i ms he
impo ance o aking a coun y’s cul u al no ms, alues, and belie s in o ac-
coun when i ing mig a ion models (Esses 2018; Ca agliu e al. 2013). In
con as , he measu es de i ed om social media da a all showed a posi i e
co ela ion wi h mig a ion lows (Fou squa e 0.41; Facebook Ads asymme -
ic 0.27 and 0.35, Facebook Ads symme ic 0.3), which means ha a high
ood and d ink simila i y was associa ed wi h la ge mig a ion lows.
E en wi h his s ingen baseline o adding geog aphic and economic
a iables, we ound ha including measu es o cul u al simila i y de i ed
om Facebook da a ocusing on ood and d ink imp o ed p edic ions be-
yond wha could be achie ed wi h all hese o he p edic o s. The coe icien s
om Model 1, including he Facebook measu es o ood and d ink simila -
i y, we e s a is ically signi ican , and he p edic i e capaci y o he model
inc eased. This sugges s ha he Facebook measu es o ood and d ink sim-
ila i y a e impo an p edic o s o mig a ion, cap u e di e en pa e ns, and
can be used o iden i y di ec ional p ocesses. This is no he case o he
model ha includes all he basic a iables and sha ed language and his-
o y (Model 2). In o he wo ds, we did no obse e imp o emen s when
adding he Facebook measu es o ood and d ink simila i y, which means
ha he e is likely an o e lap in he explana o y powe o he measu es o
ood and d ink simila i y de i ed om Facebook da a and sha ed language
and his o y. Figu e A1 in he online Appendix shows he signi ican posi i e
co ela ion be ween he measu es o ood and d ink simila i y de i ed om
Facebook da a and sha ed language and his o y. The es ima ed coe icien s
in Table T1 o he measu es o cul u al simila i y based on in e es s in ood
and d ink de i ed om Facebook da a in Model 2 we e no signi ican . Fi-
nally, e en hough he measu es o cul u al simila i y de i ed om he WVS
da a and he social media da a cap u ed di e en aspec s o cul u e, we did
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 165
no obse e signi ican coe icien s when we added hem o he model ha
included all he basic a iables and he WVS cul u al simila i y (Model 3).
Howe e , we obse ed a sligh imp o emen in he adjus ed R-squa ed com-
pa ed o Model 3 when he measu es de i ed om he Fou squa e da a and
he asymme ic measu e o ood and d ink simila i y de i ed om he Face-
book da a we e added o he model.
Despi e he small imp o emen in he p edic ion o mig a ion lows
o he ime poin conside ed, he measu es o ood and d ink simila i y de-
i ed om he Facebook da a had a p edic i e capaci y compa able o ha
o he classic a iables used in he li e a u e, such as sha ed language and
sha ed his o y. In addi ion, he measu es o ood and d ink simila i y de-
i ed om he Facebook da a con ibu ed o p edic i e models o mig a ion
by adding no jus a imely bu also an asymme ic componen . The mea-
su es o simila i y om he Facebook da a we e measu ing coun y-le el
indices o cul u al in e es s, which can shi p ecisely h ough mig a ion.
While sys ems o belie wi hin a single cul u e (e.g., he majo i y cul u e in
a coun y) should no change quickly, he a io o he majo i y cul u e o he
mino i y cul u e(s) can shi , leading o changes in coun y-le el in e es s
ha would be obse able in digi al ace da a. Pa icula ly gi en he limi-
a ions o o he measu es o cul u al simila i ies, he use o Facebook Ads
da a can p o ide an e ec i e means o cap u ing such changes in a way ha
complemen s o he measu es.
Figu e 2 shows a compa ison be ween he expec ed mig a ion lows
es ima ed by Abel and Cohen (2019) and he mig a ion lows p edic ed by
each model. The o ange line ep esen s he expec ed dis ibu ion, whe e
he p edic ed mig a ion low is equal o he expec ed mig a ion low. The
dis ibu ion o he do s, which co esponds o pai s o coun ies, changes
om one model o he o he , and he p edic ions become close o he ex-
pec ed alues o mig a ion lows. O e all, we obse ed ha he baseline
and he mo e adi ional models o e es ima ed mig a ion lows o pai s
o coun ies be ween which he e was li le mig a ion, and unde es ima ed
mig a ion lows o pai s o coun ies be ween which he e we e la ge mi-
g a ion lows. This pa e n became sligh ly less e iden wi h he inclusion o
o he a iables, including he measu es o ood and d ink simila i y de i ed
om Facebook.
Discussion
We showed he impac o adding measu es o cul u al simila i y, de i ed
om bo h su ey da a and social media da a, o g a i y models in o de
o p edic mig a ion lows. Ou esul s indica ed ha he measu e o cul-
u al simila i y de i ed om he WVS da a helped o imp o e mig a ion
p edic ion and ha he measu e o ood and d ink simila i y de i ed om
Fou squa e da a was highly co ela ed wi h mig a ion lows. Howe e ,
166 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
FIGURE 2 Compa ison be ween he expec ed mig a ion lows (x-axis) and
he mig a ion lows p edic ed (y-axis) by each one o he models using he
ull inpu da a se (240 pai s o coun ies). Bo h axes a e on a loga i hmic
scale. Each do ep esen s a pai o coun ies wi hin he 16 coun ies we
analyzed
in e ms o scalabili y and ep oducibili y, he use o hese measu es may
ha e some disad an ages. As was men ioned be o e, su eys a e cos ly and
equi e subs an ial ope a ional ime. Fo example, he WVS is ca ied ou
e e y i e yea s. The Fou squa e da a ha e a di e en se o limi a ions.
In pa icula , he Fou squa e pla o m is no as widely used as Facebook,
and i is hea ily biased om a demog aphic poin o iew. Mo eo e , he
Fou squa e da a se we conside ed was o e i e yea s olde han he
Facebook Ads da a and o e six yea s olde han he WVS da a. Du ing his
pe iod o ime, signi ican cul u al changes may ha e happened, gi en ha
he wo ld is con inuously changing in e ms o connec i i y ac oss egions.
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 167
Wi h mo e han 2.7 billion use s wo ldwide,19 Facebook cap u es a la ge
and mo e di e se popula ion han o he social media. Conside ing he
a ailabili y o Facebook’s da a, ou me hodology could be easily scaled o
conside mo e coun ies. Mo eo e , he da a om Facebook Ads a e eely
a ailable and can be con inuously upda ed and collec ed, which makes
his app oach imely, cos -e ec i e, ep oducible, and scalable. Thus, he
ele ance o hese ypes o analyses in adi ionally da a-poo con ex s, like
in low- and middle-income coun ies, will likely inc ease in he u u e.
Gi en he ad an ages o using Facebook Ads da a, we p o ided a s in-
gen es o he inc emen al e ec s o he Facebook measu es, and ou e-
sul s showed ha cul u al simila i y, as measu ed by ood and d ink in e -
es s, explained mig a ion lows o an ex en ha was compa able o ha
o s anda d p edic o s such as sha ed language and sha ed his o y. Besides
he ad an ages o using Facebook da a o measu e ood and d ink simi-
la i y, he app oach we p esen ed had addi ional ad an ages due o i s use
o an asymme ic measu e o simila i y. Mos o he g a i y models elied
on symme ic a iables o p edic mig a ion, which is i sel an asymme ic
phenomenon. Since he mig a ion lows be ween coun ies a e asymme -
ic (e.g., he e a e mo e Chileans in Spain han Spania ds in Chile), we
would expec ha he simila i y in e ms o ood and d ink in e es s would
be asymme ic as well (e.g., he e a e mo e Chileans in e es ed in Span-
ish ood han Spania ds in e es ed in Chilean ood). This phenomenon was
e lec ed in he coe icien s o he model using he Facebook asymme ic
simila i y measu e, which showed a s onge e ec on he popula i y o
he des ina ion coun y’s dishes in he coun y o o igin han ice e sa.
Fo example, Chileans’ in e es in Spanish dishes would be a s onge p e-
dic o o how many Chileans mo ed o Spain han Spania ds’ in e es in
Chilean dishes. We ound e idence o asymme ic pa e ns, such ha he
cul u al ma ke s in he coun y o des ina ion we e mo e closely associa ed
wi h mig a ion lows han he cul u al ma ke s in he coun y o o igin. We,
he e o e, ecommend ha u u e esea ch in his a ea ake asymme y in o
accoun when p edic ing mig a ion.
To he bes o ou knowledge, ou s udy is he i s o p opose a scal-
able, apidly a ailable, and asymme ic measu e o simila i y de i ed om
social media da a o p edic mig a ion. Ou indings con ibu e o he li e -
a u e by (i) showing he impo ance o cul u al simila i y, as de i ed om
ood and d ink in e es s in social media da a, o p edic ing mig a ion; and
(ii) allowing o apid p edic ions o cu en mig a ion lows ahead o he
elease o o icial s a is ics. Fo ins ance, Leasu e e al. (2023) le e aged da a
om Facebook Ads o moni o in eal- ime subna ional popula ion sizes and
in e nal displacemen in Uk aine on a daily basis, disagg ega ed by age and
sex. Simila ly o Leasu e e al. (2023)’s wo k, ou me hodology could cap-
u e apid changes in popula ions’ in e es s ac oss coun ies, o ins ance,
due o unexpec ed mig a ion, and could help in p edic ing mig a ion lows.
168 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
The p ima y objec i e o his s udy was o assess he alue o exam-
ining cul u al simila i y when s udying mig a ion. Speci ically, we aimed
o es measu es o cul u al simila i y based on ood and d ink in e es s in
social media o p edic in e na ional mig a ion lows. Measu es o cul u al
dis ance a e di icul o es ima e and hus ha e no ye been widely adop ed
in g a i y models o assessing and p edic ing mig a ion. Howe e , cul u e
plays an impo an ole in he p ocesses o mig a ion.
As was men ioned be o e, he ela ionship be ween mig a ion and cul-
u e is likely bidi ec ional, since cul u al i in e ms o language, no ms, and
alues is an impo an ac o ha people conside be o e mo ing be ween
coun ies, and mig an s ansmi cul u al elemen s om hei o igin coun-
y o hei home coun y and back again du ing he mig a ion p ocess.
In his pape , we ocused on showing how measu es o cul u al simila -
i y de i ed om Facebook use s’ ood and d ink in e es s can be used o
explain mig a ion lows be ween coun ies. Fo ins ance, imagine ha he
numbe o Facebook use s li ing in he Uni ed S a es who a e in e es ed in
some adi ional dishes om B azil has inc eased. One possible eason o
his de elopmen is ha he numbe o B azilian immig an s in he Uni ed
S a es has inc eased, and hus he numbe o Ame icans who a e exposed o
B azilian in e es s has isen. In his example, i hese B azilian immig an s
es ablished a big B azilian communi y in he Uni ed S a es, he numbe o
B azilian immig an s could inc ease e en mo e. In his case, he numbe o
Facebook use s in e es ed in B azilian ood and d ink se es as a p oxy o
he B azilian communi y es ablished in he Uni ed S a es. One o ou main
esul s shows he impo ance o he cul u al simila i y be ween coun ies,
as measu ed by Facebook use s’ in e es s in ood and d ink, o p edic ing
mig a ion lows be ween hese coun ies.20 While ou s udy has b oade
ami ica ions, he scope o his a icle is mo e limi ed, as we showed he
posi i e associa ion be ween cul u al simila i y and mig a ion lows wi h-
ou a emp ing o es ablish a causal di ec ion in his complex bidi ec ional
ela ionship. Fu u e s udies could collec addi ional da a and de elop new
me hods o add ess his issue and mo e owa d p o iding mo e causal es-
ima es o he di ec ion o he ela ionship be ween cul u e and mig a ion
lows.
Cau ion should be exe cised when in e p e ing ou esul s due o hei
limi a ions, which we would like o acknowledge. Fi s , he p esen analysis
is cons ained by da a a ailabili y: only 16 coun ies we e included in ou
analysis. The 16 coun ies selec ed o he analysis we e chosen o ma ch
he lis o coun ies included in Sil a e al. (2014) in o de o enable us o
compa e he esul s om di e en ypes o social media da a. As well as o
ensu e he compa abili y o ou indings wi h hose o o he s udies, we se-
lec ed hese 16 coun ies in o de o co e a la ge and di e se po ion o he
wo ld’s egions and o include coun ies whe e he Facebook pene a ion
a e is high, hus educing he po en ial size o he biases in he da a. The
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 169
da a collec ion could be ex ended o mo e coun ies. Howe e , while he
Facebook audiences’ in e es s o he mos cu en pe iod could be collec ed,
he lack o adequa e mig a ion da a emains a c ucial bo leneck. The las
ime pe iod o which global es ima ions o mig a ion low da a a e a ail-
able om ou main sou ce (Abel and Cohen 2019) is 2015–2019. In o he
wo ds, we do no ha e mig a ion low da a, o e en mig a ion s ock da a,
a e 2019. Mo eo e , he COVID-19 pandemic a ec ed mig a ion, and we
do no ha e upda ed da a ha we could use as a dependen a iable in ou
models. Once new mig a ion da a a e a ailable, new da a om Facebook
can be collec ed in eal- ime, and he p edic ions can be upda ed.
The measu es o simila i y ha we used elied only on da a ega d-
ing Facebook use s’ in e es s in ood and d ink. Al hough he cuisine o
a coun y is an impo an cul u al ma ke o s udying cul u al simila i y,
he p oposed me hodology could be used wi h o he ypes o a ibu es and
in e es s, which migh be ele an o s udies wi h o he goals o angles.
We expec ha a b oade ope a ionaliza ion o measu es o cul u e would
lead o he de elopmen o models wi h e en highe p edic i e accu acy. In
his sense, wha we showed is likely a lowe bound in e ms o p edic i e
capaci y.
Mo eo e , he social media da a we used, including he Facebook Ads
da a, and he in e es ca ego ies p o ided by Facebook may no be exhaus-
i e o ep esen a i e. Facebook da a include a numbe o biases, gi en ha
Facebook use s a e no necessa ily ep esen a i e o he unde lying pop-
ula ion in hei espec i e coun ies. The e is a g owing li e a u e ha has
expanded ou knowledge on how o iden i y and co ec biases in social me-
dia da a.21 In addi ion o ep esen a i i y, he classi ica ion o use s in o he
ca ego ies p o ided on Facebook Ads could be a sou ce o bias. G ow e al.
(2022) e alua ed he bias ega ding loca ion, age, and gende on Facebook.
The au ho s compa ed he in o ma ion p o ided by he pa icipan s o an
anonymous online su ey wi h Facebook Ads’ classi ica ion o he same in-
di iduals. The esul s showed ha abou 86–93% o esponden s’ answe s
ma ched Facebook’s classi ica ion. Al hough loca ion, age, and gende ap-
pea o be iden i ied mos ly co ec ly on Facebook Ads, he accu acy o he
classi ica ion o Facebook use s’ in e es s has no been es ed sys ema ically.
We hypo hesize ha Facebook co e s only a subse o use s’ in e es s, bu
u u e esea ch is needed o assess he ex en o which he ep esen a ion
o in e es s is accu a e.
We would like o emphasize he impo ance o add essing issues such
as biases in digi al ace da a, and we poin he eade s o he esou ces
men ioned abo e o a se ies o app oaches de eloped o ackle his p ob-
lem. Wi h ou a icle, we a e en e ing pa ially uncha ed e i o y in e ms
o assessing he biases ela ed o ou me hods, as ou app oaches a e no el,
and a e no ye pa o he con en ional oolbox. We hope ha ou s udy
will u he s imula e me hodological esea ch on iden i ying and co ec ing
170 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
biases when s udying cul u al dimensions using social media da a. While he
eade should be awa e ha he da a used in his a icle a e no necessa -
ily ep esen a i e o he en i e unde lying popula ions, we should also no e
ha ou decision o ocus on 16 coun ies wi h high Facebook pene a ion
a es limi ed he ex en o he biases, as i allowed o compa isons ac oss
coun ies whe e Facebook is used in ela i ely simila ways by compa able
demog aphic segmen s o he popula ion. I is no ewo hy ha , despi e he
biases, he p edic i e model pe o ms e y well. Once he biases a e ully
modeled, we expec ha he p edic i e capaci y will inc ease. We hope ha
his a icle lays he ounda ion o u he analyses ha can help us be -
e unde s and hese da a and hei po en ial, especially in coun ies and
con ex s ha ha e his o ically been da a-poo .
Conclusion
In his pape , we demons a ed ha measu es o cul u al simila i y de i ed
om su ey and social media da a can be impo an a iables in p edic ions
o mig a ion lows. We compa ed a measu e o cul u al simila i y de i ed
om he WVS wi h measu es o ood and d ink simila i y de i ed om
Fou squa e and Facebook Ads da a. By using he measu es de i ed om
he Facebook Ads da a, we in oduced a mo e nuanced iew o symme ic
and asymme ic measu es o simila i y and showed how hese measu es o
simila i y can be used o explain mig a ion lows be ween coun ies. Ou
esul s indica ed ha he Facebook measu es o ood and d ink simila i y
can play an impo an ole in p edic ing mig a ion, as hey a e compa able
o s anda d p edic o s, such as sha ed language and sha ed his o y. Finally,
while we ound ha some a iables, such as sha ed language, his o y, and
geog aphic dis ance, a e s a ic and symme ic, we also obse ed ha cul u al
a ibu es om daily li e a e sensi i e o changes in he en i onmen and
can be ep esen ed as an asymme ic measu e o simila i y be ween coun-
ies, hus adding alue o models o mig a ion om bo h a subs an i e and
a p edic i e pe spec i e.
Acknowledgmen s
The au ho s g a e ully acknowledge he esou ces p o ided by he In e na-
ional Max Planck Resea ch School o Popula ion, Heal h, and Da a Science
(IMPRS-PHDS) and he Max Planck Ins i u e o Demog aphic Resea ch
(MPIDR).
Da a a ailabili y s a emen
Acco ding o Facebook’s Te ms o Se ice, he aw da a collec ed om he
Facebook Ad e ising Pla o m canno be sha ed publicly. In his case, we
CAROLINA COIMBRA VIEIRA /SOPHIE LOHMANN /EMILIO ZAGHENI 171
do no sha e he aw da a. Ins ead, he eposi o y con ains all he da a used
in ou models, including he measu es de i ed om he Facebook Ads da a
and he code o eplica e all he analyses and gene a e he igu es (see h ps:
//gi hub.com/ca olcoimb a/g a i y- b).
No es
1 Al hough Facebook p o ides a ange
o in e es s b oadly ela ed o ood and d ink,
mos o hose in e es s do no ep esen ood
as na u ally ound in na u e (e.g., g apes,
co n). Addi ionally, Viei a e al. (2022) man-
ually alida ed he da a se by emo ing in-
e es s such as es au an s and b and names.
The majo i y o he in e es s ela ed o ood
and d ink on Facebook ep esen dishes o
any ood o d ink p ocessed by humans (e.g.,
wine, quesadilla).
2 h ps://ich.unesco.o g/en/lis s? e m[]
= ocabula y_ hesau us-10
3 h ps://www.ho s ede-insigh s.com/
models/na ional-cul u e
4 h ps://ou wo ldinda a.o g/explo
e s/mig a ion? ime=la es & ace =none
&Me ic=Ne +mig a ion+ a e&Pe iod
=To al&Sub-me ic=To al
5 h ps://jou nals.plos.o g/plosone/
a icle?id h ps://doi.o g/10.1371/jou nal.
pone.0262947
6 h ps://gi hub.com/ca olcoimb a/
cul u al-simila i y- b
7 We conduc ed addi ional analyses o
show how s able he esul s a e when we
a y he numbe o in e es s we conside in
he Facebook measu es o simila i y. The op
50 oods and d inks gene a e he bes esul s
based on all he calcula ed me ics, such as
he adjus ed R-squa ed, and signi ican coe -
icien s.
8 h ps://b andongaille.com/26-g ea -
ou squa e-demog aphics
9 h ps://www.s a is a.com/s a is ics/
814726/sha e-o -us-in e ne -use s-who-
use- ou squa e-by-age
10 h ps://99 i ms.com/blog/ ou squa e-
s a is ics/#g e
11 h ps:// inancesonline.com/
ou squa e-s a is ics
12 h ps:// ou squa e.com/p oduc s/
p icing
13 h ps://www.wo ld aluessu ey.
o g/WVSCon en s.jsp
14 The mos ecen se en h wa e o he
WVS (2017-2022) co e s 80 coun ies.
15 h ps://www.un.o g/en/de elopmen /
desa/popula ion/mig a ion/da a/es ima es2/
es ima es19.asp
16 h ps://www.un.o g/de elopmen /
desa/pd/con en /in e na ional-mig an -
s ock
17 h ps://da abank.wo ldbank.o g/
home
18 The loga i hm scale used in his s udy
is he loga i hm base 10.
19 h ps://www. acebook.com/iq/
insigh s- o-go/2740m- acebook-mon hly-
ac i e-use s-we e-2740m-as-o -sep embe -
30
20 We conduc ed addi ional analyses o
in es iga e he ole o he immig an com-
muni y in he hos coun y in shaping he
signi ican coe icien s obse ed o cul u al
simila i ies. We added mig a ion s ocks om
2019 o all he models and obse ed ha
o e all he coe icien s ega ding cul u al
simila i ies dec eased by 30% bu we e s ill
signi ican . This esul indica es ha e en
hough ood and d ink om he o igin coun-
y may ha e been in oduced o he des i-
na ion coun y by immig an s, he in e es
in hose ood and d ink canno be ully ex-
plained by he size o he immig an popula-
ion. In o he wo ds, he in e es in ood and
d ink om he o igin coun y is sp ead ac oss
he popula ion in he des ina ion coun y.
21 The e is a g owing li e a u e ha has
expanded ou knowledge on how o iden-
i y and co ec biases in social media da a.
One line o esea ch has ocused on iden i y-
ing he di e en ypes o e o s and biases in
s udies ha use digi al ace da a and on o -
172 THE VALUE OF CULTURAL SIMILARITY FOR PREDICTING MIGRATION
ganizing hem in a amewo k (Ol eanu e al.
2019; Sen e al. 2021). Fo ins ance, Sen e al.
(2021) p oposed a ca ego iza ion based on
he o al su ey e o amewo k o iden i y
se e al ypes o e o s ha may occu in s ud-
ies ha use digi al aces. As a consequence,
hese amewo ks also con ibu e o c ea ing
a common ocabula y be ween esea che s
using digi al ace da a. In addi ion, D ouho
e al. (2023) p o ided an o e iew o how
some inno a i e da a se s and me hodologi-
cal ools can en ich mig a ion esea ch. De-
spi e all he ad an ages and p omises o us-
ing digi al ace da a o mig a ion esea ch
(e.g., less ime and cos s needed o le e -
age da a o a la ge sample size), he au-
ho s poin ed ou some o he challenges
ha can a ise when wo king wi h hese
da a. Since digi al ace da a a e no gene -
a ed o esea ch pu poses, some ex a ca e
is equi ed o epu pose hei meaning, as
hey migh o he wise be oo supe icial o
inapp op ia e o add essing many cen al
esea ch ques ions. Besides conce ns abou
da a quali y, some o he key challenges in-
ol ed in wo king wi h hese da a a e e-
la ed o e hical conside a ions and selec ion
bias.Zagheni and Webe (2015) conside ed
he p oblem o selec ion bias in non ep e-
sen a i e samples, such as digi al ace da a,
and p oposed wo main app oaches o e-
duce bias: he calib a ion app oach and he
di e ence-in-di e ences app oach. In he
calib a ion app oach, he online da a a e ad-
jus ed based on eliable o icial s a is ics, in-
cluding h ough he gene a ion o co ec ion
ac o s (Zagheni and Webe 2012; Zagheni,
Webe , and Gummadi 2017; Ribei o, Ben-
e enu o, and Zagheni 2020). Fo ins ance,
Ribei o, Bene enu o, and Zagheni (2020)
compa ed da a om Facebook Ads and he
US Census and calcula ed co ec ion ac o s
o some demog aphic dimensions, such as
age, gende , educa ion, and income. How-
e e , o con ex s whe e no eliable s a is i-
cal da a a e a ailable, he au ho s sugges ed
a di e ence-in-di e ences app oach o e al-
ua e ela i e changes p e- and pos -e en
(Flo es 2017; Alexande , Polimis, and Za-
gheni 2019). Alexande , Polimis, and Za-
gheni (2019) used Facebook Ads da a and he
di e ence-in-di e ences app oach o moni-
o lows o ou mig an s om Pue o Rico
be o e and a e Hu icane Ma ia in 2017.
The di e ence-in-di e ences app oach as-
sumes a cons an ela ionship be ween es-
ima es om digi al ace da a and o icial
da a, a leas o e ela i ely sho pe iods o
ime.An eme ging line o esea ch ocuses
on using Bayesian app oaches o combine
di e en sou ces o da a o es ima e mig a-
ion ends (Rampazzo e al. 2021; Alexan-
de , Polimis, and Zagheni 2020; Hsiao e al.
2023). Fo ins ance, in a ecen s udy o-
cused on nowcas ing s ocks o mig an s in he
Uni ed S a es, Alexande , Polimis, and Za-
gheni (2020) demons a ed ha a Bayesian
hie a chical model combining da a om bo h
Facebook and he Ame ican Communi y
Su ey ou pe o ms al e na i e models ha
use only Facebook da a o ha solely ely on
ime-se ies da a om he Ame ican Commu-
ni y Su ey. Recen ly, Leasu e e al. (2023)
buil a eal- ime moni o ing sys em o es-
ima e subna ional popula ion sizes and in-
e nal displacemen in Uk aine by le e ag-
ing da a om Facebook Ads in combina-
ion wi h p e-con lic popula ion da a in
Uk aine.
Re e ences
Abel, Guy J., and Joel E. Cohen. 2019. “Bila e al In e na ional Mig a ion Flow Es ima es o 200
Coun ies.” Scien i ic Da a 6(1): 1–13.
Abel, Guy J., James Rayme , and Qing Guan. 2019. “D i ing Fac o s o Asian In e na ional Mig a-
ion Flows.” Asian Popula ion S udies 15(3): 243–265.
Alexande , Monica, Ki an Polimis, and Emilio Zagheni. 2019. “The Impac o Hu icane Ma ia on
Ou -Mig a ion om Pue o Rico: E idence om Facebook Da a.” Popula ion and De elopmen
Re iew 45(3): 617–630.
Alexande , Monica, Ki an Polimis, and Emilio Zagheni. 2020. “Combining Social Media and Su ey
Da a o Nowcas Mig an S ocks in he Uni ed S a es.” Popula ion Resea ch and Policy Re iew 41:
1–28.
