religious homophily in a secular country

21 Pages·2015·1.53 MB·English

UC San Diego
UC San Diego Previously Published Works
Title
Religious homophily in a secular country: Evidence from a voting game
in France
Permalink
https://escholarship.org/uc/item/7mw654xg
Journal
Economic Inquiry, 53(2)
ISSN
0095-2583
Authors
Adida, CL
Laitin, DD
Valfort, MA
Publication Date
2015-04-01
DOI
10.1111/ecin.12192
Copyright Information
This work is made available under the terms of a Creative Commons
Attribution-NonCommercial-ShareAlike License, availalbe at
https://creativecommons.org/licenses/by-nc-sa/4.0/

Peer reviewed
eScholarship.org Powered by the California Digital Library
University of California
RELIGIOUS HOMOPHILY IN A SECULAR COUNTRY: EVIDENCE FROM
A VOTING GAME IN FRANCE
CLAIREL.ADIDA,DAVIDD.LAITINandMARIE-ANNEVALFORT∗
Homophily—the tendency individuals have to associate with similar-others—is a
powerfuldeterminantofsocialnetworks.Yetresearchtodatedoesnotallowustodeter-
mine which dimension, e.g., ethnic, religious, gender, age, or class similarity, drives
association.Testsdemonstratinghomophilyareflawedbyrestrictingtherangeofdimen-
sions in the choice set. We introduce an experimental game in which we exogenously
expose subjects to diverse partners to determine which dimension dominates. We find
thatinasocio-demographicallydiversedistrictofParis,despiteexpectationsofsecular-
ization,religioussimilaritysignificantlypredictshomophily.Moreover,weprovideten-
tativeevidencethatreligioushomophilyistaste-based.(JELC91,D03,D72,J71,Z12)
be our coethnics2 (Currarini, Jackson, and Pin
I. INTRODUCTION
2009; Wimmer and Lewis 2010). We date those
Homophily—or the concept that individuals with whom we share similar levels of education
who are similar tend to come together—has (Hitsch, Hortacsdu, and Ariely 2010; Skopek,
long been recognized as a powerful determi- Schulz,andBlossfeld2011),anethnicity(Hitsch,
nant of social networks.1 Our friends tend to Hortacsdu,andAriely2010),areligion(Hitsch,
Hortacsdu, and Ariely 2010), a set of values
(Bearman, Moody, and Stovel 2004), and polit-
∗Theauthorsdeclarethattheyhavenorelevantormate- ical beliefs (Huber and Malhotra 2012; Klofs-
rialfinancialintereststhatrelatetotheresearchdescribedin tad, McDermott, and Hatemi 2013).3 Many of
thispaper.TheresearchwasfundedbytheNationalScience
these factors also end up predicting whom we
Foundation, “Muslim Integration into European Union
Societies: Comparative Perspectives,” Grant SES-0819635. marry(Kalmijn1998).Evenourhiringdecisions
(David Laitin, PI). The authors would like to thank Gary rely on homophily of race (Bertrand and Mul-
Cox, James Fowler, Jens Hainmueller, Seth Hill, Rahsaan lainathan 2004) and religion (Adida, Laitin, and
Maxwell, Craig McIntosh, Dan Posner, Molly Roberts,
Valfort2010).
Sebastian Saiegh, Sidney Tarrow, and two anonymous
reviewersforhelpfulfeedback.Allerrorsareourown. Similarity breeds connection, but similarity
Adida:AssistantProfessor,DepartmentofPoliticalScience, of what? Basic socio-demographic characteris-
University of California San Diego, Social Sciences ticssuchasethnicity,class,gender,age,andreli-
Building No. 389, 9500 Gilman Drive No. 0521, La
gionhaveallbeenshowntounderliehomophily
Jolla, CA 92093-0521. Phone [phone], Fax
858-534-7130,[email protected] (McPherson,Smith-Lovin,andCook2001).Our
Laitin:Department of Political Science, Stanford Univer- principal objectives in this paper are first to
sity,Stanford,CA94035-6044.Phone[phone],Fax determinewhichsocio-demographiccharacteris-
650-723-1808,[email protected]
tic emerges as the salient basis for homophily
Valfort:Paris School of Economics–Paris I Panthéon
SorbonneUniversity,Centred’EconomiedelaSorbonne
OfficeRoom318,106-112Boulevarddel’Hôpital,75647
Paris Cedex 13, France. Phone [phone], Fax 2. We follow Chandra’s (2006) definition of ethnicity
[phone], E-mail marie-anne.valfort@univ- as a subset of identity categories in which eligibility for
paris1.fr membershipisdeterminedbydescent-basedattributes;this
includesrace.
3. SeeBurgess,Sanderson,andUmana-Aponte(2011),
1. ThetermwasoriginallycoinedbyLazarsfeldandMer-
p.4forasummaryofthisliterature.
ton(1954)intheirstudyoffriendship.AsMcPherson,Smith-
Lovin,andCook(2001)reckonintheirreviewoftheliterature
ahalf-centurylater,thefocusonthepositivenetworkimpli-
cationsofhomophilygainedlessresearchattentionthanthe ABBREVIATION
focusonprejudice,thenegativesideofthesamecoin.Here
weprincipallyaddressconcernsofwhichaspectsofpeople’s FDR: FalseDiscoveryRate
characteristicsleadthemtoassociatewithoneanother.Thus,
ourfocusisonhomophily.
1
EconomicInquiry doi:10.1111/ecin.12192
(ISSN0095-2583) ©2015WesternEconomicAssociationInternational
2 ECONOMICINQUIRY
when the most common social match-ups are in By relying on an experimental framework
the mix,4 and second to get at least a partial where we impose an exogenous set of potential
understandingofthemechanismsthatsustainit. partners, our paper circumvents this identifi-
Weareamongthefirsttoinvestigatethisques- cation problem. Specifically, we introduce a
tion:researchtodatereliesonexpost(i.e.,post- votingexperimentinwhichplayershavelimited
selection)data,thatis,onpeoplewhoarealready information about, and no previous connec-
matchedoncertainsocio-demographiccharacter- tion to, each other. Experimental subjects are
istics.Forinstance,theAddHealthdataexploited assigned to a group by the investigators, and
by Currarini, Jackson, and Pin (2009) and the instructed to get to know one another one-by-
ALSPAC data exploited by Burgess, Sander- one, within a short timespan (three minutes for
son,andUmana-Aponte(2011)offerinformation each pair). They are then asked to elect a group
about established relationships: scholars mea- leader among the set of partners they have just
surehomophilybyexaminingwhatproportionof met, and which we selected. Subjects are told
anindividual’s friendsare“similarindividuals,” that the elected leader will receive a prize of
and comparing that to the hypothetical propor- €30, to be distributed among the electorate in
tion of “similar” friends individuals would have whichever way she chooses (including keeping
iftheywererandomlymatchedwithoneanother. the entire amount for herself). Our exper-
These observational data are thus not conducive iment is thus designed to capture which
toaninvestigationintowhichsocio-demographic match-up between voter and candidate best
characteristic emerges as the salient basis for predicts the voter’s choice. The most common
homophily. Indeed, individuals tend to choose socio-demographic categories—gender, class,
friends who are similar on a number of differ- ethnicity, age, religion—are in the mix. We
entdimensions,suchthatthebasiccharacteristics measure which one(s) significantly correlate(s)
enumerated above (ethnicity, class, gender, age, withplayers’votingdecisions.Additionally,our
religion)arestronglycorrelatedwithoneanother researchdesignallowsustoinvestigatewhether
inex-postdata.Suchmulticollinearitychallenges thebasisweuncoverforhomophilyisarational
ourabilitytoidentifythekeycharacteristic(s)on one.Homophilyisrationalifitisbasedontrust.
whichindividualsmatch. In this case, the voter believes that leaders with
Subsequentresearchonhomophilyhasrelied whomshesharesbasicsocio-demographicchar-
on behavioral data that are less unambigu- acteristics are more likely to be generous to her
ously ex-post. Such studies use information because of such similarity than are leaders who
from social networks or dating websites (e.g., do not share these characteristics. By contrast,
Hitsch, Hortacsdu, and Ariely 2010; Huber and homophilyisnon-rationalortaste-basedifthese
Malhotra 2012; Skopek, Schulz, and Blossfeld beliefsarenotatstake.
2011; Wimmer and Lewis 2010). These studies We implement the study in France in 2009,
analyze which characteristics of a potential the home of a larger project on immigrant inte-
partner determine a user’s decision to initiate gration and a country whose recent experience
contact. Yet these, too, suffer from some degree with immigration has fundamentally altered its
ofselectionbiasbecausethesetofpotentialpart- socio-economic,racial,andreligiouslandscape.6
ners is already the result of a selection process. Thiscontextallowsustostudyapopulationthat
Indeed, before contacting someone, a user must is diverse not only on the age, gender, and class
first browse his/her profile. This set of potential dimensions(asinmanyothercontexts),butalso
partners is thus determined either by a user’s byethnicityandreligion.
ownsearchorbyawebsite’ssuggestions—both
of which are already predicated on homophily.
studyaimstoidentifyhowindividuals’politicalpredisposi-
Hence, the set of potential partners is already
tionsdeterminetheirromanticdatingchoices,anditincludes
a set of “similar individuals.” These studies an experimental portion where subjects are presented a set
therefore suffer from the same identification of potential partners determined by the investigators. In
problemexposedabove.5 Centola (2011), participants in an internet-based network
experimentwererandomlyassignedtohomophelousversus
non-homophelous networks to observe network effects on
4. Wedrawthislistof“mostcommon”socialmatch-ups individuals’healthbehavior.Here,thedegreeofhomophily
(ethnicity, class, gender, age, and religion) from the set of isexogenouslydetermined.However,thestudydoesnotcon-
socio-demographiccharacteristicsidentifiedbyMcPherson, cernitselfwithidentifyingthebasisofhomophily.Instead,it
Smith-Lovin,andCook(2001). definesitaccordingtogender,ageandbody-massindex.
5. Centola(2011)andHuberandMalhotra(2012)arethe 6. Approximately 11% of the French population is
onlyexceptionsofwhichweareaware.HuberandMalhotra’s foreign-born(Vasileva2011).
ADIDA,LAITIN&VALFORT:RELIGIOUSHOMOPHILYINASECULARCOUNTRY 3
Our empirical results are compelling: in the falls along specific identity lines: in the case
contextofourexperimentalgameinFrance,and of a diverse region of Paris in France in 2009,
pitting the most common socio-demographic religion is significantly correlated with how
bases for homophily against one another, we individuals associate. Our results also suggest
findthatreligioussimilarityemergesastheonly that such religion-based discrimination is taste-
significant determinant of participants’ voting based. Third, by offering a direct measure of
behavior when monetary stakes are involved. the dominant social basis for homophily, we
No other social dimension, be it gender, ethnic- show that religion in France remains powerful
ity, age, or class plays as significant a role on in a society where only 13% of the population
this outcome. The voting result is remarkably explicitly claims that religion is very important
robust: it holds whether we focus on the full, in its life (World Values Survey 2006). This
socio-demographicallydiversesampleofvoters, is true even when we restrict our analysis to
or on a smaller, more homogenous sub-sample those participants most rooted in France, i.e.,
of voters with longer inter-generational ties to Frenchnationalswithallfourgrandparentsborn
France. In other words, religious homophily in metropolitan France. Even after a century of
is not imported into France by individuals of government policies to undermine the influence
recent immigrant background. Furthermore, no of religion in public life through legislative
single religious belief drives the results: Judeo- means, religion remains a significant predictor
Christians vote for fellow Judeo-Christians, ofhowtheFrenchassociatewithotherslivingin
Muslims vote for fellow Muslims, and Atheists theircommunities.7
voteforfellowAtheists.Acrossreligions,players
chooseleaderswithwhomtheyshareareligious
identity. Finally, our results do not support a II. THEEXPERIMENTALSETTING
rational calculus in this voting decision. In our
In this section, we introduce our research
games, individuals give their vote to people of
design. Specifically, we explain the recruitment
thesamereligion,thoughtheyseemnottoexpect
of our subject pool, the experimental protocol,
thatthisassuresthemamaterialadvantage.
andthevotinggameweusetomeasurethesalient
Ourresultshavethreekeyimplications.First,
basisforhomophily.
scholars have shown that homophily may con-
stitute a threat to social cohesion. It generates
segregation (Centola etal. 2007; Suen 2010), A. SubjectPool
inequality (Fernández and Rogerson 2001), and To conduct our experiment, we first brought
compromises the ability of a society to reach a together a semi-random sample of 80 partici-
consensus on major issues (Golub and Jackson pants recruited in Paris’s diverse 19th district
2012).Understandingwhichbasiccharacteristics in March 2009. This sample was collected for
are significantly correlated with homophily is a project studying Muslim immigrant integra-
thusaprerequisitetolimitinghomophily’spoten- tion in France. For that purpose, approximately
tialdeleterioussocialeffects.Ourstudydoesjust one-thirdofourparticipantswererecruitednon-
that.Second,ourresultsidentifyabasisfordis- randomly, via social networks.8 The rest were
criminationagainstcertaingroupsandnotothers, recruited from the 21 metro stations in the 19th
and thus they have implications for discrimina- district of Paris. In a fully random protocol, we
tionandprejudice(theothersideofthecoinfrom assignedaweighttoeachmetrostationbasedon
homophily). This body of research has already thedensityoftheareainwhichitislocated,with
shown that racial (Bertrand and Mullainathan thehigherdensitystationsgettingmorecardsin
2004;FindlayandReid1997;MonksandRobin- our random draw. Each recruitment team drew
son 2000) and religious (Adida, Laitin, and a metro station for each recruitment day, and
Valfort 2010) homophily are prevalent in a then a number from 1 to 10 to determine which
number of different contexts. Yet recent work passer-by to invite as a game recruit. But, for
challenges these findings by suggesting that the goals of our larger project, we sometimes
discriminationdoesnottargetanyonegroup,but deviatedfromthisprotocoltospecificallyrecruit
targets instead any and all minorities: the threat
is not Blacks or Muslims, but rather the ethnic
7. This is consistent with Berger’s analysis of voting
other (EdoandJacquemet2013;Jacquemetand
determinantsinFifthRepublicFrance(Berger1974).
Yannelis2012).Ourexperimentalresultsarecon- 8. Inourrobustnesschecks,weruntheanalysisexcluding
sistentwithresearchshowingthatdiscrimination thoseplayersrecruitedviasocialnetworks.
4 ECONOMICINQUIRY
rootedFrenchplayers,i.e.,Frenchnationalswith TABLE 1
four grandparents born in metropolitan France SummaryStatistics—CandidateCharacteristics
(in other words, participants who are at least
Standard
third-generation French, and whom we thus
Variable Mean Deviation Minimum Maximum N
refer to as FFF). When potential FFF subjects
Female .53 .50 0 1 80
walked by, recruiters were instructed to ignore Age 36.10 12.76 18 72 80
the sequence of selection, and to ask them to Education 8.30 2.17 2 10 76
participate.9 Table 1 provides basic socio- Family 4.54 2.23 1 9 74
income
demographic characteristics of our sample.10 It White .36 .48 0 1 80
indicates that our candidate pool is, on average, North .15 .36 0 1 80
African
young, well-educated, and ethnically and reli- Black .49 .50 0 1 80
giously diverse. This diversity is not surprising, Judeo- .46 .50 0 1 78
giventhatwerecruitedparticipantsfromthe19th Christian
Muslim .37 .49 0 1 78
district of Paris, a highly diverse area. Indeed, Atheist .14 .35 0 1 78
according to the 1999 French Census, the per- Believer .03 .16 0 1 78
Political 4.03 2.22 1 10 67
centage of individuals living in this district and
Ideology
born in France is 63.5—against 82.4 for all of
Paris(INSEE1999).
Notes:Theunitofobservationistheplayer.Femaleisa
binaryvariablecodedas“1”iftherespondentisawomanand
B. ExperimentalProtocol “0”iftherespondentisaman.Ageisacontinuousvariable
thatcountstherespondent’sage.Educationisanordinalvari-
Our protocol consisted of two phases: a ablerangingfrom“1”(lessthanprimaryschoolcompleted)to
“10”(higherthancollegedegreecompleted).Familyincome
registration phase and a game phase. Dur-
is an ordinal variable ranging from “1” (less than €500 a
ing the registration phase, we collected basic month)to“11”(morethan€7500amonth).Whiteisabinary
socio-demographic data about each player and variablecodedas“1”ifamajorityoftherespondent’sfour
scheduled players into a game session. During grandparentswasborninaWesternEuropeancountry,and
“0” otherwise. North African is a binary variable coded as
the game phase 2 weeks later, we brought play-
“1”ifamajorityoftherespondent’sfourgrandparentswas
ers together into groups of 10, and did so for borninaNorthAfricancountry,and“0”otherwise.Blackisa
8 sessions over 2 weekends—3 of the sessions binaryvariablecodedas“1”ifamajorityoftherespondent’s
four grandparents was born in sub-Saharan Africa or non-
were all female, 3 were all male, and 2 were
metropolitanFrance,and“0”otherwise.Judeo-Christianisa
mixed in gender.11 Each session consisted of a binaryvariablecodedas“1”iftherespondentself-identifies
seriesofgames,includingourvotinggame,and asamemberofaJudeo-Christianreligion(Catholic,Jewish,
lastedapproximately2.5hours.12 Weintroduced Orthodox,Protestant,Christian)and“0”otherwise.Muslimis
abinaryvariablecodedas“1”iftherespondentself-identifies
the experimental games to our participants as
asamemberoftheMuslimreligionand“0”otherwise.Athe-
games designed to investigate “how people istisabinaryvariablecodedas“1”iftherespondentself-
from Ile-de-France [Parisian region] make deci- identifiesasanAtheistand“0”otherwise.Believerisabinary
sions about money.”13 At the beginning of each variablecodedas“1”iftherespondentself-identifiesas“Be-
liever”or“Other”and“0”otherwise.Politicalideologyisa
categoricalvariablerangingfrom“1”(mostleft-wing)to“10”
9. Therecruitingteamreliedonethnicprofiling,infer- (mostright-wing).
ringfromdressandfacialfeatureswhetherparticipantswere
potentialFFF.
10. Table 1 indicates that Believers, e.g., those who session, players were given a name tag to be
answered “Believer” or “Other” in their religious self-
attached to their lapels on which they wrote
identification,compriseonly3%ofoursample,orjusttwo
respondents.Inourrobustnesschecks,werunaspecification their first names. The only information players
thatexcludestheseparticipants.Table1alsoindicatesthatwe had about each other at the beginning of each
havegroupedJewsandChristiansinasinglecategoryforthe sessionwasthereforetheirlooks,theirmanners,
purposeofthisanalysis.Werunarobustnesscheckwherewe
theirdress,andtheirfirstnames.Noneworeany
separateoutthesetwocategories.
11. Notethat,withsixoutofeightsessionsbeingsame-
gender sessions, our design biased our results toward the projectandhencebiastheirbehaviorduringthegames.This
emergenceofgenderhomophily.Inspiteofthis,wedonot strategy,combinedwiththefactthatweorganizedthegames
findasignificanteffectofgenderhomophilyonvotechoice. inasettingwherediversitywasseenasnatural(the19thdis-
12. Eachparticipantplayedasimultaneoustrustgame, trictofParis),worked.Indeed,inexitsurveyswherepartic-
aspeed-chattinggame,avotinggame,andadictatorgame. ipantswereasked“Quepensez-vousquenotreéquipeaura
Forthispaper,wefocusonthevotingandsimultaneoustrust apprissurvousàtraversvosdécisionsaujourd’hui?”[What
games.Spill-oversbetweengamesarenotaconcern,asgame doyouthinkourteamwillhavelearnedaboutyoufromthe
decisionsremainedprivateandanonymousatalltimes. decisionsyoumadetoday?],onlyonesubjectoutof80spec-
13. Weremainedvagueaboutourobjectivesincewedid ulatedthatreligionhadanythingtodowiththepurposesof
notwantourparticipantstosurmisethegoalofourresearch thegames.
ADIDA,LAITIN&VALFORT:RELIGIOUSHOMOPHILYINASECULARCOUNTRY 5
FIGURE1
SpeedChattingQuiz
Table 1 Player A1 Player B1
Table 2 Player A2 Player B2
Table 3 Player A3 Player B3
Table 4 Player A4 Player B4
Table 5 Player A5 Player B5
Note:Thisdiagramillustrateshowplayerswerepositionedduringthespeed-chattinggame,andthewayinwhichtheymoved
betweenrounds.
clothes or jewelry revealing religious affiliation, A players sat down, each at a table, for a total
with the exception of one player who wore a offivetablespositionedsoastomaximizespace
headscarfsignalingaMuslimidentity.14 Players andthusprivacy.GroupBplayerswereinstructed
were then assigned to one of two groups, A or tositdownacrosstheirfirstAplayer(B1across
B, and therefore did not choose from among from A1; B2 across from A2…). Players were
the available players who would be members of thengiven3minutestochatand,weemphasized,
theirgroup. to“gettoknowoneanother.”Attheendofthose
3minutes, players were given 1minute to write
C. TheVotingGame down notes about what they had just learned.
Then, all B players were instructed to stand up
We measure the salient socio-demographic
andmovetothenexttable,sothatplayerB1now
basisforhomophilywithavotinggamedesigned
satacrossA2,playerB2nowsatacrossA3,etc...,
to capture which social characteristics signifi-
andplayerB5nowsatacrossA1.Again,players
cantlypredicthowindividualsassociatewithone
weregiven3minutestogettoknoweachoftheir
another.Inourexperimentalprotocol,thevoting
partners. This process was repeated until each
game followed and built on the speed-chatting
A player had met all B players and vice versa.
game, where each player met all players from
Figure1illustratesthespeed-chattingprotocol.15
the other group (henceforth their partners) in
a speed-chatting protocol. This speed-chatting
gamewasakintoaspeed-datingscenario.Group 15. Attheendofthespeed-chattinggame,playerswere
givenaquiztestingthemontheage,religion,employmentsta-
tus,educationlevel,countryoforigin,currentlivinglocation,
14. Inourrobustnesschecks,weruntheanalysisexclud- marital status, and favorite leisure activity of each partner.
ingthisplayer. Playerswererewarded€1foreachcorrectanswer.
6 ECONOMICINQUIRY
FIGURE2
TheVoteSheet
can
Uponcompletionofthespeed-chattinggame, Then,eachplayerwasaskedhowshewoulddis-
weintroducedthevotinggamewhereeachplayer tributethe€30awardifsheweretobecomeleader
wasaskedtovoteforaleaderamongallthepart- (see Figure 3 for a sample allocation sheet). At
nerss/hehadjustmet(seeFigure2forasample the end of the game session, the allocations of
votesheet).Thisleaderwouldbeawardeda€30 the elected leaders were added to each player’s
prizetodividebetweenherselfandherelectorate. account,tobedistributedafterthecompletionof
ADIDA,LAITIN&VALFORT:RELIGIOUSHOMOPHILYINASECULARCOUNTRY 7
FIGURE3
TheAllocationSheet
the entire experimental protocol. With this vot- won€30,andplayers’allocationdecisionswere
ing game, we are able to capture whom each usedtodeterminefinalmonetaryallocationsonce
playervotedfor,aswellashoweachhypothetical votesweretalliedandleaderswereelected.These
leaderwoulddivideupthereward.Eachdecision instructionswerecarefullyrelayedtoeachgroup
was tied to monetary stakes: the elected leader ofplayers.
8 ECONOMICINQUIRY
TABLE 2
SummaryStatistics—Voter-CandidateSocialDistanceCharacteristics
Variable Mean StandardDeviation Minimum Maximum N
Differentgender .14 .35 0 1 400
Differenceinage 14.52 11.71 0 52 400
Differenceineducation 1.99 2.00 0 7 360
Differenceinfamilyincome 2.63 1.88 0 8 340
Differentethnicity .62 .49 0 1 400
Differentreligion .66 .47 0 1 380
Differenceinpoliticalideology 2.52 2.00 0 9 282
Notes:Theunitofobservationisthevoter–candidatepair.Differentgenderisabinaryvariablethattakesthevalue“1”ifthe
pairofplayersismixedgenderand“0”otherwise.Differenceinageisacontinuousvariablethatprovidestheabsolutedifference
inagebetweenthetwoplayers.Differenceineducationisanordinalvariablethatprovidestheabsolutedifferenceineducation
levelbetweenthetwoplayers.Differenceinfamilyincomeisanordinalvariablethatprovidestheabsolutedifferenceinfamily
incomebetweenthetwoplayers.Differentethnicityisabinaryvariablecodedas“1”ifthetwoplayersareofdifferentethnicity
and“0”otherwise.Differentreligionisabinaryvariablecodedas“1”ifthetwoplayersareofdifferentreligionand“0”otherwise.
Differenceinpoliticalideologyisanordinalvariablethatprovidestheabsolutedifferenceinpoliticalideologybetweenthetwo
players.
Weuseplayers’votingdecisionsasamethod We measure the socio-demographic basis for
to measure which socio-demographic character- homophilybyestimatingthefollowingmodel:
istic emerges as the salient basis for homophily
(1)
when monetary stakes are involved. This vari-
Vote=a+b′.(Distance)+b′.(Candidate)
able,weshouldnote,capturessincereratherthan 1 2
strategic voting. Indeed, two questions captured +b′.(SessionFE)+e
voters’intentinourvotinggame.Thefirstasked 3
voters to rank, in order of preference for group where Voter–Candidate pair is the unit of anal-
leader,thefivecandidatestheyhadjustmet.The ysis and Vote captures whether or not Voter
second asked voters to vote for a single candi- voted for Candidate. Distance is a vector of
dateasleader.WemeasureVoteasthecandidate variablesthatcapturethesocialdistancebetween
rankedfirstbythevoter.Itisimportanttoempha- candidate and voter on a set of sociodemo-
size that voters systematically chose to vote for graphiccharacteristics—gender,age,education,
the candidate they ranked first: voters were thus family income, ethnicity, and religion; b′ thus
1
expressing their true preference, not behaving summarizes the effect of the social distance
strategically in this decision. We then use play- between voter and candidate on vote choice. It
ers’ monetary allocations as leaders along with captures the social basis underlying homophily
data from our simultaneous trust game—which in this context. Candidate is a vector of con-
we present in a later section—to illuminate trol variables (gender, age, education, family
whether the homophily we uncover in the vot- income, ethnicity, and religion) that charac-
ing game is a rational strategy to optimize the terize the candidate; b′ thus summarizes the
2
distributivepayoff. effect of candidate characteristics on the like-
lihood of becoming a leader. Finally, Session
FE is a vector of game session fixed effects.
III. THESOCIALBASISFORHOMOPHILY
Standard errors are robust and clustered at the
A. Results individualvoterlevel,sinceindividualvoterdeci-
sions on each candidate are not independent of
Which socio-demographic characteristic(s)
oneanother.17
drives players’ vote choice in our experimen-
The critical aspect of our research design
tal context? In Table 2, we present summary
is that it allows us to identify the social basis
statistics for our social distance variables: these
summarize,foreachpairofplayersinthevoting
game,theaveragedifferenceingender,age,edu- ideology.Ourresultspersist.Discussingone’spoliticalopin-
cation level, family income, ethnicity, religion ionsisarelativelytabootopicinFrance(e.g.,Louarn2012);
hencethehighratesofmissingdataforthepoliticalideology
andpoliticalideologybetweentheplayers.16
measure.
17. We present results from logit specifications. In
16. Inrobustnesschecks,wefurthercontrolfortherole robustness tests, we check whether our results hold using
thatvaluehomophilymightplaybycontrollingforpolitical linearestimationtechniquesinstead.
ADIDA,LAITIN&VALFORT:RELIGIOUSHOMOPHILYINASECULARCOUNTRY 9
for homophily before any selection occurs: we TABLE 3
impose exogenously the group of players each TheDeterminantsofaPlayer’sVote
player must vote for, and then observe which
DV:Vote
match-up best predicts vote choice. We confirm
that the set of candidates each player must vote Model Model Model
Variable (1) (2) (3)
for is exogenously selected with a simple test
of multicollinearity across our right-hand-side (1)Differentreligion −.764* −.931** −.740*
variables. If preselection occurs, we can expect (.323) (.351) (.312)
(2)Differentgender −.467 −.328 −.155
the set of players in a single group to be sim- (.621) (.588) (.476)
ilar in a number of ways. Yet our correlation (3)Differentethnicity .058 .160 .080
tests across our right-hand-side variables indi- (.299) (.306) (.265)
(4)Agedifference .005 .002 −.000
cate that the highest correlation we find among (.014) (.014) (.012)
all social match-up variables is a correlation of (5)Educationdifference −.126* −.099 −.126^
r=.21 between the difference in education and (.063) (.068) (.066)
the difference in family income. We thus have (6)Incomedifference .023 −.004 −.018
(.082) (.081) (.079)
no reason to believe our specification suffers
Candidatecontrols No Yes Yes
frommulticollinearity. Imputedmissingdata No No Yes
Models (1) through (3) in Table 3 estimate PseudoR2 .029 .064 N/A
Equation (1) above using Stata 11.0. Model (1) Observations 307 307 391
estimates the effect of social distance between
Notes: The table reports Logit estimates. The unit of
voter and candidate on the probability of being
observationisthevoter–candidatepair.Thedependentvari-
chosen as leader. Model (2) adds controls for able,Vote,takesthevalue“1”ifthevotervotesforthecan-
candidatecharacteristics.AndModel(3)imputes didate,and“0”otherwise.Allmodelsincludesessionfixed
missing data due largely to our Education and effects. All standard errors are robust and clustered at the
individualvoterlevel.
Familyincomevariables.18 ^,*,and**denotestatisticalsignificanceatthe10%,5%,
The results in Table 3 are striking: two and1%level,respectively.
individuals who do not know each other, but
are given the chance to interact for fewer than
5 minutes, are more likely to vote for each
relying only on descriptive statistics, illustrates
other when they share the same religion. The
this pattern clearly by showing the distribution
coefficient on Different religion is consistently
of leader rankings (where Rank 1 is the voter’s
negative, and statistically significant at least at
first choice for leader) for dyads who share a
the 95% confidence level in all specifications.
religion (in light gray) and dyads who do not
No other social dimension holds a comparably
share a religion (in dark gray). For example,
robust, statistically significant effect. Figure 4,
Figure 4 clearly shows that for at least 26%
of co-religious voter-candidate pairs, the voter
18. We impute missing data relying on Stata’s “mi”
ranked the candidate as most preferred; by
command, a simulation-based statistical technique that
consists of (1) imputing the missing data by simulat- contrast, the voter ranked the candidate as
ing from a Bayesian posterior predictive distribution of most preferred in only 17% of non co-religious
the missing data under the conventional prior distribu-
voter-candidatepairs.Conversely,for14%ofco-
tion; (2) analyzing the data separately for each imputa-
tion;and(3)combiningtheresultsinstep(2)intoasingle religious voter-candidate pairs, the voter ranked
multiple-imputation result. For more information, readers the candidate as least preferred; by contrast, the
can refer to the Stata Multiple-Imputation Reference Man- voter ranked the candidate as least preferred
ual: http://www.stata.com/manuals13/mi.pdf. Additionally,
in 22% of non co-religious voter-candidate
werunanalternativemethodofaddressingmissingdatavia
Manski’smethodofbounds(Manski1989).Here,wemake pairs. To give a sense of the magnitude of the
assumptionsaboutthewaysinwhichourmissingdatavari- effect, the predicted probability of voting for a
ablescorrelatewiththeoutcomeofinterest.First,wefocus
candidate is approximately 17% when voter
ondifferencesinfamilyincome,wherewehavemissingdata
forN=60observations.Wegeneratetwovariablesfordif- and candidate do not share a religious iden-
ferences in family income, one where we assume that the tity, and close to 30%—a greater than 65%
differenceismaximalwhenthevoterdoesnotvoteforthe increase—whentheydo.19
candidateandminimalwhenthevoterdoesvoteforthecan-
didate;andonewhereweassumetheopposite.Wethendo
thesamefordifferencesineducation,wherewehavemiss-
ingdataforN=40.WeestimateModel(1)inTable3relying 19. Theseestimatesarecalculatedusingthepredictcom-
onthesenewvariables,includingthemsequentiallyandthen mandinStata,appliedtothelogitspecificationforTable3,
together.Ourresultspersist. Model(2).