university o f c o logne wor kin g paper s e ries in...
TRANSCRIPT
Department of EconomicsUniversity of CologneAlbertus-Magnus-PlatzD-50923 KölnGermany
http://www.wiso.uni-koeln.de
U N I V E R S I T Y O F C O L O G N E
W O R K I N G P A P E R S E R I E S I N E C O N O M I C S
No. 64
THE CURSE OF UNINFORMED VOTING: ANEXPERIMENTAL STUDY
JENS GROßERMICHAEL SEEBAUER
0
Thecurseofuninformedvoting:Anexperimentalstudy
JensGroßer(FloridaStateUniversityandInstituteforAdvancedStudy,Princeton)
MichaelSeebauer(UniversityofErlangen‐Nuremberg)
ABSTRACT
We studymajority voting over two alternatives in small groups. Individuals have identical preferences but are
uncertainaboutwhichalternativecanbetterachievetheircommoninterest.Beforevoting,eachindividualcanget
informed, towit, buy a valuable but imperfect signal about the better alternative. Voting is either voluntary or
compulsory. In the compulsorymode, each individual can vote for either of the two alternatives, while in the
voluntarymode they can also abstain. An uninformed random vote generates negative externalities, as itmay
overrideinformativegroupdecisionsinpivotalevents.Inourexperiment,participantsingroupsofthreeorseven
get informedmoreoftenwithcompulsory thanvoluntaryvoting, and in thiswaypartly counteract thecurseof
uninformed voting when they cannot avoid it by abstaining. Surprisingly, uninformed voting is a common
phenomenoneveninthevoluntarymode!Aconsequenceofsubstantialuninformedvotingispoorgroupefficiency
inalltreatments,indicatingtheneedtoreconsidercurrentpracticesofjuryandcommitteevoting.
Thisversion:August2013
JensGroßerisAssistantProfessorattheDepartmentsofPoliticalScienceandEconomics,FloridaStateUniversityandExperimentalSocialScienceatFloridaState,531BellamyHall,Tallahassee,FL32306‐2230,andamemberofthe Institute for Advanced Study, School of Social Science, Princeton in the academic year 2012‐13.([email protected]).MichaelSeebauerisPhDCandidateattheDepartmentofEconomics,UniversityofErlangen‐Nuremberg,LangeGasse20,90403Nuremberg,Germany([email protected]).This isasubstantiallyrevisedversionofour2006workingpaper“Votingandcostly informationacquisition:Anexperimental study,” ofwhich an abstract appeared in an article in a researchmethods volume (Großer2012).Financial support from the GEW Foundation Cologne and the German Science Foundation (DFG) is gratefullyacknowledged.WealsothanktheEconomicsLaboratoryoftheUniversityofColognefortheirhelpandhospitality,and participants at seminars at Princeton University, the University of Hawai’i, and the Institute for AdvancedStudy inPrinceton,aswell asat theGEW2005Tagung inCologne, theESA2006Meeting inTucson,SEA2008Meeting in Washington D.C., and the Experiments in Political Economy Workshop, Caltech in 2013 for theirvaluablecomments.
1
1Introduction
Groups of individuals make decisions in almost every realm of life. An important rationale for this
phenomenonistheirpotentialtoaggregateinformationthatwouldotherwiseliedispersedanddormant
among the individuals. Accordingly, it iswidely argued thatmore heads have, or can generate,more
knowledge than fewerheads,and therefore, largergroups tend tomakebetterdecisions thansmaller
groups. The importance of knowledge accumulation and aggregation in groups is evident in the
following examples: juries consider evidence to acquit or convict the accused; legislative committees
gather know‐how of experts and special interests groups in hearings; hiring committees evaluate
documents of applicants; and editors ask referees for their opinions about research papers. But do
larger groups reallyperformbetter than smaller groups in thepresenceof free rider incentives (e.g.,
whengathering information is time‐consumingandcostly)?Andhowdoes theanswerdependon the
specific information aggregation mechanism (e.g., the communication or voting procedure)? Do
individuals get informedmore oftenwith compulsory than voluntary voting?At first glance, banning
abstention, such as for jurors in criminal trials, seems to be a practical solution to improve group
decision‐making, but the impact on informational efforts and thus performance is virtually unknown.
We tackle these and similar questions using game theory and a laboratory experiment by studying
knowledge accumulation and aggregation in different sized groups where individuals have identical
preferences, information acquisition is costly, and consensus is obtained by voluntary or compulsory
majoritarianvoting.
Theaggregationofprivateinformationviavotingisakeyfunctionofdemocraticdecision‐making.1
When individuals have identical preferences but different opinions because they are uncertain about
which of several alternatives can best achieve the common interest, then pooling their private
information leadstoamutualagreementaboutthecorrectalternative(supposeonlyoneisbest).The
importantCondorcet JuryTheorem (1785)uses statistical inference to show thatmajority voting can
assume this pooling function, and therefore, the group’s decision is more likely correct than any
individual’s decision and almost surely correct in very large groups. In themost basic version of the
1Moregenerally,Piketty(1999)putstheimportanceofinformationaggregationinelectionsonalevelwiththeimportanceofinformationaggregationinmarkets.
2
modelunderlyingthistheorem,agroupchoosesbetweentwoalternativesviacostlessmajorityvoting
anduncertaintyisgivenbyacommonpriorprobabilityofone‐half,indicatingthateachalternativehas
thesamechanceofbeingthebetterone.Moreover,eachindividualgetsacostlessprivatesignal,towit,
an independent Bernoulli trial showing the correct alternative with probability ∈ , 1 and the
incorrect alternativewith probability 1 . Thus, signals are valuable but imperfect and hencemay
induce conflicting opinions. Using this simplemodel, the intuition of Condorcet’s Jury Theorem is as
follows: the pool of private information growswith each sincere vote (i.e., itmirrors the signal) and,
since , the expected difference in votes grows in favor of the correct alternative so that the
probability of a correct group decisionweakly increases in the group size.2 And, by the law of large
numbers,inverylargegroupsthefractionofcorrectsignalsapproaches withprobabilityone,sothat
thevoteoutcomealmostsurelymatchesthecorrectalternative.
Theverypromisingresultsofthetheoremrelyonseveralassumptions,chiefamongwhicharethat
informationviaprivatesignalsiscostlessandthatvotingissincereandcostless.Butitignoresthefree
rider incentives that are generated if for example, more realistically, costs of signal acquisition are
involved (Downs 1957; Olson 1965).When the incentive to rely on informational efforts of others is
strongenough, theresultcanbe“rational ignorance”where theendogenous informationpoolshrinks
ratherthanincreasesinthegroupsize,say,ifthesizeisdoubledbutaverageeffortdropsbymorethan
half due to stronger free riding in the larger group (Martinelli 2006;Mukhopadhaya 2003). In other
words,Condorcet’sJuryTheoremneednotholdinthepresenceoffreeriderincentives.
In thepresentpaper,weexaminehowcostly informationacquisitiondependsonthegroupsize
(threeandseveninourexperiment)andifcompulsoryvotinginducesindividualstogetinformedmore
often than voluntary voting.3 In our game, they first decide independently and simultaneously on
whethertobuyaprivatesignalandthereaftervote,atnocost,independentlyandsimultaneouslyover
two alternatives usingmajority rulewith random tie‐breaking (in the voluntarymode, they can also
2 According to Austen‐Smith and Banks (1996), an “informative” vote mirrors the signal and a “sincere” vote maximizesexpectedpayoffsandmayopposethesignal.Forourparameters, thetwokindsofvotesalwayscoincidesowecanusebothtermssynonymously.And,anextrasincerevoteweaklyincreasestheprobabilityofacorrectgroupdecision,becauseityieldsanincreaseifanoddnumberofvotesisreachedbutdoesnotchangetheprobabilityifanevennumberisreached.3 For amore general discussion of voluntary versus compulsory voting and the incentive effects, see for exampleAbraham(1955),Birch(2009),andLijphart (1997).Moreover,Börgers(2004)andGroßerandGiertz (2009),amongothers,comparevoluntaryandcompulsorycostlyvotingwithprivatepreferences.
3
abstain). Hence, we depart from Condorcet’s basic model in two important ways, to wit, by using
endogenous costly signals and by comparing situations where individualsinformed and
uninformedareeither freeorobligedtovote(incontrast, in thebasicmodeleveryone isassigneda
costlesssignalandthuspreferstovoteanyway).Individualsbasetheirdecisionsonlyongroupevents
that are both pivotal and informative, the latter meaning that the combined signals in the group
conveyedbysincerevotesfavoroneofthealternatives.Intheoneextreme,ifeveryoneisuninformed
andvotesrandomlyorabstains,thennosignalcostsaccrueandthegroupdecisioniscorrecthalfofthe
time.Intheotherextreme,ifeveryoneisinformedandvotessincerely,thenitiscorrectwithmaximum
possible probability, but the total signal costsmay outweigh the higher expected benefits.We utilize
quantal response equilibrium (QRE; McKelvey and Palfrey 1995), which generalizes Bayesian Nash
equilibrium (BNE) by allowing for decision‐making errors, in order to derive individual signal
acquisitionprobabilitiesinbetweenthesetwoextremes,includingforlargegroups.Wealsoderivethe
effectsoftheseprobabilitiesoncorrectgroupdecisionsandefficiency,usingcost‐benefitanalysis,and
testthe(comparativestatics)predictionsforourexperimentalparametersagainstthedata.
Whenvoting isvoluntary, in theory,uninformed individualsrationallyabstainas theirrandom
votemayoverrideaninformativegroupdecisioninpivotaleventsandlettheinformeddecide,whose
sincerevotesincreasetheprobabilityofacorrectgroupdecisionandthusexpectedbenefits.Thisform
ofdelegationisnamedthe“swingvoter’scurse”(FeddersenandPesendorfer1996).4Bycontrast,with
compulsoryvotingthecurseofuninformedvoting is tolerated. If itprevails, theexpectednetbenefits
(i.e., the increase in benefits) of voting informatively are too small compared to the costs and fewer
individualsgetinformedvis‐à‐visvoluntaryvoting.However,theanticipationofthecursecanalsoboost
signalacquisitionasthenetbenefitsareaugmentedbythevalueof“curseavoidance,”andthishappens
morelikelyinsmallergroupswherefreeriderincentivesareweaker.Suchendogeneityeffectsarevery
differentfromthesituationwithfixedinformationlevels,whereexpectedefficiencyisgreaterwiththe
opportunity toabstain(e.g.,Bhattacharya,Duffy,andKim2013;KrishnaandMorgan2011,2012)but
4 The intuition of the curse is as follows: suppose that individual is uninformed, and informed (uninformed) others votesincerely(abstain).Inpivotaleventswhereonealternativeleadsbyonevoteofothers,if abstains, thentheprobabilityofacorrectgroupdecisionis .However, if votesrandomly, thenwithprobabilityone‐halfhervoteturnsawinofthemorepromisingalternativeintoatie(whichisdecidedbyacoinflip)andwithprobabilityone‐halfhervoteonlyraisesitswinningmargin,sothegroupdecisioniscorrectwithprobability ,whichissmallerthanthe forabstaining.
4
might be further improved, for example, by a turnout‐enhancing lottery prize for a random voter
(Gerardi et al. 2009). Another important differencewith endogenous costly signals is that observing
someone’svoluntaryorcompulsoryvotedoesnotrevealwhethertheindividualhasmadeinformational
efforts,butinthevoluntarymodeaspottedabstainermaybedeemedafree‐rider.Towit,uninformed
abstainingissociallydesirable,however,beingignorantinthefirstplacemaynotbeseenlikethat.By
contrast, with exogenous signals only the favorable view of curse avoidance is valid. Hence, with
endogenouscostlysignals,thereareincentivesfortheuninformedtovoteduetothepublicgoodnature
of information and the possibility that otherswill punish abstainers in oneway or another (see also
DellaVignaetal.2013;Funk2010;andBernstein,Chadha,andMontjoy2001foroverreportingvoting).
Sincethemid‐1990s,gametheoristshavebeguntorelaxtheassumptionsofCondorcet’sJuryTheorem
andexaminedavarietyofdifferentchangestothebasicmodel.5Austen‐SmithandBanks(1996)study
the theorem as an incomplete information game and show that BNE generally involves some
uninformativevoting,notingthatrationalindividualsonlycareaboutpivotalevents(Downs1957),infer
aggregateinformationfromtheirownsignalandothers’votesinpivotalevents,andbasetheirvoteon
this information and the pivot probabilities.6 However, for our parameters, everyone voting
informatively is actually a BNE (Austen‐Smith and Banks 1996, p. 38). Feddersen and Pesendorfer
(1996)allow forabstention fromvotinganddistinguishbetween “partisans”whoalwaysvote for the
samealternativeand“independents”whowanttoselectthecorrectalternativeandareeitherinformed
oruninformed.They show that uninformed independentshave incentives to abstain even if voting is
costless(i.e.,theswingvoter’scurse)andtovoteinordertocounterbalanceanypartisanbiassoasto
maximizethevoteimpactoftheirinformedallies(seealsoFeddersenandPesendorfer1997,1999).The
Feddersen‐Pesendorfermodelfindssupportinthelaboratory(Battaglini,Morton,andPalfrey2010;and
withqualification,MortonandTyran2011)andfromobservationalstudies(e.g.,CoupéandNoury2004;
Lassen 2005;Matsusaka 1995;McMurray2010; Palfrey andPoole 1987;Wattenberg,McAllister, and
5Forsurveys,seeGerlingetal.(2005)andPiketty(1999).Forexample,Ladha(1992)studiescorrelatedvotingandfindsthatCondorcet’s Jury Theorem still holds under fairly general conditions in large groups (small groups require substantialrestrictions),andtheprobabilityofacorrectgroupdecisionisinverselyrelatedwithvotecorrelation.6SeealsoMcLennan(1998)andWit(1998).FeddersenandPesendorfer(1998)showthatwithunanimityrulethe innocentmaybeconvictedmoreoften in larger thansmaller juries,and theprobabilityof falseverdicts increases compared toothervoting rules. In the laboratory,Guarnaschelli,McKelvey, andPalfrey (2000) find fewer incorrect convictionswithunanimitythanmajorityvoting.
5
Salvanto 2000). All these studies use exogenous information. By contrast,we examine voluntary and
compulsorymajoritarianvotingwhen individualsdecideonwhetherornot toget informed.Also, the
effectsofpotential ignoranceonefficiencyarenot easilyquantifiable in the field (an exception is the
studyofBartels1996,whereuninformedvotingleadstobiasedresultsinU.S.PresidentialElections).By
contrast, our controlled laboratory study allows for amore detailed analysis of themutual effects of
endogenous costly signals and abstention, and their consequences for efficiency. Our findings with
regardtotheserelationshipshaveimportantimplicationsforthedesignofjuriesandcommittees.
Rational choice scholars (Downs 1957; Olson 1965) have pointed out that, generally, gathering
information about policy alternatives is costly. Therefore, individuals have incentives to free ride on
others’ informational efforts, which can result in rational ignorance. Mukhopadhaya (2003) revisits
Condorcet’sJuryTheorembystudyingendogenouscostlysignalsandfindsthatinBNEtheinformation
pool can shrink in the group size. In the samevein,Martinelli (2005) shows that if the signal quality
dependsontheamountinvested,individualeffortsgotozeroastheelectorategoestoinfinity.However,
if the marginal costs are near zero for nearly irrelevant information, then the vote outcome is
nonetheless representative of the majority. Feddersen and Sandroni (2006) study costly signal
acquisition in an ethical voter model and show that a significant fraction of a large electorate gets
informed, and successful information aggregation depends on the signal costs and quality. Oliveros
(2012) presents amodel inwhich costs are increasing in the signal quality and finds that BNE exist
where, forsomeindividuals,abstention ispositivelyrelatedwithquality.Moreover,GerardiandYariv
(2008), Gershkov and Szentes (2009), and Persico (2003), among others, examine costly signal
acquisitionasamechanismdesignproblem,wherethedesignerseekstoimplementasociallyoptimal
committee size and voting rule in order to overcome the free rider problem. The following recent,
independent studies are most closely related to ours. Shineman (2013) experimentally studies the
voluntary and compulsory modes as an individual decision‐making problem with costly voting and
signal acquisition, and Tyson (2013) analyzes this situation as an incomplete information game and
allows for partisans and independents. Sastro and Greiner (2010) investigate abstention and vote
invalidationwithexogenoussignalsinbothvotingmodesinthelaboratory.Finally,Elbittaretal.(2013)
experimentally study endogenous signal acquisitionwithmajority and unanimity voting for different
6
groupsizes.Toourknowledge,wewere the first to study theeffectsofvoluntaryversus compulsory
votingonendogenoussignalacquisitionanduninformedvoting,andultimatelyongroupefficiency.
2Thegameandpredictions
Thegame
Anodd‐sizedgroupof2 1individuals,labeled 1, … ,2 1,ischoosingbetweentwoalternatives,
and ,throughmajorityvoting(acoinflipbreaksatie).Weassumethateachindividualisriskneutral
andmaximizesherexpectedownpayoffs.Moreover,theyallhaveacommoninteresttoalignthegroup
decisionwithanunobservable“truestate”oftheworld,whichisalsoeitherAorB,withequalcommon
priorprobabilityforeachstate.Forexample,ajurywantstoacquittheinnocentandconvicttheguilty.
Duetotheuncertaintyaboutthetruth,however, individualsmayhavedifferentopinionsaboutwhich
alternative is correct in this regard. Formally, their payoffs are identical ex post and given by
, , 1 and , , 0, where the first argument of denotes the group
decisionandthesecondargumentdenotesthetruestate.Thus,everyoneprefersA(B)ifthetruestateis
A (B).Each individual caneitherstayuninformed(i.e., continue toknownomore than thecommon
prior), 0, or get informed (i.e., learnmore than the prior), 1, by acquiring a private signal,
∈ , at a cost 0, ∀ . A signal is an independent Bernoulli trial from a state‐dependent
distribution with Pr | Pr | ∈ , 1 and Pr | Pr | 1 .
Thus,asignal isnoisybut informativebecausetheprobabilitythat itcorrectly indicatesthetruestate
exceedstheprior,or .(Intheexperiment,wefocuson .)
The game has two stages. In the first stage (Information), individuals independently and
simultaneouslydecideonwhetherornottogetinformed.Inthesecondstage(Voting),eithervoluntary
or compulsory independent and simultaneous costless voting takes place (henceforth also labeled
, ).Withvoluntaryvoting,eachindividualcanvoteforalternativeAoralternativeB,orabstain
from voting. By contrast, with compulsory voting the option to abstain is not available. Note that
everyoneinformed and uninformedmakes a decision in this stage, and thereafter learns the vote
outcomebutnotwhetherotherswere informed. In the following,weanalyzeequilibriumbehavior in
theVotingandInformationstagesinturn,usingbackwardinduction.
7
Voting
In the Voting stage, we focus on the following symmetric BNE, which have been shown to predict
costlessvotingquitewellinmanyexperiments(e.g.,Battaglini,Morton,andPalfrey2010;McKelveyand
Ordeshook1990),thoughuninformedcompulsoryvotinghasnotbeenstudiedpreviously:
Proposition1 (Voting): In symmetricBNE, informed individuals vote sincerely in both votingmodes,
while theuninformedabstainwithvoluntaryvoting,andvote randomlywithequalprobability foreach
alternativewithcompulsoryvoting.
Proof:Seeappendix.
Signalacquisition
In the Information stage, we derive individual signal acquisition decisions using quantal response
equilibrium (QRE; McKelvey and Palfrey 1995), which has been shown to predict binary choices in
gamessimilartooursmuchbetterthanBNE(e.g.,GoereeandHolt2005;GroßerandSchram2010).QRE
generalizes BNE by including stochastic decision‐making errors that are systematic in the sense that
more lucrative decisions are made more often than less lucrative decisions (i.e., best responses are
smooth,notsharpasinBNE).Aparameter 0controlsthedegreeofnoise.Wefocusonsymmetric
QRE where everyone has the same probability of getting informed, , ∀ . Using the logit
specificationofGoereeandHolt(2005;henceforthlogitequilibrium),intheoneextremewithoutnoise,
0 andQRE turns intoBNE,while in the other extremewithpurenoise, ∞ anddecisions are
entirelyrandom,or .Wemainlyfocuson 0,implyingthat ∈ 0,1 .
The condition for a logit equilibrium ∗ to exist is derived in the appendix and given by
ln Π , , ∗ , , , , where superscript e denotes the expectation operator. The left‐
hand side (henceforth ) is identical inbothvotingmodesanddealswith theerrors. For 0, it
strictlyincreasesin ,approaches ∞( ∞)if approaches0(1),andalwaysequalszeroat .Given
thateveryonevotesàlaProposition1(denotedbythevector ∗ ),theright‐handside(henceforthRHS)
givesindividual ’sexpectednetpayoffs,orincreaseinpayoffs,ofgettinginformedifeveryoneelsegets
informedwithprobability ,whichdependsonthevotingmodeasspecifiednext.
8
Withvoluntaryvoting,theinformedvotesincerelyandtheuninformedabstain(Proposition1).The
conditionforasymmetric ∗ toexistinthismode,assumingwithoutlossofgeneralitythatindividual
receivesana‐signalwhengettinginformed,isthengivenby:
1 22
12
1 . 1
1 is described above, and 1 gives individual ’s expectednet payoffs of getting informed,
Π , , ∗ , , , ,whichconsidersthateach getsinformedwithprobability .Hence,each
votes sincerely for the correct [incorrect] alternative with probability [ 1 ], and stays
uninformed and abstainswithprobability1 . 1 includes ’s expectednetbenefits of getting
informed(i.e.,thesumterm)minusthesignalcost,c,wherethesumover 0, … , contains: the
probabilityofanevennumber2 ofothersbeinginformed, , , , ≡ 1 ;
given2 ,theprobabilityof beingpivotal, ,∗ , , ≡ 1 (i.e.,eachalternativehask
votes,sothatherA‐voteturnsatieintoawinforA);7and herexpectednetgainsinthesepivotal
events, ∗ , , 1 , , , , which considers that, if
she gets informed and uses Bayesian updating, her sincereA‐vote yields a correct [incorrect] group
decisionwithprobability [1 ],andsheonlyknowsthepriorifuninformed.
With compulsory voting, the informed vote sincerely and the uninformed vote randomly
(Proposition 1). The condition for a symmetric ∗ to exist in this mode, assuming without loss of
generalitythatindividual receivesana‐signalwhengettinginformed,isthengivenby:
1 21 1
11 1
. 2
2 is described above, and 2 gives individual ’s expectednet payoffs of getting informed,
Π , , ∗ , , , .Notethatwithcompulsoryvotingtheonlypivotaleventsarewith votesofothers
informed or uninformedfor each alternative (i.e., herA‐vote turns a tie into awin forA). IfB is
7Individual isalsopivotalifAisshortofonevoteofothers(i.e.,whenherA‐voteturnsalossintoatie).However,intheseeventsopposing signals (includinghera‐signal) eveneachotherout, so that sheknowsnomore than theprior.Hence,herinformedvotewouldonlyreplacethecoinflipandthesignalcostbewasted.
9
correct,thenherpivotalA‐voteresultsinanincorrectgroupdecision,whichyields , 0sothat
thisevent is ignored (as in 1 ). IfA is correct, 2 considers that each gets informed
withprobability andreceivesana[b]‐signalwithprobability [1 ],andstaysuninformedwith
probability1 .Hence,each votessincerelyforA [B]withprobability [ 1 ]andvotes
randomlywithprobability 1 for each alternative. 2 contains ’s expectednetbenefits of
getting informed (i.e., the sum term)minus the signal cost, . In pivotal events, for any 0, … ,
informative votes for , , theremust be uninformed random ‐votes, and for a givenpair
, ) the pivot probability is ,∗ , , , , ≡ 2 1 . Since
pivotal events also involve uninformed random votes, individual ’s expected net gains from voting
informatively in these events vary and are equal to ∗ , , , ≡ ,
namely,herBayesianupdatedprobabilitythatAiscorrectgiven 1, ),includingherownA‐vote,
times , 1andminus , ifshevotesuninformed.8
Proposition 2 (Information acquisition): In symmetric logit equilibrium, for 0 individuals get
informed with probability ∗ ∗ , , , , 0 ∈ 0,1 , , . For large enough degrees of noise,
0, this probability is unique and decreasing in the odd group size, 2 1, decreasing in the
signalcost, ,anditmovesclosertoone‐halfwhen increases.
Proof:Seeappendix.
The uniqueness requirement that is large enough is due to convex and concave properties of
1 and 2 ,respectively.Moreover,thepropositiondoesnotgivecomparativestaticsresults
fortheeffectsofthevotingmode,whichdependon and .Italsoignoreschangesintheprobabilityof
a correct signal, sincewhile raising ∈ , 1 usually increases ∗ , there is a regionof theparameter
spacewherethereverseholds.Thereasonisthatraising hastwooppositeeffects,towit,itdecreases
thepivotprobability, , . ,andincreasestheexpectednetgainsfromvotinginformatively, . .
Finally,BNE( 0)isdiscussedinfootnote30intheappendix.
8Notethatineventswithmoreb‐thana‐signalsinthegroup,individual preferstovoteagainsthera‐signal.But,exante,sheexpectsotherstohaveonaveragemorea‐thanb‐signalssovotinginformativelyisheruniquebestresponse.
10
Efficiencyandsocialoptimum
In order to compare ex ante group efficiencydefined by the group’s expected total benefits minus
expectedtotalsignalcostsacrossvotingmodesandgroupsizes,wemustderivetheprobabilityofa
correct group decision, , | , ∗ , , for , and , , where | denotes the event
that ischosenconditionalon being the truestate.Thisprobabilityuses and isdescribed in the
appendix.Given ∗ , note that increasing increases , . .Moreover, expected total signal costs
equal 2 1 in both voting modes, so ex ante expected efficiency is given by
2 1 , . , 2 1 , . , , . Finally, in the social optimum
with noise‐free decision‐making, maximization with respect to yields the implicit solution of :
, . / ,wheresuperscript denotestheoptimum.
Below,wederive ∗ , ,andtherespective , . forvarying forourexperimentaltreatmentsand
parameters(seealsoFiguresA2.1‐3intheappendix),andadditionalpredictionsfor 0and0.06and
variousgroupsizesaregiveninTablesA1andA2intheappendix.
Largegroups
Althoughwefocusonsmallgroups,ourgamehasalsointerestingimplicationsforlargeelectorates:
Proposition3(Largegroups):Asthegroupsize,2 1,goestoinfinity,if 0and ∈ , 1 ,thenthe
uniquelimitingsymmetriclogitequilibriumindividualsgetinformedwithprobability
∗ , , 01
1 ⁄ ∈ 0,1
for , andthegroupdecisionsisalmostsurelycorrect.If 0,theninthelimit ∗ , , 0
0for , andthegroupdecisioniscorrecthalfofthetime.
Proof:Seeappendix.
Theproposition states that in both votingmodes, Condorcet’s JuryTheoremalso holds for large
groupswith endogenous costly signals if there is somedecision‐making noise, 0, but not in BNE
without any noise. To see the intuition, first note that since 0 ∈ 0,1 , in large groups the
respective numbers of informed and uninformed individuals are large too. Then, by the law of large
numbers,withcompulsoryvotinganeven(odd)numberofuninformedindividualsalmostsurelycast
11
equallymanyvotesforeachalternative(onemorevoteforeitheralternative),andwithvoluntaryvoting
the uninformed abstain. But since 1 ⟺ 2 1, in the limit, noting that and 0, the
informed individuals’ fraction of correct votes exceeds their fraction of incorrect votes by an infinite
margin,thatislim → 2 1 2 1 ∞ 1for ∈ 0,1 ,souninformedvotesareirrelevantand
thegroupdecision isalmostsurelycorrect.Finally,since thepivotprobability isnil in largegroups,a
noise‐freevoterchoosesnottobuyasignal, 0 0.
3Experimentaldesignandpredictions
ThecomputerexperimentwasrunattheUniversityofCologne’sLaboratoryforEconomicResearch.9In
total,220studentsattendedeightsessionsof27or28participantsthatlastedabout1.5hours.Earnings
wereexpressedinpointsandexchangedforcashfor€1per800points.Participantsearnedanaverage
of€11.79,including€2.50forshowingup.Instructionsareinouronlineappendix.10
We employed a 2 2 treatment design, varying the voting mode (voluntary and compulsory)
within‐subjects in one dimension, and group size (three and seven) between‐subjects in the other
dimension. Treatments are labeled Voluntary voting‐3, Voluntary voting‐7, Compulsory voting‐3, and
Compulsoryvoting‐7. Each sessionhad twopartsof30decisionperiodseach,whereeachpartuseda
different voting mode and the order was changed across sessions. At the beginning of a session,
unknowntotheparticipants,subjectpoolsof27(28)wererandomlydividedintothree(two)separate
matching groups of nine (fourteen). The composition ofmatching groupswas fixed during the entire
session, and therewas no interactionbetween them.Weused a randommatching protocol:within a
matchinggroup,participantswererandomlyput intogroupsofthree(seven)atthebeginningofeach
period.All our statistical tests are run at thematching group level, forwhichwehave twelve (eight)
independentobservationsforgroupsofthree(seven),halfofthempereachorderofthevotingmodes.11
9Weusedz‐Tree(Fischbacher2007)toprogramtheexperimentandORSEE(Greiner2004)torecruitparticipants.10Theonlineappendixisavailableathttps://sites.google.com/site/jwghome/research.11Observedaverageratesofacquiredsignalsarehigher inthefirstthansecondpart inVoluntaryvoting‐3,andviceversa inCompulsory voting‐3 ( 0.021,Wilcoxon‐Mann‐Whitney tests, one‐tailed), but no difference is found for the larger groups( 0.343 and 0.557).Note that the sixmatching groups that first faced the voluntarymode have statistically significantlyhigher rates averaged over both parts than the other sixmatching groups, but no difference is found for the larger groups( 0.008and0.557,sametests).Asaresult,inthesmallergroupsaverageratesarenotdifferentbetweenvotingmodesinthefirstpart,andhigherwithcompulsoryvotinginthesecondpart( 0.120and0.004,sametests).Inlargergroups,averageratesarehigherwithcompulsorythanvoluntaryvotinginbothparts( 0.014and0.014,sametests).Importantly,theleveleffectofratesforthesmallergroupsdoesnotaffectouranalysisandmainresults.
12
Eachperiod consisted of two stages. At the beginning of eachperiod, the true state,A orB,was
selectedwithprobabilityone‐half foreachbutnotrevealedtothegroup.Inall treatments, inthefirst
stage, participants independently and simultaneously decided onwhether to acquire a private signal
aboutthetruestateandpay 25points,orstayuninformedandbearnocost.Asignalindicatedthe
true (false) state with probability (1 ). In the second stage, voluntary or compulsory
majoritarian voting with random tie‐breaking took place, where all participantsinformed and
uninformedindependentlyandsimultaneouslyvoted,atnocost, foralternativeAorBor, if feasible,
abstained.12Aftervoting,thetruestate,groupdecision,andnumberofvotesforeachalternativewere
announced. Each participant’s period earnings were determined as follows: she or he received 200
points if the group decisionmatched the true state and 0 points otherwise, independent of the own
signalacquisitionandvotedecisions,minus25pointsifsheorheboughtasignal.Formoredetailsof
theprocedures,seetheinstructionsintheonlineappendix.
Figure1:Symmetriclogitequilibriumsignalacquisitionprobabilities
Note:Thefigureshows, forourexperimental treatmentsandparameters,symmetric logitequilibrium
signalacquisitionprobabilitiesfor ∈ 0,0.1 and ∈ 1,10,100 .Thoughnotshown(butseeTableA1),
Compulsoryvoting‐3hasanadditionalBNEwherenoonegetsinformed.
Figure 1 shows, for our experimental treatments and parameters, symmetric logit equilibrium
signal acquisition probabilities per treatment for ∈ 0,0.1 and the discrete cases ∈ 1,10,100 .
12Intheexperiment,theseoptionswerelabeledYellow,Blue,andNoColorinsteadofA,B,andabstain,respectively.Notethatduetooddgroupsizes,tieswerenotpossibleinthecompulsorymode.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 1 10 100
μ (error parameter)
Lo
git
-QR
Es
ign
al a
cq
uis
itio
n p
rob
ab
ility
Voluntary voting-3 Voluntary voting-7 Compulsory voting-3 Compulsory voting-7
13
Though not shown in the figure (but see Table A1 in the appendix), Compulsory voting‐3 has an
additionalBNE( 0)withuniversalrationalignorance.Ascanbeseen,predictionsvarymarkedlyin
BNE, andmove closer to each other towards one‐halfwhen increases.With compulsory voting, for
bothgroupsizesthereisaBNEwherecursedvotingprevails(i.e.,nobodybuysasignal),andingroups
of three there is anotherBNEwhere it is counteracted by the highest predicted ∗ . In this voting
mode, 0has theeffectofpromotingsignal acquisition in the former twoequilibria (i.e.,universal
rationalignoranceisovercome)13andoppressingitinthelatterequilibrium.Moreover,giventhevoting
mode,individualsarealwaysmorelikelytogetinformedinsmallerthanlargergroups(seeProposition
2).And,giventhegroupsize,compulsoryvotingalwaystriggershigher ∗ thanvoluntaryvotingin
groups of three, but in larger groups this only holds for about 0.025while the reverse holds for
degrees of noise below this value. Using our ∗ , we can derive the groups‘ expected number of
acquiredsignals(or,informationpool),probabilityofacorrectdecision,andexpectedefficiency,which
are shown in Figures A2.1 to A2.3 in the appendix. Their comparative statics depend on and are
thereforediscussedfurtherafterestimatingthedegreesofnoiseforourdata.
Besides comparing group performances to universal ignorance, 0, it is also interesting to
confrontthemwiththesocialoptimum,whichis 1inalltreatments(i.e.,acornersolutionwhere
themarginaltotalcostsdonotexceedthemarginaltotalbenefitsevenifeveryonegetsinformed),and
withasingledelegate,D,whomakesanon‐noisygroupdecision(henceforthnoise‐freedelegate),also
always 1. These three benchmark performances are compared to our logit predictions in the
experimental results section (see also Figures A2.1 to A2.3). Moreover, since ∗ 1 in all
treatments,anysignalacquisitionprobabilitygreater than the logitprediction isefficiencyenhancing.
Thus, raising 0 strictly increases expected efficiency in all treatments but Compulsory voting‐3,
where it decreases efficiency. We can use this exception to test whether deviations from BNE are
consistenteitherwithlogitequilibriumorimplicitcooperationtowardshigherpayoffs.
13InCompulsoryvoting‐3,theBNEwithuniversalignoranceisnotrobusttoeventhesmallestnoise,soincreasing leadstoajumptotheequilibriumpathwithhighsignalacquisitionprobabilities.
14
4Experimentalresults
Inthissection,wepresentandanalyzeourexperimentalresults.Wediscussinturnsignalacquisition,
correctgroupdecisions,efficiency,uniformedvoting,andlearningaboutclosevoteoutcomes.
Figure2:SignalacquisitionAveragerates Averagenumbers
Signalacquisition
Figure 2 shows observed average rates (left panel) and average numbers (right panel) of acquired
signalsperfive‐periodblockpertreatment.TheoverallrateishigherinCompulsoryvoting‐3and‐7than
inVoluntary voting‐3 and ‐7 (0.47 and 0.38 versus 0.35 and 0.27). By contrast, the overal number is
higher inCompulsoryvoting‐7 andVoluntaryvoting‐7 than in the respective smallergroups (2.67and
1.90versus1.40and1.05).Withonlyoneexception,thesameordersofratesandnumbersofacquired
signals also hold per five‐period block. Finally, signal acquisiton tends to decline over time in all
treatments.14
Experimental result 1 (Signal acquisition): Larger groups acquire almost twice asmany signals as
smaller groups (81.3% and 90.7% more with voluntary and compulsory voting, respectively), but
individualsget informedat thesamerates.Forbothgroupsizes,compulsoryvoting inducesmoresignal
acquisitionsthanvoluntaryvoting(33.8%and40.7%more,respectively).
14Averageratesandnumbersofacquiredsignalsarehigherinperiods1‐15thanperiods16‐30inalltreatments( 0.059,Wilcoxonsignedrankstests,one‐tailed).Moreover,usingprobitregressionswithrandomeffectsatthematchinggrouplevelwithsignalacquisitionasdummydependentvariableandperiods1‐30orperiods1‐60(i.e.,overbothparts)asindependentvariable,wefindnegativeandstatisticallysignificantcoefficientsoftheperiodinalltreatments( 0.048,two‐tailed),exceptinVoluntaryvoting‐7forperiods1‐60( 0.206).
0.0
0.1
0.2
0.3
0.4
0.5
0.6
1 2 3 4 5 6
Blocks of 5 periods
Sig
na
l acq
uis
itio
n r
ate
Voluntar voting-3 Compulsory voting-3Voluntary voting-7 Compulsory voting-7
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
1 2 3 4 5 6
Blocks of 5 Periods
Nu
mb
er
of s
ign
als
ac
qu
ire
dVoluntar voting-3 Compulsory voting-3Voluntary voting-7 Compulsory voting-7
15
Support: Thenullhypothesis ofnodifference in average acquired signalsbetweengroupsof sevenand three is rejected in
favor of greaternumbers in larger groupswith voluntary and compulsory voting ( 0.0003 and 0.0002,Wilcoxon‐Mann‐
Whitney tests, one‐tailed), but not for rates for both votingmodes ( 0.140 and 0.409). And, the null of no difference in
averageacquiredsignalsbetweenthetwovotingmodesisrejectedinfavorofgreaterratesandthusnumberswithcompulsory
votingforgroupsofsevenandthree( 0.001and0.004,Wilcoxonsignedrankstests,one‐tailed).Forcompleteness,thenull
isnotrejectedforratesforVoluntaryvoting‐3versusCompulsoryvoting‐7( 0.109,Wilcoxon‐Mann‐Whitneytest,one‐tailed),
but it is rejected in favor of greater rates in Compulsory voting‐3 versus Voluntary voting‐7 and greater numbers for both
comparisons( 0.032,sametests).
Figure 3 confronts our logit predictionswith the data.We compute best predictors of the degree of
noise, ̂ ∗, usingmaximum likelihoodestimations.15This isdone for thepooleddataof all treatments,
eachweightedbyone‐fourth, forall thirtyperiods(largesymbols)andforthefirstandsecondhalfof
periodsonly(smallsymbols),whichyields ̂ ∗ 0.06respectively ̂ ∗ 0.07and ̂ ∗ 0.05.The
dropinthebestpredictorfromthefirsttothesecondhalfofperiodshintstolearningtowardsBNE(see
alsoGoereeandHolt2005;McKelveyandPalfrey1995).16AsseeninFigure3,observedratesofsignal
acquisition are predicted poorly for about 0.025 (BNE predicts worst) and much better for
∈ 0.025,0.1 .17Forthelatterinterval,wenotonlypredictgreaterinformationpoolswithcompulsory
thanvoluntaryvotingpergroupsize,butalsoinlargerthansmallergroupspervotingmode(seeFigure
A2.1).Overall,ourestimatescorrectlypredictfiveoutofsix(sixoutofsix)possiblecomparativestatics
ofobservedaveragerates(numbers)ofacquiredsignals.18Bycontrast,BNEonlypredictsfouroutofsix
(threeoutofsix)comparativestatics.Notably, itoverratesfree‐ridingandthusfailstopredictgreater
informationpoolsinlargergroups(Experimentalresult1).Finally,observedaverageratesaregreater
15Denoteby thelogitequilibriafor .Let betheobservedsignalacquisitionratefor observations.TheLog‐likelihood
that yields is ln ln 1 ln 1 ,which ismaximized for if such
exists,otherwiseacornersolutionmaximizestheLog‐likelihood.16Wealsofindadropfromthefirsttothesecondpart( ̂
∗ 0.07and ̂ ∗ 0.06,respectively).Notethattheobserved
averagesignalacquisitionrateinCompulsoryvoting‐3 isinbetweenthetreatment’stwoBNE,solearningistowardstheBNEwithuniversalignorance.WhilethismayreflectthetensionbetweenthetwoBNE,noticethatalthoughtheratepermatchinggroupvariesmoreinCompulsoryvoting‐3 thantheothertreatments,onlyoneofitsmatchinggroupsisamongthesixwithalowestrateofabout0.2(Voluntaryvoting‐3andCompulsoryvoting‐7havethreeandtwo,respectively).17Thisalsoholdspertreatment,exceptforVoluntaryvoting‐3wheretheratesarebetterpredictedbylow .Tofinda ̂ ∗pertreatment,simplymovethedatapointsinFigure3horizontallyuntiltheylieontherespectivepredictionline.18Theonlyexceptionisanunpredicted,slightlyhigheraverageratefoundinCompulsoryvoting‐7 thaninVoluntaryvoting‐3.However,thesetreatmentsarenotdirectlycomparable,andtheoneexceptionintheorderofobservedratesreportedearlieroccursinthelastblockoffiveperiods(leftpanelofFigure1),inlinewithourbestpredictors.
16
thanBNEinalltreatmentsbutCompulsoryvoting‐3(i.e.,thehigher,robustBNE),whichismoreinline
withlogitequilibriumthanimplicitcooperationtowardsthesocialoptimum.
Figure3:Signalacquisitionandestimateddegreesofnoiseinlogitequilibrium
Note:Thefigureshowsmaximumlikelihoodestimationsofthedegreeofnoise, ̂ ∗,forthepooleddataof
alltreatments,eachweightedbyone‐fourth,forallthirtyperiods(largesymbols)andforthefirstand
secondhalfofperiodsonly(smallsymbols).
Correctgroupdecisions
Next, we examine how voting translates acquired signals into group decisions. On average, groups
choose thecorrectalternative62.0%and66.0%(68.8%and69.0%)of the time inVoluntaryvoting‐3
and‐7(Compulsoryvoting‐3and‐7),respectively.
Experimentalresult2(Correctgroupdecisions):Onaverage,62.0%to69.0%ofthegroupdecisionsare
correct,wheresmallergroupswithvoluntaryvotingdoworst.
Support:Thenullhypothesisofnodifferenceinaveragecorrectgroupdecisionsbetweenvoluntaryandcompulsoryvotingis
rejectedinfavorofgreateraveragesforgroupsofthreewithcompulsoryvoting,butnot forgroupsofseven( 0.006and
0.237,Wilcoxonsignedrankstests,one‐tailed).And,thenullofnodifferenceisnotrejectedforgroupsizecomparisonsinboth
the voluntary and compulsorymode ( 0.472 and 0.179,Wilcoxon‐Mann‐Whitney tests, one‐tailed). For comparison, it is
alsonotrejected forCompulsoryvoting‐3versusVoluntaryvoting‐7,butrejected in favorofgreateraverages forCompulsory
voting‐7thanVoluntaryvoting‐3( 0.372and0.042,Wilcoxon‐Mann‐Whitneytests,one‐tailed).
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 1 10 100
μ (error parameter)
Lo
git
-QR
E a
nd
ob
se
rve
ds
ign
al a
cq
uis
itio
n r
ate
Voluntary voting-3 Voluntary voting-7 Compulsory voting-3
Compulsory voting-7 Voluntary voting-3 (Data) Voluntary voting-7 (Data)
Compulsory voting-3 (Data) Compulsory voting-7 (Data)
Per
iods
16-
30
All
perio
ds
Per
iods
1-1
5
17
Typically,inalltreatmentsaveragecorrectgroupdecisionspermatchinggroupareinbetweenthe
50% achievedwith universal ignorance and the 75% achieved by a single noise‐free delegate,19 and
henceliemarkedlybelowthesocialoptimumof84.4%ingroupsofthreeand92.9%ingroupsofseven
(seeFigureA2‐2intheappendix).Theserelativeperformancesarepredictedfor ∈ 0.025,0.1 ,albeita
delegate in Voluntary voting‐7 should do worse than the group for about 0.05. With regard to
comparative statics between treatments for ∈ 0.025,0.1 shown in Figure A2‐2, we should see on
averagefewercorrectgroupdecisionsinVoluntaryvoting‐3 than‐7(whichwedo,butnotstatistically
significantly) and about the same averages inCompulsory voting‐3 and ‐7 (whichwe do). Also, there
should bemore correct groupdecisionswith voluntary than compulsory voting for both group sizes,
whichwereject.Overall,whileinthecompulsoryvotingtreatmentsaveragecorrectgroupdecisionsare
similar to those predicted for ∈ 0.025,0.1 , in the voluntary voting treatments they are markedly
smaller than predicted. Against Proposition 1, this hints to cursed uninformed voluntary voting! For
anotherhint,groupdecisionsarenotmoreoftencorrectinVoluntaryvoting‐7thanCompulsoryvoting‐3,
even though in the former treatment the information pool was greater on average and participants
couldabstain.Wefurtherdiscussuninformedvoting,below.
Efficiency
Figure4depictspertreatment,averagedoverindividuals, theobservedsignalcosts(dashedline);net
benefits(darksolidline),whichequal200pointstimestheaveragefractionofcorrectgroupdecisions
minus100pointsfromarandomgroupdecisionwithuniversalignorance;andnetpayoffs(darkbars),
whichareequaltothenetbenefitsminusthecosts.(Thelightlineandbarsarediscussedbelow.)Ascan
be seen, individual net payoffs are smallest in Voluntary voting‐3 (15.3 points), and about equal in
Voluntaryvoting‐7 andCompulsoryvoting‐3 and ‐7 (25.3,25.9, and28.4points;note thatnetbenefits
andcostsaresomewhatsmallerinVoluntaryvoting‐7thaninthetwoothertreatments).
191(0)outof12matchinggroupshad 50%( 75%)correctgroupdecisionsinVoluntaryvoting‐3,andthesenumbersare0(3)outof12inCompulsoryvoting‐3,1(1)outof8inVoluntaryvoting‐7,and0(1)outof8inCompulsoryvoting‐7.
18
Figure4:Efficiency
Note: The figure shows per treatment, averaged over individuals, observed signal costs (dashed line), net
benefits (solid lines),andnetpayoffs(bars).Thenetvaluesaredepartingfromuniversal ignorance,where
eachindividualhasexpectedbenefitsandpayoffsof100points.Foradjustedvalues,allrealizeduninformed
votesarereplacedbyhypotheticalabstentions(ifthiscreatesatie,netvaluesaresettozeropoints).
Experimentalresult3(Efficiency):Inalltreatments,average individualnetpayoffsareinbetweenthe
benchmarksof0pointswithuniversalignoranceand41.7and46.4pointswithasinglenoise‐freedelegate
ingroupsofthreeandseven,respectively.Individualsinsmallergroupswithvoluntaryvotingearn10.0to
13.0pointslessthanthoseintheothertreatments.
Support:Thenullhypothesisofnodifference inaverageefficiencybetweenvoluntaryandcompulsoryvoting is rejected in
favorofgreateraveragesforcompulsoryvotingingroupsofthree,butnotforgroupsofseven( 0.026and0.320,Wilcoxon
signed ranks tests, one‐tailed). And, the null hypothesis of no difference for group size comparisons is rejected in favor of
greater averages for Voluntary voting‐7 than ‐3, but not in the compulsory mode ( 0.084 and 0.472, Wilcoxon‐Mann‐
Whitney tests, one‐tailed). For comparison, it is rejected in favor of larger averages forCompulsory voting‐7 thanVoluntary
voting‐3,butnotforCompulsoryvoting‐3versusVoluntaryvoting‐7( 0.062and0.472,sametests).
Note thatalmostalwaysdogroupsof threeandsevenachieve loweraveragenetpayoffs than in
socialoptimum(43.8and60.9points)andwithasinglenoise‐freedelegate(41.7and46.4points,i.e.50
minus25/groupsize),20andrecallthatnetpayoffsareequaltozerowithuniversalignorance(seealso
201(0;0)outof12matchinggroupsinVoluntaryvoting‐3earnedonaveragelessthanwithuniversal ignorance(morethanwithasinglenoise‐freedelegate;morethaninsocialoptimum),andthesenumbersare0(1;1)outof12inCompulsoryvoting‐3,1(1;0)outof8inVoluntaryvoting‐7,and0(1;0)outof8inCompulsoryvoting‐7.
‐40
‐20
0
20
40
60
80
Voluntary
voting‐3
Compulsory
voting‐3
Voluntary
voting‐7
Compulsory
voting‐7Signal costs and net benefits and payoffs in points
Net payoffs
Net payoffs(adjusted)
Net benefits
Net benefits(adjusted)
Signal costs
.
19
FigureA2.3intheappendix).Finally,asaconsequenceofsmalleraveragecorrectgroupdecisionsthan
predictedinthevoluntaryvotingtreatments,observednetpayoffsarealsosmallerthanpredicted.By
contrast,predictedandrealizednetpayoffsaresimilarinthecompulsoryvotingtreatments.
Uninformedvoting
Here,weinvestigateuninformedvotinginmoredetail.Table1showspertreatmentaveragefractionsof
sincere voting, insincere voting, and abstention of the informed, and voting and abstention of the
uninformed(notethatwecannotdistinguishbetweensincereandinsincereuninformedvoting).
Table1:Fractionsofinformedanduninformedvoting
TreatmentInformed Uninformed
VotingAbstention Voting Abstention
Sincere InsincereVoluntaryvoting‐3 0.341 0.007 0.001 0.252 0.399Compulsoryvoting‐3 0.461 0.006 ‐ 0.533 ‐Voluntaryvoting‐7 0.268 0.003 0.000 0.293 0.436Compulsoryvoting‐7 0.377 0.004 ‐ 0.618 ‐
As seen in the table,observeddecisionsmainlysupportProposition1,albeitwithone important
exception.Aspredicted, the informedalmostnever abstain in thevoluntarymodeandalmost always
votesincerelyinbothvotingmodes.21However,assuspected,wefindsubstantialvoluntaryvotingofthe
uninformed(38.7%and40.2%oftheirdecisionsinVoluntaryvoting‐3and‐7,respectively).22Figure4
alsoshowsadjustednetbenefitsandnetpayoffsaveragedoverindividuals,forwhichwereplacedeach
realizeduninformedvotewithahypotheticalabstention(graylineandbars).23Theadjustmentallows
ustoevaluatetheactualnegative impactof theswingvoter’scurseongroupdecisionsandefficiency,
butnaturally itcannotcapturehowdecisionsinthevoluntarymodewouldhavebeenmadeabsentof
uninformedvoting.Thefigurerevealsasubstantialdamageinalltreatments:theadjustmentraisesthe
21Outof3,240(3,360)decisions,weonlyfind4(1)informedabstentionsbyoneparticipantinVoluntaryvoting‐3(‐7),and22and18(9and15)insincerevotesinVoluntaryvoting‐3andCompulsoryvoting‐3(‐7and‐7),respectively.22Uninformedvoteswereslightlybiased: ineachtreatment,54%forBlueversus46%forYellow ( 0.019,binomial tests,one‐tailed).Moreover,in40.4%and21.0%(29.7%and18.8%)oftheelectionsinVoluntaryvoting‐3and‐7(Compulsoryvoting‐3and‐7),uninformedvotesreplacedacoinflipforatieofinformedvotes,in4.4%and12.1%(7.4%and13.3%)theyturnedawinofthecorrectcolorintoaloss,andin1.5%and3.3%(1.9%and3.5%)theyturnedalossofthecorrectcolorintoawin.Finally,between54.1%and65.6%(56.3%and70.6%)of these incorrect (correct) changeswere forBlue, except for38.1%correctchangesinCompulsoryvoting‐3.23Whenanadjustmentcreatesatie,weusenetbenefitsofzeropointsinsteadofacoinflip.Thischoicehasvirtuallynoeffectonthereportedaverages.
20
fractionofcorrectgroupdecisionsby3.4and8.4(3.7and9.4)percentagepointsinVoluntaryvoting‐3
and ‐7 (Compulsory voting‐3 and ‐7),which raises average individual benefits and payoffs by 6.8 and
16.9(7.4and18.8)points.Finally,wecancompareadjustedvalues in thevoluntarymode,wherethe
uninformed are predicted to abstain, to the respective realized values in the compulsory mode. In
groupsofthree(seven),adjustedaverageindividualnetbenefitsandnetpayoffswithvoluntaryvoting
are30.8and22.1(49.0and42.2),andtherespectiverealizedvalueswithcompulsoryvotingare37.6
and 25.9 (37.9 and 28.4). These numbers accentuate the generally poor performance in Voluntary
voting‐3, and suggest high forgone gains inVoluntary voting‐7 where, without cursed voting, groups
couldhavedonemarkedlybetterthaninthecompulsorymode.
Figure5:Uninformedvoluntaryvotingperparticipant
Groupsofthree Groupsofseven
Note: The figure shows for eachparticipant (dot) the fraction of periods inwhich she or hewasinformedonthehorizontalaxis,andthefractionofperiodsinwhichsheorhevotedontheverticalaxis.Largercirclesizesindicatelargernumbersofparticipantsatthesamecoordinates.
Next,weexamineuninformedvoluntaryvotingattheindividuallevel.Figure5showsscatterplots
forVoluntaryvoting‐3(leftpanel)andVoluntaryvoting‐7(rightpanel),whereeachdotrepresentsone
participant (the size of their agglomeration at the same coordinates is indicatedby the size of larger
circles). The horizontal axis shows the fraction of periods inwhich she or hewas informed, and the
verticalaxisshowsthefractionofperiodsinwhichsheorhevoted.Hence,individualsonthediagonal
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0
Fraction of periods with a signal
Fra
cti
on
of
pe
rio
ds
vo
ted
fo
r A
or
B
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0
Fraction of periods with a signal
Fra
cto
n o
f p
eri
od
s v
ote
d f
or
A o
r B
21
(gray line) never voted uninformed,24 and those above the diagonal voted uninformed to various
extents. Clearly, uninformed voluntary voting is not restricted to just a few individuals: inVoluntary
voting‐3and‐7,morethanhalfofthemvoteduninformedtosomeextent(59outof108and60outof
112participants,respectively).
Experimentalresult4 (Uninformedvoting):There isa sizablenumberofuninformedvoluntaryvotes
(25.2%and29.3%outofalldecisions ingroupsof threeandseven),nearlyhalfasmanyasuninformed
compulsoryvotes(53.3%and61.8%).Uninformedvoluntaryvotingiscommonamongindividuals.
Thereareseveralpossibleexplanationswhysomanyuninformedparticipantsdonotabstainwhen
they can. For example,people areboundedly rational inmanyaspects, and thereforemay simplynot
realize that random voting can negatively affect the group decisionwhich is less intuitive than the
positiveeffectofinformedsincerevotingorignorepivotaleventsaltogether(e.g.,EspondaandVespa
2012;Martinelli2011).Onewould thenexpect toseesome learningof theswingvoter’scurse inour
experiment, towit, a decline inuninformedvoluntary voting.However, our results in this regard are
mixed(seeFigureA3),25whichhintstoafractionofparticipantswhowereconsciousofthecursewhen
casting their uniformed voluntary vote.26Thesepeoplemayvotebecause abstaining basically reveals
their ignorance and hence can lead others to reciprocate by getting informed less often in future
periods.27Finally, in linewithoverreportingofvoting inelectionsurveys(e.g.,Bernstein,Chadha,and
Montjoy2001),somepeoplederivepridefromsignalingtoothersthattheyvotedandfeelembarrassed
to admit their failure to do so (DellaVigna et al. 2013),which toomay explain uninformed voluntary
voting.Ourexperimentisnotdesignedtofindouttheexactcausesofanomalousuninformedvoting,so
weleavethistasktofutureresearch.
24Thetwoparticipantswhoabstainedinformed(seefootnote21)virtuallyalwaysvoteduninformedotherwise.25Usingprobitregressionswithrandomeffectsatthematchinggroup,inVoluntaryvoting‐3wefindnostatisticallysignificanteffectoftheperiodonuninformedvotingforbothpartstogether( 0.956,two‐tailed),andnegativeandpositiveeffectsforthe first and second part, respectively ( 0.090). In Voluntary voting‐7, we find no significant effect for the first part( 0.490),butnegativeeffects forthesecondpartandbothpartstogether,respectively( 0.070).Recall that inthetwocompulsoryvotingtreatmentssignalacquisitiondeclinesovertime(footnote14),souninformedvotingincreasesovertime.26Participantscouldnotdirectlyobserveifothersvotedinformed.Itislikelythattheyanticipatedsomeextentofuninformedvoluntaryvoting,especiallyiftheydidsotoo.Thoughlearningthisextentisverydifficult,theycouldupdatetheirbeliefsaboutuninformedvotingaftereachperiod,usingthenumbersofcorrectandincorrectvotesofothers.Thisiswhatwestudybelow.27Inourrandommatchingprotocol,participantscouldinferthattheyencounteraparticularotherparticipantonaverage2.31and 6.67 times per part in groups of three and seven (unknown to them, theymetmore often due to thematching groupproceduredescribedearlier),soabstainingcouldindeedhavedirectandindirecteffectsonrelevantfuturegroupdecisions.
22
Next,weanalyzeuninformedvoluntaryvotingusingmodelsofboundedrationality.First,applying
individualrationalchoicetooursituation,noise‐freeplayersdecideoptimallyasiftheywerealone.By
ignoringsocialinteractionentirely,theyarealwayspivotalintheirlimitedrationalminds.Consequently,
foralltreatments,theyshouldconsistentlygetinformedandvotesincerely(asournoise‐freedelegate),
and off the equilibrium path, the uninformed randomly choose A or B or, if feasible, abstention. By
contrast,ourparticipantsareoftenuninformed,andatratesthatdependonthetreatmentwhichalso
noisybehaviorcouldnotexplain(butuninformeddecisionsinthevotingstageappeartoberandom,as
predicted;seeFigure5and footnote22).Also,belowweshowthatparticipantsconsiderclosenessof
thenextvotewhenmakingtheirsignalacquisitiondecisions(seeTable2).Overall,individual“rational”
choiceisnotaplausibleexplanationofourdata.Second,turningtologitequilibrium,itcannotimprove
on theBNEpredictionof sincerevoting (Prediction1), as this iswhat informedparticipantsvirtually
always do (Table 1). However, it can capture the observed substantial uninformed voluntary voting,
whichBNEcannot.Toseethis,notethattheuninformedsurelyabstaininthenoise‐freeextremeoflogit
equilibrium,andchooseAorBorabstentionwithprobabilityone‐third foreach in theotherextreme
withpurenoise.Then,onthelogitequilibriumpath,raisingthedegreeofnoiseshiftsprobabilitydensity
fromuninformedabstentioninequalsharestouninformedA‐respectivelyB‐votes,whichisconsistent
withthe38.7%and40.2%ofuninformedvotes(abouthalfforeachAandB;seefootnote22)foundin
Voluntaryvoting‐3and‐7.Ifoneagreesthatsincerevotingismuchmoreintuitivethantheswingvoter’s
curse, so one can justify decision‐making noise only for the uninformed, then logit equilibrium is a
plausibleexplanationofourdata(butnotethatthenthreedifferentdegreesofnoisearetobeestimated,
includingforsignalacquisitiondecisions).Thenotionof intuitivesincerevotingissupportedbyother
prominentmodelsofboundedrationality,suchask–levelreasoning(e.g.,Camerer,Ho,andChong2004),
cursedequilibrium(EysterandRabin2005),andanalogy‐basedexpectationequilibrium(Jehiel2005).As
shown by Battaglini,Morton, and Palfrey (2010),whose analysis also applies to our voting stage, all
three models predict that any informed playereven a fully unsophisticated onevotes sincerely
(differentbehaviorispredictedforuninformedplayers:thefullyunsophisticatedchooserandomly,asin
theindividualchoicemodelofftheequilibriumpath,andeveryoneelseabstains).Asmentionedearlier,
ourdataarealsoconsistentwiththenotionthatsomeparticipantsconsciouslytaketheriskofcastinga
23
cursedvoluntaryvote,whichlikelyincludesthosewhorealizethatothersdosoandthenreciprocateby
doingjustthesame.Atthispoint,wemustleaveittofutureresearchtofindouttherespectivefractions
of noisy and conscious cursedvoting.Knowing these fractions canhelp to improvepredictions, since
noisy and conscious types have different expected benefits from voting and thus different signal
acquisition probabilities in equilibrium. Ideally, these predictions also take into account imperfect
informationaboutthedistributionofthevarioustypes.Thelikelyresultisthatlogitequilibriumsignal
acquisition probabilities are in between those derived in this paper, that is, where the uninformed
always abstain in the voluntary mode and always vote in the compulsory mode. However, realizing
cursedvoluntaryvotingofothersmighttriggerstrongnegativereciprocityinindividualsanddragdown
informationpoolsbelowpredictedlevels,somethingwhichmayhaveoccurredinVoluntaryvoting‐3.
Learningaboutclosevotes
Next,weanalyzewhetherparticipantsbasetheircurrentsignalacquisitiondecisionsonclosenessofthe
nextvote,whichtheymayprojectfromothers’previousvotingdecisions.Wedosobyrunningforeach
treatment separatelyprobit regressionswith randomeffects at the level of thematching group (mg).
The dependent dummy variable, , , equals 1 if individual acquired a signal in period , and 0
otherwise. As independent variables, we use one dummy variable, two quantitative variables, and
various interactions of these variables. First, the lagged dependent dummy variable, , , separates
individualswhowereinformedinthepreviousperiod, 1,fromthosewhowerenot.Thisallowsusto
inspectwhether thosewho tend to be informed and uninformed respond to information in previous
voting decisions of others differently. Second, the first quantitative variable, # /2 ∀ , , is
definedas thepreviousnumberofinformedanduninformedvotesbyall individualsother than ,
∀ ,labeled# ∀ , ,dividedbythenumberofotherindividuals,2 .Naturally,thisvariable
isonlyusedwithvoluntaryvoting.Third,thesecondquantitativevariable, ∆ /# ∀ , ,is
definedas thenumberofcorrectvotesminusthenumberof incorrectvotesby∀ in 1, labeled
∆ ∀ , , divided by # ∀ , . The two quantitative variables are normalized using 2
respectively # ∀ , as denominators in order to account for the differences in group size.
Fourth,weuseavariablethatinteractsthetwoquantitativevariables,andthreevariablesthatinteract
24
the lagged dependent variable with each quantitative variable and their interaction variable (again,
variables using # /2 ∀ , are not applicable in the compulsorymode). Finally,we include
twoerrorterms, , and ,wherethelatterisarandomeffectusedtocorrectforthepanelstructure
in thedata.The twoquantitativevariableswere chosen, since inour randommatchingprotocol they
revealbasicinformationthataparticipantcanavailforupdatingherorhisprojectionofaclosevotein
thecurrentperiod.Tobeprecise,foranygiven,unknownstrictlypositiveprobabilitywithwhichother
individualsbuyasignal,agreaternumberoftheirvotesleadstoagreaterexpectedmarginofvictoryfor
the correct alternative (recall that informedparticipants virtually always voted sincerely).And, given
thenumberofothers’votes,agreatermarginofvictoryforthecorrectalternativemeansthattheyhave
agreaterprobabilityofbuyingasignal,onaverage.Therefore,ifanindividualpreferstoavoidthesignal
costwhenbelievingthatthenextvoteoutcomemightnotbeclose,thenboth # /2 ∀ , and
∆ /# ∀ , shouldnegativelyaffecthercurrentsignalacquisitionprobability.Moreover,
the interaction of both quantitative variables should strengthen the negative effect, since a greater
numberofothers’votesjointlywithagreaterprobabilityofthemgettinginformedandcastingasincere
votemakesacorrectgroupdecisionevenmorelikely.
Table2:Probitestimationsofsignalacquisitionswithrandomeffectsatthematchinggrouplevel
Constantandindependentvariables
CoefficientsVoluntaryvoting‐3
Compulsoryvoting‐3
Voluntaryvoting‐7
Compulsoryvoting‐7
Constant ‐1.048***(0.099) ‐1.002***(0.094) ‐1.000***(0.119) ‐1.227***(0.045)∆# ∀ ,
0.131(0.151) ‐0.141**(0.063) 0.203(0.155) ‐0.278**(0.113)#
∀ , ‐0.258**(0.114) ‐0.379*(0.204)
∆#
#
∀ , ‐0.471**(0.230) ‐0.700**(0.311)
, 1.962***(0.119) 1.983***(0.067) 1.472***(0.186) 2.182***(0.067)
,∆# ∀ ,
0.228(0.226) 0.049(0.091) ‐0.300(0.261) 0.173(0.163)
,#
∀ , 0.089(0.179) 0.706**(0.323)
,∆#
#
∀ , ‐0.099(0.344) 0.257(0.511)
Note:Thedependentvariableistheindividuals’binarydecisiononwhethertobuyasignal( 1),ornotbuy(
0).Theindependentvariables(firstcolumn)aredefinedinthemaintext.*(**;***)indicatessignificanceatthe
10%(5%;1%)level,andabsolutestandarderrorsaregiveninparentheses.
25
Table 2 reveals some interesting learning effects.28 First, in all treatments, previously informed
individualshaveamuchhigherbaselineprobabilityofgettinginformedinthecurrentperiodthanthe
previously uninformed (all constants are negative and large, and statistically significant, and all
coefficients of , are positive and large, and significant). Second, focusing first on the previously
uninformed, in both compulsory voting treatments they get informed less likely when the value of
∆ /# ∀ , increases (both coefficients are negative and statistically significant). By
contrast, in both voluntary voting treatments, they avail the extra informationof varyingnumbers of
others’ votes by getting informed less likely when the value of # /2 ∀ , increases (both
coefficients are negative and statistically significant). Moreover, they only respond to ∆ /
# ∀ , bygettinginformedwithamuchlowerprobabilitywhenthevalueoftheinteractionof
the two quantitative variables increases (both coefficients of ∆ /# ∀ , are
insignificant, andboth coefficients of the interaction variable are large andnegative, and significant).
Third, compared to theiruninformedcounterparts, thepreviously informedgenerallydonot respond
differently to information in previous voting decisions of others, except that a greater value of
# /2 ∀ , markedlyandstatisticallysignificantlyraisestheirprobabilityofgettinginformed
inthecurrentperiod,which(allothercoefficientsareinsignificantinalltreatments).Overall,itseems
that apart from the higher baseline signal acquisition probabilities of individuals who tend to be
informed,bothpreviouslyinformedanduninformedparticipantsgenerallyrespondtovotingdecisions
of others in similar ways. In particular, our probit estimations suggest thatmany of them avoid the
signalcostinanticipationofaclosevote.
Experimentalresult5 (Learning and close votes): Inall treatments,previously informedparticipants
aremore likely to get informed in the current period than those who were not informed. Moreover,
independentoftheirprevioussignalacquisitiondecision,participantsaregenerallymore(less)likelytoget
informedwhen the information inothers’previousdecisions indicates that thenextvotemight (not)be
close.
28Thoughnotincludedinthepanelmodels,uninformedvoteswereslightlybiasedbythegambler’sfallacy.Overall,afterthetruestatewasYellow,55%votedforBlue( 0.023exceptfor 0.352inVoluntaryVoting‐7,binomialtests,one‐tailed)andafteritwasBlue,55%votedforYellow( 0.036exceptfor 0.159inCompulsoryVoting‐3,sametests).
26
5Conclusions
Westudymajorityvotingovertwoalternativesinsmallgroups.Individualshaveidenticalpreferences
but are uncertain aboutwhich alternative is correct (both are equally likely correct). Prior to voting,
eachindividualcanacquireacostlysignalthatisinformativebutimperfect.Towit,itshowsthecorrect
alternative with probability ∈ , 1 , and the incorrect one with probability 1 . We investigate
voluntaryandcompulsoryvoting.Anuninformedvotecannegativelyimpactthegroupdecisionasitcan
overridean informative groupdecision.Withvoluntary voting,uninformed individuals canavoid this
cursebyabstaining,whichisnotafeasibleoptioninthecompulsorymode.
Inourexperiment,weemploygroupsofthreeandseven.Wefindthat,onaverage, largergroups
havegreaterinformationpoolsthansmallergroupsinbothvotingmodes,andthatthepoolsaregreater
with compulsory than voluntary voting. The results are predicted by logit equilibrium for degrees of
decision‐making errors commonly observed in related collective action experiments, but not byBNE.
Surprisingly, in both voluntary voting treatments there is substantial uninformed voting by many
participants, with strong negative consequences for efficiency. Overall, the fraction of correct group
decisionsandefficiencyareinbetweenthosepredictedforuniversalignorance(i.e.,thegroupdecision
is correcthalfof the time,atnosignal costs)and fora singlenoise‐freedelegatewhodecides for the
groupandwhoseexpectedperformanceislowerthaninsocialoptimum.Moreover,smallergroupswith
voluntaryvotingdoworst,whilegroupsinallothertreatmentsfareaboutequallyunsatisfactory.While
lowperformancesarepredictedby logit equilibrium in the compulsorymodewhere someuniformed
votingistolerated,theyarelowerthanpredictedinthevoluntarymodeduetotheswingvoter’scurse.
Ourresultsaredisillusioningandsuggestthatthecursemustbetakenveryseriouslyinthecurrent
practiceof juryandcommitteevoting.Forexample, inorder topreventuninformedvoluntaryvoting,
onemightconsiderintroducingsomevotingcosts(seealsoTyson2013)orrandommonitoringofthe
knowledgeofdecisionmakers.And,animportantissuesistherobustnessofourfindings.Inthisregard,
Elbittaretal (2013)andShineman(2013)providecomplementaryevidenceofuninformedvoluntary
votingindifferentsetups.Overall,oursmallgroupstudyofmajoritarianvotingwithendogenouscostly
informationhints tosevereefficiency lossesassociatedwith thecurseofuninformedvoting.Buteven
withoutthecurse,comparedtothesocialoptimum,predictedandobservedefficienciesarelowdueto
27
thecollectiveactionproblemsinvolved.Therefore,ourstudysuggeststhatthereisroomforimproving
existing designs of jury and committee voting, and hopefully it can inspire further research in this
direction.
References
Abraham,HenryJ.1955.Compulsoryvoting.WashingtonDC:PublicAffairsPress.
Austen‐Smith,David,andJeffreyS.Banks.1996.“Informationaggregation,rationality,andtheCondorcet
JuryTheorem.”AmericanPoliticalScienceReview90(1):34‐45.
Bartels, Larry M. 1996. “Uninformed votes: Information effects in presidential elections.” American
JournalofPoliticalScience40(1):194‐230.
Battaglini, Marco, Rebecca B. Morton, and Thomas R. Palfrey. 2010. “The swing voter’s curse in the
laboratory.”ReviewofEconomicStudies77:61‐89.
Bernstein,Robert,AnitaChadha,andRobertMontjoy.2001.“Overreportingvoting:Whyithappensand
whyitmatters.”PublicOpinionQuarterly65(1):22‐44.
Bhattacharya, Sourav, John Duffy, and Sun‐Tak Kim. 2013. “Compulsory versus voluntary voting: An
experimentalstudy.”Workingpaper.
Birch, Sarah. 2009. Full participation: A comparative study of compulsory voting. New York: United
NationsUniversityPress.
Börgers,Tilman.2004.“Costlyvoting”.AmericanEconomicReview94(1):57‐66.
Camerer,Colin,andTeck‐HuaHo,andJuin‐KuanChong.2004.“Acognitivehierarchymodelofbehavior
ingames.”QuarterlyJournalofEconomics119(3):861‐98.
Coughlin,PeterJ.2000.“Indefenseofunanimousjuryverdicts:Mistrals,communication,andstrategic
voting.”AmericanPoliticalScienceReview94(2):375‐93.
Coupé,Tom,andAbdulG.Noury.2004.“Choosingnottochoose:onthelinkbetweeninformationand
abstention.”EconomicsLetters84:261‐65.
DellaVigna, Stefano, John A. List, Ulrike Malmendier, and Gautam Rao. 2013. “Voting to tell others.”
Workingpaper.
Downs,Anthony.1957.Aneconomictheoryofdemocracy.NewYork:HarperCollinsPublishers.
28
Elbittar,Alexander,AndreiGomberg,CésarMartinelli,andThomasR.Palfrey.2013.“Anexperimenton
votingincommitteeswithcostlyinformationacquisition.”Workingpaper.
Esponda, Ignacio, and Emanuel Vespa. 2012. “Hypothetical thinking and information extraction:
Strategicvotinginthelaboratory.”Workingpaper.
Eyster,Erik,andMatthewRabin.2005.“Cursedequilibrium.”Econometrica73(5):1623‐72.
Feddersen,Timothy,andWolfgangPesendorfer.1996.“Theswingvoter’scurse.”AmericanEconomic
Review86(3):408‐24.
Feddersen,Timothy,andWolfgangPesendorfer.1997.“Votingbehaviorandinformationaggregationin
electionswithprivateinformation.”Econometrica65(5):1029‐58.
Feddersen,Timothy,andWolfgangPesendorfer.1998.“Convictingtheinnocent:Theinferiorityof
unanimousjuryverdicts.”AmericanPoliticalScienceReview92(1):23‐35.
Feddersen,Timothy,andWolfgangPesendorfer.1999.“Abstentioninelectionswithasymmetric
informationanddiversepreferences.”AmericanPoliticalScienceReview93(2):381‐98.
Feddersen,Timothy,andAlvaroSandroni.2006.“Ethicalvotersandcostlyinformationacquisition.”
QuarterlyJournalofPoliticalScience1:287‐311.
Fischbacher, Urs. 2007. “Z‐tree: Zurich toolbox for ready‐made economic experiments.”Experimental
Economics10:171‐78.
Funk,Patricia.2010.“Socialincentivesandvoterturnout.EvidencefromtheSwissmailballotsystem.”
JournaloftheEuropeanEconomicAssociation8(5):1077‐103.
Gerardi, Dino, and Leeat Yariv. 2008. “Information acquisition in committees.” Games and Economic
Behavior62(2):436‐59.
Gerardi,Dino,MargaretA.McConnell,JulianRomero,andLeeatYariv.2009.“Getoutthe(costly)vote:
Institutionaldesignforgreaterparticipations.”Workingpaper.
Gerling,K.,Grüner,H.P.,Kiel,A.,andSchulte,E.2005.“Informationacquisitionanddecisionmakingin
committees:asurvey.”EuropeanJournalofPoliticalEconomy21:563‐97.
Gershkov,Alex,andBalázsSzentes.2009.“Optimalvotingschemeswithcostlyinformationacquisition.”
JournalofEconomicTheory144:36‐68.
29
Goeree, Jacob,andCharlesA.Holt.2005.“Anexplanationofanomalousbehaviorinmodelsofpolitical
participation.”AmericanPoliticalScienceReview99(2):201‐213.
Greiner, Ben. 2004. “The online recruitment system ORSEE 2.0–A guide for the organization of
experimentsineconomics.”UniversityofCologne.WorkingPaperSeriesinEconomicsNo.10.
Großer, Jens.2012. “Votingmechanismdesign:Modeling institutions inexperiments.” InExperimental
Political Science, ed. B. Kittel, W. J. Luhan, and R. B. Morton. Palgrave‐Macmillan, ECPR Research
MethodsSeries:72‐91.
Großer, Jens, and Thorsten Giertz. 2009. “Candidates, voters, and endogenous group formation: A
ColonelBlottoexperiment.”Workingpaper.
Großer, Jens, and Arthur Schram. 2010. “Public opinion polls, voter turnout, and welfare: An
experimentalstudy.”AmericanJournalofPoliticalScience54(3):700‐17.
Großer,Jens,andMichaelSeebauer.2006.“Votingandcostlyinformationacquisition:Anexperimental
study.”Workingpaper.
Guarnaschelli,Serena,RichardD.McKelvey,andThomasR.Palfrey.2000.“Anexperimentalstudyofjury
decisionrules.”AmericanPoliticalScienceReview94(2):407‐23.
Jehiel,Philippe.2005.“Analogy‐basedexpectationequilibrium.”JournalofEconomicTheory123:81‐104.
Krishna, Vijay, and JohnMorgan. 2011. “Overcoming ideological bias in elections.” JournalofPolitical
Economy119(2):183‐211.
Krishna,Vijay,andJohnMorgan.2012.“Voluntaryvoting:Costsandbenefits.”Workingpaper.
Ladha, Krishna K. 1992. “The Condorcet jury theorem, free speech, and correlated votes.” American
JournalofPoliticalScience36(3):617‐34.
Lassen, David Dryer. 2005. “The effect of information on voter turnout: Evidence from a natural
experiment.”AmericanJournalofPoliticalScience49(1):103‐18.
Martinelli,César.2006.“Wouldrationalvotersacquirecostlyinformation?”JournalofEconomicTheory
129:225‐51.
Martinelli,César.2011.“Ignoranceandnaiveteinlargeelections.”Workingpaper.
McKelvey, Richard D. and Thomas R. Palfrey. 1995. “Quantal response equilibrium for normal form
games.”GamesandEconomicBehavior10:6‐38.
30
McLennan, Andrew. 1998. “Consequences of the Condorcet jury theorem for beneficial information
aggregationbyrationalagents.”AmericanPoliticalScienceReview92(2):413‐18.
McMurray,JosephC.2010.“Empiricalevidenceofstrategicvoterabstention.”Workingpaper.
Mukhopadhaya, Kaushik. 2003. “Jury size and the free rider problem.” JournalofLaw,Economicsand
Organization19(1):24‐44.
Oliveros,Santiago.2012.“Abstention,ideology,andinformationacquisition.”Workingpaper.
Olson, Mancur. 1965. The logic of collective action: Public goods and the theory of groups. Harvard
UniversityPress,Cambridge,Massachusetts
Palfrey, Thomas R., and Keith T. Poole. 1987. “The relationship between information, ideology, and
votingbehavior.”AmericanJournalofPoliticalScience31(3):511‐30.
Persico,Nicola.2003.“Committeedesignwithendogenousinformation.”ReviewofEconomicStudies70:
1‐27.
Piketty, Thomas. 1999. “The information‐aggregation approach to political institutions.” European
EconomicReview43:791‐800.
Sastro,Anne,andBenGreiner.2010.“Voterbehaviorunderasymmetricinformationincompulsoryand
voluntaryvotingsystems:Anexperimentalstudy.”Workingpaper.
Shineman,VictoriaA.2013.“Compulsoryvotingascompulsoryballoting:Howmandatoryballotinglaws
increaseinformedvotingwithoutincreasinguninformedvoting.”Workingpaper.
Tyson,Scott.2013.“Thesocialcostsofcompulsoryvoting.”Workingpaper.
Wattenberg,MartinP., IanMcAllister, andAnthonySalvanto.2000. “Howvoting is like takingan SAT
test:AnanalysisofAmericanvoterrolloff.”AmericanPoliticsQuarterly28(2):234‐50.
Wit,Jörgen.1998.“RationalchoiceandtheCondorcetjurytheorem.”GamesandEconomicBehavior22:
364‐76.
Appendix
ProofofProposition1
IntheVotingstage,notethatindividual ispivotalifotherindividuals cast equallymanyvotes
foreachalternative; onemorevoteforA;or onemorevoteforB;andsheisnon‐pivotalinall
31
otherevents.Wesaythat “strictlyprefers”adecisiontoanotheroneif,inpivotalevents,itselectsthe
alternative indicatedby strictlymore signals in the group (includingher own signal), and that she is
“indifferent”innon‐pivotaleventsorif, inpivotalevents,thetwoalternativesareindicatedbyequally
manysignalsinthegroup.Wenowanalyzevoluntaryandcompulsoryvotinginturn.
With voluntary voting, we must show that there exists a symmetric BNE where each informed
individualvotessincerelyandeachuninformedindividualabstains.Supposethateach informedother
individual votessincerelyandeachuninformed abstains.If isinformed,assumingwithoutlossof
generality thatshehasana‐signal, thensincerevoting isheronlyweaklydominantstrategybecause:
first,itweaklydominatesabstainingasinpivotalevents itisstrictlypreferred,andinpivotalevents
and and in non‐pivotal events she is indifferent; second, sincere voting weakly dominates
votingforBasinpivotalevents and itisstrictlypreferred,andinpivotalevents )andinnon‐
pivotaleventssheisindifferent.If isuninformed,thenabstainingweaklydominatesvotingforA[B]as
inpivotal events [ ] abstaining is strictlypreferred, and inpivotal events and [ and
]andinnon‐pivotaleventssheisindifferent.Thus,giventhatothersvotesincerelyifinformedand
abstain ifuninformed, ’sonlyweaklydominantstrategy is tousethesamestrategyaseveryoneelse.
ThiscompletesourproofofexistenceofoursymmetricBNEwithvoluntaryvoting.
With compulsory voting,wemust show that there exists a symmetric BNEwhere each informed
individualvotessincerelyandeachuninformedindividualvotesrandomlywithprobabilityone‐halffor
eachalternative.Aseveryonemustvoteandthegroupsizeisodd,2 1,individual isonlypivotalif
otherindividuals cast votesforeachalternative(seepivotalevents ),andsheisnon‐pivotal
inallotherevents.Supposenowthateachinformed votessincerelyandeachuninformed votes
withprobabilityone‐half foreachalternative.If is informed,assumingwithout lossofgeneralitythat
shehasana‐signal, thenneitherAnorBdominates theotherdecisionbecause: innon‐pivotalevents
andinpivotaleventswithonemoreB‐thanA‐votesofinformedotherssheisindifferentbetweenvoting
forAandforB;inpivotaleventswithatleastasmanyA‐thanB‐votesofinformedotherssheprefersto
votesincerely;andinpivotaleventswithatleasttwomoreB‐thanA‐votesshepreferstovoteagainst
herownsignal.However,usingherowna‐signal,sheexpectsthatinformedothershavemorea‐thanb‐
signals, on average, and thus her unique best response is to vote according to her signal. If is
32
uninformed, then there isalsonodominationbecause: if she ispivotal, sheprefers tovote forA [B] if
therearemoreA‐thanB‐votes[B‐thanA‐votes]ofinformedothers,andsheisindifferentifsheisnon‐
pivotalandifeachalternativehasequallymanyvotesofinformedothers.As hasnosignal,sheexpects
that informedothershaveequallymanya‐andb‐signals,onaverage,and thusanyprobabilitymixof
voting for A and for B is a best response, including , . Thus, given that others vote sincerely if
informedandvotewithprobabilityone‐halfforeachalternativeifuninformed,it isabestresponseof
individual touse the same strategy as everyone else,which showsexistenceof our symmetricBNE
withcompulsoryvoting.∎
ProofofProposition2
We first derive the symmetric logit equilibrium conditions 1 and 2 for the signal acquisition
probability ∗ , , andthereafterproveourcomparativestaticsresults.
FigureA1:Logitequilibrium
Note: The left panel shows 1 for variousdegreesofnoise.The rightpanel shows 1 for 0.06,
1 , and the signal cost (i.e., the lower bound) for our experimental treatments and parameters. In all
treatments, both sides of 1 intersect only once so that each has a unique ∗ 0.06 : 0.409 and 0.277 in
Voluntaryvoting‐3and‐7,and0.517and0.355inCompulsoryvoting‐3and‐7,respectively.
33
(Logit equilibrium condition) Anticipating equilibrium voting, ∗ (Proposition 1), individual ’s
expected payoffs of getting informed and not getting informed, 1, , ∗ , , and
0, , ∗ , , ,areaugmentedbythestochasticterms and ,respectively,where and
are iid random variables (recall that the parameter 0 controls the degree of noise). To keep
thingssimple,inthefollowingweomittheindexmwheneverouranalysisappliestobothvotingmodes
orthemodeisobvious.Then,ifeveryoneelsebuysasignalwithprobability ,individual getsinformed
ifandonly if 1 0 ⇔Π , ∗, , , / ,29whichoccurs
withprobability Π . / ,whereΠ . ≡ 1 0 denotesherexpectednet
payoffsofgettinginformedandFisthedistributionfunctionofthedifference .Takingtheinverse
of . and multiplying both sides by yields Π . . Finally, using the logistic
distribution, 1/ 1 ,givesourlogitequilibriumcondition
1Π , ∗, , , . 1
If 0,then 1 | ∈ , 0and 1 turnsintotheBNEcondition.If 0,then 1 hasthe
following properties: regarding , 0 for ∈ 0,1 , lim → 1 ∞,
1 | / 0, and lim → 1 ∞; and regarding , 0 for ∈ 0, , 0
for , and 0 for ∈ , 1 . The left panel of Figure A1 illustrates these properties for
∈ 0.01, 0.1, 0.5, 10}.Itdepicts onthehorizontalaxis,and 1 ontheverticalaxis.For 1 ,
theexpectednetpayoffsof getting informedareboundedby Π . , that is, in the lower
bound ’spivotprobabilityapproacheszero,andintheupperbounditequalsoneand 1.
(Existence)Weuse IntermediateValueTheorem to show that ∗ ∈ 0,1 , , alwaysexists
for 0.Note that 1 and 1 are continuous for ∈ 0,1 . Then, as lim → 1
∞ 1 ∈ , lim → 1 ∞,bothsidesof 1 and thusbothsidesof logit
equilibrium conditions 1 and 2 intersect at least once,which proves that for 0 there always
existsa ∗ ∈ 0,1 inbothvotingmodes.30
29Weassumewithoutlossofgeneralitythatanindifferentindividual choosestogetinformed.30ForBNE( 0), itisreadilyverifiedthat ∗ 0[ 1]ifthesignalcostistoohigh[low]sothatbothsidesof 1 donotintersectorintersectat 1.Moreover, ∗ ∈ 0,1 forintermediatecostsandbothsidesintersectatleastonce.
34
(Uniqueness) The right panel of Figure A1 shows 1 for 0.06 and 1 for our
experimentaltreatmentsandparameters.Ascanbeseen, 1 hasconvexandconcaveproperties
with voluntary and compulsory voting, respectively, which for some parameter constellations yields
multipleequilibria(thoughnot forourexperimentalparameters). Inparticular,multiplicitycanarises
when 1 is flat (i.e., is small).31 To keep things simple, we focus on unique equilibria by
assumingforbothvotingmodesthatthedegreeofnoiseislargeenough,or 0,where gives
therespectivelowerboundthatensuresuniqueness.Notethatcomputationsshowthat fallsbelow
thetypicalestimateof ̂ inbinarychoiceexperimentssimilartoours.
Next,wederivecomparativestaticspredictionsforunique ∗.
(Groupsize)Wemustshowthatincreasingthegroupsizefrom2 1to2 3decreases ∗ in
theregionoftheparameterspaceofuniqueequilibrium.Forbothvotingmodes, 1 isconstantin
sothatitsufficestoprovethat 1 | 1 | for ∈ 0,1 .If 1 strictlydecreases
with for all ∈ 0,1 , then ∗ must indeeddecrease since 1 is constant in and strictly
increasingin ∈ 0,1 .
Withvoluntaryvoting, ,theterms and on 1 areconstantinboth and .Then,
toseethat 1 decreaseswith ,itsufficestoshowfor
, ∗ , ,22
12
1 2
that 1 0 for all ∈ 0,1 . First, for individual ’s pivot probability, if 2 others are
informed, we have ,∗ , , 1 ,
∗ , , 1 1 ⇔
1 , since 1 for ∈ , 1 and ⇔ 2 2 0 for
0, … , 1. Second, the cdf of , , , 1 first‐order stochastically
dominatesthecdfof , , 1, 1 for 0. . . 1and ∈ 0,1 .Thus,
combing both results, 2 and hence 1 decrease for ∈ 0,1 if the group increases by two
individuals,sothat ∗ ∗ 1 withvoluntaryvoting.
31 For example, for 0.001 and 0.75, we compute ∗ ∈ 0.541, 0.932, 0.99999 for 1 and 0.08, and ∗ ∈0.00002, 0.035, 0.225 for 23and 0.04.
35
Withcompulsoryvoting, , the term on 2 is constant in .Then, to see that 2
decreaseswith ,itsufficestoshowfor
, ∗ , ,2
1 1 3
11 1
that 1 0forall ∈ 0,1 .First,thefractionofpivotaleventsoutofallpossibleevents
decreaseswith ,thatis 2 2 21 for ∈ 0,1 ,whichworkstodecrease 3 .Second,
let’sdenote ∗ , , , , ≡ 1 , ≡ ,andrewrite ∗ , , , ≡
as ∗ , , 0 0if 0and ∗ , , 0
0if 0,with– 1 1(i.e.,theBayesianupdatedvaluedepends
on , but not on the particular levels of and that determine ). Then, for ∈ , 1 and all
, 0, … , and , … , wehave
, , 1, , , , ;
, , 1, 1, 1 1 , , ;
, , 1, 1, 1 , , 1 ;
, , 1, , 1 1 1 , , 1 ,
where the inequalities always hold because 1 , 1 1 , 2 1 , and
2 1 1 ,respectively.Third,thecdfof 1 first‐orderstochastically
dominatesthecdfof 1 for 0, … , ,with , and ∈ 0,1 .Thus,combing
allthreeresults, 3 andhence 2 decreasefor ∈ 0,1 ifthegroupincreasesbytwoindividuals,
sothat ∗ ∗ 1 withcompulsoryvoting.
(Signalcost)Anincreasein 0decreasestheunique ∗ bythefollowingargument: 1
is constant in ,while 1 0 for ∈ 0,1 . Since 1 strictly increaseswith ∈ 0,1 ,
bothsidesof 1 mustintersectfurthertotheleftifcincreases,sothat ∗ decreaseswith .
(Degreeofnoise)Anincreasein 0, , moves ∗ 0.5closertoone‐halfbythe
followingargument:increasing yieldsasteeper 1 ,whichalwaysgoesthrough 1 |
36
0 (where it is independent of ) and 1 is constant . Thismeans that both sides of 1 must
intersectatanewequilibrium ∗ ′ ∈ ∗ , if ∗ 0.5and ∗ ′ ∈ , ∗ if ∗ 0.5.
Thus, ∗ 0.5movesclosertoone‐halfif increases.∎
Probabilityofacorrectgroupdecision
Ifthetruestateis ,theneachinformedindividualhasmorelikelya ‐thana– ‐signal,where ,
and .Withvoluntaryvoting,theprobabilityofacorrectgroupdecisionisthengivenby:
, | , ∗ , ,2 1
1
1
12
2 12
1 2
1 . 4
Thefirsttermconsidersoutrightwinsofalternative (i.e.,therearemorevotesfor than outof
informed sincere votes, , and the remaining 2 1 individuals abstain), and the second
term considers ties (i.e., there are informed sincere votes for each alternative, and 2 1 2
abstentions),which are broken by a coin flip.Moreover,with compulsory voting, the probability of a
correctgroupdecisionsisgivenby:
, | , ∗ , ,2 1 2 1
5
1 1 ,
where denotesthetotalnumberofj‐votes,and [ ]thenumberofinformedsincerej[ ]‐votes.
Alternativejwinsoutrightiftherearemoreinformedsincereanduninformedrandomvotesforjthan
, or 1.Unlike in 4 ,there isno second term in 5 because tiesare impossible inodd‐
sizedgroupswithcompulsoryvoting.
ProofofProposition3
If goestoinfinity,thelogitequilibriumcondition 1 gets
lim→
1lim→
Π , , ∗ , , , , 6
37
, , where denotes the limiting equilibrium signal acquisition probability. First, note that
6 lim → and lim → 0 in Π , . on 6 . For
0,whichimplies 0,if goestoinfinity,thenindividual ’spivotprobabilityapproacheszero,
sothatlim → Π , , ∗ , , , for , (i.e.,aslim → 1 0for ∈ 0,1
and 0, … , and,forthecompulsorymode, . 1).Therefore, 6 gets
1⇔
1⁄ ⇔ ∗ , , 0
11 ⁄ 0 7
for , , asstated in theproposition.Next, for 0and ∈ , 1 , if increaseswithoutbound,
then the difference in expected numbers of informed votes for the correct minus the incorrect
alternative is lim→
2 1 2 1 1 lim→
2 1 2 1 ∞, the expected
numberofuninformedrandomvotesequalszerowithvoluntaryvoting(i.e.,theuninformedallabstain),
and lim→
2 1 1 ∞with compulsory voting. Thus, the uninformedby themselves create a
suretieinthevoluntarymodeand,bythelawoflargenumbers,theyalmostsurelycreateatieiftheir
numberisevenandaone‐voteleadforeitheralternativeiftheirnumberisodd.Butthesetiesandone‐
vote leadsdonotmatter, as thereare infinitelymorevotes for the correct than incorrect alternative.
Therefore,for 0and ∈ , 1 ,largegroupsmakeacorrectgroupdecisionalmostsurely.Finally,for
0(i.e.,BNE),wehave 6 0.If 0 0,thenlim → Π , , ∗ , , , ,which
is a contradiction for 0, since individual would choose not to buy a signal, 0 0!
Therefore,inequilibriumitmustbethat ∗ , , 0 0(“universalrationalignorance”),andthus
thegroupdecisioniscorrecthalfofthetime,whichcompletesourproofofProposition3.∎
38
FiguresA2:Groupperformances
FigureA2.1:Expectedinformationpool
FigureA2.2:Probabilityofacorrectgroupdecision
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 1 10 100
μ (degree of noise)
Lo
git
-QR
E e
xp
ec
ted
nu
mb
er
of
ac
qu
ire
d s
ign
als
pe
r g
rou
p
Voluntary voting-3 Voluntary voting-7 Compulsory voting-3 Compulsory voting-7
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 1 10 100
μ (degree of noise)
Pro
ba
bili
ty o
f a
co
rre
ct
gro
up
de
cis
ion
Voluntary voting-3 Voluntary voting-7 Compulsory voting-3 Compulsory voting-7
Social optimum-3 Social optimum-7 Delegate-3 and -7
39
FigureA2.3:Expectedefficiency
Note:Forour logitequilibriumsignalacquisitionprobabilities(Figure1), the linesdepictpertreatment
the expected number of acquired signals per group (Figure A2.1), the probability of a correct group
decision(A2.2),andexpectedefficiencyperindividual(i.e.,groupefficiencydividedbygroupsize;A2.3).
In social optimum, as 1, the expectednumberof acquired signalsper groupequals the respective
groupsize (notshown inA2.1),andwithasinglenoise‐freedelegate,as 1, itequalsone(alsonot
shown). Theprobabilities of a correct groupdecision andaverage expected efficiencyper individual in
socialoptimum(forasingledelegate)areshownascircles(triangles)inA2.2andA2.3,respectively.
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 1 10 100
μ (degree of noise)
Ex
pe
cte
d a
ve
rag
e in
div
idu
al
eff
icie
nc
y
Voluntary voting-3 Voluntary voting-7 Compulsory voting-3 Compulsory voting-7
Social optimum-3 Social optimum-7 Delegate-3 Delegate-7
40
TableA1:BayesianNashequilibrium
BNE predictions ( ) Voluntaryvoting Compulsoryvoting
Groupsize
∗ Efficiency ∗ Efficiency
1 1 .750 .625 1 .750 .6253 .321 .675 .635 .615
0.723.500
.647
.5005 .181 .662 .639 0 .500 .5007 .125 .657 .641 0 .500 .5009 .096 .654 .642 0 .500 .50011 .078 .652 .642 0 .500 .50013 .065 .651 .642 0 .500 .50015 .056 .650 .643 0 .500 .50017 .049 .649 .643 0 .500 .500
TableA2:Logitequilibrium
Logitpredictions ( . ) Voluntaryvoting Compulsoryvoting
Groupsize
∗ Efficiency ∗ Efficiency
1 .889 .722 .611 .889 .722 .6113 .409 .705 .654 .517 .690 .6255 .322 .732 .692 .408 .686 .6357 .277 .754 .719 .355 .688 .6449 .250 .771 .740 .323 .692 .65111 .230 .786 .757 .301 .696 .65813 .216 .799 .772 .284 .700 .66515 .205 .810 .784 .272 .705 .67151 .140 .914 .896 .193 .756 .732
41
FigureA3:Fractionsofinformedvoting,abstention,anduninformedvotingoverperiods
Bothparts Part1 Part2
Voluntaryvoting‐3
Voluntaryvoting‐7
Compulsoryvoting‐3
Compulsoryvoting‐7
Periods
0.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
0.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
0.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
0.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 300.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30