nolongerdependsontheirvaleneparameters. Thequotient
a(c)a(c ′ )/a(c, c ′ )
an be interpreted as a measure of orrelation between the andidates. The
more strongly the eletorate of two andidates overlap, the smaller will be
the distane between them.
Obtaining the Spatial Representation
The key for obtainingaspatial representation of the alternatives is
Multidi-mensionalSaling and PrinipalComponent Analysis. Let
y c be k
points in
R k entered at zero. A priori, one an assume k
tobe equal tothe number
k
tobe equal tothe number10
Unfortunately, this formula was misprinted in Laslier (2006). This is the orreted
version.
of available alternatives, sine
k
points in a Eulidean spae of a higherdi-mensionan alwaysbeembeddedin
R k. Theinertia ofy = (y) c∈C is dened
as the sum of the squareddistanes,
I(y) = X
c,c ′ ∈ C
|| y c − y c ′ || 2 .
Let
y E denotetheprojetionofthesystem ofpointsy
onalinearsubspaeE
of
R k. I (y E )
istheinertiaexplainedbyE
and theratioI(y E )/I(y)
measures
the qualityof the representation of
y
byy E. The goalis nowtonda
lower-dimensional representation of
y
suh that the explainedinertia ismaximal.Let
D
bethe(k × k)
symmetrimatrixofsquareddistanes,i.e.D 12isthe
squared distane between alternative 1 and 2 alulated aording to (2.3).
The representation isbased ona matrix
Γ
derived fromD
, alled matrix ofinertia, whih is dened as
Γ = (D · J + J · D − J · D · J − D)/2
(2.4)where
J
isa(k × k)
matrix with allentries equalto 1. Letµ 1 , ..., µ k be the
eigenvalues of
Γ
, ordered from largest to smallest, and letν 1 , ..., ν k be the
orresponding normed eigenvetors. The best projetion of the initialpoint
system into a linear,
i
-dimensional subspaeE i of R k ontains the enter,
P
P
c y c
, and is spanned by the rsti
eigenvetors. Under the speiations above,theperentageofexplainedinertiainthei
-dimensionalrepresentation is given byThe quality of representation for eah single point an be measured by the
osine of the projetion, whih is given by
r(c, i) = || y E i ||
|| y || .
(2.5)That is, the loser
r(c, i)
is to1, the better andidatec
isrepresented.As mentioned above, the model leaves unspeied the poliy saliene
parameter
α
, and needs tobealibrated by nding apropervalue ofα
, inthe following sense. First,
α
needs to be suh that the estimated distanesare non-negative. From Equation2.3, we obtainalowerbound on
α
. Forallpairwise omparisonsof alternatives, it must be that 11
2.6 Appendix: Auray of the Spatial
Repre-sentations
The spatial representation relies on the eigenvetors and eigenvalues of the
matrixofinertiagivenin(2.4),whihisderivedfromtheestimateddistanes.
The positions of the parties along the rst axis of the representation are
given by
√ µ 1 ∗ ν 1, (see Laslier, 1996) where µ 1 is the largest eigenvalue
and
ν 1 is the orresponding eigenvetor. Only the positive eigenvalues and their orresponding eigenvetors an be used for the spatial representation
of the parties, in the sense that the dimensionality of the representation is
onstrainedby thenumberofpositiveeigenvalues. The smallerweset
α
,themorelikelyitisthatweobtainnegativeeigenvaluesandhene,themorelikely
it is that weare not able to nd agoodrepresentation inthe
k
-dimensional spae. An aurate Eulidean representation is possible for larger values ofα
. However, inreased auray of the representing point system omes at a ost as the fration of explained inertia by the rst eigenvalues dereaseswith larger values of
α
. Hene, a reliable, low-dimensional representation requires a `proper' alibrationofα
.Party Vote, Messel
Let us start with the ase of the Party Vote in Messel. If
α = 3.69
, whihis the (rounded) lower bound derived from (2.3) in order for all estimated
11
Duetothemisprintmentionedpreviously,theversionofthisformulagivenin Laslier
(2006) does not allow for a omputation of the threshold for
α
. This is the orretedversion.
dimensional representation, Messel,Party Vote.
Parties SPD CDU GRE FDP Left APP FAM REP NPD
SPD 0.00 0.21 0.01 0.40 0.03 1.72 2.97 1.32 0.29
CDU 0.21 0.00 0.16 0.03 0.09 0.62 1.40 0.74 0.03
GRE 0.01 0.16 0.00 0.33 0.01 1.67 2.81 1.18 0.22
FDP 0.40 0.03 0.33 0.00 0.21 0.39 1.02 0.44 0.00
Left 0.03 0.09 0.01 0.21 0.00 1.49 2.41 1.12 0.20
APP 1.72 0.62 1.67 0.39 1.49 0.00 0.20 0.00 0.42
FAM 2.97 1.40 2.81 1.02 2.41 0.20 0.00 0.17 1.29
REP 1.32 0.74 1.18 0.44 1.12 0.00 0.17 0.00 3.39
NPD 0.29 0.03 0.22 0.00 0.20 0.42 1.29 3.39 0.00
SPD= Soial Demorati Party, CDU= Christian Demorati Union, GRE=
Greens, FDP= Liberal Demorati Party, Left= The Left, APP= Animal Pr
ote-tionParty, FAM= FamilyParty,REP= Republians, NPD=NationalDemorati
Party.
distanes to be positive, the rst seven eigenvalues are real and positive.
The distane aording to equation (2.3) between REP and NPD is 0.0058.
Given the orresponding set of eigenvalues and vetors, the distane in the
seven-dimensional representation would be 0.11536. Although we an then
explain about 65
%
of the inertia with a projetion on three dimensions, the pointsystem thatweprojetisaninaurateimageoftheoriginaldistanes.In order to provide anaurate piture with enough explanatory power,
we hose
α = 4
. The rst eight eigenvalues are then real and positive and the eight-dimensionalrepresentation seems reliable. Inorder toquantify theauray ofthe representation,we onsider perentual dierenes asfollows.
Let
d 1 be the distane between two parties aording to equation (2.3) and
let
d 2 be the distane we observe in the eight-dimensional representation.
The perentual dierene isgiven by
(d 2 − d 1 )/d 1.
With our hoie of
α = 4
and an eight-dimensional representation, the perentual dierene between the two distanes is never larger than 3.39%Cosines 1 dim 2dim 3 dim 4 dim
SPD 66.83 67.45 80.09 97.44
CDU 40.59 75.58 76.25 78.98
GRE 70.42 73.28 77.65 81.43
FDP 20.47 69.81 72.63 83.91
Left 58.16 74.89 81.29 89.97
APP 15.89 29.61 73.20 73.20
FAM 2.19 30.97 53.20 64.56
REP 76.88 84.17 88.51 88.55
NPD 91.87 97.42 98.03 98.03
SPD= Soial Demorati Party, CDU= Christian Demorati Union, GRE=
Greens, FDP= Liberal Demorati Party, Left= The Left, APP= Animal Pr
ote-tionParty, FAM= FamilyParty,REP= Republians, NPD=NationalDemorati
Party.
(see Table 2.5). In fat, for most of the pairwise omparisons, it is stritly
smaller than 1%. Further, the inertia explained by the lower-dimensional
projetion that we represent is relatively high. With three dimensions, we
explain about 63% of the inertia.
12
The osines of the projetion omputed
aording to(2.5) indiatethat most of the partiesare well-represented (see
Table 2.6). The only exeption is the Family Party, whih would require
more dimensions.
Party Vote, Konstanz
WeproeedanalogouslyforthePartyVoteinKonstanz. Equation(2.3)
deliv-ersaboundof
α ≥ 3.71
forallmodeldistanestobepositive. Withα = 3.85
,the rst ten eigenvalues are real and positive. The largest perentual
dier-ene between the model distane and the ten-dimensional representation is
12
Onedimension: 34.38%. Twodimensions: 51.20%Three dimensions: 62.89%. Four
dimensions: 73.34%.
then 3.05%, for the CDU and the Violet Party (see Table 2.7). All other
dierenes are stritly smaller, hene the overall ten-dimensional
represen-tation is an aurate piture of the model distanes. As mentioned in the
text, with three dimensionswe onlyexplain about 56% ofthe inertia.
13
The
osinesof projetionforthe Greens,the PirateParty,the AnimalProtetion
Party and the ödp are rather low (see Table 2.8). The Pirate Party and the
ödp require even more than ve dimensions.
Table 2.7: Perentual dierene of the distanes between the model and the
ten-dimensional representation, Konstanz, Party Vote.
Parties GRE SPD CDU FDP PIR Left APP ödp VIO REP NPD
GRE 0.00 0.20 0.30 0.21 0.45 0.12 0.61 1.06 2.01 0.51 1.03
SPD 0.20 0.00 0.02 0.00 1.16 0.60 1.44 2.02 2.97 1.13 1.78
CDU 0.30 0.02 0.00 0.01 1.29 0.57 1.64 2.37 3.05 2.15 2.41
FDP 0.21 0.00 0.01 0.00 1.20 0.49 1.45 1.95 2.49 1.83 2.53
PIR 0.45 1.16 1.29 1.20 0.00 0.12 0.01 0.12 0.57 0.04 0.27
Left 0.12 0.60 0.57 0.49 0.12 0.00 0.21 0.44 1.23 0.29 0.80
APP 0.61 1.44 1.64 1.45 0.01 0.21 0.00 0.07 0.48 0.01 0.13
ödp 1.06 2.02 2.37 1.95 0.12 0.44 0.07 0.00 0.20 0.02 0.02
VIO 2.01 2.97 3.05 2.49 0.57 1.23 0.48 0.20 0.00 0.37 0.09
REP 0.51 1.13 2.15 1.83 0.04 0.29 0.01 0.02 0.37 0.00 1.48
NPD 1.03 1.78 2.41 2.53 0.27 0.80 0.13 0.02 0.09 1.48 0.00
GRE= Greens, SPD= Soial Demorati Party, CDU= Christian Demorati
Union, FDP= Liberal Demorati Party, PIR= The Pirates, Left= The Left,
APP= Animal Protetion Party, ödp= Eologi Demorati Party, VIO=The
Vi-olets, REP= Republians, NPD=National Demorati Party.
Candidate Vote, Messel
Foralldistanestobepositivethe poliysalieneparameterhastobelarger
thanorequalto3.84. AnalogouslytothedisussionforPartyVoteinMessel,
13
Onedimension: 23.58%. Twodimensions: 43.95%Three dimensions: 56.23%. Four
dimensions: 65.67%.
Cosines 1 dim 2dim 3 dim 4 dim
GRE 55.48 61.70 61.88 85.83
SPD 66.18 68.97 73.07 93.77
CDU 29.49 79.14 81.63 81.97
FDP 23.14 77.54 80.71 81.63
PIR 1.39 19.78 41.87 56.78
Left 9.21 55.21 75.42 76.96
APP 17.08 30.28 60.50 60.89
ödp 2.77 18.66 69.74 69.98
VIO 41.15 63.01 74.79 74.82
REP 87.36 94.47 95.20 96.03
NPD 94.13 94.40 95.65 95.68
GRE= Greens, SPD= Soial Demorati Party, CDU= Christian Demorati
Union, FDP= Liberal Demorati Party, PIR= The Pirates, Left= The Left,
APP= Animal Protetion Party, ödp= Eologi Demorati Party, VIO=The
Vi-olets, REP= Republians, NPD=National Demorati Party.
weareledtoadoptahighervalue. With
α = 4.45
,therstseven eigenvalues of the analysis are real and positive. The piture of the seven-dimensionalrepresentationisanaurateimageofthedistanes alulatedbythe model.
Thediereneformostof theomparisonsisbelow1% (seeTable2.9). With
three dimensions, we explain about 69% of the inertia.
14
Further, all eight
andidates are well-represented with only three dimensions, as indiated by
the osines of the projetion (see Table 2.10).
Candidate Vote, Konstanz
The threshold valuefor
α
suh that allmodeldistanes are positiveis equalto 2.55. With
α = 3.35
, the rst four eigenvalues are real and positive.As indiated by Table 2.11, inthis ase four dimensions are already enough
14
Onedimension: 32.43%. Twodimensions: 54.18%Three dimensions: 68.76%. Four
dimensions: 80.59%.
dimensional representation, Messel,Candidate Vote.
Parties SPD CDU GRE FDP Left FRV REP NPD
SPD 0.00 0.11 0.01 0.23 0.45 1.42 2.44 0.48
CDU 0.11 0.00 0.07 0.03 0.05 0.70 2.07 0.26
GRE 0.01 0.07 0.00 0.18 0.37 1.39 2.20 0.39
FDP 0.23 0.03 0.18 0.00 0.01 0.54 1.60 0.10
Left 0.45 0.05 0.37 0.01 0.00 0.41 1.31 0.03
FRV 1.42 0.70 1.39 0.54 0.41 0.00 0.40 0.13
REP 2.44 2.07 2.20 1.60 1.31 0.40 0.00 4.43
NPD 0.48 0.26 0.39 0.10 0.03 0.13 4.43 0.00
SPD= Soial Demorati Party, CDU= Christian Demorati Union, GRE=
Greens, FDP= Liberal Demorati Party, Left= The Left, FRV= Free Voters,
REP= Republians, NPD= NationalDemorati Party.
to nd aurate representation of the modeldistanes. Further, with three
dimensions,weobtainanexplainedinertiaof91%.
15
Allandidatesare
well-represented with only three dimensions, as indiated by the osines of the
projetion (see Table 2.12).
15
Onedimension: 44.34%. Twodimensions: 77.74%Three dimensions: 90.75%. Four
dimensions: 100.00%.
Table 2.10: Cosinesof projetions, Messel,Candidates.
Cosines 1 dim 2dim 3 dim 4 dim
SPD 65.78 66.89 84.09 95.06
CDU 46.44 70.77 76.08 76.91
GRE 71.72 72.61 72.82 96.24
FDP 23.25 77.79 80.33 81.57
Left 44.03 81.82 81.94 81.94
FRV 18.49 18.76 87.28 97.51
REP 70.91 87.13 88.72 88.83
NPD 79.60 90.08 95.47 97.89
SPD= Soial Demorati Party, CDU= Christian Demorati Union, GRE=
Greens, FDP= Liberal Demorati Party, Left= The Left, FRV= Free Voters,
REP= Republians, NPD= NationalDemorati Party.
Figure2.3: Three-dimensionalprojetion oftheseven-dimensionalandidate
rep-resentation, Messel.
(a) Axes1and 2 (b) Axes1and 3
Figure 2.4: Three-dimensional projetion of the four-dimensional andidate
rep-resentation, Konstanz.
(a) Axes1and 2 (b) Axes1and 3
Table 2.11: Perentual dierenes of the distanes between the model and the
four-dimensional representation, Konstanz, Candidate Vote.
Parties GRE SPD CDU FDP Left NPD
GRE 0.00 0.37 2.05 0.22 0.70 0.33
SPD 0.37 0.00 1.02 0.30 0.66 0.04
CDU 2.05 1.02 0.00 2.82 0.97 0.89
FDP 0.22 0.30 2.82 0.00 0.02 0.92
Left 0.70 0.66 0.97 0.02 0.00 2.06
NPD 0.33 0.04 0.89 0.92 2.06 0.00
GRE= Greens, SPD= Soial Demorati Party, CDU= Christian Demorati
Union, FDP= Liberal Demorati Party, Left= The Left, NPD=National
Demo-rati Party.
Table 2.12: Cosinesof projetions, Konstanz,Candidates.
Cosines 1 dim 2 dim 3 dim 4dim
GRE 32.48 64.46 97.41 100.00
SPD 8.78 82.11 99.20 100.00
CDU 89.73 90.56 91.05 100.00
FDP 72.11 88.07 88.59 100.00
Left 95.86 96.10 96.54 100.00
NPD 39.08 99.65 99.76 100.00
GRE= Greens, SPD= Soial Demorati Party, CDU= Christian Demorati
Union, FDP= Liberal Demorati Party, Left= The Left, NPD=National
Demo-rati Party.
Preferene Aggregation Under
Unertainty
3.1 Introdution
Attheveryheartofeverydividedmajorityproblemliesaoordination
prob-lem. Anabsolutemajorityofvotersisuniedintheirdislikeofonepartiular
andidate they jointly onsider to be inferior. However, their support is
di-vided between two(or more)otherandidates. Theyhavetoombineeorts
and onentrate their supportbehind one andidate. Otherwise, their worst
option, baked up by a weak majority him/herself, will emerge as the
win-ner of the eletion. The reognition of these kindof problems follows a long
tradition in the study of eletoral systems and an at least be dated bak
to de Borda (1781). He used a divided majority problem to illustrate the
drawbaks of the voting method the Frenh Aademy of Sienes employed
for the seletion proess of new board members at that time. The method
was Plurality Voting (or PV) and de Bordaexpressed his onerns that the
method an selet new board members a majority onsiders to be inferior,
even if everybody votes sinerely (inaordane with their preferene).
Inthe meantime, real world instanes ofdivided majorityproblems have
been observed, espeially in eletions based on or involving PV. `One man,
one vote' systems make it partiularly easy for suh `paradoxial' results
to appear. One infamous example ourred during the Frenh presidential
eletion in2002, where right-extremistJean-Marie LePen, the leader of the
extreme Right party Front National, took seond plae in the rst round of
the eletion although 82% of the voters in the seond round voted against
him. Le-Pen did not win, but him entering the seond round aused
in-ternational mass media overage shedding bad light onthe Frenh politial
landsape.
1
Another example is the 1998 governor eletion in Minnesota
where former professional wrestler Jesse `The Body' Ventura won although
64% of the voters preferred any of the twoother majorandidates.
1
The Frenh president is eleted via a plurality vote run-o system. If a andidate
reeivesanabsolutemajorityofvotesintherstround,he/sheiseleted. Ifnoandidate
is ableto reeivean absolutemajorityof votes, the twoandidateswith the mostvotes
meetin aseond round. Inthe rstround, JaquesChirareeived19.88%of thevotes,
Jean Marie Le Pen 16.86%, and Linoel Jospin 16.18%. Le Pen, together with Chira
enteredtheruno. Chirawontheseondround withmorethan82%ofthevotes.
The very fat that a andidate an win an eletion despite being
on-sidered inferior by an absolutemajority poses a major and grave threat. It
an be argued that suh results ompromise the wishes of the voters. They
derail the demorati proess by undermining the legitimay of appointed
leaders. The ulprit inase of PV eletionsis easilyidentied. The method
respets alleged maxima only and disregards any other type of preferene
information. Any eletion in whih the ombined vote share of andidates
promoting similar poliies exeeds that of the winner raises suspiion and
doubt about PV aurately reeting voters' wishes. Eetive attempts in
replaing the `standard' eletoralsystem must addressthis ore diulty.
Multi-votesmethodspossessthepotentialtosuessfullyreplaePV.
Ap-proval Voting or AV and the Borda Count or BC an be onsidered to be
among the most promisingof suh methods both in terms of sienti work
devoted totheir study and the many positive(desirable) properties they
ex-hibit. Both methods give voters additional opportunities to express their
wishes and desires. AV wasrst analytially desribed by Brams and
Fish-burn (1978). It merely requires voters to reveal whih andidates they nd
`aeptable', i.e.,eah voter onlyneeds tomarkthe namesof the andidates
he approves of. Arguments have been put forward inthe literature that the
methodprovides anaurate reetion of voters' wishes and is not
vulnera-bletovotermanipulation(seeBramsandFishburn,1978;Fishburn,1978a,b;
Brams andFishburn,2005; Wolitzky,2009). In ontrast, BC requiresvoters
toprovideaompleterankingof andidates. Therst mathematialtreatise
on BC dates bak to de Borda (1781), however, the history of the method
an be traed bak to as early as the 15th entury in the work of Niolas
Cusanus.
2
Asshown bySaari(1994),BCminimizestheamountofparadoxes
withinthe set ofpositionalvotingmethods(inludingAV and PV).Further,
the BC-winner(thealternativewinningBCeletionswhenvotersusesinere
strategies) is the alternativeranked highestonaverage in the eletorate.
Reformresolution proposing a hange in the eletoral system in favor of
AV and BC must take intoaount that both methods have identied
de-2
Seee.g.MLean(1990)foranearlyhistorialaountofBordaandCondoret
prin-iples.
ienies too. The main theoretial ritiques against AV in the literature is
the multipliity of potential outomes for the same set of preferenes (see
Saari and Van Newenhizen, 1988b,a). AV and BC respet neither the
Con-doret nor the majority riterion: Condoret-Winners and alternatives that
are preferredto any otheralternativeby anabsolutemajority donot
nees-sarily win.
3
In terms of strategi inentives, BC is often ritiized as being
easy to manipulate and vulnerable to strategi misrepresentation of
prefer-enes. It is relatively easy for voters to prevent an undesirable outome by
insinerely raising (`ompromising') or lowering (`burying') the positions of
alternatives. Eletoralsystemreformsmustguaranteethatthesystemavoids
aws of the past system and that it does not introdue new problems of a
dierent kind.
In order to fully understand the impliations of suh alternative
meth-odsfor soiety,it is essentialto olletempirial data ontheir performane.
Field experiments have been a rst step in this diretion providing us with
invaluabledata and with lear evidene ontheir atual feasibility(see
Alós-Ferrer and Grani¢,2012b;Baujardand Igersheim,2009;Laslierand Vander
Straeten, 2008). Thereare,however, anumberoflimitationsassoiatedwith
eld experiments. First, due to legal onerns (ensuring voter anonymity in
atual eletions) aswell asmethodologialonsiderations, one isnot able to
fully identify the partiipants' atual preferenes. Seond, self-seletion
bi-ases may our, whihannot fully be aounted for. Hene, without
know-ing either the preferenes of the partiipants or whether the samples are
truly unbiased, ertain properties of voting methods annot be tested. As
a omplement to eld experimentation, the ontrolled laboratory
environ-ment typial of experimental eonomis allows usto indue any preferenes
overandidatesby modelingthe partiipants'expetedpayoonditionalon
the eletionoutomes. Hene, we an study theoretial properties of voting
methodsin the laboratory that annotbe tested inthe eld.
This paper reports on the results of series of experimental (laboratory)
4-alternative eletions. The experimental design is based on Forsythe,
My-3
ACondoret-Winner(Loser)isanalternativeableto beat(isbeatenby)everyother
alternativeinapairwiseomparison.
erson, Rietz, and Weber (1993) and Forsythe, Rietz, Myerson, and Weber
(1996) andthe indued preferene proleorresponds toanadapted divided
majority problem. We assess and ompare the Condoret-Eieny (how
often the Condoret-Winner is eleted) of three voting rules: AV, BC, and
PV. In addition,we independently vary the underlyinginformational
stru-ture fromafull-informationframeworkwherepartiipantsarefullyinformed
about the preferene struture in the whole eletorate to an
inomplete-information framework where partiipants only know their own payos and
the eletion histories and analyze its impat on the Condoret-Eieny of
the voting methods.
4
One feature of our design is the interation of the
partiipantsoverseveral roundsexpliitlyenablingthemtoaquire
familiar-ity with the voting method and allowing for learningeets. This repeated
environment iswell suited for analyzing the partiipants' adaptivebehavior
and the relevane of strategi onsiderations as it almost begs for strategi
behavior onpart of the voters.
We ontribute to the existing literature in the following ways. First, we
provide evidenethat multi-votes systems inrease the frequeny of
suess-ful oordination in divided majority settings in omparison to PV. A
de-parture from the full-informationframework to a more realisti
inomplete-informationframeworkaninreasethesuperiorityofthesemethodsoverPV.
Information imperfetions, in our ase inomplete-information,in
ombina-tion with the relatively low information provided by a PV ballot message,
an hange eletion results dramatially. Coordination on the
Condoret-Winner breaks down and almost never ours with inomplete-information
under PV.At the sametime, the underlyinginformationstruture has little
to no eet on the eletion outomes generated under AV and BC. Our
re-sults suggest that, in general, the responsiveness towards the informational
struture may serve as another important dimension to evaluate the merits
of voting methods.
Seond, the analysis of the individual voting behavior reveals interesting
4
For atheoretial aount on how information imperfetions an inuene the
Con-doret Eieny of voting methods see, e.g., Laslier (2009); Bouton and Castanheira
(2012).
patterns. In line with, e.g. Ballester and Rey-Biel (2009), we observe that
unertainty redues the amountofstrategivotingand inreases sinerity to
a great extent under PV. Interestingly, this is also true for AV, but not for
BC. We have lear evidene that, althoughAV and BC are roughly equally
BC. We have lear evidene that, althoughAV and BC are roughly equally