Divided Majority and Information Aggregation - Theory and Experiment

Divided Majority and Information Aggregation: Theory and Experiment Abstract This paper both theoretically and experimentally studies the properties of plurality and approval voting when the majority is divided as a result of information imperfections. The minority backs a third alternative, which the majority views as strictly inferior. The majority thus faces two problems: aggregating information and coordinating to defeat the minority candidate. Two types of equilibria coexist under plurality: either voters aggregate information, but this requires splitting their votes, or they coordinate but cannot aggregate information. With approval voting, expected welfare is strictly higher, because some voters multiple vote to achieve both goals at once. In the laboratory, we observe both types of equilibrium under plurality. Which one is selected depends on the size of the minority. Approval voting vastly outperforms plurality. Finally, subject behavior suggests the need to study asymmetric equilibria. 1 Introduction Elections are typically expected to achieve better-informed decisions than what a single individual could achieve alone (Condorcet 1785, Austen-Smith and Banks 1996, Feddersen and Pesendorfer 1997, 1998, etc). The rationale is that each single individual only has a fraction of the overall information held by the electorate. If each voter can convey her privately-held information through her ballot, voting results will reveal the aggregate information held by the electorate. However, this is a big “if”: in plurality, for instance, rational voters are typically expected to coordinate their ballots on only two alternatives, independently of the number of competing alternatives (Duverger’s Law). Therefore, unless the number of candidates is exactly two, information aggregation is jeopardized. This impossibility resonates with centuries of scholarly research on how to design an electoral system that would be better able to aggregate heterogeneous preferences and information in an e¢ cacious way (see e.g. Condorcet 1785, Borda 1781, Arrow 1951, Cox 1997, Myerson and Weber 1993, and Myerson 2002). Frustration with plurality is also apparent in civil society: a large number of activists lobby in favor of reforming the electoral system1 and many reform proposals have been o¢ cially introduced.2 One of the most supported alternatives to plurality is approval voting (AV).3 Yet a major hurdle stands in the way of any reform of the electoral system: the substantial lack of knowledge surrounding the capacity of AV (or other systems) to outperform plurality.4 We need a better understanding of the properties of each potential electoral system to identify and implement meaningful reforms. With this purpose in mind, this paper studies the properties of plurality and AV when voters are strategic but imperfectly informed. Our analysis features two main novelties: …rst, we study these systems both theoretically and experimentally. We focus on the case 1 See e.g. the Electoral Reform Society (www.electoral-reform.org.uk) and the Fair Vote Reforms initiative (www.fairvote.org). 2 Two examples are North Dakota in 1987, where a bill to enact approval voting in some statewide elections passed the Senate but not the House and, more recently, the U.K., which held a national referendum in 2011 on whether to replace plurality voting with alternative voting. 3 Under approval voting, voters can “approve of” as many candidates as they want, each approval counts as one vote and the candidate that obtains the largest number of votes wins (Weber 1977, 1995, Brams and Fishburn 1978, 1983, Laslier 2009). 4 For instance, the 2011 referendum in the U.K. arguably rejected the implementation of the proposed substitute, alternative voting, largely because of the uncertainty surrounding its capacity to outperform plurality. 2 in which a majority needs to both aggregate information and coordinate ballots to defeat a minority alternative, the Condorcet loser. Second, instead of focusing on the limiting properties of these systems when the electorate is arbitrarily large, we study them for any electorate size. This means that our conclusions are equally valid from committees to general elections. A …rst theoretical …nding is that, in plurality, the need to aggregate information produces an equilibrium in which voters vote informatively (that is, their ballot conveys information about their beliefs, e.g. because they vote sincerely), despite the need to coordinate against the minority. This equilibrium is not “knife edge”, and may rationalize the oft-observed pattern that strictly more than two candidates receive positive but different vote shares, despite the prediction of Duverger’s Law (Duverger 1963, Palfrey 1989, Myerson and Weber 1993, Cox 1997, and Fey 1997). When the minority is small, this equilibrium supports information aggregation, in the sense that the alternative with the largest expected vote share is the full information Condorcet winner. In contrast, when the minority is large, the alternative with the largest vote share is the Condorcet loser, in which case this equilibrium is highly ine¢ cient. This informative equilibrium may thus exist despite the fact that majority voters would bene…t from collectively deviating towards a two-alternative equilibrium. That is, Duverger’s Law may or may not be observed in equilibrium, depending on how majority voters coordinate their ballots. In our divided majority setup, we also …nd that AV can always produce strictly higher welfare than plurality: for any equilibrium in plurality, allowing voters to approve of more than one alternative produces a two-pronged bene…t: …rst, it reduces the threat posed by the minority alternative. Second, voters in the majority can better aggregate information. While characterizing the exact equilibrium in approval voting is not possible for any electorate size,5 we are able to formulate two substantiated conjectures: (i) the equilibrium is unique, and (ii) the equilibrium strategy is such that majority voters always approve of the candidate they deem best and sometimes also approve of the other majority candidate. These conjectures …nd support in one formal result and many numerical simulations. First, we prove that voters must adopt this voting pattern in any interior equilibrium (in 5 In contrast, in a related setup, Bouton and Castanheira (2012) fully characterize the equilibrium for arbitrarily large electorate sizes. The equilibrium is then unique and it implies that the full information Condorcet winner always has the largest expected vote share. 3 which voters play nondegenerate mixed strategies). Second, for all the parametric values we checked, the equilibrium was indeed unique. These theoretical results pose an interesting trade-o¤ between these two electoral mechanisms. On the one hand, AV is inherently more complex than plurality since it extends the set of actions that each voter can take.6 A risk exists then that actual voters make more mistakes at the time of voting, which could wash out the favorable theoretical properties of AV. On the other hand, our theoretical …ndings are that AV reduces the number of equilibria and therefore simpli…es strategic interactions amongst voters when they have imperfect information. In other words, AV should facilitate the voters’two-pronged goal of aggregating information and coordinating ballots to avoid a victory of the Condorcet loser. To assess the validity of these theoretical …ndings, we have run a series of laboratory experiments. Through these, we can evaluate the e¤ective performance of each system in a controlled environment. Equally important, these experiments allow us to bring new light to the debates on voter rationality: determining whether voters behave strategically and respond to incentives is a central issue in the quest for better political institutions. Our setup, which combines the need to (i) aggregate information, and (ii) coordinate ballots, is an ideal testing ground for such questions. First, each of these two problems produce di¤erent –often opposite–voting incentives. The literature has extensively studied each of these problems, but typically in isolation (see subsection 1.1). Combining the two allows us to test whether and in which proportion voters react to a change in incentives when we modify the relative value of coordinating ballots versus aggregating information. Second, studying multicandidate rather than two-candidate elections widens the set of electoral systems (and thus voter incentives) that can be analyzed. In our case, the predicted behavior of voters is substantially di¤erent between plurality and AV. The experiments reveal interesting patterns and support most of our theoretical predictions. We …rst study setups in which information is symmetric across states of nature. Under plurality, we observe the emergence of both types of equilibria: when the minority is su¢ ciently small, all groups stick to playing the informative equilibrium. By contrast, 6 With three alternatives, plurality o¤ers four possible actions: abstain, and vote for either one of the three alternatives. AV adds another four possible actions: three double approvals, and approving of all alternatives. Saari and Newenhizen (1988) argue that this may produce indeterminate outcomes, and Niemi (1984) argues that AV “begs voters to behave strategically”, in a highly elaborate manner. 4 when the minority is “large”, in the sense that the informative equilibrium leads to a high probability of victory for the Condorcet Loser, all groups eventually decide to give up information aggregation and coordinate their ballots on the same alternative, as predicted by Duverger’s Law. However, all groups begin by voting sincerely, and convergence to one of the two potential Duverger’s Law equilibria is quite slow. Our theoretical model identi…es a force explaining their behavior. Under AV, and in line with theoretical predictions, some voters double vote to increase the vote shares of both majority candidates. As predicted, the amount of double voting also increases with the size of the minority. However, the absolute level of double voting is lower than predicted. Comparing the two systems, we observe that voters make fewer strategic mistakes under AV than under plurality. Moreover, when the minority is large, voters need more time to reach equilibrium play in plurality than in AV. This suggests that voters can more easily handle the larger set of voting possibilities o¤ered by AV than the need to select an equilibrium under plurality. In contrast with the theory (which focuses on symmetric equilibria), individual behavior displays substantial heterogeneity among subjects in AV: many subjects always double vote, whereas many other subjects always single vote their signal. The observation that double-voting increases with the size of the minority is mainly driven by a switch in the relative number of subjects in each cluster. This pattern points to the need to extend the theory and consider equilibria in asymmetric strategies. Extending the model in this direction, we …nd that this type of behavior is indeed an equilibrium which performs particularly well in explaining the level of double-voting observed in the laboratory. We then turn to those treatments in which the quality of information varies across states of nature. In line with theoretical predictions, subjects eventually adjust their behavior to better aggregate information. In the case of plurality, the data provides further evidence to con…rm our theoretical prediction that three-candidate equilibria are a natural focal point when majority voters have common values. In the case of AV, the results are even stronger, in the sense that voters converge faster toward the theoretical prediction. Finally, we analyze the welfare properties of both electoral systems. A valuable feature of our common value setup among majority voters is that it allows us to make clear welfare 5 predictions: in equilibrium, majority voters’payo¤ should be strictly higher with AV than in plurality. This is exactly what we observe in all di¤erent treatments. 1.1 Related literature This paper relates to several strands of the literature on strategic voting and the comparison of electoral systems. We organize our discussion in three subsections: …rst, we review the literature on two-alternative elections in which voters have common values and private information about the quality of the two alternatives, i.e. there is an information aggregation problem. We then review the literature on multi-alternative elections with private valued voters, which focuses essentially on coordination problems. Lastly, we review the recent literature on multi-alternative elections with common values, in which our paper is rooted. This literature highlights the trade-o¤ that majority voters face between aggregating information on the one hand and coordinating to defeat a weak candidate on the other. In all subsections, we review both the theoretical and experimental literature (we constrain ourselves to laboratory experiments). 1.1.1 Two-Alternative Elections: Information Aggregation Theory The literature on the role of elections as a way of aggregating information dates back at least to Condorcet (1795).7 This desirable property of non-unanimous electoral systems is actually reinforced by the presence of strategic voters (Austen-Smith and Banks 1996, Feddersen and Pesendorfer 1996, 1997, 1999, Myerson 1998, and McMurray 2012).8;9 The robustness of this information aggregation property is discussed extensively in the literature. On the one hand, it is robust (to some extent) to the introduction of private values, i.e. all voters do not necessarily agree on the which alternative is best conditional on the state of nature (Gerardi 2000), costly voting (Krishna and Morgan 2011, 2012), and costly acquisition of information (Martinelli 2006, Oliveros 2011). On the other hand, it 7 See Piketty (1999) for a review of the information-aggregation approach to political institutions. Two exceptions are Feddersen and Pesendorfer (1998), who show the inferiority of the unanimity rule, and Mandler (2012) who studies the case in which the probabilities to receive each signal are themselves random. Austen-Smith and Feddersen (2006) have shown that this conclusion is robust to pre-voting deliberations (on this topic, see also Coughlan 2001, and Eraslan and Bond 2009 ). 9 Ladha et al. (1996) have identi…ed situations in which there exists an asymmetric equilibrium in which voters who receive the same signal behave di¤erently. This equilibrium leads to more e¢ cient information aggregation than the sincere voting equilibrium in which voters vote their signal. Our section 6.3.1 about asymmetric equilibria in AV builds on that idea. 8 6 is not robust to adversarial preferences (Fey and Kim 2007, Bhattacharya 2012), to voters preferring to vote for the winner (Callander 2008), and to voters being ambiguity averse (Ellis 2012). Experiments The main message of the experimental literature on information aggregation in two-candidate elections is that subjects vote strategically and adapt to the di¤erent rules and parametrizations used in the experiments (although not to the extend predicted by theory). Guarnaschelli, McKelvey and Palfrey (2000) experimentally compare behavior under majority and unanimity rule. They …nd that, in a setting where information is symmetric across states, subjects vote sincerely when using majority but frequently vote against their signal when the rule in place is unanimity. Ladha, Miller and Oppenheimer (1996), Bhattacharya, Du¤y and Kim (2012) and Bouton, Llorente-Saguer and Malherbe (2012) show that such strategic behavior is also present with majority rule in situations where information precision di¤ers across states.10 Finally, Battaglini, Morton and Palfrey (2008, 2012) show that uninformed voters strategically abstain in order to delegate the decision to the informed ones, and even compensate for partisan biases.11 1.1.2 Multi-Alternative Elections: Coordination Problems Theory Coordination problems are arguably the central issue in the literature on multialternative elections with private value voters. As shown repeatedly (see e.g. Myerson and Weber 1993, Cox 1997, Myerson 2000, 2002, Myatt 2007, Bouton 2012, Bouton and Gratton 2012), when a divided majority is facing a uni…ed minority block, electoral systems produce (i) bad equilibria, in which the minority candidate gets elected, and (ii) equilibrium multiplicity, which leaves elections open to focal manipulations and coordination failures. In a purely private values environment, designing an electoral system exempt from such problems has so far proved impossible.12 The two systems under consideration in this paper, i.e. plurality and AV, have been 10 Goeree and Yariv (2010) provide further evidence that voters adjust their behavior to the electoral rule and show that pre-vote deliberations signi…cantly diminishes institutional di¤erences and uniformly improves e¢ ciency. 11 This feature has been also tested with observational data (McMurray 2012). 12 Another interesting type of multi-alternative elections are those in which decisions on multiple binary issues are taken simultaneously, the so-called combinatorial voting. See Ahn and Oliveros (2012) for a thorough analysis. 7 analyzed in private values environments. For plurality, the most famous result is the so-called Duverger’s Law (Duverger 1963, Riker 1982, Palfrey 1989 and Cox 1997): “the simple-majority single-ballot system [the plurality electoral system] favors the two-party systems”. Duverger’s intuition is that voters have incentives to abandon their most preferred candidat if she has no chance to win the election to instead rally behind their most preferred serious candidate. In a setup including only strategic voters, Duverger’s Law implies that only two candidates can obtain a positive fraction of the votes. Perhaps surprisingly, there is a least one equilibrium under plurality in which a Condorcet winner obtains zero vote. Myerson and Weber (1993) prove the existence of a non-Duverger’s law equilibrium, in which three candidates obtain a positive fraction of the votes. Yet, as discussed by Fey (1997), the existence of such a non-Duverger’s law equilibrium relies on quite demanding conditions on the structure of information in the electorate. Moreover, it is not expectationally stable.13;14 In contrast, in our setup, we prove the existence of an expectationally stable equilibrium in which three candidates receive a substantial fraction of the votes. This strand of the literature also suggests that AV resists coordination problems better than many other electoral systems (Myerson and Weber 1993, Myerson 2000, 2002, Laslier 2010, Nunez 2010). In terms of voting behavior, the theoretical prediction is quite clear: there is no “hole” in the ballots cast by voters, i.e. voters identify a cuto¤-candidate an only approve of the candidates yielding a utility at least as high as that one. Experiments The experimental literature on multicandidate elections with private value voters is surprisingly small.15 The seminal papers of Forsythe et al (1993, 1996) are closest to our paper. They consider three-candidate elections in which a divided majority is opposed to a uni…ed minority. Voters are perfectly informed about the distribution of types in the electorate. Forsythe et al (1993) …nds that, in elections without polls or shared history, plurality rule frequently leads to a victory of the Condorcet loser. However, both polls and shared histories (i) decrease the frequency of such coordination failure among 13 This non-Duverger’s Law equilibrium might also be deemed conterintuitive since it requires the supporters of the strongest majority candidate mix between their most-preferred candidate and the other majority candidate. 14 See Myatt (2007) for a discussion of the existence of non-Duverger’s law equilibria in a setup with aggregate uncertainty. 15 See Rietz (2008), Laslier (2010) or Palfrey (2012) for more detailed reviews of that literature. 8 majority voters, and (ii) favor the emergence of Duverger’s law.16 Forsythe et al. (1996) analyze alternative voting procedures. They …nd that, under AV, it is easier for majority voters to overcome the coordination problem. Granic (2012) successfully replicates these …ndings and extends them to settings where subjects are imperfectly informed about the distribution of types in the electorate.17 1.1.3 Multi-Alternative Elections: Information Aggregation and Coordination Problems Theory Few papers analyzing multi-alternative elections with common valued voters. This strand of the literature is divided into two parts: analyses considering (i) setups with pure common values voters, and (ii) setups with both private and common values voters. McLennan (1998) shows that, in games of pure common values, any strategy that maximizes utility is an equilibrium. Ahn and Oliveros (2011a) exploit this to prove that in a pure common value setup, “the maximal equilibrium utility under approval voting is greater than or equal to the maximal equilibrium utility under plurality voting or under negative voting.” (p. 3).18 Analyses including both private and common values voters capture an essential tradeo¤ between two goals in multi-alternative elections: aggregating information on the one hand and coordinating against a weak candidate on the other.19 Goertz and Maniquet (2011) …nd a numerical example in which approval voting fails to aggregate information. That example crucially relies on a large fraction of the voters having assigning a zeroprobability to one state of nature. Bouton and Castanheira (2012) show that, in large elections, AV satis…es Full Information and Coordination Equivalence. That is, it solves both the information aggregation problem and the coordination problem at once. A nec16 Their …ndings are consistent with later evidence on other setups by Gerber et al. (1998), Bassi (2008) or Hizen et al (2010). 17 Van der Straeten et al. (2010) or Bassi (2008) also study AV experimentally although in substantially di¤erent settings. Van der Straeten et al. (2010) show that in a setting where alternatives can be ordered in a one dimensional policy line, voters are strategic and the Condorcet winner wins substantially more compared to plurality, runo¤ elections or single transferale vote. Bassi (2008) shows that strategic voting is substantial in AV, but less than with plurality and more than with Borda rule. 18 As mentioned above, another interesting type of multi-alternative elections are those in which decisions on multiple binary issues are taken simultaneously, the so-called combinatorial voting. Ahn and Oliveros (2011b) analyze such elections in a pure common value setup. 19 In such a setup, Piketty (2000) analyzes the incentives of voters to communicate between two di¤erent elections or between the two rounds of a majority runo¤ election (see also Castanheira 2003). See Martinelli (2002) for a more detailed analysis of information aggregation in three-candidate majority runo¤ elections. 9 essary and su¢ cient condition for this result is that su¢ ciently many voters may need additional information to identify their best alternative (i.e. they have a “doubt” that information can dispel). From a theory standpoint, the paper closest to ours is Bouton and Castanheira (2012). There are several important di¤erences between their model and ours. The most important is that instead of focusing on limiting properties when the electorate grows large, we consider electorates of any size. Relaxing this assumption does not come at no cost: we are indeed forced to focus on a more stylized structure of preferences. For instance, we cannot analyze situations in which majority voters have heterogenous preferences (and may not agree on which candidate is best when information is perfect). Moreover, we can neither provide a general proof of uniqueness, nor determine the equilibrium strategy in all cases. Yet, as highlighted in the introduction, we prove several novel results that we then bring to the laboratory. Experiments To the best of our knowledge, our paper is the …rst laboratory experiment which explores multi-alternative elections with common value voters. 2 A common value model We consider a voting game with an electorate of …xed and …nite size who must elect a policy P out of three possible alternatives, A; B and C. The electorate is split in two groups: n active voters who constitute a majority, and nC voters who constitute a minority. There are two states of nature: ! = fa; bg, which materialize with probabilities q (!) > 0. While these probabilities are common knowledge, the actual state of nature is not observed by the time of the election. Active voters’utility depends both on the policy outcome and on the state of nature: utility is high (U = V ) if A is elected and the state is a, or if B is elected and the state is b. It is intermediate (U = v 2 (0; V )) if A wins and the state is b or if B wins and the 10 state is a. Finally, utility is low (normalized to zero) if C is elected: U (P j!) = V if (P; !) = (A; a) or (B; b) = v if (P; !) = (A; b) or (B; a) (1) = 0 if P = C: For the sake of simplicity, minority voters are passive in the game: they always vote for C, which means that C receives nC ballots independently of the state of nature or the electoral system. This implies that active voters must cast at least nC ballots in favor of either A or B to avoid the victory of C. We focus on the interesting case in which C-voters represent a large minority: n 1 > nC > n=2: Hence, C is a Condorcet loser (it would lose against either A or B in a one-on-one contest), but it can win the election if active voters split their votes between A and B. Timing. Before the election (at time 0), nature chooses whether the state is a or b. At time 1, each voter receives a signal s 2 fsA ; sB g ; with conditional probabilities r (sj!) > 0 and r (sA j!) + r (sB j!) = 1: Probabilities are common knowledge but signals are private. We set signal sA to be more likely in state a than in state b: r (sA ja) > r (sA jb) ; and therefore r (sB ja) < r (sB jb) : The distribution of signals is unbiased if r (sA ja) = r (sB jb). Note that r (sA ja)+r (sA jb) = 1 in this case. The distribution of signals is biased if r (sA ja) 6= r (sB jb) and, by convention, we will focus on the case in which the “more abundant”signal is sA : r (sA ja)+r (sA jb) > 1. Having received her signal, the voter updates her beliefs about each state through Bayes’rule, to: q (!js) = q(!) r (sj!) : q(a) r (sja) + q(b) r (sjb) (2) Like Bouton and Castanheira (2012), we assume that signals are su¢ ciently strong to create a divided majority: q (ajsA ) q (ajsB ) >1> : q (bjsA ) q (bjsB ) 11 (3) That is, conditional on receiving signal sA , alternative A yields strictly higher expected utility than alternative B, and conversely for a voter who receives signal sB . The election is held at time 2, when the actual state of nature is still unobserved, and payo¤s realize at time 3: the winner of the election and the actual state of nature are revealed, and each voter receives her utility U (P; !). Strategy space and equilibrium concept. The alternative winning the election is the one receiving the largest number of votes, with ties being broken by a fair dice. Still, the action space, i.e. which ballots are feasible, depends on the electoral rule. We consider two such rules: plurality and approval voting. In plurality, each voter can cast a ballot on one alternative or abstain. The voters’ action set is thus: P lu = fA; B; C; ?g ; where, by an abuse of notation, action A (respectively B, C) denotes a ballot in favor of A (resp. B, C) and ? denotes abstention.20 In approval voting, each voter can approve of as many (or as few) alternatives as she wishes: AV = fA; B; C; AB; AC; BC; ABC; ?g ; where, by an abuse of notation, action A denotes a ballot in favor of A only, action BC denotes a joint approval of B and C, etc. Each approval counts as one vote: when a voter only approves of A, then only alternative A is credited with a vote. If the voter approves of both A and B, then A and B are credited with one vote each, and so on. As in plurality, the alternative with the most votes wins the election. The only di¤erence between approval voting and plurality is that a voter can also cast a double or triple approval. While a single approval ( = A; B and C) can be pivotal against any other alternative, double approvals ( = AB, BC and AC) can only be pivotal against one precise alternative. For instance, if the voter plays AC; she is voting against B: her ballot can only be pivotal against that alternative, either in favor of A or of C. Finally, a triple approval (ABC) can never be pivotal: it is strategically equivalent to abstention. 20 Abstention will turn out to be a dominated action in both rules. Hence, removing abstention from the choice set would not a¤ect the analysis. 12 Let x denote the number of voters who played action number of votes received by an alternative 2 at time 2. The total is denoted by X : Under plurality the total number of votes received by alternative A; for instance, is simply: XA = xA :Under approval voting, it is: XA = xA + xAB + xAC + xABC : A symmetric strategy is a mapping : S ! 4( R ), where s( ) denotes the proba- bility that some randomly sampled voter who received signal s plays action , S is the set of possible signals, and ( R) is the simplex over the actions possible under the electoral rule R 2 fP LU; AV g. Given a strategy , the expected share of voters playing action in state ! is thus: ! ( )= X s( ) E [x ( ) j!; ] = ! The expected number of ballots s r (sj!) : (4) is: ( ) n: Since voters do not observe the state of nature, expected vote shares only vary across states because the expected fraction of active voters receiving each signal is di¤erent. Let an action pro…le x be the vector that lists, for each action of ballots ; the realized number : Since we focus on symmetric strategies for the time being, and since the conditional probabilities of receiving a signal s are iid, the probability distribution over the possible action pro…les is given by the multinomial probability distribution: Pr (xj!) = n! Y 2 ! ( )x( x ( )! ) ; R = P lu; AV R For this voting game, we analyze the properties of Bayesian Nash equilibria that (1) do not involve weakly dominated strategies and (2) satisfy what we call sincere stability. That is, we impose the re…nement that the equilibrium must be robust to the presence of an arbitrarily small fraction " of sincere votes (that is: sA (A) ; sB (B) " > 0), and we look for sequences of equilibria with " ! 0: The reason why we introduce sincere stability is to get rid of equilibria that would only be sustainable when all pivot probabilities are exactly zero, and all voters are then indi¤erent between all actions. Imagine for instance that all active voters play A. In that 13 case, the number of votes for A is n and the number of votes for C is nC , with probability 1. Voters are then indi¤erent between all possible actions, since a ballot can never be pivotal. In contrast, sincere stability imposes that a small fraction of the voters votes for their preferred alternative. This implies that at least some pivot probabilities become strictly positive, and hence that indi¤erence is broken. There are two main arguments in favor of our sincere stability re…nement. First, it captures the essence of properness in a very tractable way.21 ,22 Second, it is behaviorally relevant since experimental data (both in our experiments and others) suggest that some voters vote for their ex ante most preferred alternative no matter what. 3 Plurality This section analyzes the equilibrium properties of plurality voting, in which voters can either vote for one alternative or abstain. We …nd that two types of equilibria coexist: in one, all active voters play a same (pure) strategy independently of their signal: they all vote either for A or for B. This type of equilibrium is known as a Duverger’s Law equilibrium, in which only two alternatives receive a strictly positive vote share (Palfrey 1989, Cox 1997). In the second type of equilibrium, an active voter’s strategy does depend on her signal. Depending on parameter values, this equilibrium either features sincere voting, that is voters with signal sA (resp. sB ) vote A (resp. B) or a strictly mixed strategy in which voters with the most abundant signal (sA by convention) mix between A and B. Importantly, these three-party equilibria exist for any population size and are robust to signal biases. In other words, they do not rely on A and B’s vote shares being identical. To the contrary: they imply that vote shares di¤er. 3.1 Pivot Probabilities and Payo¤s When deciding for which alternative to vote, a voter must …rst assess the expected value of each possible ballot. This value depends on the possible pivot events: unless the ballot 21 We are grateful to Eric Van Damme for insightful discussions about re…nement concepts in voting games. 22 We do not use more traditional re…nement concepts such as perfection or properness because, in the voting context, the former does not have much bite since weakly dominated strategies are typically excluded from the equilibrium analysis. The latter is less tractable since it requires a sophisticated comparison of pivot probabilities for totally mixed strategies. 14 is pivotal and a¤ects the outcome of the election, it leaves the voter’s utility unchanged. We denote by pivQP the pivot event that one voter’s ballot changes the outcome of the election from a victory of P towards a victory of Q. In our setup, the comparison between the three potentially relevant actions, A, B and C, is simpli…ed by two elements: …rst, voting for C is a dominated action. Hence, we can set ! C ( ) equal to zero. Second, a vote for A or for B can only be pivotal against C; since we impose that nC > n=2. This implies that abstention is also a dominated action, and simpli…es the other computations without a¤ecting generality. Voter i’s ballot can only be pivotal, say in favour of A if, without i’s ballot, the number of A-ballots (xA ) is either the same as or one less than the number of C-ballots, nC , and i votes for A. To assess the probability of such an event, each active voter must identify the distribution of the other n the vote shares ! 1 voters’ballots, given the strategy . Dropping from for the sake or readability, the probability of one vote being pivotal, respectively in favour of A and B is: p!AC Pr (pivAC j!; Plurality) = p!BC Pr (pivBC j!; Plurality) = 1)! ( !A )nC 1 ( (n ) 1)! ( 2 ! nC B ) 1 n nC ( !A ) ! A 1 (nC 1)!(n nC 1)! 2 (n ! n nC B nC ! B 1 (nC 1)!(n nC 1)! nC + + ! B n nC ! A n nC ; (5) ; (6) where the two terms between brackets represent the cases in which one breaks and makes a tie. Note that the pivot probabilities p!AC and p!BC are continuous in Let G ( js) denote the expected gain of an action ! A and !: B 2 fA; Bg over abstention, ?: G (Ajs) = q (ajs) paAC V + q (bjs) pbAC v (> 0); (7) G (Bjs) = q (ajs) paBC v + q (bjs) pbBC V (> 0): (8) It is obvious that both actions yield higher payo¤s than abstention, which is thus dominated. The pay-o¤ di¤erence between actions A and B is: G (Ajs) G (Bjs) = q (ajs) [V paAC 15 vpaBC ] + q (bjs) [vpbAC V pbBC ]: (9) 3.2 Duverger’s Law Equilibria The game theoretic version of Duverger’s Law (Duverger 1963) states that, in plurality elections, only two alternatives should obtain a strictly positive fraction of the votes when voters play strategically (Palfrey 1989, Myerson and Weber 1993, Bouton and Castanheira 2009): De…nition 1 A Duverger’s Law equilibrium is a voting equilibrium in which only two alternatives obtain a strictly positive fraction of the votes. Our …rst proposition shows that: Proposition 1 In plurality, Duverger’s law equilibria exist for any electorate size, prior probabilities of the two states, and distribution of signals. Proof. Consider sA (A) = " and p! AC = p! BC sB ! A ! B Hence, from (9), we have that G (Ajs) (B) = 1: From (5) and (6) ; we have: 2nC n ! A (n !n A C nC ) + + ! B (n ! B nC nC ) ! 0: "!0 G (Bjs) < 0 for any " in the neighborhood of 0. These equilibria are such that either all active voters vote for alternative A or they all vote for alternative B: Duverger’s Law equilibria feature pros and cons. On the one hand, they ensure that either A or B receives strictly more votes than C, and hence that C cannot win the election. On the other hand, they prevent learning. That is, the winner of the election cannot vary with the state of nature. The reason why Duverger’s Law equilibria exist in plurality elections is the classical one: voters do not want to waste their ballot on an alternative that is very unlikely to win. Consider for instance the strategy pro…le (BjsA ) = 1 " and (BjsB ) = 1 with " strictly positive but arbitrarily small. In that case, an A-ballot is much less likely to be pivotal against C than a B-ballot.23 Therefore, the value of a B-ballot is larger than that of an A-ballot, both for sA - and sB -voters. 23 For (BjsA ) = 1 = (BjsB ) ; all pivot probabilities are equal to zero. In this case, voters are indi¤erent between all actions. Sincere stability means that we identify incentives for (BjsA ) ! 1: They imply that G (BjsA ) > G (AjsA ) in the neighborhood of this Duverger’s Law equilibrium. 16 3.3 Informative Equilibria In Duverger’s Law equilibria, all active voters play the same (pure) strategy independently of their signal. Information about voter preferences is therefore lost. Still, this type of equilibrium is typically considered as the only reasonable one if voters are short-term instrumentally rational, in Cox’s (1997) terminology. Indeed, in a classical private value setup, equilibria violating Duverger’s Law require that the vote shares of the second and third alternatives are (almost) equal, a “knife-edge” case (Palfrey 1989). Therefore, empirical research associates strategic voting with the propensity for a voter to abandon his or her most-preferred candidate in order to vote for a second-best candidate who is a more serious contender (see Blais and Nadeau 1996, Cox 1997, Alvarez and Nagler 2000, Blais et al. 2005). Observing that only relatively low fractions of the electorate switch to their second-best alternative in this way is thus interpreted as evidence that few voters are instrumental or rational. Yet, as shown by Propositions 2 and 3 below, common values among majority voters give rise to other equilibria in which “short-term instrumentally rational voters” should actually deviate from either Duverger’s Law equilibria or “knife-edge” three-candidate equilibria. The key di¤erence is that, in our setup, voters value the information generated by their own and by other voters’ballots. Like in Austen-Smith and Banks (1996) and Myerson (1998), they compare their probability of being pivotal in each state of nature. In what we call an informative equilibrium, voting strategies imply that these pivot probabilities are su¢ ciently close to one another. Such an equilibrium entails that (a) all alternatives receive a strictly positive vote share, (b) in a given state, the expected vote shares of each alternative are di¤erent, and (c) A is the strongest majority contender in one state of nature, and B in the other state.24 When information is close to being symmetric across states, voters vote sincerely in an informative equilibrium: a voter who receives signal sA then votes for A, whereas a voter who receives signal sB votes for B. That is, abandoning one’s preferred candidate would not be a best response when one expects other voters to vote sincerely: 24 If, in addition, the expected vote share of A and of B in their respective state is su¢ ciently larger than C’s, then the informative equilibrium is also expectationally stable in the sense of Fey (1997). See also Bouton and Castanheira (2009, Propositions 7.3 and 7.4). 17 Proposition 2 In the unbiased case r (sA ja) = r (sB jb) ; the sincere voting equilibrium exists 8n; nc : Moreover, there exists a value (n; nc ) > 0 such that sincere voting is an equilibrium for any asymmetric distributions satisfying r (sA ja) Proof. sA We start with the symmetric case, i.e. (A) = 1 = sB r (sB jb) < (n; nc ) : r (sA ja) = r (sB jb) : Under sincere voting, (B) (note that sincere stability is not a binding condition in this case), (5) and (6) imply paAC = pbBC > pbAC = paBC : Then, from (9): G (Bjs) = [V paAC G (Ajs) Since q (ajsA ) G (AjsB ) q (bjsA ) > 0 > q (ajsB ) vpaBC ] [q (ajs) q (bjs)] : q (bjsB ) ; this implies G (AjsA ) G (BjsA ) > 0 > G (BjsB ) : Sincere voting is thus an equilibrium strategy. By the continuity of pivot probabilities with respect to ! A and ! B; it immediately follows that there must exist a value (n; nC ) > 0 such that sincere voting is an equilibrium for any jr (sA ja) r (sB jb)j < (n; nc ). The intuition for the proof is simply that, in the unbiased case, sincere voting implies that the likelihood of being pivotal against C is the same with an A-ballot in state a as with a B-ballot in state b. Therefore, sA -voters strictly prefer to vote for A and sB -voters strictly prefer to vote for B. The pros and cons of sincere voting are the exact ‡ipside of the ones identi…ed for Duverger’s Law equilibria: as illustrated by the following example,25 it allows for learning, but does not guarantee the defeat of the Condorcet loser. Example 1 Consider a case in which n = 12; nC = 7; and r (sA ja) = r (sB jb) = 2=3. In this case, sincere voting implies that the best alternative (A in state a or B in state b) has the highest expected vote share and wins with a probability of 73%. C then has the second largest expected vote share and wins with a probability of 23% in either state. The alternative with the lowest –but strictly positive– vote share is B in state a and A in state b. When nC is 9, the alternative with the largest expected vote share is C. In this informative equilibrium, C thus wins with a probability above 71%, whereas the best alternative wins with a probability below 29%. Based on Proposition 2 and Example 1, one may be misled into thinking that infor25 Each numerical example reproduces the parameters used in one of the treatments of our laboratory experiments (see section 5). In all examples, the payo¤s are: V = 200; v = 110 and the value of C is 20. Normalizing the latter to 0 would also reduce the other payo¤s by 20. 18 mative equilibria require signals to be unbiased. Two remarks are in order. First, voting sincerely is still an equilibrium when the signal bias exists but is not too large. In that case, the di¤erence in beliefs between sA and sB -voters keeps dominating the di¤erence in pivot probabilities. Yet, for any given bias r (sA ja) r (sB jb) > 0, sincere voting is only an equilibrium if electorate size is su¢ ciently small: as electorate size increases to in…nity, given the biased signal structure, the ratio of pivot probabilities would either converge to zero or in…nity if voters kept voting sincerely. Second, the fact that the signal structure becomes too biased to sustain sincere voting does not imply that voters switch to a Duverger’s Law equilibrium: Proposition 3 instead shows that, in an informative equilibrium, sA -voters adopt a mixed strategy and vote for B with strictly positive probability. This allows them to lean against the bias in the signal structure. In other words, the short-term instrumentally rational voter should partly abandon the strongest contender and lend support to the weakest majority alternative: Proposition 3 Let r (sA ja) equilibrium with sA r (sB jb) > (A) 2 (0; 1) and sB (n; nc ). Then, there exists a mixed strategy (B) = 1, such that alternative A receives strictly more votes in state a than in state b, and conversely for alternative B. Proof. See Appendix A2. The intuition for this result is that strong biases in the signal structure imply that the di¤erence in pivot probabilities between states a and b becomes too large if voters were to vote sincerely. To correct for this bias in the informative equilibrium, sA -voters must strictly mix between A and B. In this way, they partially compensate the gap in pivot probabilities caused by the bias. The proof of the proposition establishes that there always exists one such mixture that is an equilibrium. It is such that sA -voters are indi¤erent between voting A and B, whereas sB -voters strictly prefer the latter. The intuition for the proof of this result is best conveyed with the help of a second example: Example 2 Electorate size is n = 12 and nC = 7, and the signal structure is r (sA ja) = 8=9 > 2=3 = r (sB jb). The two states of nature are equally likely: q (!) = 1=2. For these parameter values, an sA -voter would prefer to vote for B if all the other voters were to 19 vote sincerely. Indeed, sincere voting implies:26 V pbBC 8 q (ajsA ) = < 13:5 = q (bjsA ) 11 V paAC vpbAC : vpaBC (10) That is, the weighted probabilities of being pivotal in favour of B in state b is much larger than the pivot probabilities in favour of A, which implies G (AjsA ) mixed-strategy equilibrium is reached when sA (A) = 0:9153 and G (BjsA ) < 0: The sB (B) = 1: by reducing the expected vote share of A and increasing the vote share of B, the relative probability of being pivotal in favour of A in state a has increased to the point in which: V pbBC q (ajsA ) 8 = = q (bjsA ) 11 V paAC vpbAC q (ajsB ) 1 > = : a vpBC q (bjsB ) 3 Expressed in terms of payo¤ s, this means that sA -voters are now indi¤ erent between voting A and B, whereas sB -voters strictly prefer to vote B.Also interesting is that this equilibrium shares relevant features with a Condorcet-Jury type of equilibrium: even though vote shares are substantially di¤ erent across states and alternatives: a A = 0:81 > b B = 0:69 > nC = 0:58 > n b A = 0:31 > a B = 0:19; A leads in state a and B leads in state b. Their winning probabilities are respectively 96% and 79%. Yet, this informative equilibrium gives C a strictly positive probability of victory (18% in state b and 3% in state a). Importantly, though, expected utility is higher in this equilibrium than in Duverger’s Law equilibria. The example vividly illustrates that neither the existence nor the stability of this equilibrium relies on some form of symmetry between vote shares. Also, as proved by Bouton and Castanheira (2009) for large Poisson games, this mixed-strategy equilibrium also exists in very large electorates (i.e. for n ! 1), with the di¤erence that the gap between a A and b B decreases to zero (i.e. limn!1 a A = limn!1 b ), B and that stability relies on r (sA ja) being su¢ ciently larger than nC =n. 26 Note that, by (9), G (Ajs) G (Bjs) > 0 i¤ the left-hand side of (10) is larger than the right-hand side. 20 4 Approval Voting 4.1 Payo¤s and Dominated Strategies Under approval voting, voters have access to a larger choice set, which makes their choice potentially more complex. Single approvals (A; B, C) have exactly the same e¤ect as in plurality. Double or triple approvals instead ensure that one selectively abstains between the approved alternatives. For instance, an AB-ballot can only be pivotal against C. The following lemma shows that the set of undominated strategies is more restricted: Lemma 1 Independently of a voter’s signal, the actions weakly dominated by some action in s (A) + 2 fC; AC; BC; ABC; ?g are 2 fA; B; ABg. Hence, in equilibrium: s (B) + s (AB) = 1; 8s 2 fsA ; sB g : (11) Proof. The proof is straightforward: consider a majority-block voter and compare actions AB and ABC. While the latter can never be pivotal, an AB-ballot can be pivotal against C, either in favor of A or in favor of B: In both cases, and independently of the true state of nature, utility can only increase. Hence, AB weakly dominates ABC. All other dominance relationships are obtained by performing similar two-by-two comparisons: AB weakly dominates ABC; ? and C; A weakly dominates AC; and B weakly dominates BC. The intuition for the lemma is that abstaining or approving of C can only increase C’s probability of winning. In contrast, the actions in the undominated set (A; B; and AB) can only reduce it. The remaining question is how a voter may want to allocate her ballot across these undominated actions. This depends on the value of each undominated action, which itself depends on the pivot events. Let ! QP denote the probability that a single-Q ballot by some voter i is pivotal in favor of Q at the expense of P in state of nature ! 2 fa; bg and the voting rule is approval voting. The precise derivation of these pivot probabilities is detailed in Appendix A1. For a voter who received signal s, the value GAV of a single-A ballot under approval voting is then: GAV (Ajs) = q (ajs) [ a AC V + a AB (V v)] + q (bjs) [ 21 b AC v + b AB (v V )]: (12) Note that the probability of being pivotal between A and B is no longer zero. This is due to the possibility of double voting, which can increase the total score of both A and B above that of C. Similarly, the value of a single-B ballot is: GAV (Bjs) = q (ajs) [ a BC v + a BA (v V )] + q (bjs) [ b BC V + b BA (V v)]: (13) The value of a double ballot follows almost immediately from (12) and (13). Double voting cannot be pivotal between A and B; while adding up the chances of being pivotal against C, either in favor of A or in favor of B: GAV (ABjs) = q (ajs) [ where a and b a AC V + a BC v a ] + q (bjs) [ b AC v + b BC V b ]; (14) are correcting factors for three-way ties (see Appendix A1 for a precise de…nition). Consider the case in which, without voter i’s ballot, both alternatives A and B lose to C by one vote (that is, both obtain nC 1 votes). A single-A ballot creates a tie between A and C. Thus, the ballot is pivotal in favor of A with probability 1=2. Likewise, a single-B ballot is pivotal in favor of B with probability 1/2. Yet, a double vote AB creates a three-way tie, which still allows C to win with probability 1/3: the winning a AC probabilities of A and B are 1/3 instead of 1=2. Summing up the probabilities a BC thus overestimates the value of the double ballot by (V + v) =6. The terms a and and b correct for these overestimations in that and three other cases: when A trails behind both B and C by one vote, when B trails behind both A and C by one vote, and when A; B and C have the same number of votes. These correcting factors become vanishingly small when the population size increases towards in…nity, because the probability of three-way ties also becomes vanishingly small compared to other two-way pivot events. Yet, our purpose in this paper is to assess the properties of plurality and approval voting both for small-committee and for large-population elections, which implies that we need to take account of such three-way ties.27 27 These correcting factors actually prove extremely relevant in small committees, in particular for the characterization of the asymmetric equilibria that we analyze in Section 6.3.1. 22 The payo¤ di¤erential between actions A and AB is therefore: GAV (Ajs) GAV (ABjs) = q (ajs) [ +q (bjs) a AB h (V b AB a BC v v) (v b BC V V) a + ] (15) b + i : With straightforward, although tedious, manipulations, one …nds that the …rst term in (15) may either be positive or negative, whereas the second is strictly negative. Similarly, the …rst term in (16) must be strictly negative: GAV (Bjs) GAV (ABjs) = q (ajs) [ +q (bjs) a BA (v h a AC V V) b BA (V + b AC v v) + a ] b (16) i : Finally, the payo¤ di¤erence between actions A and B is: GAV (Ajs) GAV (Bjs) = q (ajs) [ q (bjs) 4.2 a AC V h b BC V a BC v +( b AC v + a AB + b AB a BA ) (V + b BA v)] (V (17) i v) : Equilibrium Analysis The action set under approval voting is an extension of the action set under plurality. Therefore, in a common value setting as ours, there is always an equilibrium in AV for which the welfare is (weakly) higher than for any equilibrium in plurality (Ahn and Oliveros 2011, proposition 1).28 Furthermore, our setup imposes that the size of the minority is large which, as we observed in Section 3, implies that the probability of being pivotal between A and B is always zero under plurality. Theorem 1 directly follows from that fact and from (15 17): Theorem 1 There always exists a sincerely stable equilibrium in AV for which expected welfare is strictly higher than for any equilibrium in plurality. In that equilibrium, some voters must double vote, and sA (A) ; sB (B) > 0: 28 Ahn and Oliveros (2011) exploit McLennan (1998) to show that, in a common value setup as ours, one can rank equilibrium outcomes under approval voting as opposed to some other plurality and negative voting. By revealed preferences, since the action set in the two other rules is a strict subset of the action set under AV, “the maximal equilibrium utility under approval voting is greater than or equal to the maximal equilibrium utility under plurality voting or under negative voting.” (p. 3). 23 Proof. See Appendix A3. The intuition for this result is as follows: when one compares the set of undominated actions in plurality and in AV, one sees that the only relevant di¤erence is the possibility to double vote AB. As emphasized above, double voting AB can be reinterpreted as abstaining between A and B while cumulating the chances of being pivotal against C. Now, any strategy pro…le in plurality imposes that none of the “other voters” plays AB. Thus, any voter must realize that she can never be pivotal between A and B with a single vote. In contrast, her expected utility increases strictly by double voting AB rather than mimicking the other voters’strategy pro…le.29 It thus follows that: Corollary 1 The strategies that are an equilibrium in plurality cannot be an equilibrium in AV. In particular, Duverger’s Law equilibria do not exist under AV. Moreover, since voters have common valued preferences, if such a deviation is bene…cial for one voter, it must also increase all the other voters’ expected utility. Double voting has pros and cons in terms of the election outcome. On the one hand, it reduces the risk that C wins the election. On the other hand, a voter who double votes does not exploit her signal, which may reduce the probability that A wins in state a and B wins in state b. Yet, there cannot be so much double voting that information aggregation would be made impossible: Corollary 2 Pure double voting is never an equilibrium in AV. The reason is straightforward: if all the other voters double vote, then voter i knows (a) that her vote cannot be pivotal against C and (b) that she is as likely to be pivotal in state a as in state b. Hence, her preferred reaction is to vote sincerely and reveal her signal. Pure double voting has been termed the Burr dilemma by Nagel (2007), who argues that approval voting is inherently biased towards such ties. He documents this with the “[approval] experiment [that] ended disastrously in 1800 with the infamous Electoral College tie between Je¤ erson and Burr ”. Lemma 2 shows why such a “disaster” cannot happen in equilibrium of an election which is not dictated by party discipline. 29 To repeat, this is due to the fact that we focus on the case of large minorities. If the size of the minority, nC , falls towards zero, then the propensity to double vote may well drop to zero as well (see Bouton and Castanheira, 2012). 24 Together, Corollaries 1 and 2 show that a voter’s best response is to double vote if the other voters single vote “excessively” and to single vote sincerely if the other voters double vote “excessively”. In a large Poisson game setup, Bouton and Castanheira (2012) shows that this pattern is monotonic, and that the relative value of the double and single votes cross only once. In other words, approval voting displays a unique equilibrium. Our setup di¤ers in two respects. First, we are not focusing on arbitrarily large electorates as they do. Second, in our setup the size of both the majority and the minority (n and nC ) are …xed. The small-sample and large-sample properties of such voting games can be quite di¤erent from one another. For instance, we already emphasized that three-way ties can be neglected in large samples, but not in small samples. This implies that one cannot establish a proof of equilibrium uniqueness for any electorate size and signal distribution. Yet, our next theorem can pinpoint precise voting patterns for any interior equilibrium: Theorem 2 Whenever both sA - and sB -voters adopt a nondegenerate mixed strategy, then it must be that voters with signal sA only mix between A and AB, and voters with signal sB only mix between B and AB. Proof. See Appendix A3. This theorem builds on the comparison between the preferences of sA and sB voters: conjecture for instance a case in which the former play B with strictly positive probabilities. Since a voter with signal sB values B even more, it must only play B, which contradicts the very nature of an interior equilibrium. To extend this result to equilibria in which (one of the two groups of) voters play degenerate strategies, we would have to focus on larger electorates, which is not the purpose of our analysis. Indeed, Bouton and Castanheira (2012) show that the equilibrium is unique and achieves Full Information and Coordination Equivalence. On the other hand, in small electorate, we need to rely on numerical simulations to study the properties of any such game for a given set of parameters. For all the parametric values we checked, the equilibrium was unique and such that voters with signal sA never play B (i.e. they mix between A and AB), and voters with signal sB never play A. This held both for interior equilibria and for equilibria in which (one of the two sets of) voters play a degenerate strategy. Two additional examples are useful to better understand the features and comparative statics of voting equilibria in AV: 25 Example 3 Consider the same set of parameters as in example 1: n = 12; nC = 7 or 9; and r (sA ja) = r (sB jb) = 2=3. The two states of nature are equally likely, and the payo¤ s are: V = 200; v = 110; while the status quo yields a utility of 20. As emphasized above, there is a unique equilibrium under approval voting for these parameter values (as for all the other parameter values we checked).30 It is such that: sA (A) = sB (B) = 0:64 and sA (AB) = sB (AB) = 0:36 when nC = 7; sA (A) = sB (B) = 0:30 and sA (AB) = sB (AB) = 0:70 when nC = 9: These equilibrium pro…les imply that A wins with a probability of 82% in state a, as does B in state b, whereas C’s probability of winning is below 1% when nC = 7. When nC = 9, A wins with a probability of 73% in state a (as does B in state b), whereas C’s probability of winning remains as low as 1.5%. These values should be contrasted with the sincere voting equilibrium in plurality (see example 1), in which the probability of selecting the best outcome was substantially lower, whereas the risk that C wins was substantially larger. Comparing equilibrium behavior with nC = 7 and nC = 9 in Example 3 shows that the larger nC , the more double voting in equilibrium. This pattern was found to be monotonic and consistent across numerical examples for any value of n and signal structures. Example 4 Consider the same set of parameters as in Example 3, except for r (sA ja) = 8=9: This reproduces the biased signal setup of Example 2. Like in the previous example, the equilibrium is unique. It yields: 0:74 > sB sA (A) = 0:26 < sB (B) = 0:52 and sA (AB) = (AB) = 0:48: This equilibrium pro…le implies that A wins with a probability of 87% in state a, whereas B wins with a probability of 90% in state b. C’s winning probabilities are 0.5% in state a and 2.8% in state b. The equilibrium with biased information has the property that the voters with the most abundant signal should single vote less than the voters with the least abundant signal. The rationale for this result might be obvious to the readers knowledgeable about the Condorcet Jury Theorem: if sA - and sB -voters were to single vote with the same 30 In the strategy space ( sA (A) ; sB (B)), there is a unique cuto¤ for which G (AjsA ) = G (ABjsA ), and the same holds for G (BjsB ) = G (ABjsB ). The equilibrium lies at the intersection between these two reaction functions. 26 intensity, A’s winning probabilities would be disproportionately higher than B’s. Moreover, the pivot probabilities between A and B would be lower in state a than in state b; which should induce all voters to put more value on being pivotal in favour of B. An empirical question is whether actual voters would best respond in this way or whether this type of reasoning is too abstract. 5 Experimental Design and Procedures 5.1 Experimental Design To test our theoretical predictions we ran controlled laboratory experiments. Subjects were introduced to a game that had the very same structure as the one presented in Section 2. All participants were given the role of an active voter, whereas passive voters were simulated by the computer.31 Following the experimental literature on the Condorcet Jury Theorem initiated by Guarnaschelli et al. (2000), the two states of the world were called blue jar and red jar, whereas the signals were called blue ball and red ball. The red jar contained six red balls and three blue balls. Depending on the treatment, the blue jar contained either six blue and three red balls (unbiased signals) or eight blue and one red ball (biased signals). One of the jars was selected randomly by the computer, with equal probability. The subjects were not told which jar had been selected, but they received a ball randomly drawn from the selected jar. After seeing their ball, each subject could vote from a set of three candidates: blue, red or gray.32 Blue and red were the two majority candidates and gray was the Condorcet loser. Subjects were told that the computer would cast nC votes for gray in each election (nC varied across treatments). The subjects’ payo¤ depended both on the color of the selected jar and that of the winner of the election. If the color of the winner matched the color of the jar, the payo¤ to all members of the group was 200 euro cents. If the winner was blue and the jar red or the other way around, the payo¤ to all members was 110 cents. Finally, if gray won, the 31 Morton and Tyran (2012) show that preferences in one group are not a¤ected by preferencs of an opposite group. Therefore, we conjecture that having computerized rather than human subjects did not change the behavior of majority voters in a signi…cant way. Having partisans (who are equivalent to our passive voters) simulated by the computer has been used in previous studies – see Battaglini et al. (2008, 2010). 32 The colors that we used in the experiments were blue, red and black. Throughout the paper, however, we refer to blue, red and gray respectively. 27 Treatment PL7 PL9 AV7 AV9 PL7B AV7B Voting rule Plurality Plurality Approval Approval Plurality Approval Minority size (nc ) 7 9 7 9 7 7 Precision Blue State Precision Red State 2=3 2=3 2=3 2=3 8=9 8=9 2=3 2=3 2=3 2=3 2=3 2=3 Sessions / Ind. Obs. 3/6 3/6 3/6 3/6 3/6 3/6 Group numbers 1-6 7-12 13-18 19-24 25-30 31-36 Table 1: Treatment overview. Note: ind. obs. stands for “individual observations”. payo¤ to all members was 20 cents. We consider three treatment variables, which leads to six di¤erent treatments. The …rst variable is the voting mechanism: in PL treatments, the voting mechanism was plurality. In this case, subjects could vote for only one of the three candidates. In AV treatments, the voting mechanism was approval voting. In this case, voters could vote for any non-empty subset of candidates.33 With either mechanism, the candidate with the most votes wins, and ties were broken with equal probability. The second variable is the size of the minority, nC , which was set to either 7 or 9. We will refer to them as small and large minority. Finally, the third variable is whether the signal structure is unbiased or biased. In unbiased treatments, signal precision was identical across states and set to r (blue ball j blue jar) = r (red ball j red jar) = 23 . In biased treatments (which we indicate by B ), the precision of the signal was the same in state red, whereas r (blue ball j blue jar) was increased to 8 9. Table 1 provides an overview of the di¤erent treatments. 5.2 Experimental Procedures Experiments were conducted at the BonnEconLab of the University of Bonn between July 2011 and January 2012. We ran a total of 18 sessions with 24 subjects each. No subject participated in more than one session. Students were recruited through the online recruitment system ORSEE (Greiner 2004) and the experiment was programmed and conducted with the software z-Tree (Fischbacher 2007). All experimental sessions were organized along the same procedure: subjects received 33 As in Guarnaschelli et al. (2000), abstention was not allowed (remember that abstention is always a strictly dominated action). In a similar setting, Forsythe et al (1993) allowed for abstention and found that the abstention rate was as low as 0.65%. 28 detailed written instructions, which an instructor read aloud (see Appendix B). Each session proceeded in two parts: in the …rst part, subjects played in …xed groups one of the treatments described for 100 periods.34 Before starting, students were asked to answer a questionnaire to check their full understanding of the experimental design. In the second part, subjects received new instructions, and then made 10 choices in simple lotteries, as in Holt and Laury (2002). We ran this second part to elicit subjects’risk preferences. To determine payment at the end of the experiment, the computer randomly selected four periods from the …rst part and one lottery from the second part.35 In total, subjects earned an average of e13.47, including a showup-fee of e3. Each experimental session lasted approximately one hour. 6 Experimental Results We present our experimental results in four distinct sections. Section 6.1 presents the outcomes when information is unbiased, and Section 6.2 when it is biased. Section 6.3 turns to individual behavior and extends the model to asymmetric equilibria. Finally, section 6.4 summarizes the main …ndings regarding outcomes and welfare. 6.1 6.1.1 Unbiased Treatments Plurality As shown in section 3, two types of equilibria coexist under plurality when information is unbiased: Duverger’s Law and sincere voting equilibria. In the former type of equilibria, participants should disregard their signal and coordinate on voting always blue or always red. In the latter instead, participants should vote their signal. Table 2 shows the average frequencies with which subjects voted sincerely (we call this voting the signal ), for the majoritarian party opposite to the signal (we will call this voting opposite) or for gray.36 34 In the setup of the Condorcet Jury Theorem, Ali et al (2008) …nd no signi…cant di¤erence between random matching (which they call ad hoc committees) and …xed matching (which they call standing committees). 35 In the …rst round of experiments (the seven sessions with the groups 1, 2, 7, 8, 9, 10, 13, 14, 15, 16, 19, 20, 21 and 22), we selected seven periods to determine payment. We reduced this to four periods after realizing that the experiment had taken much less time than expected. We …nd no di¤erence in behavior between these two sets of sessions. 36 The …gures identi…ed with a ‘*’report the predicted voting pattern for the last 50 periods, conditional on the color (blue or red) on which the group coordinated. For instance, if the group coordinated on blue, 29 Treatment PL7 Minority Size Small PL9 Large Signal Opposite Gray Signal Opposite Gray Periods 1-50 91.80 7.78 0.42 68.47 31.11 0.41 Periods 51-100 90.94 8.89 0.17 59.25 40.67 0.08 Equilibrium Sincere Voting Duverger’s Law 100.00 50.00 50.00 100.00 48.92 51.08 - Table 2: Aggregate voting behavior in plurality treatments with unbiased information, separated by …rst and second half, and equilibrium predictions. In the case of Duverger’s Law in PL9, the prediction is adjusted to the color that each group converged to. In the presence of a small minority, the participants’voting behavior is consistent with the sincere voting equilibrium: taking an average across all groups and periods, 91.38% of the ballots were sincere in PL7. This …nding holds across groups: there were at least 86.42% of sincere ballots in each independent group. As shown in Table 3, this behavior is quite stable over time: we regress the frequency of “voting the signal” on the period number, and …nd that the coe¢ cient is not signi…cantly di¤erent from zero. Most deviations from sincere voting behavior consisted of “voting opposite”. Part of this might be related to subjects misunderstanding the fact that random events are independent across periods. Such misunderstanding has been observed frequently in the literature.37 Finally, less than 0.5% of the observations were votes for gray. Voting behavior is substantially di¤erent in the presence of a large minority (PL9). First, only 63.86% of the observations are consistent with sincere voting behavior. Second, there is a clear temporal trend: as shown in Table 3, the frequency of voting the signal signi…cantly decreases over time: the predicted frequency of sincere voting drops from 73.90% in the …rst periods to 53.70% in the last period. That is, participants begin the experiment by voting sincerely (94.44% of them voted their signal in the …rst period), and then they adjust their behavior by increasingly voting against their signal. This pattern is and if 40% of the voters obtain a blue ball in a given draw, then 40% should play “signal” and 60% play “opposite”. 37 The misperception of the independence between independent events is known in the lituerature as the gambler’s fallacy. More precisely, the gambler’s fallacy is the mistaken notion that the likelihood of an event that occurs with a …xed probability increases or decreases depending upon recent occurrences. The gambler’s fallacy has been documented extensively in laboratory experiments, survey data and …eld experiments (see Tversky and Kahneman 1971 or Clotfelder and Cook 1993). In our experiment, the gambler’s fallacy might lead subjects to disregard signals on the ground that the perceived likelihood of the signal being wrong is higher than the likelihood of the signal being right after some particular histories. 30 Period Constant Observations R2 PL7 -0.0000983 (0.000202) PL9 -0.00204 (0.000182) 0.919 (0.0150) 7200 0.0001 0.741 (0.0248) 7200 0.0150 Table 3: Linear Regression of the frequency of voting the signal on the period number and a constant, clustering by independent group, for each treatment with plurality and unbiased information. Standard errors in parentheses. p < 0:10, p < 0:05, p < 0:01 fully consistent with the progressive shift from a sincere voting equilibrium to a Duverger’s Law equilibrium. Figure 1 illustrates this shift by plotting the observed frequency of voting blue, red and gray (irrespective of the signals subjects receive) for each group in the PL9 treatment. The horizontal dashed line displays the minimal vote share required to defeat gray (in case nobody plays the dominated strategy of voting gray). As one can see, all six groups did converge to a Duverger’s Law equilibrium: groups 7 and 9 converged to voting blue regardless of one’s signal, whereas the other groups converged to voting red. Two additional elements emerge from these data on PL9. First, which Duverger’s Law equilibrium they converge to is actually decided very early: all but one group ended up coordinating on the …rst color obtaining strictly more than six of the majority votes (it happened in period 1 for four groups and in period 2 for one group).38 This is in line with the …ndings of Forsythe et al (1993, p235): “a majority candidate who was ahead of the other in early elections tended to win the later elections, while the other majority candidate was driven out of subsequent races”. Second, even though this pattern is obvious with hindsight, the groups’shifting from sincere voting towards a Duverger’s Law equilibrium was fairly slow. This is exempli…ed by the minimum period from which either blue or red repeatedly obtained enough votes to win: the period number was respectively 50, 59, 83, 63, 21 and 26 for groups 7–12. This raises two interlinked questions: (1) starting at the same sincere voting equilibrium as “small minority”groups, why do all “large minority”groups switch to a Duverger’s Law equilibrium? (2) why is that switch slow? Answering the …rst question is relatively 38 The exception is group 11 in which blue got seven votes in the …rst period and then red received more votes from period 2 onwards. 31 PL9, 8 PL9, 9 PL9, 10 PL9, 11 PL9, 12 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 PL9, 7 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 Period (Groups of five) Blue Gray Red Min Percentage to Win Figure 1: Frequence of voting blue, red and gray irrespective of the signal in groups of treatment PL9U. The dashed line indicates the minimum frequence of vote share required to defreat gray (in case nobody from the majority votes for the condorcet loser). 32 straightforward: facing a large minority means that voters face an expected utility of 69.76 in PL9, instead of 152.76 in PL7. This should be compared to an expected utility of 155 in a Duverger’s Law equilibrium. The incentive to get away from the sincere voting equilibrium is thus substantially higher in PL9. Addressing the second question is more di¢ cult. Going back to theory, we can compute a voter’s best response for any strategy pro…le by the other 11 voters. The phase diagram in Figure 2 illustrates this graphically. The horizontal axis displays the other blue voters’ propensity to vote blue, and the vertical axis displays the other red voters’propensity to vote red. The solid curve represents the locus of the others’strategies for which a voter is indi¤erent between playing blue or red if she receives a blue ball. To the left of that curve, the payo¤ of red is higher than that of blue, and conversely to the right of the curve. The dashed curve represents the equivalent locus for a voter who receives a red ball. Above the curve, she prefers the payo¤ of playing red is higher than that of blue, and conversely below the curve. The arrows in Figure 2 display the attraction zones of each of the three equilibria mentioned: sincere voting in the top right corner, and the two Duverger’s Law equilibria in the bottom right and top left corners. As it appears clearly, the attraction zone of the sincere voting equilibrium is much larger in PL7 than in PL9, but it is still pretty large even in PL9. Within this attraction zone, after receiving her signal, a voter thus faces two opposite incentives: a static incentive to reveal her signal, and a dynamic incentive to coordinate towards one of the two Duverger’s Law equilibria. This helps explain why convergence is slow, despite the dynamic incentive. Moreover, the larger dynamic incentive and the smaller attraction zone in PL9 help explain why all groups ended up moving away from sincere voting. 6.1.2 Approval Voting Table 4 summarizes behavior in AV treatments. It displays the frequencies with which subjects single vote for their signal, double vote for the two majority candidates, single vote opposite to their signal, and vote for gray (possibly in combination with another candidate). These two treatments reproduce the parametric cases covered in Example 3, which we found to display a unique symmetric equilibrium. In that equilibrium a voter should only 33 Figure 2: Phase diagram of treatments PL7 and PL9. The horizontal axis displays the probability of sincere voting by blue voters while the vertical axis displays the probability of sincere voting by red voters. The solid line indicates the indi¤erence curve for the blue voters, while the dashed line indicates the indi¤erence curve for the red voters. Treatment AV7 Minority Size Small AV9 Large Signal Double Vote Opposite Gray Signal Double Vote Opposite Gray Periods 1-50 70.92 22.22 6.50 0.36 47.08 45.67 6.86 0.39 Periods 51-100 71.94 24.36 3.69 0.00 43.33 51.64 4.97 0.06 Equilibrium 64.00 36.00 30.00 70.00 - Table 4: Aggregate voting behavior in approval voting treatments with unbiased information. Gray refers to voting for gray or a combination of gray and others. 34 single vote for her signal or double vote Blue and Red. A huge majority of actions were in line with this theoretical prediction: 94.73% in the case of AV7 and 93.86% in the case of AV9. One could think that AV involves higher complexity and therefore higher frequency of mistakes, but we actually observed the opposite. De…ne mistakes as playing an action that is not a best response to the equilibrium. That is, in PL7, AV7 or AV9, a subject made a mistake when she voted opposite to her signal or for gray; in PL9, she made a mistake when voting for another color than the one the rest of the group converged to. We …nd that subjects made more mistakes under plurality than under approval voting. In treatments with a small minority, the percentages of mistakes in the second half (where equilibrium selection is clearer) were 3.69% in AV7 as opposed to 9.06% in PL7. In treatments with a large minority, they represented 5.03% of the ballots in AV9 as opposed to 11.06% in PL9. These di¤erences are signi…cant in both cases (Mann-Whitney, z = 2.082, p = 0.0374 with c = 7 and z = 1.925, p = 0.0542 with c = 9).39 The second theoretical prediction drawn from Example 3 refers to the e¤ect of minority size: there should be an increase in the frequency of double voting. The rationale is that voters need to double vote more to contain the risks of election of the minority candidate, gray. Table 4 shows that this is indeed the way in which the subjects adapted their voting behavior: the percentage of double voting was multiplied by more than two, from 23.29% in treatment AV7 to 48.66% in treatment AV9. This di¤erence is signi…cant at 1% (Mann-Whitney, z = 2.722, p < 0.01). Although the comparative statics go in the direction predicted by theory, one should notice that there is a considerable di¤erence between the data and the theoretical predictions: the amount of double-voting was well below the prediction. In treatment AV7, the observed amount of double-voting was 24.36 instead of 36.00 in theory. In treatment AV9 it was 51.64 instead of 70.00. These di¤erences are signi…cant at 5% in both cases (Mann-Whitney, z = 2.201, p < 0.05). Section 6.3 returns to this discrepancy to show that asymmetric equilibria help explain this gap. 39 One could also de…ne mistakes as voting for gray (or gray with something else). In this case as well, plurality leads to more mistakes than AV, but these di¤erences are not signi…cant (Mann-Whitney, z = 1.000, p = 0.3173 with c = 7 and z = 0.561, p = 0.5751 with c = 9). 35 6.2 The E¤ects of Biased Information In PL7, we observed that all independent groups coordinated on the sincere voting equilibrium. One reason might be the symmetry between the blue and red signals, which made coordination challenging for the subjects. In treatment PL7B, we instead made the signal structure strongly biased in favor of the blue signal by setting r (blue ball j blue jar ) = 8=9. So, if the voters were to keep playing sincere, blue would win disproportionately more often than red. Together with Example 2, Propositions 1 and 3 show that voters may still coordinate on either of two potential equilibria. First, they may coordinate on the Duverger’s Law outcome. We may even expect that coordination was made simpler: the higher expected vote share of blue may have fostered coordination by red voters towards voting blue. The second possible outcome is the informative equilibrium. Example 2 showed that, to aggregate information, blue voters should then adopt a strictly mixed strategy and vote red with strictly positive probability (8.43% according to Example 2). In the data, we observe that one independent group (group 28) out of six coordinated on the “blue”Duverger’s Law equilibrium, whereas the other …ve adopted a strategy coherent with the informative equilibrium. Let us analyze each in turn: in group 28, almost all voters casted a blue ballot as of period 31: from that period onwards, blue obtained at least seven votes in all cases, and strictly more than seven in 94.29% of the cases.40 Table 5 below summarizes the behavior of the other 5 independent groups. Focusing on the last …fty periods to account for convergence towards an equilibrium, we observe that blue voters indeed voted red with a probability close to the theoretical prediction, whereas red voters mainly voted sincerely. The di¤erence between these two behaviors is statistically signi…cant (Mann-Whitney, z = 2:023, p < 0:043). Moreover, we do not …nd any signi…cant di¤erence between the theoretical prediction and the observed frequency of voting blue when getting a blue ball (Mann-Whitney, z = 0:405, p = 0:6858). The transition from sincere voting in the …rst periods towards either of the two types of equilibria illustrates the di¢ culty of coordinating on one equilibrium when several are available, each with its pros and cons. The main advantage of Duverger’s Law equilibria is 40 The fact that the only group that converged to a Duverger’s Law equilibrium did coordinate on blue is coherent with the idea that biased signals foster coordination. However, this intuition cannot be tested since there is only one such group to study. Moreover, the results in PL9 o¤er an alternative rationale, which is that voters coordinate on the color that won in the …rst period(s). The investigation of which factors dominate equilibrium selection seem an interesting topic for future research. 36 Signal if blue Opposite if blue Signal if red Opposite if red Gray Periods 1-50 92.99 6.89 96.39 3.13 0.27 Periods 51-100 90.75 8.38 97.48 1.74 0.83 Equilibrium 91.53 8.47 100 0 0 Table 5: Aggregate voting behavior in treatment PL7B. Group 28 was excluded given that it converged to a Duverger’s Law equilibrium. that it removes uncertainty about the outcome of the election, and ensures that gray can never win. Its main drawback is that there is no aggregation of information. Conversely, the informative equilibrium maximizes the probability of electing the best candidate in each state but does not o¤er the “safety”of Duverger’s Law equilibria: all candidates have a strictly positive probability of winning. The model helps identify two reasons why the latter equilibrium is more likely to be selected. First, it yields a higher expected payo¤ than Duverger’s Law equilibria (178.37 instead of 155). Second, as identi…ed by the phase diagram displayed in Figure 3, when starting from sincere voting (the top-right corner), the local dynamics of individual best responses point towards the informative mixed strategy equilibrium (the point MSE on the graph) rather than towards either Duverger’s Law equilibria. These data also provide ample evidence to con…rm our theoretical prediction that threecandidate equilibria are a natural focal point when voters have common values. It is also noteworthy that voters do not simply play “sincere”. In the …ve groups that selected the informative equilibrium, blue voters actually adjusted their behavior to better aggregate information. While Bhattacharya et al (2012) already con…rm that such voting behavior is empirically robust in a two-candidate setting, we believe that our experimental results are the …rst ones to identify this pattern in a three-candidate setting. Approval voting o¤ers the opportunity to double vote to achieve the twin objectives of information aggregation and coordination. Treatment AV7B is the same as PL7B with the only di¤erence that subjects can exploit this opportunity. Thus, in this treatment, voters face the more complex challenge of having to deal with a broader choice set but, as identi…ed in Example 4, their task is simpli…ed by the fact that the equilibrium is unique in this setup. Like in the informative equilibrium of PL7B, blue voters should play blue 37 Figure 3: Phase diagram of treatment PL7B. The horizontal axis displays the probability of sincere voting by blue voters, while the vertical axis displays the probability of sincere voting by red voters. The solid line indicates the indi¤erence curve for the blue voters; the dashed line indicates the indi¤erence curve for the red voters. Signal if blue Double Vote if blue Signal if red Double Vote if red Opposite Gray Periods 1-50 66.95 29.87 74.56 20.52 2.94 0.89 Periods 51-100 61.16 37.16 80.52 17.98 1.56 0.06 Equilibrium 50.1 49.9 92.6 7.4 0 0 Table 6: Aggregate Voting Behavior in treatment AV7B. less often than red voters should play red. The di¤erence is that blue voters should double vote instead of voting red. Table 6 con…rms that the voters’behavior is in line with this prediction. Actually, the results are even stronger since the di¤erence between the blue and red voters is signi…cant not only for the second half of the sample but for the whole experiment (Mann-Whitney, z = 2.201, p = 0.028). 6.3 Individual Behavior We begin by describing individual behavior in plurality treatments with unbiased information. These cases do not allow for much variation among players: in treatment PL7 most 38 AV9 .8 .6 .4 .2 0 Frequency of Double-Voting 1 AV7 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Frequency of Voting the Signal Individual averages Symmetric Equilibrium Average Asymmetric Equilibrium Figure 4: Individual behavior in AV treatments with unbiased information. Each hollow circle in the graph corresponds to the observed frequence of play for one or more subjects. Its size represents the number of subjects who actually adopted that frequence of play. The red circle represents the average frequency of play observed, the orange triangle represents the symmetric equilibrium prediction and the green square represents the asymmetric equilibrium prediction. subjects voted sincerely throughout the entire experiment: 43.06% of the participants always did and 88.89% of the subjects voted sincerely in more than 75% of the occasions. As we saw, behavior is somewhat di¤erent in PL9, since voters always converged to a Duverger’s law equilibrium, although slowly. In the last half of the experiment, 88.94% voted for the color their group converged to. The case of approval voting is somehow more interesting. Figure 4 disaggregates behavior at the individual level in the last …fty periods of treatments AV7 (left panel) and AV9 (right panel). The horizontal axis plots the frequency of voting the signal and the vertical axis plots the frequency of double voting. Each circle in the graph corresponds to the observed frequency of play for one or more subjects. Its size represents the number of subjects who actually adopted that frequency of play: the bigger the circle, the higher the number of subjects. According to Theorem 2, subjects should only mix between voting the signal or double- 39 voting. If all subjects voted in this way, all the circles should be on the diagonal between (0,1) and (1,0). We observe that most circles are indeed on this diagonal but, instead of observing a large number of voters playing the predicted mixed strategy, we observe signi…cant heterogeneity with two opposite clusters: one that plays the pure strategy of always double voting and another one with subjects who always single vote their signal. The treatment e¤ect observed in Section 6.1.2 is mainly driven by a switch in the relative number of subjects in each cluster. This pattern pointed at the need to consider asymmetric strategies. Pushing the line of reasoning of Ahn and Oliveros (2011) further, allowing for asymmetric strategies can be interpreted as an extension of the group’s choice set, which may only increase expected welfare. Allowing some voters to specialize in double or single voting may produce signi…cant advantages. The challenge is to identify potential equilibria when allowing for asymmetric strategies. We deal with this issue in the next subsection. 6.3.1 Asymmetric Equilibria with Approval Voting In this subsection, we relax the assumption of symmetric strategies, ubiquitous as it is in the voting literature.41 That is, we do not force voters who receive the same signal to play the same strategy. The following proposition proves, for a broad set of parameter values (including the ones used in the experiment), the existence of at least one asymmetric equilibrium, i.e. in which voters play asymmetric strategies. We also characterize this asymmetric equilibrium: voters specialize independently of their signal in either single voting or double voting. That is, some voters always single vote and others always double vote. The “single voters” condition their behavior on their signal. If the signal structure is su¢ ciently unbiased, they all vote sincerely, i.e. A if signal sA and B if signal sB : If the signal structure is not su¢ ciently unbiased, then the voters receiving the less abundant signal (in expectation) vote sincerely whereas those who receive the more abundant signal mix between A and B in order to compensate for the bias. As discussed below, it appears that this equilibrium helps rationalize the behavior observed in the laboratory. Proposition 4 Suppose that q (a) = q (b), r (sA ja) r (sB jb) and V 2v. Any strategy pro…le satisfying the following conditions is an equilibrium: 41 There are noticeable exceptions in the literature on the Condorcet Jury Theorem. McLennan (1998), Ladha, Miller and Oppenheimer (2000) or Dekel and Piccione (2000) consider asymmetric equilibria. 40 1. 2nC n + 1 voters always double vote; 2. The rest of the voters single vote informatively with the following probabilities: 1v sA 1v sB where (A) = 8 > > > > > < > > > > > : r (sA ja) n nC n nC 1 r (sA jb) n nC r (sA ja) n nC 1 r(sA ja) r (sA jb) 1 if 1 if r(sA jb) r(sA ja) r(sA jb) r(sA ja) r(sA jb) n nC n nC 1 n nC n nC 1 > r(sB jb) r(sB ja) ; r(sB jb) r(sB ja) (B) = 1; 1v s ( ) is the probability that a single voter of type s plays action : Proof. See Online Appendix. Such an asymmetric behavior can be sustained in equilibrium because voters who specialize in single voting perceive the expected e¤ect of any given ballot on the …nal outcome di¤erently than voters specializing in double voting. In particular, “single voters” are pivotal only when A, B, and C receive exactly the same number of votes, whereas “double voters” are pivotal if either A is trailing behind by one vote or if it is leading by one vote. The best responses of these two groups of voters are thus di¤erent. We illustrate Proposition 4 as well as some properties of the asymmetric equilibrium through a numerical example: Example 5 Assume again (as in example 3, p26): n = 12; nC = 7; and q (sA ja) = q (sB jb) = 2=3. In the asymmetric equilibrium, 2nC n + 1 = 3 voters double vote, and the other 9 single vote their signal: Compared with the symmetric equilibrium, the aggregate level of double voting decreases from 36% to 25%, but this is enough to ensure that the Condorcet loser cannot win the election. Indeed, with three double votes, and nine single votes, one of the two majority alternatives must receive at least eight votes, i.e. strictly more than the Condorcet loser. Finally, the likelihood of choosing the best candidate increases from 82% in the symmetric equilibrium to 85:5%. The better aggregation of information holds even if we abstract from the cases in which the Condorcet loser wins in the symmetric equilibrium. This is because the (expected) number of voters who reveal their information, i.e. the single voters, is larger in the asymmetric equilibrium than in the symmetric one (9 vs. 7:68). 41 Treatment AV7 Minority Size Small AV9 Large Signal Double Vote Opposite Gray Signal Double Vote Opposite Gray Periods 1-50 70.92 22.22 6.50 0.36 47.08 45.67 6.86 0.39 Periods 51-100 71.94 24.36 3.69 0.00 43.33 51.64 4.97 0.06 Symmetric Equilibrium 64.00 36.00 30.00 70.00 - Asymmetric Equilibrium 75.00 25.00 41.66 58.33 Table 7: Aggregate voting behavior in AV treatments with unbiased information (Revisited). Gray refers to voting gray or a combination of gray and others. As shown in Table 7, such asymmetric equilibria appear to organize laboratory data better than the symmetric equilibrium. In treatment AV7, the predicted level of double voting in the asymmetric equilibrium is 25% compared to the observed 24.46%. The di¤erence is not signi…cant (Wilcoxon, z = -0.524 p = 0.6002). In the case of AV9, the predicted level of double voting is 58.33% compared to the observed 51.64%. The di¤erence is signi…cant (Wilcoxon, z = 2.201, p = 0.0277). The equilibrium described in Proposition 4 also makes an interesting prediction for the biased treatment AV7B. In the asymmetric equilibrium, the level of double voting is independent of the signal structure. This is not what we observe in the data (see Table 6). As shown in Section 6.2, we observe signi…cantly di¤erent behavior depending on the signal received. However, as with treatment AV7, there seem to be signi…cant di¤erences in behavior between subjects. Although it is beyond the scope of this paper, it might be useful to explore di¤erent types of asymmetric equilibria in this type of setting. 6.4 Welfare and Outcomes Previous multicandidate election setups used in laboratory experiments were based on theories which were inconclusive when it came to comparing welfare across voting mechanisms. The theoretical predictions of Myerson and Weber (1993) used in Forsythe et al. (1996), for instance, do not make a clear-cut prediction when comparing plurality and approval voting.42 A valuable feature of our common value setup is that it allows us to 42 Forsythe et al. (1993) does make the prediction that approval voting performs weakly better than the Borda Rule. In one of the equilibria, approval voting can replicate the Borda rule. In the other two equilibria, AV dominates Borda. In contrast, they cannot make clear predictions concerning plurality 42 Treatment PL7 PL9 PL7B AV7 AV9 AV7B Periods 1-50 136.70 101.15 169.85 167.00 146.75 188.90 Periods 51-100 138.50 147.80 171.95 183.95 168.95 192.50 Equilibrium Symmetric Asymmetric 154.87 156.20 178.76 179.65 189.80 164.70 181.10 193.58 ? Table 8: Average payo¤ and theoretical predictions. In the case of Plurality, equilibrium predictions refer to the equilibrium where experimental groups converged to. make clear predictions in terms of welfare: in equilibrium, the active voters’payo¤ should be strictly higher with approval voting than in plurality (Theorem 1, p.23). Table 8, columns 2 and 3, displays the average payment obtained by the subjects in each treatment, respectively for the …rst and second …fty periods. Comparing PV and AV treatments two by two, one can observe that realized payo¤s are systematically higher in AV treatments than in the associated PV treatment. All these di¤erences are signi…cant at a 1% con…dence level.43 It is also interesting to see the e¤ect of the size of the minority. The comparative statics are not clear in the case of plurality, given equilibrium multiplicity. The expected payo¤ should be strictly decreasing in nC in an informative equilibrium, whereas it is independent of nC in a Duverger’s Law equilibrium. In the case of our experiment, the expected payo¤ of Duverger’s Law equilibria is 155. Table 8 shows an interesting reversal when comparing PL7 and PL9: in the …rst half, the average payo¤ is higher in treatment PL7 than in treatment PL9, while the opposite is true for the second half. This is explained by the progressive switch towards a Duverger’s Law equilibrium in all independent groups under PL9, and the selection of the sincere voting equilibrium by all groups in PL7. The latter happens in spite of the fact that Duverger’s Law equilibria payo¤ dominate sincere voting in both treatments. On the other hand, the slow convergence process in the …rst half of PL9 treatments explains why payo¤s are higher in PL7 than in PL9 treatments for the …rst 50 periods. Across the entire experiment session, payo¤ is lower under PL9 than under PL7 (Mann Whitney test, z = 1.922, p-value = 0.0547). versus AV. 43 Mann-Whitney tests are: z = 2.882 and p-value 0.0039 for AV7-PV7, z = 2.722 and p-value = 0.0065 for AV9-PV9, and z = 2.913 and p-value = 0.0036 for AV7B-PV7B. 43 In the case of approval voting the theoretical prediction is unambiguous since there is a unique equilibrium. Both in the symmetric and asymmetric equilibria, the voters’payo¤s is predicted to be strictly decreasing with the size of the minority. As one can see from Table 8, this is what we observe in the data. These di¤erences are statistically signi…cant (Mann Whitney test, z = 2.242, p-value = 0.0250). Looking more closely at actual outcomes (see Appendix A3), one can check that this is due to a remarkable increase in the frequence of victory of the best candidate, combined with a drop in the frequence of victory of the Condorcet loser. This observation can be made in all two-by-two comparisons, including PL9 against AV9, because Gray won 10% of the times even in the second half of the PL9 experiment, due to slow convergence towards the equilibrium. 7 Conclusions In this paper we studied the properties of plurality and approval voting both theoretically and experimentally. We considered a case in which the majority is divided between two alternatives as a result of information imperfections and the minority backs a third alternative, which the majority views as strictly inferior. The majority thus faced two problems: aggregating information and coordinating to defeat the minority candidate. In plurality, we identify two distinct types of equilibria. First, in Duverger’s Law equilibria, all majority voters coordinate their ballots on a same majority alternative. This means that they can fully insure against a victory of the minority alternative, but they have to give up any possibility of aggregating information. Second, there also exists an informative equilibrium, in which majority voters split their ballots between the two majority alternatives. In equilibrium, the majority alternative with the largest expected voter share is the full information Condorcet winner. The drwaback of this this strategy is that the Condorcet loser may win the election. Interestingly, this equilibrium is not “knife edge”: strictly more than two candidates receive di¤erent but substantial vote shares. This theoretical …nding helps rationalize some empirical regularities in the literature that are oft-considered as supporting evidence for the lack of a “rational-instrumental” voting behavior. In approval voting, the structure of incentives is quite di¤erent: the threat posed by the minority candidate provides an incentive to double vote in order to reduce the relative 44 size of the minority. However, this does not produce indeterminate results: there is no equilibrium in which all majority voters double vote. Hence, voters in the majority keep aggregating information. We actually …nd that, in comparison with any potential equilibrium in plurality, there exists an equilibrium in AV that produces strictly higher expected welfare. Next, we characterize the equilibrium with approval voting in any interior equilibrium: each voter must adopt a mixed strategy of either single voting for her preferred candidate only, or double voting for both majority alternatives. We then tested our predictions with laboratory experiments. They revealed very interesting patterns and supported most of our theoretical predictions. We …rst studied setups in which information is symmetric across states: under plurality, we observed the emergence of both informative and Duverger’s Law equilibria. The former materialized when the minority was su¢ ciently small. By contrast, when the minority was large, all groups eventually converged to a Duverger’s Law equilibrium. Under AV, and as predicted, some voters double voted to increase the vote shares of both majority candidates but they kept aggregating information. We observed extremely strong welfare performance of AV in the lab: the subjects’ behavior allowed them to elect the full information Condorcet winner with a probability very close to what a social planner would have achieved after observing all available signals. However, to achieve this, subjects did not behave exactly as predicted by theory: some voters specialized in double or single voting, and the amount of double voting was lower than predicted. This led us to analyze the existence of asymmetric equilibria, which we found to have superior theoretical performance, and to better organize the data than symmetric equilibria. This might point to the need of further studying asymmetric equilibria in the information aggregation literature. The experimental results also o¤er systematic support for strategic voting behavior that is consistent with the model’s prediction and inconsistent with an assumption of sincere or naïve voting. Indeed, with both voting mechanisms, subjects learned how to aggregate information: they adopted an “insincere”strategy when we biased signals across states of nature. Secondly, we observed that, in line with theory, the subjects’ behavior adjusted quite di¤erently in plurality than in approval voting, leading to substantially higher realized payo¤s in AV treatments. We believe that this paper opens up many interesting theoretical and experimental 45 questions about multicandidate elections. Investigating how each voting rule performs in such a setup, how plurality and approval voting perform in a setup where majority voters have mix of private and common values, or investigate deeper what induces individual voters to adapt their strategy and collectively select one or another equilibrium are just some of the potential lines of research to pursue. 46