You are on page 1of 38

Playing for Your Own Audience: Extremism in Two Party Elections

Gbor Virg September 2004

Abstract This paper considers a two party election with costly voting and a singledimensional policy space. First, the parties choose platforms and then the voters decide whether to vote and who to vote for. Each party has its own constituency who does not observe the other partys action. The paper shows that in such a setting the parties always locate away from the median, since the voters who dislike the parties platform do not observe its policy choice and its own constituents like a policy choice that caters to their taste. As the number of potential voters increases the parties adopt more extreme platforms and as it becomes innitely large the parties locate at the two endpoints of the policy interval, full extremism occurs. As the cost of voting increases parties also tend to occupy more non-centrist locations, while if a voter is more likely to obtain more information of the other party he does not vote for, the opposite eect appears. Making voters more symmetrically informed about the two parties platforms increases the welfare of society, while asymmetric information acquisition by the voters is worse than no information acquisition at all in a private value framework.

I would like to thank Dan Bernhardt and Jozsef Molnar for their useful suggestions. University of Rochester, Department of Economics, 228 Harkness Hall, NY14627, e-mail: gvirag@troi.cc.rochester.edu

Introduction

An article in The Wall Street Journal 1 focused on George W. Bushes strategy in the 2004 presidential elections. The following is a quotation from the Journal article: " Since the advent of television brought presidential candidates into the voters living rooms, the general-election of both major parties have been targeted toward winning swing voters at the political center. Now, more than any modern campaign, the Bush eort, led by White House advisor Karl Rove, downplayed that goal in favor of a drive to wring more voters from the presidents committed core of supporters. Mr. Rove calls it a mobilization election." The article emphasizes that when voting is costly then it is important for a candidate to convince his own constituentcy to vote. However, it does not take it into account that such a policy may backre by mobilizing the voters of the opposing party hurt by a noncentrist platform of a candidate. To illustrate this point consider a standard model of policy choice model with costly voting. First, the two parties choose their platforms, then the voters decide whether to vote (incurring a cost K ) or stay at home.2 Assume that a voters loss function is the square of the distance between his ideal point and a candidates position. In such a model one cannot explain how extremism might occur. While a more extreme platform mobilizes the core supporters of a party, more is lost by mobilizing the supporters of the opponent. Thus the two parties still locate at the political center. Interestingly, no voter participates in equilibrium, since the two positions are the same and voting is costly. One way to avoid this result is to assume that a party is able to make its core supporters believe that he has taken an extremist platform, while making other voters believe that he is still at the center. While a party may certainly try to follow this strategy it may only work if voters are not rational. This paper oers a model with rational voters where the parties take extreme positions without being able to mislead the voters. The starting point is that if a voter is sure not to vote for a candidate, then he might not pay attention to what this candidates platform exactly is, while he may follow more closely what "his candidate" does when he decides whether to participate or not. An indirect evidence
1 2

"Bushs Unusual Campaign Plan: Focus on Base, Not Swing Vote", August 30, 2004 This setup was rst considered by Ledyard (1984).

supports this starting assumption. According to a New York Times article3 Fox News s news coverage achieved a higher rating than all of the three big network channels (ABC, NBC, CBS) for the rst time during the 2004 Republican National Convention. A report issued by the Reuters4 attribute it to the fact that conservative voters are more likely to watch Fox News than the general electorate and that they are also more likely to watch the Republican Convention than others: "Fox beneted from the fact that its audience is "ideologically very much aligned with the Republican Party"5 and apparently turned out in large numbers to watch the GOP convention than the Democrats, said independent network news analyst Andrew Tyndall." The same report observes that CNN with a more liberal audience compared to Fox News did much better even in absolute terms than Fox News during the Democratic National Convention: "In the 10 p.m. hour Thursday, which included President Bushs speech, ..., Fox averaged 7.3 million viewers... Atlanta-based CNN, ... which won the cable-ratings battle during the Democratic convention, attracted just 2.7 million viewers in the 10 p.m. hour on Thursday." If we assume that ones decision of which channel to watch is independent of what campaign event is exactly covered then the data supports our main assumption that voters are observing their own party more closely. Under this modelling assumption a party may take an extreme position, since it knows that the voters of the other party do not observe that position, while its core supporters like this deviation from the political center. When voters are more polarized those tendencies for extremism are even stronger since there are less voters in the center, so an extreme position carries less punishment from voters in the center. The most surprising result of this paper is that in a large electorate the two parties locate at the two extremes of the policy space. The reason is that in such a large election only a small fraction of the voters participate, otherwise the probability of being pivotal would converge to 0, and abstaining would be the best strategy. But a more extreme type has more incentive to vote if losses are quadratic (in the distance between the implemented
Bill Carter: "Networks Left to Reect on Weeks Poor Ratings," September 3, 2004 Steve Gorman: "Bush, Republicans Outpoll Kerry, Democrats on TV", available at http://story.news.yahoo.com/news?tmpl=story2&u=/nm/20040904/tv_nm/television_convention_dc 5 The article quotes research data published by the Pew Research Center this June: "Twenty-ve percent of Republicans say they watch CNN regularly, while 41 percent are Fox watchers, according to the study. Meanwhile, 44 percent of Democrats watch CNN, while 29 percent watch Fox."
4 3

policy and a voters ideal point).6 Then only the most extreme voters participate in large elections and a party has every incentive to cater to their needs by moving to the extremes. The model may not be the most appropriate to apply to large elections since - similarly to the literature initiated by Ferejohn and Fiorina (1974) - the model predicts that the participation rate converges to zero if the number of voters become innitely large. Then our result should be only interpreted as a possibility: If there are no counterbalancing eects the two parties locate at the very extremes of the policy space, depending on who participates in the elections. If the voters who show up are mostly extremist then the model predicts that the parties take extreme positions as well. In such a setting welfare losses readily arise. It is important to note that rational voters do not take it on face value what their own party claims about the position of the opponent, since it is always in a partys interest to picture the opponent as the most extreme. One instance when this problem does not arise is when the two candidates participate in debates and their positions can be compared. However, it is not clear that voters who are relatively far from the center actually watch those debates or take the time to study the program of the party they do not vote for. If that is indeed the case then the conclusions of the model become relevant in two party elections. There are three more comparative statics results worth mentioning. First, as the participation costs increase less voters participate, only the more extreme ones and thus the parties take more extreme positions. Second, if we allow a voter to obtain with probability p the action of the other party as well, then we show that the parties move toward the center, and when p = 1 (perfect information is acquired about the platforms) the parties locate near the median. Third, if we let moderate voters (located between t and 1 t) to observe the platforms of both parties then the parties move their platform closer to the center. More interestingly, the parties locate in such a way so that these moderate (and symmetrically informed) voters never vote in equilibrium. It might be argued that it is welfare improving to have news sources that are each individually unbiased compared to a situation where the dierent biases of the dierent news sources counterbalance each other. Indeed, if each voter subscribes to at most one
The assumption of a convex loss function may also capture that voters at the extremes are more likely to participate, because they have a lower cost of participation or equivalently they care more about what policy is implemented.
6

newspaper and that newspaper adequately covers only the position of one of the candidates then it creates a situation where voters are asymmetrically informed. Note again, that if it is well known that this newspaper supports one of the candidates then a rational reader does not pay attention to what it says about the other candidate, since it always has the incentive to claim that the other candidate is an extremist. Analyzing how media bias, costly voting and information acquisition may relate to political extremism is a topic of future research.

Literature review

The seminal works on voting are Downs (1957) and Hotelling (1929), who consider spatial voting models in a two-party system and establish the result oce oriented candidates move to the center of the political spectrum. Ferejohn and Fiorina (1974) analyze large voting games with two parties, xed platforms and costly voting. They show that the participation rate is converging to zero as the electorate becomes very large, the paradox of voting. The paper most closely related to this is Ledyard (1984), who analyzes a game of costly voting with strategic choice of platforms where the voters perfectly observe the actions of both parties and the cost of voting is stochastic. Under the assumption of concave utility functions (an assumption we maintain throughout the analysis) and other technical assumptions he shows that parties choose the same platform and no voter participates. As a special case of our model when voters have perfect information about the two platforms we obtain similar results. On the other hand, as we move further from this perfect information case in our model these results no longer hold. Borgers (2004) analyzes costly voting with two candidates and concludes that compulsory participation yields inferior welfare consequences compared to the voluntary voting case. The reason is that by making voting compulsory the probability that a certain voter is pivotal decreases and thus his utility from voting goes down as well. In our model there is an extra eect that may make compulsory voting benecial. If voting is voluntary then extreme types vote only and the two parties move away from the median. If voting was compulsory then both parties locate in the center, since now they have to convince all voters to vote for them. If the electorate is large and voting costs are small then we can show that compulsory voting is welfare increasing. 4

Turner and Weninger (2001) conducts an empirical analysis7 and shows that in the case of industry lobbying rms more moderate preferences are less likely to participate in a public meeting. This result is somewhat similar to the nding of the this model that more extreme voters are more likely to vote. Palfrey and Rosenthal (1983) analyze two party elections, where voting costs are known and the candidates have xed positions (like in Borgers (2004)). They nd multiple equilibria some of them are inecient. Our model nds only inecient equilibria in the sense that if voters observe the positions of the parties imperfectly then the parties locate away from the median imposing welfare costs on the society. The ineciency we concentrate on derives from the locational choices of the parties, while in Palfrey and Rosenthal (1983) this issue did not arise. This paper is also related to the literature on political extremism in two-party elections. Several papers show that if the two parties are at least partially policy motivated then they do not necessarily locate in the political center, if there is uncertainty about which voters will particiapte in the voting. Some examples are Calvert (1985), Roemer (1997) and Wittman (1983). A more recent paper is Ghosh (2002) who studies the behavior of policy motivated parties who choose their own candidates. In this case political extremism occurs in equilibrium as well. The starting point of this paper is rather dierent. We assume that parties are purely oce motivated and assume that voters do not observe policy positions perfectly in an asymmetric manner, i.e. voters tend to observe the action of their own party better. Then the factor that induces extremism is informational rather than preference determined.

The model

There are two parties, L and R who compete for votes. There are n bidders who decide whether to participate or not and who to vote for. Bidders are ex-ante identical and their types are independently and uniformly distributed on the [0, 1] interval. First, the two parties make their decisions about their locations on the same unit interval, i.e. vL , vR [0, 1]. After that the voters on the left (yL 0.5) observe party Ls decision and they decide whether to vote or not and who to vote for. Voters on the other side observe
7

Using the theoretical model of Osborne, Rosenthal, Turner (2000).

only the action of party R before voting. The party with the more votes win, in case of tie they employ a symmetric tie-breaking rule. The two parties are committed to implement their announced platform. We assume that the parties maximize the probability with which they win. Equivalently we can assume that they maximize their vote share or the expected number of votes they receive. Voters care only about the policy outcome, v. The utility of a voter with type y is (v y )2 K if he votes and (v y )2 if he does not. The expected utility of type yL from voting is 1 ((vR yL )2 (vL yL )2 ) Pr(piv ) K, 2 while the expected utility from abstaing is normalized 0. In the above formula Pr(piv ) denotes the probability that (given the strategies of the other players) a player is pivotal, i.e. without his vote there would be a tie or his own party, party L would be one vote down. We need to multiply this probability by
1 2

because a voter can change the probability of

his party winning only by 50% in those cases. Note, that the actual policy of his own party, vL appears in the formula, while vR denotes his belief of what the other party does. We analyze pure strategy, symmetric perferct Bayesian equilibria. The parties and the voters follow pure strategy. Symmetry amounts to the following: i) The two parties occupy a position at the same distance from the median: vL + vR = 1. ii) Voters with the same type use the same strategy (i.e. who to vote for if at all). iii) A voter with type y votes for the opposite party as type 1 y, if that type participates, otherwise he abstains as well. Finally, we require that voters play a Bayesian Nash-equilibrium in the voting phase of the game.

Preliminary analysis

In this Section we discuss some features of the model that makes it easier to follow the formal analysis in the next section. We make the assumption that voters who observe party Ls decision (those with low types) do not infer anything from that about the location of party R.
8

Note, that party L would never locate at vL > 0.5, because high types do not

observe this choice, so they do not respond, while low types punish such behavior. Thus in a symmetric equilibrium a voter is always closer to his own party (except the median who is at the same distance), and he only has to decide between voting for his own party or abstaining. The median voter never participates in a symmetric equilibrium, since he is indierent between the two platforms and voting is costly. Before moving on to the formal analysis a remark about the parties incentives is in order. A party cannot change the behavor of the constituent of the other party. The only thing it can change is the amount of participation his own constituents engage in9 . So, the strategic component between the two parties is completely elimiated in the sense that no matter what the opponent actually does the same decision is optimal for a given party. Thus a party chooses its location to maximize the number of votes it can get. But that decision rule maximizes its vote share and probability of winning as well. A party cannot induce too high level of participation, since then voters would not be pivotal often enough. As a consequence, no matter how well a party is located compared to its opponent (or rather compared to the belief about that location) a party cannot induce more than a certain share of the voters to participate. In a sense a too successful candidate is a victim of his own succes, the more voters vote for him the less likely that a certain voter is pivotal, so the less incentive an extra voter has to participate. This problem cannot be eliminated by moving closer to where the voters are, i.e. moving to the center.
Otherwise party L would always have an incentive to signal that his opponent is an extremist. If we allowed such inference of the voters then many equilibria would arise, because if a party deviates then its constituency might think that the opponent is at the middle and so they do not vote for their party. Also, these equilibria do not seem reasonable, since a given party always want to signal the same thing about its opponent, a not very satisfactory case for signaling. 9 Stricly speaking by locating at some unfavorable location the party can induce his own constituents to vote for the oher party, but since it is clearly suboptimal to do this we can rule this case out.
8

Analysis

We will concentrate on pure strategy, symmetric equilibria. Let us analyze the decision problem of the voters on the left side of the unit interval, the voters of the other party face similar problems. Unless otherwise mentioned the probability for being pivotal is for a voter who considers voting for party L. Let vL ( 1/2) be the position that party L occupies and let the voters believe that party R occupies position vR ( 1/2). A voter to the left of the median does not vote for R, because it is better for him to abstain. Then the only decision to make is whether to vote or not to vote. The next Lemma shows that extreme types have more incentives to vote under the assumed (convex) utility specication.
0 0 Lemma 1 If yL < yL 1/2 and yL votes for L then yL votes for L as well.

Proof. The utility from participating is the product of the probability of being pivotal and the gain that is made conditional on being pivotal. The probability component is the same for dierent types of the same player, only the gain component is dierent:
2 2 g = (vR yL )2 (vL yL )2 = (vR vL ) 2yL (vR vL ) 0 2 0 2 2 2 0 ) (vL yL ) = (vR vL ) 2yL (vR vL ) = g0 . < (vR yL 0 . So, if type yL has an incentive to vote then so does type yL

Thus in any such equilibrium the types who vote can be characterized by a cuto value, y such that types less than y vote for L, while types between y and 1/2 do not vote. A similar rule applies for types above 1/2. Let us study rst the case where there is only one voter to gain some intuition. In this case Pr(piv | vL , vR ) = 1 regardless of the cutpoints, (vL , vR ). First, we show that the parties do not move to the position of the median voter if there is positive participation costs. Indeed, suppose that vL = vR = 0.5 was true in an equilibrium. Then voing is not protable for any type, since the two positions are the same and voting costs preclude voting then. If K is not too big then 8

there is a protable deviation for both parties. Party L can move as close to 0 as needed to induce some participation from the extreme types.
10

The reason is that when a party

moves to the extreme it is noticed only by its own constituents, so costs of extremism are not imposed, since voters on the other side do not observe this deviation and so they are not motivated to vote. Let us nd the equilibrium in this simple one-voter game. Lemma 2 In any symmetric, pure strategy equilibrium if n = 1 and K is not too big, vL and vR are such that the left and righ cuto values satisfy the equations that yL = vL and yR = vR . Proof. First, we can show that full participation (i.e. yL = yR = 1/2) cannot be an equilibrium, since in a symmetric equilibrium type 1/2 does not gain from participation, so he always abstains. Also, if K is not too big, it is not an equilibrium if no type
0 participates. Suppose that vL 6= yL < 1/2. Then if party L changes his location to vL in a

way that it is closer to yL , then type yL achieves higher utility from participating, while the utility from non-participation remains the same. Then that type strictly prefers to
0 participate. But then some type higher than type yL will satisfy the indierence condition 0 2 0 0 2 ) (vL yL ) ] = 2K. [(vR yL

(1)

0 , which This means that party L can increase the number of his voters to include type yL

means that he will receive more votes, so the original location was not optimal. Then equation (1) and the above Lemma implies that (vR vL )2 = 2K. In a symmetric equilibrium vR = 1 vL . Then using the last two equations yields 1 2K 1 + 2K vL = and vR = . 2 2
10

This is possible if (0.5 0)2 = 1/4 > 2K.

We will make this assumption throughout, to rule out equilibria when noone participates.

For concreteness let K = 0.02. Then vL = 0.4 and vR = 0.6. To obtain a comparative statics as the number of voter changes we turn now to a two player game. The only dierence in the derivation involves the calculation of the probability of being pivotal. A voter (on the left) is pivotal if and only if the other voter does not vote for party L, which is with probability 1 yL : Pr(piv | L, yL , yR ) = 1 yL . Lemma 3 In any symmetric, pure strategy equilibrium if n = 2 and K is not too big, vL and vR are such that the left and righ cuto values satisfy the equations that yL = vL < 1/2 and yR = vR > 1/2. Proof. First, we can show that full participation (i.e. yL = yR = 1/2) cannot be an equilibrium, since in a symmetric equilibrium type 1/2 does not gain from participation, so he always abstains. Also, if K is not too big, in equilibrium some types must participate. The utility of type yL from participation if the (left) cuto is also yL is [(vR yL )2 (vL yL )2 ](1 yL ). Note, that this expression is strictly decreasing in yL , so the indierence condition [(vR yL )2 (vL yL )2 ](1 yL ) = 2K has a unique solution. This means that given the beliefs about yR and the observation of vL the cuto type, yL is uniquely determined.
0 Suppose that vL 6= yL < 1/2. Then if party L changes his location to vL in a way

that it is closer to yL , then type yL achieves higher utility from participating (if the other player employs cuto startegy yL ), while the utility from non-participation remains the same. Then that type strictly prefers to participate. But then some type higher than
0 type yL will satisfy the indierence condition 0 2 0 0 2 0 [(vR yL ) (vL yL ) ](1 yL ) = 2K.

(2)

10

This means that party L can increase his share of votes, which means that he will receive more votes, so the original location was not optimal. Then equation (2) and the above Lemma implies that (vR vL )2 (1 vL ) = 2K. In a symmetric equilibium vR = 1 vL . Combining the last two equations yields: (1 2vL )2 (1 vL ) = 2K. It is easy to show that vL is lower now than in the case with only one voter. For example when K = 0.02 then we have now vL = 0.38. This suggests, that the more voters there are the more extreme the parties political position become. The intuition is that with more voters each possible voter is less likely to participate (since he is pivotal less often). But then only extremists participate, so the parties need to cater to their tastes. With more than two voters calculations become more involved. The main diculty is that there might be complemetarities between the strategies of the dierent voters, i.e. if the others participate more then a voter might have more incentive to do so. These complementarities may give rise to multiple equilibria and it is also not uniquely pinned down what happens in the voting stage of the game. For this reason we introduce the next assumption, which helps avoiding the problems of multiple equilibria in the voting phase of the game. We assume that voters coordinate and vote in a manner that the maximum amount of participation is reached for the given observations and beliefs: Assumption: If the constituency of L observes policy position vL and believes that voters on the right side use cuto strategy yR and party R locates at vR then all types less than yL participate, such that i) yL = 0.5, if it is a Nash-equilibrium for all types to participate, i.e. 1 [(vR 0.5)2 (vL 0.5)2 ] Pr(piv | 0.5, yR ) K. 2 11

ii) or if the last condition does not hold then 1 yL = max[(vR z )2 (vL z )2 ] Pr(piv | z, yR ) = K z 2 or iii) yL = 0 if even this equation cannot be satised. For voters on the right side of the interval the appropriate corresponding assumptions are made. We can easily show that there exists a cuto value, yL that satises this assumption. While this seems a strong restriction we can show that even this does not lead to high participation levels if n becomes large, so the assumption can be viewed as a technical one that simplies the analysis. The reason we need the above assumption is to prevent voters from coordinating to a dierent voting equilibrium when the party changes its location slightly. Since it is actually the only Pareto ecient outcome for the voters (and the party prefers it obviously as well) in the voting stage (conditional on vL and their beliefs about vR , yR ) this assumption seems a reasonable starting point. The main bite of it is indicated by the following Lemma that is exactly the same as the previous ones, but now using also the assumption of maximum participation: Lemma 4 In any symmetric, pure strategy equilibrium that satises the assumption of maximum participation, vL and vR are such that the left and righ cuto values satisfy the equations that yL = vL and yR = vR . Proof. We start with the same steps as before applied to the equilibrium that has the highest participation in the voting stage. Note, that in a symmetric equilbrium type 0.5 never participates, so we use the second part of the assumption here.11 Again, we can show that as we move the locational choice closer to the original cuto value it becomes true that if everyone else uses cuto level yL then type yL has strict incentive to participate: 1 [(vR yL )2 (vL yL )2 ] Pr(piv | yL , yR ) > K 2
0 Then by continuity there there is yL > yL such that 0 2 0 2 1 0 ) (vL yL ) ] Pr(piv | yL , yR ) = K. [(vR yL 2
11

Case iii) when no type participates is ruled out if K is not too large.

12

But then by our assumption at least this new, higher level of participation is induced
0 when party L deviates to vL , which makes it a protable deviation.

We now prove that in general n voter games an equilibrium exists under mild restrictions. The main diculty lies in establishing that the choice of the left cuto values are strategic substitutes in the sense that if the other voters employ a higher cuto value, yL then a given voter has less chance for being pivotal and thus less incentive to participate. This feature eliminates the problems of multiple equilibria in the voting phase and guarantees that an equilibrium indeed exists: Theorem 1 If n is even or n is odd and it is large enough then there exists a unique equilibrium in the class of symmetric, pure strategy equilibria satisfying the assumption of maximum participation. Proof. Step 1 : First, we construct a candidate for such an equilibrium and in Step 2 we verify that it is indeed such an equilibrium. We will assume throughout that K is not too big, so that there is no equilibrium in which no types participate. Let party L and its constituency believe that party R uses strategy vR . Then to have a candidate for an equilibrium it must hold that yR = vR , so this is the belief we assign to party L and its voters. We are looking for a strategy, vL that satises two conditions: i) Type yL = vL is indierent between participating and abstaining if the other voters use cuto level yL . Formally, 1 [(vR vL )2 (vL vL )2 ] Pr(piv | yL = vL , yR = vR ) = K. 2 ii) Symmetry is satised: vL = 1 vR . Putting these conditions together yields T (vL ) = (1 2vL )2 Pr(piv | vL , 1 vL ) = 2K. To prove that there exists a solution to this equation note that T (0.5) = 0 < 2K and T (0) = 1 > 2K. 13 (3)

Moreover, we can show that there is a unique solution of the above indierence condition. To see this it is sucient to prove that for all vL Pr(piv | vL , 1 vL ) < 0. vL We can verify this inequality by exploiting the fact that if participation rates are higher then more other voters participate, which makes it less likely that one is pivotal. The details are in the Appendix. Step 2 : Now, we prove that the solution of equation (3) is an equilibrium satisfying the condition of maximum participation. We need to verify two conditions:
0 a) There is not a cuto level, yL > vL , s.t. 0 2 1 0 [(vR vL )2 (vL yL ) ] Pr(piv | yL , vR ) K. 2 0 0 , and cuto strategy yL > vL such that b) There is not a strategy vL 0 2 0 0 2 1 0 L[(vR yL ) (vL yL ) ] Pr(piv | yL , vR ) K. 2

Part a) rules out that the voters can coordinate on an equilibrium (in the voting stage) that achieves higher participation levels. Part b) rules out that party L can achieve higher turnout by deviating from vL . The following claim helps establishing the Theorem :
0 Claim: If for all yL > yL = vL = 1 yR = 1 vR 0 Pr(piv | yL , vR ) 0 0 yL

then conditions a) and b) are satised applied to the candidate equilibrium found in Step 1. Proof: By denition of the solution found in Step 1 : 1 [(vR vL )2 (vL vL )2 ] Pr(piv | vL , vR ) = K. 2
0 > vL then If yL 0 2 0 2 [(vR vL )2 (vL vL )2 ] > [(vR yL ) (vL yL ) ].

Also, by assumption
0 , vR ). Pr(piv | vL , vR ) Pr(piv | yL

14

Then
0 2 1 0 [(vR vL )2 (vL yL ) ] Pr(piv | yL , vR ) < K 2 follows, which establishes part a). To verify condition b) we prove that if all the other 0 0 and party L uses strategy vL , then it does not worth it even voters use cuto strategy yL 0 strictly prefers abstaining for type yL to participate. But then certainly a higher type, yL

as well. To see this consider the problem of such a type yL . His utility from participation is 1 0 0 [(vR yL )2 (vL yL )2 ] Pr(piv | yL , vR ). 2 Since vL = yL it follows that
0 yL )2 ]. [(vR yL )2 (vL yL )2 ] > [(vR yL )2 (vL

By assumption
0 Pr(piv | vL , vR ) Pr(piv | yL , vR ).

Thus K= 1 [(vR yL )2 (vL yL )2 ] Pr(piv | vL , vR ) > 2 2 0 2 1 0 [(vR yL ) (vL yL ) ] Pr(piv | yL , vR ), 2 which concludes the proof of the claim. Finally, we show that the assumption of the previous claim is valid:
0 > yL = 1 yR Claim: For any n that is even or odd and is large enough n for all yL 0 Pr(piv | yL , vR ) 0. 0 yL

Proof: The probability of being pivotal if the other voters use cutpoint strategies
0 , yR ) is characterized by (yL 0 0 0 , yR ) = Pr(piv | yL , yR , 0 voted) Pr(0 voted | yL , yR ) + ... + Pr(piv | yL 0 0 , yR , n 1 other voted) Pr(n 1 other voted | yL , yR ). + Pr(piv | yL

Let
0 0 , yR , i other voters voted) Pr(i other voted | yL , yR ) qi = Pr(piv | yL

15

and
0 pi = Pr(i other voted | yL , yR ).

Then
0 Pr(piv | yL , yR ) = p0 q0 + ... + pn1 qn1

and
0 Pr(piv | yL , vR ) p0 pn1 q0 qn1 = (q0 0 + ... + qn1 ) + (p0 0 + ... + pn1 ). 0 0 0 yL yL yL yL yL

In the Appendix we show that A = (q0 Showing that q0 qn1 + ... + pn1 )<0 0 0 yL yL is more straightforward. If k other voters voted and k is even then a voter is pivotal if B = (p0 and only if there is tie among the others. For k = 0 this has probability 1, i.e. q0 = 1. For the case when k > 0 the probability that some other voter votes for party L iconditional on voting at all is = Then the probability of being pivotal is qk =
0 yL . 0 yL + 1 yR

p0 pn1 + ... + qn1 ) < 0. 0 0 yL yL

k! k/2 (1 )k/2 , (k/2)!(k/2)!

0 0 which decreases in yL under our assumption that yL > 1 yR and thus > 1/2. A similar

calculation shows that the same result holds when k is odd. Then we have that q0 qk = 0 and < 0 for k > 0, 0 0 yL yL which implies that B < 0 must hold. As we argued earlier the more voters there are the smaller chance for any given voter to be pivotal given a cuto value, vL . This suggests then that as the number of voters increases the participation rate must go down. The following Theorem states this result formally: 16

Theorem 2 In the unique pure strategy, symmetric equilibrium that satises the assumption of maximum participation, the participation rate, vL is decreasing in n. Proof. Suppose that the solution of equation (3) is indeed an equilibrium (the conditions of the previous Theorem are sucient for this to happen). If it was true that Pr n (piv | vL , 1 vL ) > Pr for all n then it would follow that vL (n) > vL (n + 1). To verify condition (4) we just need to note that in a symmetric case (i.e. when yL + yR = vL + (1 vL ) = 1) it holds that q0 > q1 = q2 > q3 = q4 > .... = qk > ... and that those values do not depend on n. On the other hand as n goes up then (holding vL constant) it is more likely that more voters participate. More precisely we can show that for all k n
n n+1 +1 pn + ... + pn 0 + ... + pk > p0 k . n+1 (piv

| vL , 1 vL ),

(4)

Thus there is a rst order stochastic dominance relationship between the two probability distributions and since the sequence of q s is descending (and does not depend on the number of voters) this last inequality implies that Pr Pr
n n = pn 0 q0 + ... + pn1 qn1 + 0qn > +1 +1 = pn q0 + ... + pn 0 n qn .

n+1

Now, we characterize the equilibrium when n becomes arbitrarily large. The following Theorem shows that in the limit only types at the extremes participate: Theorem 3 In any pure strategy equilibrium as n we have yL 0, yR 1. 17

Proof. Step 1: In equilibrium the probability of being pivotal cannot converge to 0, since then there would be no incentive to participate. Then it is sucient to show that if the result of the above Theorem was false then the probability of being pivotal converged to 0. It is easy to see that if exactly one of the two cutos converged to the extreme then the probability of being pivotal certainly coverges to 0. So, we will suppose below that neither of them converges to the extremes. Moreover, yL and yR cannot converge to asymmetric values (i.e. yL + yR = 1 must hold), otherwise the probability of being pivotal would converge to zero as well. Step 2 : Then take a sequence of equilibria. We can nd a subsequence, such that
n yL y bL > 0.

Then for any k

Pr(number of other votes to L k) 0. By Step 1 there is symmetry between the two cuto levels, thus the following holds: Pr(number of other votes to R k) 0. Step 3 : If 2k other voters voted12 then a voter is pivotal if there is a tie if he did not vote. By symmetry (from Step 1 ) the probability of this is Pr k (piv ) = It is easy to show that Pr k (piv ) 0 as k . Step 4: It holds that Pr(piv ) = Pr(number of votes to each k) Pr(piv | number k) + + Pr(number of votes to each > k) Pr(piv | number > k). By Steps 2 and 3 as n we can nd suciently big values for k, such that Pr(number of votes to each k ) 0
12

(2k)! k!k! 22k

1 3 5 2k 1 ... . 246 2k

The case of odd number of voters can be handled similarly.

18

and Pr(piv | number of votes to each > k) 0. Step 5: Then Using the last step we conclude that the probabilit of being pivotal converges to 0 in contradiction to Step 1. The result that only very few voters participate in large voting games with costly voting is standard, see e.g. Borgers (2004). In addition to reproducing this feature, our model predicts that exactly the most extreme types participate in the limit. Putting together the last Theorem and Lemma and using that in the limit any equilibrium is close to a symmetric one (since yL + yR 1) yields the following suprising result: Proposition 1 In any pure strategy equilibrium that satises the assumption of maximum participation, as n becomes arbitrarily large the two parties locate at the extremes: vL 0, vR 1. Since in the limit only the most extreme types participate, catering to them means that the parties move also to the extremes. Thus if the constituents of the other party do not pay attention then the extreme voters will hijack the elections, inducing high welfare losses. Recent years have seen a polarization of the US media, new television channels and radio programs appeared that now more perfectly serve their audience with the appropriate slant in their coverage. This phenomenon may decrease the publics knowledge about the viewpoint of the other party. So, rather paradoxically the appearance of more news sources may lead to an electorate that is less precisely informed. Although the politicians themselves may present what they claim is the opinion of the other candidate, rational voters do not take these sources at face value, since they are avare that the politicians always have the incentives to make the voters believe that the opponent party is an extremist. An appealing feature of the equilibrium is that in equilibrium voters know the position of the party they do not vote for. Consequently, if they were given the opportunity to acquire information about the platform of that party it would not yield any benet to them, so they would not pursue that option even if there is a very small cost attached to it.

19

As the cost of voting, K increases the policies become more extreme: vL < 0. K The intuition is that as the cost goes up only more extreme voters participate and the parties trying to convince these voters on the margin to participate adopt more extreme platforms. This leads us to the observation that a higher level of participation is desirable for non-extreme voters (and the public in general), because in that case the marginal voter who participates is closer to the median and thus the parties adopt centrist platforms. This result is in contrast with the ndings of Borgers (2004) who shows that in a twoparty model with xed party platforms a higher participation probability has negative externality on the other voters, since it decreases their probability of being pivotal. Then making voting compulsory reduces welfare. In our model this eect is opposed by the one that higher participation rates induce the parties to choose more moderate policies. Indeed, suppose that K is very small and n goes to innity. If voting is compulsory then the parties locate in the median but the voters all have to incur the voting costs. Nonetheless, if these costs are small then it is not a big loss for them. On the other hand, if voting is voluntary then the parties choose extreme platforms and although the voting cost, K is avoided it does not make up for the loss induced by the implementation of an extreme platform. Compulsory voting is superior in this case.

Robustness of the results

In this Section I propose some modications of the above game to assess the consequences of more symmetric information structures. First, suppose that each voter now observes the policy of the other party with probability p > 0. When p = 1 we are back to the standard model with perfect information and positive cost of voting. Ledyard (1984) analyzed this game and proved that in every equilibrium both parties locate near the median and no voter participates. This suggests that as p increases our extremism result is weakened and the participation level goes down. We illustrate these results in a simple one-voter game, but all of the results hold for the general n-voter case, if p is small enough.13 Assume that each party maximizes its
13

The calculations for that case appear in Appendix 2.

20

probability of winning the elections.14 The case when p is high induces an equilibrium with no participation and centrist platforms, the results of Ledyard (1984). Focusing on the case when p is not too high we obtain the following result: Theorem 4 In a one-voter model if p p =
12K 1+2K

then in the unique symmetric, pure

strategy equilibrium with positive participation the following results hold: vL >0 p and yL < 0. p Proof. See Appendix. Making voters more informed in the sense that p goes up has a positive eect on welfare, since it induces parties to move closer to the median, thereby reducing the eciency loss induced by extreme policies adopted. It is very important that it is not more knowledge itself, but the reduction of the asymmetry in a voters information about the two parties that contributes to this favorable outcome. Maybe more interestingly, making voters more informed about their own parties leads to a reduction in welfare. Let us consider one more modication of the baseline model. Assume that p = 0, but that voters located on interval [t, 1 t] are moderates, who observe the positions of both parties. First, we consider the model with only one voter. Let (x, 1 x) denote the location of the two parties in the baseline model, where no voter observes the platforms of both parties, i.e. x satises (1 2x)2 = 2K. There are three cases, depending on the proportion of moderate voters: Case 1, t x: In this case there are not enough moderates to change the voting outcome. In the equilibrium of the baseline model (with locations (x, 1 x)) no moderate voter votes and one can easily show it remains an equilibrium even in the modied game. The key observation is that a party cannot increase the number of its voters in a way that was not available in the baseline model. So, if there are not enough moderate voters, who observe both platforms, then they are ompletely neglected, since extremist voters are more likely to vote.
14

We would obtain qualitatively similar results if a party maximized its share of the votes.

21

Case 2, x > t > 1 2

that both voters locate at the median, (1/2, 1/2). The equilibrium of the baseline model, (x, 1 x) is not an equilibrium either, since now some moderate voters vote and actually they are the voters on the margin who the parties strive to convince to vote or to abstain (e.g. party L tries to convince "marginal voters" on the left to vote and "marginal voters" on the right to abstain). Since, the marginal voters observe now both platforms there is an incentive for both parties to move close to the median. This argument bites, as long as the type on the left, who is indierent between participating and abstaining, vL is a moderate. Then it follows, that vL t must hold. But if vL < t in a symmetric equilibrium then it implies that vL < x and also that the marginal voter of the opponent is an extremist. In that case it is a good idea to take a slightly more extreme position for party L, since that does not induce more participation of voters on the right hand side, while it increases participation among voters on the left hand side.15 So, it cannot be an equilibrium either. Thus it must hold in equilibrium that vL = t, (5)

1 4

2K : In this case one can check that it is not an equilibrium

which means that the parties move away from the median exactly so much that no moderate voters participate, but all extremists do. Moving more to the extreme would start mobilizing some of the (moderate) voters of the other party (R). Moving towards the center would decrease the amount of votes party L receives, without inducing more abstantion among voters of party R, since in order to increase abstantion among those voters, some extremists must be convinced now to abstain, which is not feasible, because they do not
The reason it increases participation among voters of party L is that in this case y > vL and thus by moving closer to vL party L can increase the willingness of this type to participate. Since there is no threat of increased participation among voters of party R, so the argument of the baseline model applies to show that decreasing the distance between y and vL increases the winning probability of party L.
15

22

observe the position of party L. Then the only symmetric equilibrium, (y, 1 y ) is such that condition (5) holds, i.e. (1 y t)2 (y t)2 = 2K or y= K 1 > x. 2 1 2t

For concreteness take K = 0.02. Then x = 0.4 holds as we have shown in the baseline case. Now, let t = 0.3, that is there are enough moderate voters to change the equilibrium of the game. Then the last equation implies that y = 0.45, so parties move closer to the political center than in the baseline case where noone observes the platforms of both parties and party L locates at x = 0.4. One can show that the more voters there are who observe both platforms, the less extreme positions will be taken by the two parties: y 2K < 0. = t (1 2t)2
1 4

Case 3: t

1 2

platforms that it becomes an equilibrium of the voting game that both parties locate at the median and all voters abstain. The proof that it becomes an equilibrium is very simple: If the other party occupies that centrist position then party L by moving away from the center induces type 1 t, a voter of party R to participate before it can induce his most loyal voter, type 0 to do so. But then the mobilization eort backres and it is better not to induce any participation at all. Locating at (1/2, 1/2) is not the only equilibrium in this case, but it is more important for our purposes that if t is close enough to 0 (i.e. most voters observe both platforms) then all equilibria will converge to one where both parties are close to the median and no voter paticipates.16 Let us consider the general n-voter case. Let xn denote the (unique) equilibrium when no voters observe the platforms of both parties: (1 2xn )2 Pr(piv | xn , 1 xn ) = 2K. The results below can be proven by extending the above argument:
Other equilibria are in the form considered in case 2. The dierence is that in case 3 there are full abstantion equilibria as well.
16

2K : In that case there are so many voters observing both

23

Theorem 5 If there are n voters, p = 0 then the following results hold: i) If t xn then the equilibrium of the baseline model remains an equilibrium of the modied model, moreover this is the unique pure strategy, symmetric equilibrium satisfying the assumption q 1 of maximum participation. ii) If xn > t > 1 2K then there is a unique symmetric, 2 4

pure strategy equilibrium satisfying the assumption of maximum participation, (y, 1 y ) moderate voters there are, the more centrist positions the parties take y < 0. t
1 4

and it holds that vL = t, i.e. exactly the extreme voters participate. Moreover, the more

iii) If t

min( 1 2

1/2 and no voter participates. An interesting conclusion of the model is that moderate voters, who observe both platforms never participate in equilibrium, since a party makes it sure that the moderate voters of the other party reach a minimum utility that induces them not to participate. One can say then that parties move to the extremes to please their only own base only so much so that they do not mobilize the moderate (and informed) voters of the other party. Finally, consider a model more in the spirit of probabilistic voting. Suppose that the ideal points of the bidders are still distributed uniformly, but the cost of voting is not constant, but a random variable uniformly distributed on [c, c], independently of the ideal point of the voter. Let c (0, 1/2) and c be large enough, so that for no voter with this cost it is ever optimal to participate. The other assumptions are as in the baseline model. One can again show that extreme types are more likely to vote and there is a highest type, xL 1/2 such that only voters with ideal points less than this value vote with positive probability for parrty L. A similar cuto value exists among voters of party R. In the characterization of the equilibrium we take the strategy of part R as given and characterize the best reply of party L. Theorem 6 For any n that is even or odd and large enough, under the assumption of maximum participation there exists a unique symmetric equilibrium of the voting game, characterized by equation 1 (1 2vL )(1 4vL ) Pr(piv | xL = 1 2vL , xR = 2vL ) = c. 2 24 (6)

2K, x) then it becomes an equilibrium that both parties locate at

Proof. See Appendix 3. Similar results can be established as in the baseline case. In particular, the parties do not locate at the median in equilibrium. Also, as the number of voters increase the parties locate closer to the extremes of the unit interval and from (6) vL <0 c follows. Intuitively, the higher voting costs are the more extreme the participating voters are, and the parties have incentives to move closer to the participating voters, yielding a result of heightened extremism.

Conclusion

We considered a model of a costly voting with two parties, who choose their political platforms strategically. The main departure from the previous literature is that a partys constituency has better information about this party than about the opponent. Then each partys action is primarily observed by its own voters, which induces them to locate away from the median to motivate these voters to participate in the election. As the electorate becomes larger only more extreme voters participate and the parties to please them occupy more extreme positions. Also, as the cost of voting increases or the probability that voters observe the platform of the party they do not vote for decreases the same eect arises. The welfare implications of such extremism are negative if voters are symmetrically distributed and our metric for welfare is the average loss a voter faces by not being able to determine the policy himself. In this setup one can show that contrary to some previous results mandatory voting might yield higher welfare than voluntary, since in the mandatory system the parties face the usual logic of voting, which induces them to move to the center. An interesting implication of the model is that making voters more informed in an asymmetric manner can reduce welfare, so symmetric ignorance might be preferable in such a private value framework. An important future research avenue is to analyze how the bias in dierent media sources can contribute to this extremism result by providing less reliable information about the position of one candidate than about the other.

25

Appendix
Proof. Proof that
Pr(piv|vL ,1vL ) vL

< 0: In a symmetric case (i.e. when yL + yR = vL +

(1 vL ) = 1) it holds that q0 > q1 = q2 > q3 = q4 > .... = qk > .... On the other hand as vL goes up then it is more likely that more voters participate. Let
0 vL > vL and let pk (p0k ) denote the probability that k other voter participates if the cuto 0 value is vL (vL ). We can establish that for all k n 2

p0 + ... + pk > p00 + ... + p0k . Thus there is a rst order stochastic dominance relationship between the two probability distributions and since the sequence of q s is descending this last inequality implies that Pr = p0 q0 + ... + pn1 qn1 > Pr = p00 q0 + ... + p0n1 qn1 .

Proof. Proof that A < 0 : First note that p0 + ... + pn1 = 1, so pn1 p0 + ... + = 0. 0 0 yL yL The probability that k other voted is pk = Let
0 0 ), s = 1 r = (1 yR + yL ). r = (yR yL

(7)

(n 1)! 0 n1k 0 k (yR yL ) (1 yR + yL ) . k!(n 1 k)!

Then pk 1 1 = (kr (n 1 k )s) = (k (n 1)s) 0 yL pk pk Then pk 0 k k. 0 yL 26

If k is even then as we have shown above qk = while for k 1 odd: qk1 = Then k+1 qk+2 = 4(1 ) <1 qk k+2 and similarly qk+1 k+1 = 4(1 ) < 1. qk1 k+2 qk+1 k+1 < 1, = 2(1 ) qk k+2 but qk qk1 Let k be the largest k such that pk 0. 0 yL First, assume that k is even and that n 1 is odd. Let A = H + G, such that H = (q0 and G = (q1 Also, let =( pk pk+2 p0 p2 pn1 + 0 + ... + 0 ) + ( + ... + ) = 1 + ( 1 ) 0 0 0 yL yL yL yL yL 27 pk1 pk+1 p1 p3 pn2 + q3 0 + ... + qk1 ) + (qk+1 + ... + qn2 ). 0 0 0 0 yL yL yL yL yL pk pk+2 p0 p2 pn1 + q2 0 + ... + qk 0 ) + (qk+2 + ... + qn1 ) 0 0 0 yL yL yL yL yL = 2 1. (k 1)! k/21 (1 )k/2 . (k/2 1)!(k/2)! k! k/2 (1 )k/2 , (k/2)!(k/2)!

0 Also under the assumption that yL > 1 yR and thus > 1/2,

and =( Since q0 > q2 > ...qk > qk+2 ... > qn1 and pk pk 0 for k k and 0 > 0 for k > k 0 yL yL it follows that H qk ( pk pk+2 p0 p2 pn1 + 0 + ... + 0 ) + qk+2 ( + ... + )= 0 0 0 yL yL yL yL yL = qk 1 + qk+2 ( 1 ) = (qk qk+2 )1 + qk+2 . pk1 pk+1 p3 pn2 p1 + 0 + ... + )+( + ... + ) = 1 + ( 1 ). 0 0 0 0 yL yL yL yL yL

Similarly, G (qk1 qk+1 ) 1 + qk+1 . Using the two upper bunds from above and that + = 0 it follows that A = H + G (qk qk+2 )1 + (qk1 qk+1 ) 1 + (qk+2 qk+1 ). After using the above results about variables {qk } it is now sucient to show that 0 to obtain that the upper bound on the right hand side is negative and thus conclude n e = n 1 denote the number of other bidders. Then the following formula follows after simple calculations: =
h 1 = n er n + 0 0 the proof. Recall that r = (yR yL ), s = 1 r = (1 yR + yL ). For convenience let

h1 = n e [r n

(n e 1)! n (n e 1)! 2 n sr h2 + s r h3 .... 1!(n e 2)! 2!(n e 3)! (n e 1)! (n e 1)! hk hk1 + + ...] sk1 rn sk rn (k 1)!(n e k)! k!(n e k 1)!
h1 < 0. = n e(r s)n

n e! hk hk1 (ksk1 rn (n e k )sk rn )= k!(n e k)!

n e! h2 h3 (n e 2)s2 rn ) + ....+ (2srn 2!(n e 2)!

28

The last inequality follows because we assumed that n e = n 1 was odd. If n was odd, but large enough then it is easy to show that yR 1 must hold in our candidate equilibrium. participate, since then the probability for its constituents to be pivotal would converge

0 Then party L cannot induce non-extreme types (yL separated fom 0 in the limit) to

to 0, which cannot be the case. So, r 1 and s 0 for large enough n for any possible deviation, which guarantees that < 0 holds in this case as well. Finally, we check the case when k 1 is even. Let H, coresspond to the even components and G, to the odd ones as before. Following identical steps, but now using that k, the critical value is odd we obtain the following upper bound for A : A (qk1 qk+1 )1 + (qk qk+2 ) 1 + (qk+1 qk+2 ). After observing that qk+1 > qk > qk+2 we can follow the same argument as in the other case to complete the proof. Proof of Theorem 4: Proof. Suppose that in the equilibrium the parties take positions (vL , vR ). Then in this simple one-voter case the following cuto values apply in equilibrium17 : yL = and yR =
2 2 vL 2K vR 2(vR vL ) 2 2 vR vL + 2K . 2(vR vL )

0 Suppose that party L deviates to position vL . Then the cuto value for his constituents

become
0 yL =

The cuto value for the other voters if they do not observe this deviation (with probability 1 p) remains yR . The cuto value for those who observe it changes to
0 zR
17

2 0 2 vR (vL ) 2K . 0 2(vR vL )

(8)

2 0 2 (vL ) + 2K vR = . 2(vR vL )

(9)

Note, that in equilibrium the voters are correct about the positions, so it does not cahnge their decision if they actually observe the position of the other party or not.

29

The probability for party L to win is Pr(L wins) =


0 yL 0 0 + pzR (1 p)yR 1 yL 0 0 0 + . + [p(zR yL ) + (1 p)(yR yL )] = 2 2 2

0 0 0 + pzR with respect to vL . Using equations (8) and (9) Then party L has to maximize yL

and that in equilibrium vR + vL = 1 the solution to this problem is characterized by (1 2vL )2 = 2K which implies that vL > 0. p Then substituting vL from equation (10) and using that vR = 1 vL implies
0 yL = yL = 2 (vL )2 2K vR + vL K vR = = 2(vR vL ) 2 vR vL K K 1 1 = . = 2 1 2vL 2 1 2vL

1p , 1+p

(10)

Then yL vL yL = < 0. p vL p We can check that under our condition that p p it holds that yL 0 in equilibrium, but if p becomes larger then the above equations would yield a negative participation rate, yL , so in that case there would not be any symmetric equilibrium with positive turnout.

Appendix 2

We analyze symmetric, pure strategy equilibria in the case when p > 0 as well. The following assumption is made as an equilibrium selection device: Assumption (Help the deviator): Suppose that in the candidate equilibrium party L locates at vL , party R locates at vR and the voters choose cuto strategies (yL , yR ). Then if
0 party L deviates to policy position vL then the voters all choose the Bayesian equilibrium

in the voting stage such that it maximizes the vote share of the deviator. We call this assumption the Assumption of Helping the Deviator. While with a positive p it is no longer true that this is the only Pareto-optimal equilibrium for the voters who
0 observe vL , we can employ a forward induction type of argument to justify it. Party L

30

presumably deviated to increase his vote share, so it might be reasonable to require that whenever possible the voters should coordinate to satisfy the partys needs. Let us derive now a necessary condition for such a symmetric equilibrium under our last assumption. So, consider a candidate equilibrium with vR = 1 vL and yR = 1 yL . Since in a pure strategy equilibrium beliefs are correct for sure, so the cuto strategies used do not depend on whether one observes the action of the other party. Suppose that
0 0 party L deviates to position vL . Then its constituents would choose position yL . Other 0 0 voters use strategy zR if they observe vL and stick to strategy yR if they do not. It is easy 0 0 0 0 to see that if vL is close to vL then there exist (unique) values (yL , zR ) close to (yL , yR )

such that these strategies together with yL constitue an equilibrium in the voting game, that does not necessarily satisfy the Assumption of Helping the Deviator. If such an equilibrium increases the expected vote share of party L above 1/2 (the vote share in the symmetric situation) then the Assumption of Helping the Deviator imply that voters coordinate on an equilibrium that gives at least this big expected vote share for L, so we found a protable deviation. Thus, any voting equilibrium nearby must give expected vote share that is not more than 1/2. Moreover, one can prove that such a nearby equilibrium values are dierentiable functions of the position of party L. Then if this equilibrium would bring an expected vote share strictly less than 1/2 then choosing the opposite deviation would yield an expected vote share above 1/2, a contradiction. So, in that neighbourhood the expected vote share must have a 0 derivative with respect to vL . This condition can be formalized as follows:
0 0 [p(1 zR ) + (1 p)(1 yR )] yL = . 0 0 vL vL

Thus p

0 0 yL zR + = 0. 0 0 vL vL

(11)

By denition of an equilibrium it follows that Pr(piv |


0 0 2 0 0 2 L, vL )[(vR yL ) (vL yL ) ]=

0 0 0 2 0 )[(zR vL ) (zR vR )2 ] = 2K. = Pr(piv | R, vL

Since
0 0 Pr(piv | L, vL ) = Pr(piv | R, vL )

31

locally, since expected vote shares remain 1/2 for each, it follows that in this neigbourhood:
0 2 0 0 2 0 0 2 0 ) (vL yL ) ] = [(zR vL ) (zR vR )2 ]. [(vR yL

After some algebra we obtain


0 0 0 yL + zR = vR + vL .

Thus

0 0 yL zR + 0 = 1. 0 vL vL

(12)

Using equations (11) and (12) we obtain


0 1 zR = 0 vL 1p

(13)

and

After dierentiating equation

0 yL p = . 0 vL 1p

(14)

0 0 2 0 0 2 )[(vR yL ) (vL yL ) ] = 2K, Pr(piv | L, vL 0 evaluating it at vL = vL = 1 vR and using formulas (13) and (14) we obtain

(15)

Pr 0 (piv | L, vL )(1 2vL )(1 2yL )+Pr(piv | L, vL )[2vR

p p 2(vL yL ) 2vL ] = 0. 1p 1p

Let H (y ) be the probability of being pivotal if fraction y votes for L, fraction y votes for R and (1 2y ) abstains.18 By equation (14) it holds that this fraction behaves in such a way that y p . = 0 vL 1p

With this new notation we can write

Pr(piv | L, vL ) = H (yL ) and Pr 0 (piv | L, vL ) = H 0 (yL )


18

p . 1p

This can be written as (assuming n is even) H (y ) = y 0+0 (1 2y )n + y 1 (1 2y )n1 n! n! + n(n 1)y 2 (1 2y )n2 + ... + yn. (n 1)! (n/2)!(n/2)!

32

Then our necessary condition for having a symmetric equilibrium becomes: pH 0 (yL )(1 2vL )(1 2yL ) = H (yL )[2(1 p)(vL yL ) 2p(1 2vL )]. From this it is obvious that if p = 0 then vL = yL , but when p > 0 then vL yL > 0 must hold. Equation (15) can be rewritten as H (yL )(1 2vL )(1 2yL ) = 2K. Using this equation and equation (16) yields pH 0 (yL )2K = H 2 (yL )[(1 p)(1 Note that at p = 0 this simplies to 2K 2K 2yL ) 2p ]. (18) (1 2yL )H (yL ) H (yL )(1 2yL ) (17) (16)

H (yL )(1 2yL )2 = H (yL )(1 2vL )2 = 2K, our previous equation for the baseline model. Let (yL ) = H 2 (yL )[(1 p)(1 It is easy to show that 2K 2K 2yL ) 2p ]. (1 2yL )H (yL ) H (yL )(1 2yL ) 0 (yL ) < 0. For p = 0 condition (18) has only one solution then. Using continuity arguments it follows that if p is close enough to 0 then it is still true. Moreover, <0 p implies that for small p (where the left hand side of (18) is nearly constant in yL ) it holds that yL <0 p and using equation (17) and the fact that H 0 (yL ) < 0 implies that vL > 0. p Since second order conditions held strictly for the p = 0 case a continuity argument implies that if p is small enough then party L does not have a protable non-local deviation either, i.e. the (unique) solution of (18) is indeed an equilibrium of our game. 33

Appendix 3
Proof. Fix vR and analyze the optimal choice for party L. The denition of xL implies

Proof of Theorem 6: that this type with cost level c is indierent between voting and abstaining: 1 ((vR xL )2 (vL xL )2 ) Pr(piv | vL , vR ) = c.19 2 (19)

Let P be the probability of voting for the constituent of party L and introduce the variable Pr(piv | P ) to denote the probability that a voter of L is pivotal taking vR as given. Formally, P =
xL Z 0

((vR x)2 (vL x)2 ) 1 Pr(piv | P ) c 2 dx. cc

(20)

At the optimal choice (vL ) of L the rst order condition implies that P = 0. vL Note, that at such a point Pr(piv | P ) = 0. vL Using equations (19), (20) and (22) equation (21) becomes
xL Z 0

(21)

(22)

(vL x)dx = 0 xL . 2

or vL =

Using this last equation and symmetry, vR = 1 vL and xR = 1 xL implies together with equation (19) that in equilibrium 1 (1 2vL )(1 4vL ) Pr(piv | xL = 1 2vL , xR = 2vL ) = c. 2
19 Implicitly we use the assumption of maximum participation here to rule out cases in which voters of party L coordinate on abstention, which is always an equilibrium in the voting phase of the game.

34

This equation has a unique solution in vL under our assumptions, so there is one candidate equilibrium.20 We only need to show that the global second order conditions are satised for party L to conclude the proof. Let e = P (c c). P e Pr(piv | P ) P K, vL P

Dierentiating (20) yields e P = vL


xL Z 0

(vL x) Pr(piv | P )dx +

where K is a positive number. Then Pr(piv | P ) 1


xL Z

e P = vL

(vL x)dx .

0 Pr(piv|P ) K P

If we start at a candidate equilibrium and want to induce a higher participation of the voters of L then we have shown in the proof of existence in the baseline model that Pr(piv | P ) < 0, P which implies that P sgn = vL
20

xL Z 0

(vL x)dx = xL (

xL vL ), 2

To show this it is sucient to have that Pr(piv | xL = 2vL , xR = 1 2vL , vL , vR = 1 vL ) < 0. vL

This is equivalent to (as seen in the proof of the baseline case): P (xL = 2vL , xR = 1 2vL , vL , vR = 1 vL ) > 0. vL After writing P up we have: P =
2 vL Z 0
R +vL 2x c( vv 1) R +vL 2xL

cc

dx =

2 vL Z 0

2x c( 11 4vL 1)

cc

dx,

which after dierentiation yields the result.

35

where the sign = means that the two expressions have the same sign. Then it is sucient to prove that if vL is increased above its level at the candidate equilibrium then vL > xL 2

sgn

holds. (A similar argument will work when we decrease the level of vL .) Let us start now from the candidate equilibrium where Pr(piv | P ) =0 vL and thus (because of (19)) ((vR xL )2 (vL xL )2 ) = 0. vL Using that vL = we obtain xL xL vL xL /2 = = < 2. vL vR vL 1 xL Thus if we increase vL then we locally have vL > and thus P < 0. vL So, P is decreasing in vL to the right of the level of vL of the candidate equilibrium. Suppose that for some larger vL P = 0. vL There vL = and thus (using that xL < 1/2 < vR ) xL vL xL /2 xL = = < 2. vL vR vL vR xL /2 xL 2 xL 2 xL xL , vR = 1 2 2

But this means again that

P <0 vL must hold to the right of such vL , and consequently the global second order conditions are satised. 36

References
[1] Borgers, T. (2004): "Costly Voting," American Economic Review, 94(1), pp. 57-66 [2] Calvert, R. L. (1985): "Robustness of the Multidimensional Voting Model: Candidate Motivation, Uncertainty and Convergence," American Journal of Political Science, 29, pp. 69-95 [3] Downs, A. (1957): An Economic Theory of Democracy, New York, Harper and Row [4] Ferejohn, J. and M. Fiorina (1974): "The Paradox of Not Voting: A Decision Theoretic Analysis," American Political Science Review, 68(2) pp. 525-536 [5] Ghosh, P.(2002): "Electoral Competition, Moderating Institutions and Political Extremism," Mimeo, University of British Columbia [6] Hotelling, H. (1929): "Stability in Competition," Economic Journal, 39, pp, 41-57 [7] Ledyard, J. O. (1984): "The Pure Theory of Large Two-Candidate Elections," Public Choice, 44(1), pp. 7-41 [8] Osborne, M., Roshental, J. and M. Turner (2000): "Meetings With Costly Participation", American Economic Review, 90(4), pp. 927-43 [9] Palfrey, T. and H. Rosenthal (1983): "A Strategic Calculus of Voting," Public Choice, 41(1), pp. 7-53 [10] Palfrey, T. and H. Rosenthal (1985): "Voter Participation and Strategic Uncertainty," American Political Science Review, 79(1), pp. 62-78 [11] Turner, M. and Q. Weninger (2001): "Meetings With Costly Participation: An Empirical Analysis," Review of Economic Studies, forthcoming [12] Roemer, J. (1997): "Political-economic Equilibrium When Parties Represent Constituencies: The Unidimensional Case," Social Choice and Welfare, 14, pp. 479-502 [13] Wittman, D. (1983): "Candidate Motivation: A Synthesis of Alternatives," American Political Science Review, 77, pp. 142-157

37

You might also like