Full text of DOI:10.1162/003355300554836

ELECTORAL COMPETITION UNDER THE THREAT
OF POLITICAL UNREST*
MATTHEW ELLMAN AND LEONARD WANTCHEKON
We study elections in which one party (the strong party) controls a source of
political unrest; e.g., this party could instigate riots if it lost the election. We show
that the strong party is more likely to win the election when there is less
information about its ability to cause unrest. This is because when the weak party
is better informed, it can more reliably prevent political unrest by implementing a
‘‘centrist’’ policy. When there is uncertainty over the credibility of the threat,
‘‘posturing’’ by the strong party leads to platform divergence.
I. INTRODUCTION
This paper examines how voting and the outcomes of majority
rule elections are affected by factors beyond the control of the
winning party. We present a model that helps to predict how
threats against the electoral process influence (1) platform choice
by political parties, (2) which party wins the election, and (3) the
policy outcomes finally implemented. In a democracy where the
majority winner sets policy, dissatisfied groups, including groups
outside the electoral process, may still be able to interfere with
that policy. A losing party may organize a coup; voters may riot;
unions may go on strike; investors may take their capital abroad;
terrorists and foreign powers may threaten disruption and loss of
life; foreign powers may withdraw aid or even impose a trade
embargo. We analyze the effect of these nonelectoral factors on
electoral outcomes.
We consider a majority-rule electoral competition between
two political parties with ideal policies lying to either side of the
median voter ’s ideal policy. We allow for only a one-dimensional
policy space, but the disruption possibility leads to a two-
dimensional outcome space.
As evident in the examples given above, the actors threaten-
* We thank Timothy Feddersen, Geoffrey Garrett, Oliver Hart, Mathias
Hounkpe, Nien-heˆ Hsieh, Elena Krasnokutskaya, Eric Maskin, Steve Morris,
Roger Myerson, Martin Osborne, Ben Polak, Andrei Shleifer, Christopher Udry,
Asher Wolinsky, two anonymous referees, and an editor of this J ournal for
comments and suggestions. We are also grateful for helpful comments from
seminar participants at the University of Guelph, the University of Michigan,
Harvard University, New York University, the University of Western Ontario, and
Yale University. Matthew Ellman would like to thank the Alfred P. Sloan
Foundation for its generous fellowship support. We are responsible for any
remaining errors.
2000 by the President and Fellows of Harvard College and the Massachusetts Institute of
Technology.
௠
The Quarterly J ournal of Economics, May 2000
499

500
QUARTERLY J OURNAL OF ECONOMICS
ing to interfere with the winner ’s policy can be inside or outside
the electoral process. We focus on the case where there is only one
such actor and where the policy preference of the threatening
actor is aligned with one of the political parties. That party’s
policy proposal faces no threat of disruption and is called the
‘‘strong’’ party. The other party will be called the ‘‘weak’’ party. We
will say that the ‘‘strong’’ party is ‘‘directly strong’’ if it controls the
unrest itself (e.g., when the threatening actor and the strong
party are one and the same and ‘‘inside’’ the political process). It is
‘‘indirectly strong’’ if the threatening actor is a distinct agent, an
‘‘outsider ’’ to the political process.1 The distinction between direct
and outsider control is important only in Sections V and VI where
we allow parties to commit to a platform.2
‘‘Outsider control’’ is the most common case. Examples in-
clude terrorist activities by private militias, labor strikes by
unions, and capital flight by investors responding to high taxes or
restrictive regulations. Direct control is applicable when one
political party controls the military or has a private militia. For
instance, in the 1970 and 1980 elections in Zimbabwe, voters and
political parties knew that Robert Mugabe controlled the military
and could organize a coup if he lost an election. Under both direct
control and outsider control, the threatening actor accepts the
winner ’s policy if it is close enough to its own preferred policy, but
otherwise rejects the policy by initiating political unrest or
causing disruption.
Our first result is surprising: when parties cannot commit to
the policy, they will implement upon winning, the ‘‘weak party’’
may benefit from its own weakness. When the weak party wins, it
will always choose policy that is a compromise with the threaten-
ing actor (be it a party or an outsider). A winning strong party will
always choose to implement its ideal policy: this incurs no risk of
interference even under outsider control. However, when the
weak party does not know how much compromise is necessary,
voters may not trust the weak party to limit the risk of disruption.
Our second result demonstrates how uncertainty captures
distrust: when the threatening party’s propensity to interfere is
1. In this ‘‘outsider control’’ case, the ‘‘strength’’ derives from the fact that the
threatening actor has the same ideal point and its threats ‘‘support’’ the strong
party’s interests.
2. Note that there are two reasons why a party might cause disruption: to
obtain direct benefits of office and to shift the policy outcome. In the outsider
control case, only the policy shifting reason is valid, but our analysis applies
equally given the assumption we make.

ELECTORAL COMPETITION
501
privately known by the strong party, voters may support an
extremist ‘‘strong party’’ to avoid the risk of disruption. This is
illustrated by El Salvador ’s first post-civil war election (see
Section VII and Wantchekon [1999]).
When political parties can commit to their platforms prior to
elections, we ask whether the platforms will converge, and which
party will win the election. We find that parties converge to the
median voter ’s ideal point when the threat is minor. When the
threat is serious, they converge at the strong party’s reservation
policy; that is, the policy outcome at which the strong party is
indifferent between disrupting the political process and accepting
the winner ’s policy proposal. Furthermore, in the case of direct
control with a direct benefit to winning the election, our model
predicts that the strong party wins if the threat is serious. This is
so because the strong party can offer a policy that is a little closer
to the median voter ’s ideal point than the weak party can offer
while credibly avoiding the risk of political unrest. The strong
party would lose the direct benefits from winning if it creates
unrest to undo its own policy. Thus, the strong party can credibly
promise a more centrist undisrupted policy than can the weak
party.
Our model also allows us to predict when platforms will
diverge in an election. When there is uncertainty over the
credibility of the interference threat, ‘‘posturing’’ by a strong party
with private information about the risk of unrest can lead to
platform divergence. Such platform divergence occurs because the
strong party knows that voters are willing to trade off policy
preference against political unrest.
The paper is organized as follows. Section II describes the
basic model, which involves two competing parties that differ in
their ability to interfere in the political process. Section III
analyzes the equilibrium behavior of the basic model, assuming
that the two parties’ costs of political unrest are exogenous and
known by everyone. Section IV shows the equilibrium outcome
when parties’ costs of disrupting the political process are private
information. Sections V and VI extend the basic model to the case
where parties can choose policy platforms before the election.
With known party costs of disruption, platforms tend to converge.
However, reintroducing private information in Section VI, we can
explain divergent platforms even when the direct benefits from
winning the election are arbitrarily large. Section VII discusses

502
QUARTERLY J OURNAL OF ECONOMICS
four applications of the model. Section VIII concludes, and all
proofs are in the Appendix.
II. THE MODEL
We consider an electoral game involving two political parties
and a large number of voters. One set of actors can induce political
unrest as a response to electoral defeat. The threat of unrest or
‘‘interference’’ may be controlled by a political party or by an
outsider. The game has four stages. First, parties choose plat-
forms. Then, there is an election. After the election, the winner
sets a policy. Finally, the losing party or the outsider either
accepts that policy or responds with disruptive interference. We
will first study the case where parties cannot commit to their
platforms (in which case the first stage of the game is irrelevant),
and then we will move to the case where they can commit to their
platforms.
Preferences
We assume that voters are risk neutral and have single-
peaked political preferences represented by an ideal point, ␪ ʦ
[Ϫ2,2], with constant marginal disutility for deviation from this
ideal point. We assume that the median voter denoted M has an
ideal point at ␪ ϭ 0. Furthermore, a voter of type ␪ has a fixed
negative payoff, Ϫc , whenever political unrest takes place.
␪
Under direct control, one of the two parties, s, is strong and
earns Ϫcs if it initiates political unrest. The other party, w, is weak
and gains a lower payoff Ϫcw if political unrest occurs (Ϫcw Ͻ Ϫcs).3
When unrest is controlled from outside the electoral system, the
outsider, denoted z, gets a payoff Ϫcz by initiating unrest. The
weak party and the strong party then get lower payoffs, Ϫcw and
Ϫcs, respectively.
Like voters, parties have single-peaked policy preferences.
The weak party has its ideal point at ϩ2, while the strong party
and the threatening outsider have their ideal points at Ϫ2.4 We
denote by i, the identity of the winning party and by y its policy
proposal. We denote by r the probability of interference by the
3. In the case of a military coup by the strong party, one could argue that Ϫcs
would be positive because this party takes power and gets the direct benefits of
office. However, we will assume that ‘‘interference’’ generally involves some costs
for all parties.
4. This assumes that no voter is more extreme than either party, but this
simplification can easily be removed.

ELECTORAL COMPETITION
503
losing party or the outsider. Thus, the outcome of the election can
be characterized by the triple: ( y,r,i).
Payoffs
Players’ utility functions depend on (1) the distance between
their ideal points and the policy outcomes, (2) the cost of political
unrest and (3) the probability of political unrest. The payoff for a
voter with ideal point ␪, U ( y,r,i) is given by
␪
Ϫ[(1 Ϫ r) · 0 y Ϫ ␪0 ϩ r · c ].
␪
Political parties have preferences of the same form, but they
may also value winning per se. That is, we allow for the addition of
a benefit, R, to the winning party’s payoff, conditional on the
winner not facing interference.5 But this direct benefit is irrele-
vant when parties cannot commit to a platform, so we set R ϭ 0
without loss of generality until Section V. Noting that 0 y Ϫ
(Ϫ2)0 ϭ 2 ϩ y and 0 y Ϫ 2 0 ϭ 2 Ϫ y,6 the strong party’s payoff,
Us( y,r,i), is given by
Ϫ[(1 Ϫ r) · ( y ϩ 2 Ϫ R · Is) ϩ r · cs],
where Is ϭ 1 if s wins the election and Is ϭ 0 if it loses. Meanwhile,
the weak party’s payoff, Uw( y,r,i), is
Ϫ[(1 Ϫ r) · (2 Ϫ y Ϫ R · Iw) ϩ r · cw].
Again, Iw ϭ 1 if w wins, and Iw ϭ 0 otherwise. For z, the payoff
Uz( y,r,i) is
Ϫ[(1 Ϫ r) · ( y ϩ 2) ϩ r · cz].
Note that if there is no unrest, each party is guaranteed a
payoff of at least Ϫ4, and the median voter is guaranteed a payoff
of at least Ϫ2: in the extreme case when one party wins the
election and implements its ideal point, its payoff is R, its
opponent’s Ϫ4, and the median voter ’s Ϫ2.
We assume that (1) political unrest is more costly to the weak
party than to the threatening actor, and (2) interference by the
threatening actor imposes such large costs on all other parties
that they prefer any policy to one which always leads to interfer-
5. To simplify, we assume that payoffs from interference do not depend on
which party won.
6. Note that 0 y Ϫ (Ϫ2)0 ϭ 2 ϩ y and 0 y Ϫ 2 0 ϭ 2 Ϫ y hold for y ʦ [Ϫ2,2]. y ʦ
[Ϫ2,2] can be justified by using the fact that neither party uses dominated
strategies.

504
QUARTERLY J OURNAL OF ECONOMICS
ence. This assumption simplifies the analysis and is appropriate
to major threats such as coups and widespread strikes, terrorist
activity, and capital flight,7 and (3) the threatening actor has costs
sufficiently low that it interferes with the electoral outcome if the
weak party does not compromise toward Ϫ2; that is,
ASSUMPTION A1. cw Ͼ 4, cs ʦ (0,4);
ASSUMPTION A1Ј. cw,cs Ͼ 4, cz ʦ (0,4).
The first part of each assumption captures the fact that the
weak and the indirectly strong parties have a vested interest in
maintaining peace and in avoiding disruption. The second part
captures s or z’s incentive to interfere after the election. We
further assume that the cost of violence on voters is independent
of their ideal points and that political unrest imposes such a large
cost on voters that they prefer any policy to one that always leads
to political unrest. That is,
ASSUMPTION A2. ᭙␪, c ϭ c Ն 4.
␪
The median voter M is said to be decisive if and only if the
strong party s (respectively, the weak party w) wins whenever M
strictly prefers s (respectively, w). Assumption A2 enables us to
derive conditions under which the median voter is decisive. We
solve for the Subgame Perfect Equilibrium extending the solution
concept to Perfect Bayesian Equilibria when private information
is present. We rule out noncredible threats by adding the mild
assumption that no player follows a weakly dominated strategy.
In particular,
ASSUMPTION AA. s, w, and z never adopt a weakly dominated
strategy.
This allows us to use A1 and A2 to prove our key lemma
showing that the median voter is decisive in the case of direct
control, even though the postelectoral outcome has two dimen-
sions, y and r. For the case of outsider control (A1 replacing A1Ј),
we need to be sure that s can always predict what z will do. We
therefore assume throughout that s knows cz and to avoid having z
follow a mixed strategy which s cannot predict, we assume that z
is peaceful when indifferent.
7. A small subset of voters who are not so hurt by the interference is plausible
but should not undo our results.

ELECTORAL COMPETITION
505
ASSUMPTION AB. s observes cz, and z never creates unrest when
indifferent.
The lemma is stated below. In order to prove the decisiveness
of the median voter (part (ii)), we will use the fact that voters
anticipate that political unrest never occurs when s wins the
election (part (i)).
LEMMA 1. Under both direct (A1, A2, and AA) and outsider (A1Ј,
A2, AA, and AB) control, (i) r ϭ 0 at any policy potentially set
by s; (ii) the median voter is decisive.
The proof of Lemma 1 uses the fact that the utility functions
and the costs of political unrest to the players are separable. In
Ellman and Wantchekon [1999] we prove the lemma under less
restrictive assumptions on the cost of unrest, c .
␪
In the next two sections we solve for equilibrium when
neither party can commit to a platform.
III. NO PLATFORM COMMITMENT AND NO PRIVATE INFORMATION
We first describe the outsider control case and briefly explain
the direct control case. Suppose that y is the policy chosen at stage
3 by the majority winner. At stage 4, the outsider z initiates
political unrest if Ϫcz Ͼ Ϫ2 Ϫ y. We define y ϵ cz Ϫ 2 so that the
two payoffs are equated at y ϭ y. Given Assumption AB, z’s best
response is to initiate political unrest if y Ͼ y and not to initiate
political unrest if y Յ y.8
When w wins and chooses y at stage 3, w anticipates a payoff
of Ϫcw from setting y Ͼ y and y Ϫ 2 from setting y Յ y. It is optimal
for w to avoid unrest with the least possible policy compromise by
setting y ϭ y. Using an asterisk to identify w’s equilibrium
*
strategies, we have yw ϭ y. In contrast, when s wins and chooses
the policy, s can choose its ideal point. Since s and z have the same
ideal point, z will not cause unrest against this policy. Thus, the
policy outcome is ys ϭ Ϫ2 if s wins, and yw ϭ y if w wins. At stage 2,
voters anticipate these electoral outcomes. By A1Ј, y is closer to
the median voter ’s ideal point than Ϫ2, so the median voter
prefers w to win. So, by Lemma 1, w wins. The policy implemented
is y and the probability of unrest, r ϭ 0. Note that there is no risk
8. Note that there is only one threatening actor: neither w nor s ‘‘interferes’’by
Assumption A1Ј.

506
QUARTERLY J OURNAL OF ECONOMICS
of unrest because w dislikes unrest and is perfectly aware how
much compromise is required.
The result is equally true under A1 (the direct control case).
We only need to note that s still sets ys ϭ Ϫ2 at stage 3 because
this guarantees its highest possible payoff. The outcome is exactly
as above, except that y ϭ cs Ϫ 2. We will refer to y as the
reservation policy of the threatening actor. It is the minimal policy
compromise which w (or w and s) must make to prevent s (or z)
from causing trouble.
PROPOSITION 1. Under A1 (respectively, A1Ј), there is a unique
Subgame Perfect Equilibrium: the weak party wins the
election, implements the strong party’s (respectively, threat-
ening actor ’s) reservation policy, and there is no political
*
unrest, i.e., i ϭ w; yw ϭ y ϵ cs Ϫ 2 (respectively, cz Ϫ 2) and
r ϭ 0.
Proposition 1 shows that the threat of collapse in the political
process can prevent the winner from ‘‘taking all’’ even in a
majoritarian system. Under complete information, the median
voter prefers the weak party because she anticipates that the
weak party will be moderated toward her ideal point by the fear of
political unrest.9
The median voter is decisive in the one-dimensional context
where political unrest never occurs because preferences are single
peaked and identical up to translations of the ideal point. If the
strong party or threatening outsider could credibly commit against
creating unrest after the election, the median voter would antici-
pate that, if elected, the weak party would implement its ideal
policy. In this case, the median voter would vote for the strong
party. Thus, the strong party’s electoral loss comes from its
inability to commit not to interfere after the election.
Robustness. We note two possibilities that reverse this sur-
prising result. If cs Ͻ 0, w would be unable to compromise enough
to avoid political unrest, and voters would vote for s instead of w.
Second, as mentioned above, one might want to consider the case
in which the order of stages 3 and 4 is reversed (for example, if
9. This is an example of (ex post) moderation of policy outcomes. The
moderation is independent of voter behavior in contrast to Alesina and Rosenthal
[1995] in which moderation arises because voters select a president and then select
an opposing group of legislators. We also note that our result will apply even for a
multiparty election as long as the threat of unrest forces the winning party to
moderate its policy proposal.

ELECTORAL COMPETITION
507
policy choice becomes fixed only some time after the election, and
the electoral winner can neutralize the threat of s or z in this
interim since the interference decision must be made before the
policy choice is observed).10 Unrest must preempt ‘‘disarmament.’’
Now that the electoral winner sets the policy after the unrest
decision has been made, w will set y ϭ 2 at stage 4. Anticipating
this, s (or z) would create unrest as soon as w wins (before being
‘‘disarmed’’). Voters then prefer to vote for s rather than w because
y ϭ 2 is better than r ϭ 1. In this case, w suffers from being unable
to commit to a policy, and w would benefit when platform
commitment becomes possible at stage 1.11 We focus on the timing
stated in Section II because a party’s policy plan often becomes
clear before the party is able to neutralize interference threats.
However, the alternative timing is an interesting avenue for
further research.
Proposition 1 can help make sense of electoral outcomes in
new democracies where the most ‘‘peaceful’’ party or the party
with the weakest military support has been elected. For instance,
in Chile in 1990 Patricio Alwin’s Center-Left coalition won against
Herna´n Bu¨chi’s strong right-wing party which had close ties to the
armed forces. Upon winning, Alwin implemented relatively conser-
vative economic policies and continued the market reforms started
under General Pinochet. We suggest that sixteen years of Pino-
chet made his party’s reservation policy relatively predictable and
that this policy compromise was necessary to prevent a military
coup.12 For this reason voters could trust Alwin’s party to compro-
mise enough to avoid a coup. As we show in the next section, when
the reservation policy is private information, the weak party may
lose because voters know that it would risk causing unrest.
IV. NO PLATFORM COMMITMENT AND THE STRONG PARTY
HAS PRIVATE INFORMATION
In this section we allow the strong party to have private
information over the ability of the threatening actor to disrupt the
10. Sometimes militias can be disarmed, and new legislation can constrain
unions or prevent capital flight.
11. Thanks to Oliver Hart for suggesting further investigation of this
commitment issue.
12. More recently (November 1998), the Chilean government, particularly its
socialist foreign minister, strongly supported Pinochet in his attempt to avoid
extradition from England to Spain where he would face charges of human rights
violations during his presidency.

508
QUARTERLY J OURNAL OF ECONOMICS
political process. We show that the strong party may win because
of people’s fear of such disruption. For the direct control case, it is
assumed that the strong party alone knows the true value of cs
and hence its reservation policy y. This assumption makes good
sense when cs is reinterpreted as the strong party’s subjective
expectation (at stage 4) of its cost from interference. The strong
party chooses whether or not to initiate political unrest based on
this subjective expectation.
For the case of outsider control, we have to assume that the
outsider shares its information with the strong party, that is to
say, only s and z know the true value of cz and hence y. There is no
difference between the results under direct control and outsider
control so we simply describe the case of direct control. In order to
simplify the analysis, we assume that
2 ϩ a
52 ϩ b,
with Prob. p
cs ϭ
with Prob. 1 Ϫ p,
where 5a,b6 ʦ [Ϫ2,2]. Thus, y ϭ a, with probability p, and that y ϭ
b, with probability 1 Ϫ p.13 In order to focus our analysis on the
case where all types of the strong party have reservation policies
relatively close to their ideal point of Ϫ2, we assume that the
‘‘softer ’’ type of the strong party has its reservation policy b,
located to the left of the median voter. That is,
ASSUMPTION 3. Ϫ2 Յ a Ͻ b Ͻ 0.
The probability of political unrest if w is elected is now given
by
0
p
1
if yw ʦ [Ϫ2,b]
if yw ʦ [a,b]
if yw ʦ [b,2]
r ( yw) ϭ
.
5
*
*
ʦ
Define Y w ϵ arg maxy 5Ϫ02 Ϫ y 0(1 Ϫ r ( y)) Ϫ cw · r ( y)6 and yw
*
*
Y w. We claim that Y w ϭ 5a,b6. To prove this claim, note that any
policy y ʦ (b,2] leads to a payoff of Ϫcw and is strictly dominated by
y ϭ a. This is due to the fact that y ϭ a gives a payoff of at least
Ϫ4 Ն Ϫcw. Similarly, y ʦ (a,b) is strictly dominated by y ϭ b, and
y ʦ [Ϫ2,a) is strictly dominated by y ϭ a.
e
*
Let r * ϭ r ( yw) and r denote the equilibrium probability of
13. For an analysis of the continuous types case, see Ellman and Wantchekon
[1999].

ELECTORAL COMPETITION
509
political unrest. We show that when w wins the election, r e ϭ r *.
When w wins, w trades off policy against risk of political unrest.
We define by cˆw the cost of unrest at which w is indifferent between
choosing a ‘‘safe’’ policy, y ϭ a and a ‘‘risky’’ policy y ϭ b.
Proposition 2 presents a sufficient condition for the occurrence of
political unrest in equilibrium, and Proposition 3 describes the
electoral outcomes. Proposition 2 has two parts: 2(i) points out
that if w wins, the risk of unrest is a decreasing step function of cw.
2(ii) points out that even if w risks to cause political unrest, the
equilibrium risk of unrest is 0 when c0 is sufficiently large. This is
because voters will then vote for s which always sets y ϭ Ϫ2 so
that r ϭ 0. We define by cˆ0 the cost of unrest at which the median
voter would switch from voting for w (at y ϭ b and r ϭ p) to voting
for s.
DEFINITION 2.
2 ϩ b(1 Ϫ p)
(2 Ϫ a) ϩ (b Ϫ 2)(1 Ϫ p)
cˆ0 ϵ
and
cˆw ϵ
.
p
p
PROPOSITION 2. (i) When the weak party wins, the probability of
political unrest, r * ϭ p if cw Յ cˆw and r * ϭ 0 as cw Ն cˆw. (ii)
Furthermore, r e ϭ r * for c0 Ͻ cˆ0, and r e ϭ 0 for c0 Ͼ cˆ0.
When the cost of political unrest for the weak party is
sufficiently large, that is if cw Ն cˆw, then the weak party will win
the election by credibly promising to set yw ϭ a, leading to r * ϭ 0
(and r e ϭ r * ϭ 0). When cw Ͻ cˆw, the weak party will set yw Ͼ a.
This leads to a positive probability of interference when the weak
party wins, r * Ͼ 0. As a result, if cw Ͻ cˆw, and c0 Ͻ cˆ0, then the
weak party wins. Otherwise, the strong party wins. When c0 ϭ cˆ0,
each party wins with probability 1/2. For the sake of emphasis, we
state the corollary as a proposition.
PROPOSITION 3. The strong party wins when both cw Ͻ cˆw and
c0 Ͼ cˆ0.
The result shows that if cw is sufficiently large, the weak party
never risks causing political unrest and will always win exactly as
in Proposition 1. If c0 is small, the median voter is not overly
concerned by political unrest. As a result, the median voter will
vote for the weak party, even though this choice can lead to
political unrest. So for the strong party to win, c0 must be large
relative to cw.
The strong party has an incentive to scare voters (make them

510
QUARTERLY J OURNAL OF ECONOMICS
feel that c0 is large) but also to obfuscate its willingness to create
unrest. The uncertainty surrounding its own (or the threatening
actor ’s) militancy creates an atmosphere of insecurity if cw is
sufficiently small that voters believe w will then be willing to risk
unrest. This compels voters to lean toward the strong party if they
believe that c0 Ն cˆ0.
Figure I presents the equilibrium probability of unrest r * as a
function of cw.
Proposition 3 reflects on electoral behavior in the first demo-
cratic elections of a country trying to move forward after a period
of civil war. For instance, in Liberia in 1997 restoration of civil
order was the main motive behind the massive vote for the former
warlord, Charles Taylor. The same can be said about ARENA’s
victory in the 1994 presidential elections in El Salvador (see
Section VII). Even in Western democracies faced with serious
outside challenges (such as France in 1958) or the threat of
internal collapse (Weimar in the 1930s), concerns about the
survival of the democratic process may lead the electorate to
prefer politicians who have strong ties to the armed forces and can
enforce some form of civil order. The rise of fascism or military-
style government in some Western democracies before and imme-
diately after the Second World War can be seen as cases where
c0 Ͼ cˆ0 led to the electoral victory of a strong party.
V. PLATFORM COMMITMENT AND NO PRIVATE INFORMATION
In Sections III and IV parties were unable to commit to a
policy platform. Voting was determined by the parties’ ideal
FIGURE I
Equilibrium Probability of Unrest as a Function of cw

ELECTORAL COMPETITION
511
points, and the distribution of the threatening actor ’s reservation
policy. Here and in the next section, we assume that while parties
cannot commit against initiating political unrest, they can commit
to a policy platform. Platform commitment in this case means the
winning party will implement the policy proposed in its campaign
platform, so long as there is no political unrest. This assumption is
more plausible for stable countries such as the United Kingdom,
Sweden, Norway, and Chile since the 1980s.
In order to motivate platform commitment, we argue that in
stable democracies, repeated elections allow parties to try to build
a reputation for honoring their campaign promises (see Alesina
[1988]). For the case of directly controlled threats, we further
argue that it is harder to build a reputation for creating unrest.14
Alternatively, we can assume that, while choice of party leader
provides a credible commitment to the policy platform repre-
sented by the leader ’s specific ideology, either someone else
controls the disruption decision or the leader ’s specific willingness
to disrupt is not observable to the voters and the weak party.
We define by xs and xw, the strong party and the weak party’s
platforms. There are two stages of backward induction left to be
solved. Even after parties choose their platforms, voters have to
anticipate how the strong party would respond if the weak party
won the election. Additionally, the parties choose their platforms
in anticipation of the electoral and political outcomes that will
follow. Now that policy commitment is possible before the election,
the direct benefit from winning the election, R, cannot be ‘‘normal-
ized’’ to 0. We denote the probability that s wins by ␲(xs,xw) ϵ
Pr(i ϭ s 0xs,xw).
For simplicity, we focus on the outsider control case (assump-
tion A1Ј) and briefly describe the insider control case afterwards.15
Since there is no private information in this section, the reserva-
tion policy of the outsider, y ϵ cz Ϫ 2 is known by all the players.
We find that the equilibrium platforms converge at min (0,y).16
PROPOSITION 4. Under outsider control of political unrest with no
private information, party platforms converge at min (0,y).
14. The winner of an election in a stable democracy might be better controlled
than the loser. For instance, an executive who violates the constitution can be
impeached, while an opposition leader who breaks the law might have less to lose.
15. As mentioned before, we assume that z, the outsider, has an ideal point of
Ϫ2, but a less extreme outsider would give similar results.
16. This is a version of the well-known median voter theorem. As Shepsle
[1991] shows, the theorem does not hold if the election involves more than two
parties.

512
QUARTERLY J OURNAL OF ECONOMICS
There is no political unrest, and either party may win the
election.
Proposition 4 shows that in an environment where parties
can commit to their platforms, we should observe more platform
convergence. The result is driven by the fact that both parties can
now credibly make enough policy compromise in order to win the
election, in contrast to the no-commitment case (Section III)
where only the weak party can make any credible policy compro-
mise. Under direct control of political unrest, platform conver-
gence still occurs at min (0,y) only now, for the case where y Ͻ 0,
the strong party always wins in equilibrium. This is because the
strong party can deviate to y ϩ ⑀ for ⑀ Ͻ R and pick up votes
whereas the weak party cannot win with such a deviation because
voters know that s would initiate political unrest against y ϭ xw ϭ
y ϩ ⑀.17
Our results help explain platform choices in the 1990 Chilean
elections mentioned above, in which the two candidates adopted
essentially the same platform on the main policy issues, especially
tax reform. The same logic applies to the 1992 and 1995 elections
in Sweden, where increased capital mobility owing to liberaliza-
tion and integration of international capital markets led social
democrats and conservatives to converge toward fiscal discipline
and less generous welfare and wage policies (see Moses [1994] for
a description of this ‘‘ideological abdication’’ of the Swedish social
democrats). As the next section shows, when voters are ignorant of
the risks of capital flight or other forms of disruption, such
‘‘ideological abdication’’ becomes less likely, and platform diver-
gence will occur.
VI. PLATFORM COMMITMENT WHEN THE STRONG PARTY
HAS PRIVATE INFORMATION
In this section we show that the strong and weak parties may
choose divergent policy positions. This is in contrast to the
well-known result that platform commitment leads to conver-
gence. Roemer [1994] and Calvert [1985] have shown that incom-
plete information over voter preferences can explain some polar-
ization. However, they point out that convergence of equilibrium
platforms is a relatively robust result when parties value winning
per se (R Ͼ 0), and uncertainty about the median voter ’s location
17. The proof of this result is available upon request.

ELECTORAL COMPETITION
513
is not great. In the case of outsider control, we derive a strong
polarization result which holds even when the benefit from
winning is arbitrarily large and voter preferences are known.
While other explanations of polarization are compelling in other
situations, our novel result has a useful predictive power when-
ever key factors threatening some form of unrest can be discerned.
For the sake of clarity, we continue to focus on the more plausible
case of outsider control. We explain how our results change for the
‘‘insider ’’ control case at the end.
Parties can commit to their platforms as in the previous
section, but now the level of intransigence of the outsider is
imperfectly observed by the weak party and the voters. We show
that polarization arises on account of posturing by the (indirectly)
strong party. Even when the outsider is relatively weak, the
strong party can pretend to know that the outsider is militant in
the hope that voters will be fooled. The strong party signals this
claim of militancy by adopting an extremist platform (it is less
willing to converge when the outsider threat really is militant).
When the voters are swayed by this posturing, they accept the
strong party’s extreme platform. The weak party will not converge
to that platform because it thinks that voters may be less afraid of
the strong party than the strong party had hoped. The weak party
then gains from having offered a platform closer to its own ideal
point.
As in Section III, we model the uncertainty over the threaten-
ing actor ’s intransigence by allowing cz and therefore the outsid-
er ’s reservation policy to be private information. We assume that z
can be one of two types: z1 is militant and has y ϭ a; z2 is weaker
and has y ϭ b. Only s (and z) are assumed to learn the actual
realization of cz before the election is held. Since we are in the
outsider control case, we have assumption A1Ј with stochastic cz.
Note that while cs is fixed for this case, we write s ϭ s1 when s
learns that z is ‘‘militant’’ (has low cost cz ϭ a ϩ 2) and s ϭ s2 when
z is ‘‘soft.’’
If voters are sufficiently afraid of a militant threat ( p is high
or c0 is high or both are high), s1 and s2 might pool on x Յ a and
always win so that w might as well converge. The following
assumption helps to rule out the pooling equilibria which allow
convergence through voters’ fear.
ASSUMPTION 4. p Ͻ pˆ ϵ (b Ϫ a)/(b ϩ c0).
If the weak party is too averse to unrest (high cw), it may

514
QUARTERLY J OURNAL OF ECONOMICS
avoid all risk of unrest by offering only the highest no risk
platform, x ϭ a. Assumption A5 is used to ensure that the weak
party is willing to risk political unrest by not converging to the
militant outsider ’s reservation policy, a, in the hope that z is only
type z2. Note that this assumption gets monotonically weaker as R
increases and is reasonable even for R close to 0.
ASSUMPTION 5. cw Ͻ c0 ϩ 2 ϩ (c0 Ϫ 2/b ϩ 2)R.
We can now prove that platforms must diverge and that this
divergence prediction is robust to an arbitrarily large direct
benefit of winning. Formally, we have
PROPOSITION 5. Under outsider control of political unrest, all PBE
(in undominated strategies) exhibit platform divergence of at
least b Ϫ a, a fraction no less than min ( p · (R/(2R ϩ b ϩ 2)),
1 Ϫ p) of the time. In particular, this divergence is robust to
arbitrarily large R.
The proof (see the Appendix) shows that, given our assump-
tions, s always takes an extremist position when s knows that z is
militant (a policy to the right of a would lead to unrest with
probability 1). Furthermore, in equilibrium, w (and perhaps s
when aware that z is relatively soft) will move to platform b with
probability bounded away from 0. This is explained in the proof of
Proposition 5 (see the Appendix).
The equilibrium in which s1 and s2 pool at a while w offers b
and always wins is one of many which cannot arise in the case of
direct control. This is so because s2 wants to win and can achieve
this goal by offering b ϩ ⑀ when s controls unrest directly. In the
case of direct control, the equilibria can be pinned down more
tightly. Unfortunately, when R = ϱ, we cannot rule out diver-
gence so easily as explained below.
The Case of Direct Control
Platform divergence can also occur when the strong party
directly controls political unrest. For instance, for arbitrarily
large R, there exists a Perfect Bayesian Equilibrium in which the
stronger type of the strong party chooses xs1 ϭ a, the weaker type
of strong party randomizes between xs ϭ a and xs ϭ b, and the
2
2
weak party chooses xw ϭ b. In this equilibrium, the weak party
sometimes wins when the platforms diverge but always loses
when the platforms converge at b.
However, we can no longer rule out the existence of conver-

ELECTORAL COMPETITION
515
gent equilibria as R = ϱ. The problem is that s1 can now offer x to
the right of a because s1 will not revolt against itself for any x Ͻ
a ϩ R even though it revolts against w for any x Ͼ a.18 As R = ϱ,
a ϩ R = ϱ too, so this upper bound on s1’s platform becomes too
high to guarantee divergence. Even without this, divergence can
arise as a result of private information about the strength of a
threat. We now use the model to help understand platform
divergence under the threat of disruption without worrying
exactly how much control the strong party has over that threat.
Before we turn to some applications of the model, we present
a summary table (Table I) of the equilibrium policies and out-
comes discussed in Sections III–VI.
VII. APPLICATIONS
A. Platform Divergence
Great Britain, 1974. A central question in the February 1974
elections in the United Kingdom was, ‘‘who could control the
unions?’’ The Labour party presented itself as the party most
capable of negotiating an end to serious labor disruption. On the
other hand, the Conservative Party and the incumbent prime
minister Edward Heath wanted a mandate to fight inflation and
union militancy by limiting wage increases [Butler and Kavanagh
1974, p. 265]. According to Butler and Kavanagh, the electorate was
ambivalent about the unions. The median voter appeared to favor the
Conservative Party which could ‘‘stand up to the unions which push
for large wage claims,’’ yet the median voter did not want ‘‘the
inconvenience which would attend dispute’’[p. 256]. In the election the
Labour Party narrowly won by promoting a pact with the unions.
The evidence compiled by Dorfman [1978] and Butler and
Kavanagh [1974] suggests that public ambivalence regarding the
unions resulted in a platform divergence between the two parties
on the election’s main issue: how to fight inflation. While the
Conservatives stressed the need to control wages, Labour pressed
for price controls and limits on profits [Butler and Kavanagh, p.
55]. Our model provides the rationale for this platform divergence
between the two parties.19 Because Labour thought that voters
18. This is because when s wins, it stands to lose the political benefit R which
is contingent on winning power without resorting to extra-political means.
19. We use the outsider control case. The following evidence shows that the
unions were not fully controlled by the Labour Party. On February 17, 1974,
Labour Party leader Harold Wilson, declared the creation of a ‘‘social contract’’

516
QUARTERLY J OURNAL OF ECONOMICS
TABLE I
SUMMARY OF THE MAIN RESULTS
Complete information
Private information
No commitment ● The weak party wins the
election, implements a ‘‘cen-
trist’’ and ‘‘safe’’ policy, y,
and there is no unrest
● When c0 and cw are ‘‘moder-
ate,’’ the weak party wins,
implements a ‘‘centrist’’
policy, and there is a risk of
unrest (Propositions 2, 3).
● When c0 is ‘‘high,’’ the strong
party wins, implements an
‘‘extremist’’ policy and there
is no unrest (Propositions
2, 3).
(Proposition 1).
Commitment
● Both parties choose the
same platform. The strong
party wins if it directly con-
trols the source of the
threat. Otherwise, either
party wins. There is no
political unrest
● When cw is ‘‘moderate’’ and p
is ‘‘low,’’ the strong party
and the weak party choose
‘‘significantly’’ different plat-
forms, either party wins,
and there is a risk of unrest
(Proposition 5).
(Proposition 4).
● When either cw is p is
‘‘high,’’ both parties choose
the same platform, the
strong party wins, and there
is no unrest (Proposition 5).
Definition (1) y is the reservation policy of the strong party, that is, the minimal policy compromise that
the weak party must make to prevent political unrest, (2) c0 is the cost of unrest to the median voter, (3) cw is
the cost of unrest to the weak party, and (4) p is the probability that the strong party is ‘‘very’’ strong.
were fearful enough of disruptive strikes, it had no incentive to
adopt a moderate position. In contrast, the Conservatives hoped
that voters would doubt the threat of strikes, and showed no
leniency toward the unions by confirming their uncompromising
position on controlling wage increases [Butler and Kavanagh,
p. 98].
Taiwan, 1996. The 1996 presidential election in Taiwan took
place, not under the pressure of domestic violence, but under the
pressure of an external power. The main candidates were Lee
Teng-hui from the Kuo Min Tan (KMT), Peng Ming-min from the
with the union leadership. The following day, however, the leader of the Trade
Unions Council, denied having reached any such an agreement [Butler and
Kavanagh, p. 98]. In fact, Wilson was referring to the 1973 agreement between
Labour and the unions which ‘‘provided that, in return to various social policies
and the repeal of the Industrial Relations Act, the unions would show voluntary
restraint’’ [p. 55].

ELECTORAL COMPETITION
517
Democratic Progressive Party (DPP), and Lin Yang-kang from the
New Party (NP). Prior to the election, China staged a nine-month
military campaign with the clear intent of threatening and
intimidating Taiwan’s electorate and stopping its move toward
independence. However, while the New Party maintained its
anti-independence platform, the KMT and the DPP grew more
defiant of China in light of the threats.20 Both the KMT and the
DPP shifted their campaign platforms in favor of more pro-
independence policies. Proposition 5 suggests that such platform
polarization is the result of voters’ uncertainty about the real
intentions of the ‘‘outsider,’’ the Chinese government.21
B. Electoral Victory of an Extremist and ‘‘Strong’’ Party
El Salvador, 1994. In the 1994 presidential election in El
Salvador, the issue of land reform was the most polarizing and the
most important. Land reform was discussed in the peace accords
between the FMLN and ARENA, but at the time of the 1994
election there was still much uncertainty as to how the issue
would be addressed [Stahler-Stolk 1995]. According to a 1994
survey, 51 percent of the rural population had no land, and 2.9
percent of the landowners held 46 percent of the land [Montgom-
ery 1995]. The peasants had consistently and unequivocally
favored a comprehensive land reform policy (see Montgomery and
Stahler-Stolk). In light of this evidence, we conclude that the
median voter in rural areas was a landless or near-landless
peasant who favored land reform. Nonetheless, in the 1994
election this voter preferred ARENA, a party opposed to land
reform.
Given that uncertainty over the implementation of the peace
accords was the deciding factor in the peasants’ decision to
support ARENA, voters must have perceived violence under an
ARENA government to be less likely than violence under an
FMLN government (i.e., ARENA is the strong party and there is a
risk of violence if the FMLN wins). According to 1994 polls, a
plurality of voters (31.1 percent) thought the peace accords would
be implemented if ARENA were elected and 65.6 percent of the
electorate believed that this party was backed by the military
[Instituto Universitario de Opinio´n Pu´blica 1994]. Even some top
FMLN officials thought that a victory by their party could
20. See Newsweek, April 1, 1996. Also, Financial Times, March 8, 1996, wrote:
‘‘Beijing Plays to Weaken Lee’s Hand: But the Reaction Has Been Defiant’’ [p. 33].
21. See Wantchekon and Lam [1996] for more details.

518
QUARTERLY J OURNAL OF ECONOMICS
endanger the country’s stability [Vickers and Spence 1994]. From
1992 to 1994, six top-ranked FMLN leaders were assassinated by
right-wing militias. In this environment, rural poor voters be-
lieved that FMLN victory would jeopardize the peace accords and
lead to a collapse of the democratic process. These concerns about
stability and order (c0 Ͼ cˆ0) led them to prefer ARENA, even
though ARENAwould implement policies that hurt their interests
regarding land reform as arises in the model when c0 Ͻ cˆ0.
Our model suggests that as threats of violence become less of
an issue (c0 Ͻ cˆ0), ARENA should lose its decisive advantage
vis-a`-vis the FMLN and the political process should become more
competitive. Consistent with our analysis is the outcome of the
March 1997 congressional elections in the country. The FMLN
won 32.1 percent of the vote, as compared with 33.3 percent for
ARENA. In the mayoral elections the FMLN more than qua-
drupled the number of municipalities under its control (54),
including the city of San Salvador and its suburbs where more
than one-fifth of the country’s population lives.
A competing explanation for the FMLN’s semi-victory could
be a possible backlash against ‘‘Washington consensus’’ policies.
But this explanation would not be valid since the country enjoyed
a remarkable 4 percent growth rate from 1996 to 1997 despite or
because of those policies. In other words, the fact that the ARENA
lost the 1997 parliamentary and municipal elections in spite of an
excellent economic record confirms our thesis: (1) the threat of
violence was the decisive factor in ARENA’s 1994 victory and (2)
after the implementation of the peace accords, as the threat
declined, a large proportion of the electorate would vote for the
FMLN and against the ARENA.
Liberia, 1997. The 1997 presidential elections in Liberia took
place after eight years of civil war that killed over 200,000 people
and destroyed most of the economic infrastructure of the country.
The two major candidates were the former World Bank economist
Ellen J ohnson Sirleaf and the former warlord Charles Taylor.
Throughout the pre-electoral campaign, Taylor presented himself
as the candidate who holds the key for stability and peace. In
order to signal to voters what might happen if he were to lose the
elections, he threatened the election commission with large-scale
violence if the election were postponed. These threats and con-
cerns for stability and security became the decisive factor in the
electoral outcome. According to a J uly 1997 survey, ‘‘voters wanted

ELECTORAL COMPETITION
519
to put end to the war and to elect a strong leader who can
maintain itself in office’’ [Economist, J uly 26, 1997]. In the end,
nearly 75 percent of the electorate preferred Taylor over Sirleaf.
Proposition 3 and the empirical evidence suggest that Liberians
voted in such a great number for Taylor because they were
convinced that a new civil war will break out if they were to vote
against him. In other words, the outcome of the 1997 elections in
Liberia was the result of a strategic move by the electorate to put
power in ‘‘strong’’ hands in order to secure peace and security.
VIII. CONCLUSION
This paper analyzes electoral incentives and outcomes when
parties face threats of political unrest. We find that without
platform commitment, the weaker party is moderated toward the
center and wins the election. But when the risk of disruption is
privately known by the strong party, the strong party becomes
more likely to win as voter fear of disruption grows. With platform
commitment and no private information, platform convergence
occurs, and either party can win unless (1) the strong party
directly controls the threat and (2) the threat is serious. In that
case, the strong party always wins on the convergent policy.
Finally, when the strong party has private information, we predict
platform divergence as a result of signaling. When the threat is
controlled by an outsider, this divergence is robust to arbitrarily
large direct benefits of office. The equilibrium strategies of parties
depend on the cost of political unrest to voters only when the
threat is private information.
We have illustrated the main results of the model with the
cases of the United Kingdom, Taiwan, El Salvador, and Liberia.
For instance, in the 1994 Presidential elections in El Salvador,
peasants who favored land reform voted strategically for a party
opposed to this reform in order to minimize the risk of postelection
violence. In the 1974 parliamentary election in the United King-
dom, Labour took a high-wage position, hoping that voters would
believe this redistribution to be necessary to avoid further strikes.
Meanwhile, the Conservatives took a divergent position hoping
that voters would doubt the credibility of the strike threats. The
recent convergence of the Labour and Conservative Parties under
Blair ’s leadership coincides with a weakening of union power in a
manner not inconsistent with our model.

520
QUARTERLY J OURNAL OF ECONOMICS
APPENDIX
Notation
When the letter j denotes party s (respectively, w), Ϫj will
denote party w (respectively, s). Recall that i denotes the winner of
the election. We will write sMw (respectively, s␪w) to denote ‘‘the
median voter (respectively, voter ␪) strictly prefers s to w.’’ We will
write ‘‘ϩ⑀’’ when we want to suppress the phrase, ‘‘for sufficiently
small, positive ⑀.’’
For Sections V and VI, ␲(xs,xw) is the probability that s wins
the election when s and w have committed to that pair of
platforms. Xs and Xw denote the supports for s and w’s strategies;
for Section VI, X1, X2 are the supports for s when s ϭ s1, s ϭ s2,
respectively. Recall that for the outsider control case, s ϭ s1
implies that (s knows that) z is militant and s ϭ s2 implies that (s
knows that) z is soft. Denote by ␩, the common posterior weight on
s ϭ s1 in the beliefs of voters and w after observing xs.
Proof of Lemma 1
We define y to be the policy implemented by the winning
party, and x to be a pre-electoral platform. We also define by y the
reservation policy of the outsider or the strong party: y ϵ cz Ϫ 2 for
the case of outsider control and, y ϵ cs Ϫ 2 for the case of direct
control. For the case of direct control when platform commitment
is possible, we further define x ϵ y ϩ R to be the greatest platform,
xs, against which s would not create unrest against itself after
winning. µ( · ) is the measure function on the space of voters with
µ(␪ : ␪ ʦ [Ϫ2,2]) ϭ 1. We will prove the lemma for the no-
commitment case. The proof for the commitment case is derived
by replacing y by x and y by x.
(i) We first prove that r ϭ 0 whenever s wins. s knows y in all
cases. Setting y (or x) above y leads to r ϭ 1 (by definition of y), and
this gives Ϫcs to s.
(a) Under outsider control, Ϫcs Ͻ Ϫ4 (by A1Ј) and is s’s lowest
possible payoff so any strategy containing y (or x) above y is
weakly dominated by the strategy obtained by replacing any such
policies (platforms) by y ϭ Ϫ2 (x ϭ Ϫ2). By Assumption AB, r ϭ 0
for any strategy with all policies weakly less than y. Hence, r ϭ 0
when s wins by Assumption AA.
(b) Under direct control, Ϫcs Ͼ Ϫ4 (by A1), but s can always
create unrest at stage 4 so s’s payoff is bounded below by Ϫcs, even

ELECTORAL COMPETITION
521
if s loses. Replacing any policies (platforms) that ever lead to
unrest, by y ϭ Ϫ2 (x ϭ Ϫ2), therefore creates a weakly dominating
strategy. Hence, r ϭ 0 for any policy in s’s strategy set by
Assumption AA.
Henceforth, we only use r to refer to the probability of unrest
when w wins. By part (i), when s wins, the policy/political unrest
outcome is of the form ( ys,0). When w wins, the outcome is of the
form ( yw,r). (For the case of platform commitment, substitute x
for y.)
(ii) We can now prove that M is decisive by considering all
six possible orderings of ys, yw, and 0. To prove that M is decisive,
we prove that sMw
w␪s) Ն 1/2.
µ(␪ : s␪w) Ն 1/2 and similarly wMs
µ(␪ :
In the case with platform commitment, the voters know that
y ϭ x; so only r needs to be estimated, and we look at deterministic
policy choices. We begin with this case.
In each of the six cases (A–F), we prove that sMw
µ(␪ : s␪w) Ն 1/2 and wMs
µ(␪ : w␪s) Ն 1/2 using the fact that
the two sets ␪ Ͼ 0 and ␪ Ͻ 0 each constitute a measure 1/2 of
voters. We define ⌬ ϵ U (s) Ϫ U (w), ⌬0 ϵ U0(s) Ϫ U0(w), and ⌬ ϵ
␪
␪
␪
⌬ Ϫ ⌬0. It is clear that sMw
␪
⌬0 Ͼ 0 and s␪w
⌬ Ͼ 0 so, for
␪
example, ⌬ Ͼ 0
(sMw
s␪w).
We characterize values of ⌬ and ⌬0 in cases (a)–(d) that will
␪
be needed in the proof:
(a) ys,yw Յ ␪
(1)
⌬ ϭ ys Ϫ ␪ Ϫ ( yw Ϫ ␪) · (1 Ϫ r) ϩ r · c
␪
ϭ ys Ϫ yw · (1 Ϫ r) ϩ r · (c Ϫ ␪).
ys,yw Յ 0
(2)
(3)
(4)
⌬0 ϭ ys Ϫ yw · (1 Ϫ r) ϩ r · c
(b) ys Յ ␪ Յ yw
⌬ ϭ ys Ϫ ␪ Ϫ (␪ Ϫ yw) · (1 Ϫ r) ϩ r · c
␪
ys Յ 0 Յ yw
⌬0 ϭ ys ϩ yw · (1 Ϫ r) ϩ r · c
(c) ␪ Յ ys,yw

522
(5)
QUARTERLY J OURNAL OF ECONOMICS
⌬ ϭ ␪ Ϫ ys Ϫ (␪ Ϫ yw) · (1 Ϫ r) ϩ r · c
␪
ϭ Ϫys ϩ yw(1 Ϫ r) ϩ r (c ϩ ␪)
0 Յ ys,yw
(6)
(7)
(8)
⌬0 ϭ Ϫys ϩ yw · (1 Ϫ r) ϩ r · c
(d) yw Յ ␪ Յ ys
⌬ ϭ ␪ Ϫ ys Ϫ ( yw Ϫ ␪) · (1 Ϫ r) ϩ r · c
␪
yw Յ 0 Յ ys
⌬0 ϭ Ϫys Ϫ yw · (1 Ϫ r) ϩ r · c
Ca se A: ys Յ yw Յ 0.
First, we verify that sMw
µ(␪ : s␪w) Ն 1/2 by showing that
s␪w for all ␪ Ͻ 0 in this case. We investigate in turn the voting
behavior of the three sets of voters, ␪ Ͻ ys, ys Յ ␪ Յ yw, and yw Յ
␪ Յ 0.
(i) It is clear that all ␪ Յ ys prefer s because by (5),
⌬ ϭ Ϫys ϩ yw(1 Ϫ r) ϩ ␪ · r ϩ r · c Ն 0,
␪
since Ϫys ϭ 0 ys 0 Ն 0 yw(1 Ϫ r)0 and c Ͼ 4 Ͼ ␪. (ii) For ys Յ ␪ Յ yw Յ 0,
by (3) and (2)
⌬ ϭ 2( yw Ϫ ␪)(1 Ϫ r) Ϫ r␪ Ն 0
as both terms are positive.
(iii) For ys Յ yw Յ ␪ Յ 0, we use (1) and (2) to compute ⌬ ϭ
Ϫr␪ Ն 0. Thus, sMw
Second, wMs
µ(␪ : s␪w) Ն 1/2.
µ(␪ : w␪s) Ն 1/2 since all ␪ Ͼ 0 prefer w
whenever M prefers w. This follows from the fact that ⌬ ϭ Ϫr␪ by
(1) and (2) and this is clearly less than or equal to 0 for ␪ Ͼ 0. Thus,
M is decisive when ys Յ yw Յ 0.
Ca se B. ys Յ 0 Յ yw.
(i) All ␪ Յ ys prefer s since
⌬ ϭ Ϫys ϩ yw(1 Ϫ r) ϩ ␪r ϩ rc Ͼ 0;
␪
(ii) for ␪ ʦ ( ys,0), we use (3) and (4) to compute
⌬ ϭ ϩ␪ Ϫ ␪(1 Ϫ r) ϭ Ϫ␪(2 Ϫ r) Ͼ Ϫr␪ Ն 0.
Therefore, sMw
µ(␪ : s␪w) Ն 1/2.

ELECTORAL COMPETITION
523
If wMs, all voters with ideal points ␪ such that ␪ ʦ [0,yw] will
vote for w, since
⌬ ϭ Ϫ␪ ϩ r␪ Յ 0.
In addition, all voters with ideal points ␪ such that ␪ ʦ [ yw,2] will
also vote for w, since by (4)
⌬ ϭ Ϫ2yw(1 Ϫ r) ϩ r␪ Յ 0.
Therefore, wMs
0 Յ yw.
µ(␪ : w␪s) Ն 1/2. Thus, M is decisive when ys Յ
Ca se C: 0 Յ ys Յ yw.
First, if sMw, then all voters with ideal points ␪ Յ 0 will prefer
s since by (5) and (6) ⌬ ϭ r␪ Ն 0. Second, because ys is closer to the
median voter ’s ideal point and because there is no political unrest
when ys is implemented, wMs can be ignored because it cannot
arise. So, M is decisive when 0 Յ ys Յ yw.
Ca se D: yw Յ ys Յ 0.
All ␪ Ͼ 0 clearly prefer s since ys is closer to their ideal point
and there is no risk of political unrest when s implements ys. This
is also true for ␪ ϭ 0 which means that both sMw and s␪w. wMs
never arises. So, M is decisive when yw Յ ys Յ 0.
Ca se E: yw Յ 0 Յ ys.
All ␪ Ͼ ys clearly prefer s since ys is closed to their ideal points
and never leads to political unrest. For all ␪ ʦ [0,ys], we use (7)
and (8) to compute
⌬ ϭ 2␪ Ϫ r␪ Ն 0.
Thus, sMw
s␪w for all ␪ ʦ (0,2] or µ(␪ : s␪w) Ն 1/2.
If wMs, all ␪ Յ yw prefer w since ⌬ ϭ 2yw(1 Ϫ r) ϩ r␪ Յ 0. In
␪
addition, all ␪ ʦ [ yw,0] also prefer w since by (7), (8),
⌬ ϭ 2␪ Ϫ r␪ Յ Ϫr␪ Յ 0.
Thus, wMs
w␪s for all ␪ ʦ [Ϫ2,0) or µ(␪ : w␪s) Ն 1/2.
Therefore, M is decisive when yw Յ 0 Յ ys.
Ca se F : 0 Յ yw Յ ys.
sMw implies that s␪w ᭙␪ Ͼ 0. First, all ␪ Ͼ ys clearly prefer s
since ys is closer to their ideal points and there is no risk of unrest
when s implements ys. Second, all ␪ ʦ [ yw,ys] also vote for s since
by (7), (6),
⌬ ϭ (2␪ Ϫ 2yw)(1 Ϫ r) ϩ r␪ Ն 0.

524
QUARTERLY J OURNAL OF ECONOMICS
Finally, all ␪ ʦ [0,yw] also prefer s since by (5), (6),
⌬ ϭ r␪ Ն 0.
Conversely, if wMs, then all ␪ Յ 0 prefer w since by (5), (6),
⌬ ϭ r␪ Յ 0.
Thus, M is decisive when 0 Յ yw Յ ys.
In the case without platform commitment, voters may now
have nontrivial uncertainty over yw, but we assumed common
beliefs. Furthermore, ys ϭ Ϫ2, so we only have to treat cases A and
B, without the subcases A(i) and B(i). Taking the expectation of ⌬
over the remaining subcases, for ␪ Ն 0 and ␪ Յ 0 in turn, verifies
that M is decisive exactly as before. This completes the proof that
M is decisive.
ᮀ
Proof of Proposition 1
When s wins, s chooses its ideal point, y ϭ Ϫ2: w never causes
unrest. When w wins, w always seeks to avoid political unrest
because cw Ն 4. Therefore, yw Յ y and since w gains by increasing
yw on the range [Ϫ2,y), the equilibrium must have yw ϭ y and
r ( y) ϭ 0. With r ( y) ϭ 0, w prevents unrest by implementing
exactly y. All voters with ␪ Ͼ ( y Ϫ 2)/2 prefer w. Since y ϭ cs Ϫ 2 Ͻ
2 by A1, this includes all voters with ␪ Ն 0, and hence at least 50
percent of the electorate so w wins. (We could have used Lemma 1
here, but we wrote out the above proof to show that c is
␪
irrelevant.) This is the unique SPE.
ᮀ
Proof of Proposition 2
First, we show by contradiction that r (cw) is decreasing in cw.
Let cЈ and cw be two levels of the cost unrest to the weak party
w
such that cЈ Ͻ cw. Define by r Ј ϵ r (cЈ ) and r ϵ r (cw). Define also by
w
w
yЈ ϭ y(cЈ ) and yw ϭ y(cw). We have
w
w
[U ( yЈ ,cw) Ϫ U ( yw,cw)] Ϫ [U ( yЈ ,cЈ ) Ϫ U ( yw,cЈ )] Յ 0
w
w
w
w
because the first bracket is nonpositive by optimality of yw given cw
and the second bracket is nonnegative by optimality of yЈ given
w
cЈ . But, after substituting and simplifying, the right-hand side
w
equals
(9)
(cw Ϫ cЈ )[r ( yw) Ϫ r ( yЈ )] ϭ (cw Ϫ cЈ )(r Ϫ r Ј) Յ 0.
w w w
Since cЈ Ͻ cw, (9) implies that r Ј Ͼ r. This means that r * is a
w

ELECTORAL COMPETITION
525
decreasing (step) function of cw. (r * is a step function since r * ʦ
50,p,16).
(ii) cw,c˜s (respectively, c˜z for outsider control case) determine r,
the risk of disruption when w wins the election. When s wins, r ϭ
0. Note that r is independent of c0. When r ϭ 0, there is nothing to
prove. When r Ͼ 0, we will now show that r e is a step function
which shifts down from r to 0 at a critical level of c0,cˆ0. In the
Subgame Perfect Equilibrium, M is decisive and chooses between
( ys,rs) ϭ (Ϫ2,0) and ( yw,rw) ϭ ( yw,r). When s wins (i ϭ s), M has
payoff Ϫ2 and r e ϭ 0. When w wins (i ϭ w), M has payoff
U0( yw,r) ϭ Ϫ0 yw 0(1 Ϫ r) Ϫ r · c0. If r Ͼ 0, [ѨU0( yw,r)]/Ѩc0 ϭ Ϫr Ͻ 0
while u0(s) is independent of c0. Hence, M may prefer w for low
values of c0 but whenever r Ͼ 0, M prefers s for sufficiently high
values of c0 and the shift in preference occurs at some value cˆ0 ʦ
[2,ϱ). This proves that when r Ͼ 0 there exists
i ϭ s and r e ϭ 0,
c0 Ͼ cˆ0
c0 Ͻ cˆ0
cˆ0 ʦ [2,ϱ) :
5i ϭ w and r e ϭ r,
In particular, r e declines as c0 rises, and r e ϭ 0 as c0 Ͼ cˆ0.
ᮀ
Proof of Proposition 4
In proving Lemma 1, we showed that s never causes r Ͼ 0 so
sup Xs Յ y. Similarly, if w offers a platform to the right of y, w
knows that if it wins it gets its lowest possible payoff of Ϫcw. By
deviating to platform y, it does strictly better if it ever wins (as
R Ͼ 0 and y is the best policy it can hope for), and it can do no
worse (since Ϫ(2 Ϫ y) Ն Ϫcw). Thus, sup Xw Յ y. Using these two
implications of Assumption AA (that weakly dominated strategies
are avoided), we separate the proof into two cases.
In order to rule out uninteresting knife-edge equilibria in
which one party never wins (equally ruled out by trembling hand
perfection arguments), we assume that voters mix between s and
w when indifferent (Assumption AC).
Case y Յ 0 : sup Xs Յ y
w’s best response is y because
against xs Ͻ y, this uniquely gives w its highest possible payoff
given the threat, while against xs ϭ y the policy cannot be
improved for w and no deviation could increase w’s chance of
winning. If w can never win, we must have Xs ϭ 5 y6 and ␲( y, y) ϭ 1
(s cannot offer any policy closer to 0 than y either), but this
contradicts Assumption AC. Since w can win, Xw ϭ 5 y6. The only

526
QUARTERLY J OURNAL OF ECONOMICS
best response for s in turn is also y unless ␲( y, y) ϭ 0 in which case
s is willing to offer any platform. We designed AC to ensure that
␲( y, y) Þ 0. Using AC, we have a unique best response of y so that
there is a unique equilibrium with convergence of platforms at y.
Without using AC, there would simply be an additional range of
equilibria in which only w wins and s adopts any strategy it likes
or only s wins and w randomizes over strategies (which must have
significant weight on y). In these equilibria, the winner always
offers y so the policy outcome is still at the ‘‘convergence point,’’ y.
Case y Ͼ 0. Given that all strategies use only platforms x Յ y,
the problem reduces to platform competition on the restricted
policy space, [Ϫ2, y] and where the extra dimension created by the
risk of unrest can be neglected since no party ever offers a
platform with r Ͼ 0. The median voter ’s ideal point, 0, lies strictly
inside this range and we can apply the well-known median
convergence theorem—see Roemer [1994] for a proof that treats a
case in which parties are ideological but also value winning the
election for its own sake and therefore encompasses our model. ᮀ
Proof of Proposition 5
We assume that s always knows y and so, exactly as in the
proof of Proposition 4, undominatedness (Assumption AA) re-
quires that sup X1 Յ a and sup X2 Յ b; meanwhile, w only knows
that y Յ b, so we can only know that, sup Xw Յ b: any platform
above b guarantees w its lowest payoff when it wins and is weakly
dominated by offering b which gives w the highest payoff it can
hope for when it wins. Thus, Assumption AA gives sup Xw Յ b and
sup X1 Յ a while sup X2 Յ b. We now derive further restrictions on
equilibrium form.
1. Suppose that s pools on a single platform x. Then x Յ a, else
s1 will never choose x. Furthermore, ␩(x) ϭ p since voters learn
nothing. We now consider what w’s responses can give: (i) if xw ϭ
b, then w wins because p Ͻ pˆ by A4 implies that
p(Ϫc0) ϩ (1 Ϫ p)b Ͼ Ϫpˆ · c0 ϩ (1 Ϫ pˆ)b ϭ b Ϫ (b Ϫ a) ϭ a Ն x
and so w gets a payoff of Uw ϵ Ϫp · cw ϩ (1 Ϫ p)(b Ϫ 2 ϩ R); (ii) if
xw Ͻ x, then s wins (r ϭ 0 either way), and so w gets a payoff of
u Ј ϵ x Ϫ 2 Յ a Ϫ 2. Now p Ͻ pˆ, cw ϩ a Ϫ 2 Ͼ 0 and since
w
((c0 Ϫ 2)/(b ϩ 2)) Յ ((c0 ϩ a)/(b Ϫ a)) ϭ (1/pˆ) Ϫ 1, A5 guarantees

ELECTORAL COMPETITION
527
that cw Յ c0 ϩ 2 ϩ R((1/pˆ) Ϫ 1) so,
Uw Ϫ U Ј Ն (1 Ϫ p)(b Ϫ a ϩ R) Ϫ p(cw ϩ a Ϫ 2)
w
Ͼ (1 Ϫ pˆ)(b Ϫ a ϩ R) Ϫ pˆ (c0 ϩ 2 ϩ R(1/pˆ) Ϫ 1)
ϩ a Ϫ 2).
ϭ b Ϫ a Ϫ pˆ (b ϩ c0) ϭ 0.
(iii) For x Ͻ a, xw ʦ [x,a) is dominated by xw ϭ a as R ϩ a Ϫ 2 Ͼ x Ϫ
2 (when x ϭ a, [x,a) is empty); (iv) For xw ʦ (a,b), w gets a convex
combination of Uw Ϫ (b Ϫ xw) and U Ј both of which are worse
w
than Uw so that would be a dominated response.
It follows that w’s best response is restricted to xw ʦ 5a,b6. Now
if xw ϭ a with probability 1, then, either ␲(x,a) Ͻ 1 in which case s2
would deviate to a ϩ ⑀ or ␲(x,a) ϭ 1, but then w would deviate to b
to get uw which is greater than a Ϫ 2. So we must have b ʦ Xw.
Suppose that xw ϭ b with probability 1 and ␲(b,b) Ͼ 0. Then s2
would deviate to b to increase its payoff by ␲(b,b)R. There is a set
of equilibria in which Xw ϭ 5b6, Xs ϭ 5x6 (with x Յ a) and ␲(b,b) ϭ 0.
In these equilibria, w always wins, but the platforms must exhibit
divergence of at least b Ϫ a so we store this result.
2. Now we must study nonsingleton strategies for s. We begin
with a corollary of the above: w cannot always lose: an equilibrium
in which w always loses would have to have pure pooling (if
X1 ʜ X2 contained more than one platform, we can derive a
contradiction because both types of s would strictly prefer offering,
and winning on, the more extreme of any two such platforms) but
the above argument proved that w must always win in any pure
pooling equilibrium.
3. We can show that w’s best response lies in 5a,b6 in any PBE.
*
We define X w to be the set of xw ʦ Xw for which w wins with strictly
*
positive probability in the hypothesized PBE. We know that X w Þ
л, the null set, because w cannot always lose.
*
For xw ʦ X w ʝ [Ϫ2,a), we want to consider a deviation to xw ϭ
a. We have to consider the consequences of the deviation for all
possible xs ʦ Xs (of course, some of the cases may not arise) and
integrate using the weights in s’s strategy to compare the
expected return to xw with that from the deviation. (i) Against xs Ͼ
a, w loses both at xw and at the suggested deviation (s’s platform
will be closer to 0 given that xs Յ b and s wins because s never
causes unrest by Lemma 1 (i)). (ii) Against those xs Ͻ a to which w

528
QUARTERLY J OURNAL OF ECONOMICS
was losing at xw, w will win by deviating because r ϭ 0 when xw ϭ a
and 0a 0 Ͻ 0xs 0, this gives a ‘‘benefit gain’’ of R and a ‘‘policy gain’’ of
a Ϫ xs. (iii) Against those xs Ͻ a at which w was winning at xw, w
still wins and has a policy gain of a Ϫ xs. (iv) Against xs ϭ a, w
would have been losing and a deviation to a implies no policy
change and possibly some benefit gain. Now by the definition of
*
X w, we know that w must have been winning sometimes so the
expectation places positive weight on case (iii) where xw ϭ a brings
a strict improvement and since this deviation causes no losses in
the other cases, it is strictly optimal for w to deviate. This
*
contradiction proves that X w ʝ [Ϫ2,a) ϭ л.
*
With a little more difficulty, we can also prove that X w
ʝ
*
(a,b) ϭ л. Take xw ʦ X w ʝ (a,b) and consider the deviation to xw ϭ
b. (i) Against xs ʦ (a,b), ␩ ϭ 0, and w wins at b with a policy gain of
b Ϫ xs (and benefit gain of R if w was losing). (ii) Against xs ϭ b, w
was losing and may now win R with no policy change. (iii) Against
those xs Յ a to which w was losing, either w continues to lose or w
now wins at b and the key question is whether w wants to start
winning. To answer this, note that w only starts winning if ␩(xs) ϭ
␩ satisfies
xs Ͻ ␩(Ϫc0) ϩ (1 Ϫ ␩)b
␩ Ͻ (b Ϫ xs)/(b ϩ c0).
Using assumption A5 to substitute for cw,
␩(Ϫcw) ϩ (1 Ϫ ␩)(b Ϫ 2 ϩ R) Ϫ (xs Ϫ 2)
Ͼ (b Ϫ xs)(Ϫc0 Ϫ 2 Ϫ ((c0 Ϫ 2)/(b ϩ 2))R)
ϩ (c0 ϩ xs)(b Ϫ 2 ϩ R) Ϫ (xs Ϫ 2)(b ϩ c0)
b ϩ c0
Ϫ ((c0 Ϫ 2)/(b ϩ 2))b ϩ xs(1 ϩ ((c0 Ϫ 2)/(b ϩ 2))) ϩ c0
ϭ
.
b ϩ c0
R, which, noting that xs Ն Ϫ2,
((c0 Ϫ 2)/(b ϩ 2))(Ϫb Ϫ 2) ϩ (c0 Ϫ 2)
b ϩ c0
Ն
. R ϭ 0,
from which we see that w only starts to win when w is strictly
better off by winning. (iv) Against those xs Յ a at which w was
winning, w makes a policy gain of b Ϫ xs. Again, we have that the
deviation never hurts and because case (iii) cannot be the only