Domestic Political Survival and International Conflict

8/7/2019 Domestic Political Survival and International Conflict

1/29

Review o Economic Studies (2011) xx, 129 doi: 10.1093/restud/rdq027

The Author 2011. Published by Oxord University Press on behal o

The Review o Economic Studies Limited.

Domestic Political Survival andInternational Conict:

Is Democracy Good or Peace?SANDEEP BALIGANorthwestern University

DAVID O. LUCCAFederal Reserve Board

and

TOMAS SJSTRMRutgers University

First version received December2008; fnal version accepted June 2010 (Eds.)

We build a game-theoretic model where aggression can be triggered by domestic political concernsas well as the ear o being attacked. In the model, leaders o ull and limited democracies risk losingpower i they do not stand up to threats rom abroad. In addition, the leader o a ully democratic countryloses the support o the median voter i he attacks a non-hostile country. The result is a non-monotonicrelationship between democracy and peace. Using Polity data, we classiy countries as ull democracies,

limited democracies, and dictatorships. For the period 18162000, Correlates o War data suggest thatlimited democracies are more aggressive than other regime types, including dictatorships, and not onlyduring periods when the political regime is changing. In particular, a dyad o limited democracies is morelikely to be involved in a militarized dispute than any other dyad (including mixed dyads, where thetwo countries have dierent regime types). Thus, while ull democratization might advance the causeo peace, limited democratization might advance the cause o war. We also fnd that as the environmentbecomes more hostile, ully democratic countries become more aggressive aster than other regime types.

Key words: Schellings dilemma, Limited democracy, Democratic peace

JEL Codes: D74, D78

1. INTRODUCTION

The idea that democracy promotes peace has a long history. Thomas Paine argued that monar-chs go to war to enrich themselves, but a more democratic system o government would leadto lasting peace: What inducement has the armer, while ollowing the plough, to lay asidehis peaceul pursuit, and go to war with the armer o another country? ( Paine, 1985 p. 169).Immanuel Kant agreed that i the consent o the subjects is required to determine whether thereshall be war or not, nothing is more natural than that they should weigh the matter well, beoreundertaking such a bad business (Kant, 1795, 1903, p. 122). More recently, the democratic

1

The Review of Economic Studies Advance Access published March 3, 2011


2/29

2 REVIEW OF ECONOMIC STUDIES

peace hypothesis has inuenced the neoconservative view o international relations (Kaplanand Kristol, 2003). U.S. policy makers o dierent political persuasions have invoked it in sup-port o a policy to seek and support the growth o democratic movements and institutions inevery nation and culture.1 But some anecdotal observations seem to support a more realistviewpoint. For example, ater the breakup o Yugoslavia, democratic reorms were ollowed by

war, not peace. When given a chance in the legislative elections o 2006, the Palestinians votedor Hamas, which did not have a particularly peaceul platorm. Such anecdotes suggest that de-mocratization does not always promote peace. Even ully democratic countries such as the U.S.sometimes turn aggressive: under perceived threats to the homeland, the democratically electedPresident George W. Bush declared war on Iraq.

We develop a simple game-theoretic model o conict based on Baliga and Sjstrm (2004).Each leader can behave aggressively or peaceully. A leaders true propensity to be aggressive,his type, is his private inormation. Since actions are strategic complements, the ear thatthe other leader might be an aggressive type can trigger aggression, creating a ear spiral we callSchellings dilemma (see Schelling, 1960; Jervis, 1976, 1978; Kydd, 1997). Unlike Baliga andSjstrm (2004), we assume a leader may be removed rom power. Whether a leader can stayin power depends on the preerences o his citizens, the political system, and the outcome o the

interaction between the two countries. The political system interacts with Schellings dilemmato create a non-monotonic relationship between democracy and peace.

Like the leaders, citizens have dierent types. By hypothesis, the median type preers to livein peace. This imposes a dovish bias on a dyad o two ull democracies (whose leaders can bereplaced by their median voters). Thus, a dyadic democratic peace is likely to obtain. However,when acing a country that is not ully democratic, the median voter may support aggression outo ear and may replace a leader who is not aggressive enough. (For example, Neville Chamber-lain had to resign ater appeasing Hitler.) This gives rise to a hawkish bias. Thus, in a ullydemocratic country, a dovish bias is replaced by a hawkish bias when the environment becomesmore hostile. In contrast, a dictator is not responsive to the preerences o his citizens, so thereis neither a hawkish nor a dovish bias. Accordingly, a dyad o two dictators is less peaceul than

a ully democratic dyad, but a dictator responds less aggressively than a democratically electedleader to increased threats rom abroad.

In the model, the leader o a limited democracy risks losing power i hawks in his popula-tion turn against him. For instance, the German leaders during World War I believed signing apeace agreement would lead to their demise (Asprey, 1991, pp. 486487, 491). Conversely, thesupport o the hawkish minority trumps the opposition o more peaceul citizens. Thus, a lim-ited democracy experiences a hawkish bias similar to a ull democracy under threat rom abroadbut never a dovish bias. On balance, this makes limited democracies more aggressive than anyother regime type. In a ull democracy, i the citizens eel sae, they want a dovish leader, buti they eel threatened, they want a hawkish leader. In dictatorships and limited democracies,the citizens are not powerul enough to overthrow a hawkish leader, but the leader o a lim-

ited democracy risks losing power by appearing too dovish. This generates a non-monotonicrelationship between democracy and peace.Our empirical analysis reassesses the link between democracy and peace using a exible

semiparametric unctional orm, where fxed eects account or unobserved heterogeneity acrosscountry dyads. We use Polity IV data to classiy regimes as dictatorships, limited democraciesor ull democracies. Following the literature on the democratic peace hypothesis, we defne a

1. The quote is rom President George W. Bushs second inaugural address. President Clinton, in his 1994 Stateo the Union address, noted that the best strategy to ensure our security and to build a durable peace is to support theadvance o democracy elsewhere. Democracies do not attack each other.


3/29

BALIGA ET AL. DOMESTIC POLITICAL SURVIVAL & INTERNATIONAL CONFLICT 3

conict as a militarized dispute in the Correlates o War data set. The data, which span over theperiod 18162000, contain many military disputes between limited democracies. In the nine-teenth century, Britain had a Parliament, but even ater the Great Reorm Act o 1832, onlyabout 200,000 people were allowed to vote. Those who owned property in multiple constituen-cies could vote multiple times.2 Hence, Britain is classifed as a limited democracy or 58 years

and becomes a ull democracy only ater 1879. France, Italy, Spain, and Germany are alsolimited democracies at key points in the nineteenth and early twentieth centuries. These coun-tries, together with Russia and the Ottoman Empire, were involved in many militarized disputesin Europe and throughout the world. For much o the nineteenth century, Britain and Russiahad many skirmishes and outright wars in the Great Game or domination o Central Asia(Hopkirk, 1990). France is also involved in many disputes and is a limited democracy during theBelgian War o Independence and the Franco-Prussian War. Germany is a limited democracy atthe start o the First World War.

Over the ull sample, the data strongly support a dyadic democratic peace hypothesis: dyadsconsisting o two ull democracies are more peaceul than all other pairs o regime types. This isconsistent with previous empirical studies (Babst, 1972; Levy, 1988; Maoz and Russett, 1993;Russett and Oneal, 2001). Over the same period, limited democracies were the most aggressive

regime type. In particular, dyads consisting o two limited democracies are more likely to experi-ence militarized disputes than any other dyads, including mixed dyads where the two countrieshave dierent regime types. These results are robust to changing the defnitions o the three cat-egories (using the Polity scores) and to alternative specifcations o our empirical model. Theeects are quantitatively signifcant. Parameter estimates o a linear probability model specifca-tion, suggest that the likelihood that a dyad engages in a militarized dispute alls roughly 35% ithe dyad changes rom a pair o limited democracies to a pair o dictatorships. We also fnd thati some country j aces an opponent which changes rom a ull democracy to another regimetype, the estimated equilibrium probability o conict increases most dramatically when countryj is a ull democracy. This suggests that as the environment becomes more hostile, democraciesrespond more aggressively than other regime types, which is also consistent with our theoretical

model.A more nuanced picture emerges when we split the data into subsamples. Beore World War

II, the data strongly suggest that limited democracies were the most conict prone. It is harderto draw conclusions or the post World War II period, when very ew countries are classifedas limited democracies, and ull democracies have very stable Polity scores. The Cold Warwas a special period where great power wars became almost unthinkable due to the existenceo large nuclear arsenals (Gaddis, 2005). Did the weakening and demise o the Soviet Unionbring a return to the pre-1945 patterns? Although the time period is arguably short, in thepost-1984 period, it does seem that dyads o limited democracies are again the most prone toconict.

It is commonly argued that a process o democratization, e.g. in the Middle East, will lead to

peace (Bush, 2003). But both theory and data suggest that the relationship between democracyand peace may be complex and non-monotonic. Replacing a dictatorship with a limited democ-racy may actually increase the risk o militarized disputes. Even i a dictatorship is replaced by aull democracy, this may not reduce the risk o militarized disputes i the region is dominated byhostile non-democratic countries. In the data, only dyads consisting otwo ull democracies arepeaceul. Democratic countries such as Israel and India, with hostile neighbours, do not enjoy alow level o conict.

2. The inamous rotten borough o Old Sarum sent two representatives to Parliament. In 1831, it had only 11eligible voters, all o whom were landowners living elsewhere. See Paine (1985); Thorne (1986).


4/29


The paper is organized as ollows. Related literature is discussed in the next section. Thetheoretical model is presented in Section 3. Section 4 describes the empirical results. Section 5concludes.

2. RELATED LITERATURE

Theoretical and empirical work in economics and political science has investigated the rela-tionship between political systems and war. Jackson and Morelli (2007) ormalize the idea thatleaders start wars when their preerences are sufciently biased away rom their citizens pre-erences. Levy and Razin (2004) provide a theory o the democratic peace based on incompleteinormation. They assume the representative citizen is less well inormed about the beneft oconcessions than the leader and show that democratically elected leaders are more likely to re-veal inormation truthully. In Bueno De Mesquita et al. (1999), political leaders must bribe keysupporters to stay in power when oreign policy ails. A dictator has to bribe ewer supportersand is thereore more likely to go to war than a democratically elected leader. On the other hand,in order to avoid being replaced, a leader may gamble or resurrection with an aggressiveoreign policy (Downs and Rocke, 1994; Bueno De Mesquita and Silverson, 1995; Hess and

Orphanides, 1995). Fearon (1994) assumes leaders suer audience costs i they back downduring a war o attrition. I audience costs are higher in democracies, then democracies are morecommitted to a conict and may be more reluctant to enter into one. Tangeras (2008) assumesthat leaders have private inormation about the probability o winning a war. Democraticallyelected leaders are more reluctant to start a war because they will lose power i the war endsbadly. According to Leeds (1999), democratic leaders are more able to commit to honouringagreements and thus more able to cooperate.

These theories provide underpinnings or the democratic peace hypothesis, but it is not obvi-ous how they can be extended to explain the non-monotonicity we fnd in the data. For example,a natural extension oFearon (1994) model would be to assume the audience costs o limiteddemocracies lie between those o dictatorships and ull democracies, but this would not produce

non-monotonicity. Similarly, i the leader o a limited democracy has less biased preerencesthan a dictator, then the Jackson and Morelli (2007) model would predict that limited democra-cies go to war less oten than dictatorships.

Our theory incorporates an important eature oBueno De Mesquita et al. (1999): the supportor the leaders action is derived rom heterogeneous preerences among the citizens. In ourmodel, leaders o ull and limited democracies suer audience costs (as in Fearon, 1994) i theyare dovish when the opposing leader is hawkish; in addition, a leader o a ull democracy acesaudience costs (rom the median voter) i he is hawkish against a dovish opponent; a dictatoraces no audience costs at all. The result is a non-monotonic relationship between democracyand peace.

Mansfeld and Snyder (2005) argue that increased nationalism can cause conict during a

period o transition when a regime is being democratized. However, in our baseline empiricalmodel, dyads o limited democracies are the most conict ridden even when controlling orregime transitions (using Mansfeld and Snyders, 2005, transitional dummies). This suggeststhat limited democracies are not only prone to conict during periods o transition.

Several articles have investigated the hypothesis that dyads consisting o countries with sim-ilar regime types, and thus perhaps shared values are relatively peaceul. Peceny, Beer andSanchez-Terry (2004) classiy autocratic regimes as personalist, military and single-party dic-tatorships and fnd evidence that dyads consisting o two autocracies o the same type are rel-atively peaceul. Bennett (2006) analyses plots o conict probabilities or dyads with dierentPolity scores. He fnds that the hypothesized relationship between similarity and peace holds


5/29


or dyads with either very high or very low Polity scores, but not in the intermediate range.This is consistent with our fnding that dyads o two limited democracies (which have interme-diate Polity scores by defnition) are relatively conict prone. However, it is challenging withinBennetts pooled logit specifcation to ormally test or non-monotonicity and to assess robust-ness within higher-order parametric specifcations because the unctional orm is bidimensional

and marginal eects are non-linear unctions o explanatory variables. In addition, his specif-cation cannot include dyadic fxed eects. Our dummy variable non-parametric approach hasdyad-specifc fxed eects, and non-monotonicity can be assessed through simple tests on co-efcients. Unlike Bennetts continuous specifcation, we defne limited democracies by cut-oPolity scores, but we veriy the robustness o our results by varying the cut-o points.

Other authors have analysed limited democracies along other dimensions and ound reasonsor why such regimes might experience conicts. Fearon and Laitin (2003) fnd that limiteddemocracies are more prone to civil wars, as insurgencies are more likely to succeed in weakerpolitical regimes. Epstein et al. (2006) fnd that political transitions rom limited democraciesto other political regimes are harder to explain than political transitions o autocracies and ulldemocracies.

Determining the underlying motives behind conicts, based on a subjective reading o his-

tory, will always leave scope or disagreement. Our theoretical model, building on Baliga andSjstrm (2004), assumes that conicts can be sparked by ear (Schellings dilemma). Histo-rians have uncovered many examples o such ear spirals.3 For example, Thucydides (1972,1.23, p. 49) argued that the Peloponnesian War was caused by the growth o Athenian powerand the ear which this caused in Sparta. The period that preceded World War I was charac-terized by mutual distrust and ear (Sontag, 1933; Tuchman, 1962; Wainstein, 1971). A spiralo ear was evident during the Cold War arms race (Leer, 1992). The IndiaPakistan armsrace is a current example o escalation uelled by mutual distrust, and Bobbitt (2008, p. 10)suggests a similar logic will continue to operate in the wars o the twenty-frst century: Wethink terrorists will attack; so they think we think the terrorists will attack; so they think weshall intervene; so they will attack; so we must. Nevertheless, there is disagreement about the

number o large-scale wars that can be said to have been triggered by ear (see Van Evera, 1999;Reiter, 2000). Reiter (1995) argues that leaders who understand the spiraling logic can preventconict by communicating. Baliga and Sjstrm (2004) veriy that, in theory at least, cheaptalk can sometimes prevent a conict, but it cannot always do so. Our current model assumesthat leaders are partly motivated by domestic political concerns and may behave hawkishly inorder to maintain political support. Thus, ear is not the only reason or starting a war, and theargument by Reiter (1995) that World War I was not a pure ear spiral is consistent with ourmodel:

Domestic politics in a number o nations set the stage or war, though some . . . have goneurther to argue that Germany sought war . . . to shore up the threatened domestic political

order at home (Reiter, 1995, p. 22).

3. A SIMPLE MODEL OF SCHELLINGS DILEMMA

3.1. Basic assumptions

There are two countries, i {1, 2}. Each country i has a leader, leader i , and a continuum o cit-izens. The two leaders play a game that is similar to the arms race game o Baliga and Sjstrm

3. OFlaherty and Sethi (2010) argue that ear spirals can explain (changes in) murder rates in U.S. cities.


6/29


(2004). Each leader can choose an aggressive (hawkish) strategy A or a peaceul (dovish) strat-egy P. The aggressive strategy may represent building new weapons, preparing or war, orattacking the other country. Each citizen has a cost type, a cost o aggression c, drawn rom adistribution F with support [c, c]. We assume F is continuous, strictly increasing and concave.The median cost type is denoted cmed, so F(cmed) = 1/2. Each leader i has a cost type ci , which

is independently drawn rom the same distributionF

. Each leaders type is his private inorma-tion. Everything else is common knowledge. To study the pure impact o political institutionson the incentive to go to war, we assume that there is no ex ante dierence between the twocountries: the distribution F is the same in both.

The pay-o or a citizen o country i who has cost type c is given by the ollowing matrix,where the row represents the choice o his own leader (leader i), and the column represents thechoice o the other leader (leader j).

A P

A c c

P d 0(1)

Thus, i leader i chooses A, the type c citizen o country i suers the cost c (whatever leaderj does). In addition, i leader j chooses P when leader i chooses A, each citizen o country iderives a beneft . Conversely, i leader i chooses P when leader j chooses A, each citizen ocountry i suers a cost d. I both leaders choose P, pay-os are normalized to 0. Notice that and d are not type-dependent.

The parameter can be interpreted as the gain rom being on the oensive, while the param-eter d represents the loss rom being on the deensive. For example, i the aggressive strategyA is to attack, then might represent a frst mover advantage, net o any cost imposed onthe aggressor by the international community, while d is the opponents cost o being attacked.I instead A represents developing a new missile, then might represent the utility gain rom

increased bargaining power, net o the cost o sanctions brought about by a missile test, while drepresents the corresponding loss o bargaining power or the opponent. We assume 0 < < d,so the marginal incentive to choose A is highest when the opponent chooses A. This strategiccomplementarity captures the intuition, undamental to Schellings dilemma, that conictscan escalate. We are interested in how political systems mitigate or aggravate the tendency to-wards escalation.

A citizen o cost type c is a hawkish type ic < . For the hawkish citizen, A is a dominantstrategy because c > d and c > 0. The raction o citizens who are hawks is F(). Acitizen o cost type c is a dovish type ic > d. For the dovish citizen, P is a dominant strategybecause d > c and 0 > c. Notice that a dove is an extreme pacifst who wants his leaderto be peaceul even when the opponent is aggressive. The raction o citizens who are doves is

1 F(d). A citizen o cost type c is a coordination type i < c < d. For the coordination type,the best response to A is A, and the best response to P is P. Coordination types capture theidea that behaviour may be driven by ear: although they preer the outcome P P to the outcomeA A, they want their leader to choose A i they ear the opponent will choose A. The raction ocitizens who are coordination types is F(d) F().

Assumption 1 0 < c < < cmed < d < c.

Assumption 1 implies that the median citizen is a coordination type. Thus, i the represen-tative (median) citizen in each country could directly choose either A or P, the resulting game


7/29


would be a coordination game with two Nash equilibria A A and P P. We will assume that P P isthe risk-dominantequilibrium o this game, in the sense that the representative citizen thinks thegain rom choosing P when the opponent is peaceul (cmed ) exceeds the loss rom choosingP when the opponent is aggressive (d cmed). 4

Assumption 2 cmed >

d cmed.

War is oten thought o as a prisoners dilemma where, in our terminology, all decision mak-ers are dominant strategy hawks. See the discussion o the Peloponnesian War by Nye (2007),pp. 1819), the discussion o World War I by Snyder (1971), Snyder and Diesing (1977) orother historical examples, and Axelrod (1984) or a general discussion.5 More recent contribu-tions, such as Baliga and Sjstrm (2004), allow coordination types as well. Here, we go onestep urther and also allow the existence o pacifstic doves, who do not avour going to war eveni they are sure the other country will attack. However, the generalization is modest: our fnalassumption states that the doves do not outnumber the hawks.

Assumption 3 F() > 1 F(d).

In our model o political regimes, the inuence o a group o citizens will be proportional toits size. By Assumption 3, hawks will be more likely to be pivotal than doves. More generally, therelative importance o hawks and doves could be derived rom a model where dierent citizenshave dierent ability to inuence or coerce others. A political bias akin to Jackson and Morelli(2007) would result i hawks had disproportionate political power. Our current model is simpler,and in view o the previous literature (which emphasizes hawks and, sometimes, coordinationtypes), Assumption 3 does not seem unreasonable.

3.2. Political regimes

Ater the two leaders have chosen their strategies, each citizen decides whether or not to supporthis leader. The decision is retrospective, as in Barro (1973) and Ferejohn (1986). In eect, thecitizen acts as a principal who rewards or punishes the agent (the leader). Thus, a citizeno country i supports leader i i and only i leader is action was a best-response to leader j saction according to the citizens own preerences (as given by the matrix (1)).

Following Bueno De Mesquita et al. (1999), a political system is characterized by a criticalraction o support i

12 . This is defned as the raction o the citizens o country i that must

support leader i in order or him to stay in power. The value o staying in power is the rentsrom ofce, denoted R > 0. Our theory assumes that all political regimes generate the same Rin order to ocus on the impact o political institutions on the incentives to go to war.6 To rule out

4. The concept o risk dominance is amiliar rom the global games literature ( Carlsson and van Damme, 1993).In global games, types are highly correlated, a reasonable assumption i the uncertainty concerns the value o a contestedresource such as a piece o land. Chassang and Padr-i-Miquel (2009) and Chassang and Padr-i-Miquel (2010) use thisapproach to study conict. In contrast, we assume that types are independent, a reasonable assumption i types representidiosyncratic preerence shocks (private costs and benefts rom going to war). As shown in Section 3.3, Assumption 2implies the democratic peace when types are independent.

5. More recent theories o war emphasize that a player may preer to fght a war rather than making concessions,i concessions lead to adverse shits o power (Fearon, 1995, 1996; Powell, 2006).

6. Debs and Goemans (2008) argue that the cost o losing power diers across regime types, and they fndevidence that these costs inuence the probability o conict. Conconi, Sahuguet and Zanardi (2008) link term limits ondemocratic leaders to the probability o war. These arguments are complementary to ours.


8/29


corner equilibria, where even the most aggressive leaders (cost type c) choose P, we assumeR < c. 7

I both leaders choose the same action, then each leader is supported by his median citizen.That is, each leader has the support o at least hal the population, which is sufcient to remainin power in any political regime (as the critical raction satisfes i 1/2). Suppose instead

that the two leaders miscoordinate, say Leader 1 choosesA

and Leader 2 choosesP

. Then,neither leader is supported by his median citizen. Indeed, Leader 1 is supported by a ractionF() < 1/2 o his citizens (the hawks), while Leader 2 is supported by a raction 1 F(d) < 1/2o his citizens (the doves). Leader 1 remains in power i F() 1 , and Leader 2 remains inpower i 1 F(d) 2 .

Assumption 3 implies 1 F(d) < F(). It ollows that the regime o country i belongs toone o three categories, depending on the size o i . First, i

i > F(), then miscoordina-

tion always causes leader i to be replaced. In other words, the support o the median citizenis necessary or leader i to remain in power. Such a regime is a ull democracy. Second, atthe other extreme, i i 1 F(d), then leader i stays in power whatever happens. Such aregime is a dictatorship or autocracy. The third and fnal case is the intermediate situation,where 1 F(d) < i F(). In this case, leader i loses power i he chooses P while the oppo-

nent chooses A, but he stays in power otherwise. Since this regime type is intermediate betweendictatorship and ull democracy, we label it limited democracy.

In a ull democracy, since the median voter is a coordination type, leader i enjoys rents romofce i and only i he matches the action o the opponent. Thereore, i country i is a ulldemocracy, leader i s pay-o matrix is

A P

A R ci ciP d R

(2)

where ci is his own cost type. (The row represents leader i s own choice, the column leader

js choice.) In a limited democracy, leader i can stay in power except when he chooses P andthe opponent chooses A. Thereore, i country i is a limited democracy, then leader is pay-omatrix is

A P

A R ci R + ciP d R

(3)

I country i is a dictatorship, then leader i gets R whatever happens. Thus, R does not inuencehis behaviour and can be dropped, so dictatorial leader is pay-o matrix is simply given by thematrix (1), where c = ci is his own cost type.

3.3. Equilibrium

Country is regime type is denoted Ti {De, Di, Li}, corresponding to ull democracy (De),dictatorship (Di), and limited democracy (Li). Leader i knows the regime type o country j but

7. There is a unique, interior, equilibrium as long as F is concave and R < c. I F is concave but R > c,then equilibrium or a dyad o two ull democracies will be at a corner, where even the most aggressive type plays P ,because large rents rom ofce trump all other concerns. This is an extreme version o the democratic peace. Leaderso other regime types could still play A, however, so replacing one o the democracies by a dierent regime wouldincrease the probability o conict.


9/29


FIGURE 1

Hawkish limited democracies

not the cost type o leader j . The two leaders move simultaneously. Leader is optimal decisiondepends on his own cost type, his own regime type, and the probability p j that leader j plays A.

First, i country i is a dictatorship, then the pay-os o leader i are given by the matrix (1).Hence, leader i preers A i

ci + (1 p j ) d p j . (4)

The leader o country i must ollow a cut-o strategy, playing A i and only i ci +(d ) p j . Thereore, the probability that leader i chooses A is pi = h( p j , Di), where

h( p j , Di) F( + p j (d )). (5)

The unction h(, Di) can be thought o as a dictators best-response unction.Second, i country i is a limited democracy, then leader i s pay-os are given by the matrix(3). Hence, leader i preers A i

R ci + (1 p j ) p j d+ (1 p j )R, (6)

which is true i and only i ci + p j (d ) + p j R. Thereore, the probability that leader ichooses A is pi = h( p j , Li), where

h( p j , Li) F( + p j (d ) + p j R). (7)

This is the best-response unction or the leader o a limited democracy.

Third, i country i is a ull democracy, then leader is pay-os are given by the matrix (2).Hence, leader i preers A i

p j R + (1 p j ) ci p j d+ (1 p j )R, (8)

which is true i and only i ci + p j (d ) + p j R (1 p j )R. Thereore, the probabilitythat leader i chooses A is pi = h( p j , De), where

h( p j , De) F( + p j (d ) + p j R (1 p j )R)). (9)

This is the best-response unction or the leader o a ull democracy.


10/29


Since F is concave, the best-response unctions h( p2, T1) and h( p1, T2) are concave and in-tersect only once. The intersection represents the unique Bayesian Nash equilibrium in the gamebetween the two leaders. In equilibrium, leader i chooses A with probability pi . Our assump-tions guarantee that 0 < pi < 1. There is a conict between countries i and j i at least one leaderchooses A. Thus, the probability o conict is wi j = pi + (1 pi ) p j , which is increasing in pi

andp

j . For a pair o regime typesT

i ,T

j {

De, Di, Li}, we denote the equilibrium probabilityo conict by wTi Tj .

The equilibrium, or various regime types, is illustrated in Figure 1. Leader 1s probabilityo playing A is given on the horizontal axis and leader 2s on the vertical axis. Notice thath( p1, Li) > h( p1, T2), or any T2 {De, Di} and any p1 (0, 1). That is, or any p1, Leader 2 ismore likely to choose A i country 2 is a limited democracy rather than some other regime type.

Figure 1 reveals that turning country 2 into a limited democracy shits leader 2s best-response curve up. Since leader 1s best-response curve h( p2, T1) is upward sloping, in thenew equilibrium both leaders will be strictly more likely to choose A. For example, i country2 changes rom a dictatorship to a limited democracy, the best-response unction changes romh( p1, Di) to h( p1, Li), and the ear spiral will produce higher levels o both p1 and p2. Thus,we have the ollowing result (which does not require Assumption 2):

Proposition 1 (Hawkish Limited Democracy). Replacing any other regime type in coun-try i with a limited democracy increases the equilibrium probability o conict, whatever the

regime type in country j.

We can interpret this proposition in terms o Schellings dilemma. Since F is strictly increas-ing, R > 0 and d > , the best-response unction o leader i is always increasing in p j . That is,actions are strategic complements. First, suppose both countries are dictatorships, so domesticpolitical support is irrelevant. For a hawk (with cost type less than ), A is the dominant strat-egy. Eliminating his dominated strategy P, we conclude that A is played with at least probabilityF(). Ater this frst round o elimination o dominated strategies, P becomes dominated or

some other types. Specifcally, consider any type ci such that

ci < + F()(d ). (10)

Using inequality (4), this type o dictator must play A, knowing that p j (the probability theopponent plays A) is at least F(). Eliminating P or all types (o both dictators) such thatinequality (10) holds makes P dominated or yet more types. This process o elimination odominated strategies or ear spiral causes more and more high-cost (peaceul) types to play A.This is Schellings dilemma (see Baliga and Sjstrm, 2004).8

Now suppose the regime in country i changes rom dictatorship to limited democracy. Theleader o a limited democracy is ousted i he plays P when the opponent plays A. To see how

this hawkish bias reinorces Schellings dilemma, again notice that a hawkish type o leader j

8. I F is not concave, there may be multiple Bayesian Nash equilibria, but Schellings dilemma still applies:dominant strategy hawks choose A, causing other types to choose A, etc. Since actions are strategic complements, therewould be a lowest and a highest equilibrium in terms o probability o aggression (c. Vives, 2001). Replacing anyother regime type with a limited democracy increases the probability o conict both at the lowest and highest equilib-rium. In this sense, the theory can be extended to the case o multiple equilibria. The theory can also be generalizedalong the lines o Xue (2006) to allow sequential moves. Suppose A is an irreversible act, such as the test fring o amissile or an actual invasion. In equilibrium, the most hawkish types are the frst to choose A, creating a dynamic earspiral.


11/29


surely plays A, so p j F(). But now, the second round o elimination o dominated strategiesinvolves more types. Specifcally, consider any type ci o leader i such that

ci < + F()(d ) + F()R. (11)

Using inequality (6), this type must play A, knowing that p j F(). Thus, in the second round,we eliminate P or all types such that inequality (11) holds. Comparing the inequalities (11) and(10), we fnd that P is eliminated or more types in a limited democracy. The dierence is due tothe term F()R, which represents the rents rom ofce the leader o a limited democracy loses ihe plays P when leader j plays A. By the same argument, in each round o elimination, moretypes eliminate P when country i is a limited democracy (and by strategic complementarity, thesame holds or country j). This exacerbates Schellings dilemma.

Now suppose the regime in country i changes rom limited democracy to ull democracy. A-ter this transition, P can be eliminated by ewertypes in each round o elimination o dominatedstrategies. This is due to the rents rom ofce the leader o a ull democracy loses i he plays Awhen leader j plays P. Thereore, in each round, ewer types eliminate P when country i is aull democracy (and by strategic complementarity, the same holds or country j ). This mitigates

Schellings dilemma.Having identifed the least peaceul dyad in Proposition 1, we now consider, which dyad is

most peaceul. Clearly, by Proposition 1, we may ocus on dyads that do not include any limiteddemocracy. It can be checked that h( p, De) and h( p, Di) have a unique intersection at p = 1/2(see Figure 2). I p > 1/2, then h( p, De) > h( p, Di). Thus, when acing an aggressive opponentwho is likely to play A, the leader o the ull democracy is more likely to choose A than adictator (hawkish bias) because he loses power i he responds to A with P. On the other hand,i p < 1/2, then h( p, De) < h( p, Di). Thus, when acing a peaceul opponent who is likely toplay P, the leader o the ull democracy is more likely to choose P than a dictator (dovish bias)because he loses power i he responds to P with A. The intersection o h( p, De) and h( p, Di)lies below the 45 degree line because

h

1

2, Di

= h

1

2, De

= F

d+

2

< F(cmed) =

1

2,

FIGURE 2

Dyadic Democratic Peace


12/29


where the inequality uses Assumption 2. Because o this, as can be verifed using Figure 2, eachleader in a dyad that excludes limited democracies chooses A with probability less than 1/2. Inthis region, the ully democratic leader has a dovish bias, so wDeDe < wDeDi < wDiDi. Thus, wehave the ollowing proposition:

Proposition 2(Dyadic Democratic Peace

). A dyad o ull democracies is more peaceul

than any other pair o regime types.

Despite the dyadic democratic peace, there are many historical examples o democraciesgoing to war against less democratic states. In our model, the equilibrium probability o conictincreases dramatically when a dyad o two ull democracies changes to a mixed dyad with justone ull democracy, or two reasons. First, the less democratic regime will not have a dovish bias,making it more likely to choose A. But this triggers the second eect: the democratic leadersdovish bias disappears because the median voter wants him to respond to A with A.

In Figure 3, pT T denotes the equilibrium probability that regime type T chooses A whenplaying against regime type T. Thus, i country 2 is a ull democracy, with best-response curveh( p1, De), then the equilibrium probabilities are ( pDeDe, pDeDe) i country 1 is also a ull democ-

racy (with best-response curve h( p2, De)), but they are ( pDiDe, pDeDi) i country 1 is a dicta-torship (with best-response curve h( p2, Di)). I country 2 changes rom a ull democracy to adictatorship, then leader 2s best-response unction shits up, rom h( p1, De) to h( p1, Di). Theequilibrium probabilities increase rom ( pDeDe, pDeDe) to ( pDeDi, pDiDe) i country 1 is a ulldemocracy, and rom ( pDiDe, pDeDi) to ( pDiDi, pDiDi) i country 1 is a dictatorship. The increasein both p1 and p2 is larger in the ormer case, as illustrated in Figure 3. This ollows rom com-paring the slopes o the best-response unctions, obtained by dierentiating equations (5), (7),

FIGURE 3

Democracies Turn Hawkish


13/29


and (9): by the concavity o F,

h( p, De)

p>

h( p, Di)

p>

h( p, Li)

p,

where the frst inequality holds as long as p < 1/2 (which is true in the absence o limited

democracies). Since the best-response unction o a ull democracy has the steepest slope, ulldemocracies react most aggressively to an increased threat level. Thus, when country 2 becomesmore aggressive (due to a change away rom ull democracy), the probability o conict in-creases more when country 1 is a ull democracy than when it is a dictatorship. A similar fgurereveals that the probability o conict increases more when country 1 is a ull democracy thanwhen it is a limited dictatorship and the same holds or the case where country 2 changes roma ull democracy to a limited democracy. Thus, we have the ollowing proposition:

Proposition 3 (Democracies Turn Hawkish). As country j changes rom a ull democracyto any other regime type T {Di, Li}, the equilibrium probability o conict increases more icountry i is a ull democracy than i it is any other regime type T {Di, Li}. That is,

wDeT wDeDe > wT T wTDe. (12)

4. EMPIRICAL ANALYSIS

We frst describe the data, then the empirical model, and fnally discuss the estimation results.

4.1. Data

4.1.1. Conict data. We use data on militarized interstate disputes (MIDs) rom theCorrelates o War (Jones, Bremer and Singer, 1996) project (COW, hereater). This data set is

an unbalanced panel that provides inormation on MIDs among approximately 190 states, atannual requency, starting in 1816 and ending in 2000. Militarized disputes include interstatewars, but also threats, demonstrations o orce, and other hostile interstate actions. This broadinterpretation o conict, which increases the requency o conicts in the data, is consistentwith our theoretical model. The COW has been the predominant data source in the empiricalliterature on the democratic peace hypothesis (e.g. Oneal and Russett (1997)). In its monadicorm, the COW records, or each year, whether a country is involved in a MID. But in ourtheory, the incentive to be aggressive depends on the regime types o pairs o countries. Wethereore use the data set in a dyadic orm, which contains conict inormation or each possiblepair o countries in the system.9

Our theoretical model is a static game. It does not identiy one country as having initi-

ated the conict (both may have chosen A), neither does it predict the duration o conictsor coalition ormation in multilateral disputes. Thereore, we estimate the probability o conictor country pairs using the so-called undirected orm o the data, and we exclude all dyad-year observations corresponding to either an ongoing dispute or a country joining an ongoingdispute.

9. Data or the historical period 18161992 in the COW are available in monadic orm, and orming the dyadicdata requires additional inormation not reported in the original data set. Maoz (1999) has augmented the standardmonadic COW data set and constructed a dyadic data set or the years 18161992. The COW v 3.02 contains militarizeddispute data in dyadic orm or the remaining years 19932000 included in our sample.


14/29


4.1.2. Classiying regime types. Data on political regime characteristics are rom thePolity IV data set (Marshall and Jaggers, 2002). Indexes measuring the competitiveness o polit-ical participation, competitiveness o the process or selecting the chie executive, regulation opolitical participation, openness o executive recruitment and constraints on the chie executiveare used to construct democracy and autocracy scores ranging rom 0 to 10 or each regime.

We ollow previous literature (Oneal and Russett, 1997 among many others) and use the dier-ence between the democracy and the autocracy indexthe combined Polity or net democracyscoreto classiy countries as dictatorships, limited democracies or ull democracies. Very highvalues o the score signal strong democratic institutions with strong checks on the leaders power.Very low values, instead, suggest the absence o any controls on the leader. Intermediate valueso the score correspond to regimes in which some limits on the leaders power exist, but notenough or the regime to qualiy as ully democratic. We will use such intermediate values todefne limited democracies. This approach has also been used in other studies that ocus onregimes in this middle range, sometimes known as anocracies or mixed regimes (see Gurr,1974; Goemans, 2000; Mansfeld and Snyder, 2005).10 The net democracy index rom Polity IVranges rom 10 to 10, taking 21 possible values in all. In the baseline model, we divide therange o possible net democracy scores into three subintervals o equal length. Thus, a dictator-

ship corresponds to values smaller than 3, a limited democracy to values between 3 and 3,and a ull democracy to values greater than 3.

A ew examples illustrate the category o limited democracies. Germany in the late nine-teenth and early twentieth century is a limited democracy. Power was concentrated in the handso the kaisers, but they could not repress the views o the population entirely. Under Wilhelm II,the Socialist party ormed a strong voting block in the Reichstag and even won the general elec-tions in 1913 (Craig, 1978, Chapter 8). Louis-Philippe, King o the French, was appointed byFrances Chamber o Deputies ater the July Revolution o 1830 (Howarth, 1961). During mucho his reign, France is a limited democracy. Napoleon III initially ruled dictatorially, but romthe 1860s, he gave the French Parliament more power (Wetzel, 2001). By 1870, France is clas-sifed as a limited democracy in our data. Latin American countries such as Argentina, Ecuador,

Nicaragua and Peru are heavily represented in our sample o limited democracies in both thenineteenth and twentieth centuries. Ecuador is classifed as a limited democracy or the longestamount o time, 114 years between 1854 and 1971. Middle Eastern countries begin to appear inthis classifcation in the post-war period. Anecdotal evidence, and the results in Mansfeld andSnyder (2005), suggest that these limited democracies may be highly prone to conict.

Mansfeld and Snyder (2005) argue that conicts are likely to occur during transitions todemocracy. In act, our empirical fndings will suggest that limited democracies are inherentlymore aggressive and not just during periods o political regime transitions. In the nineteenthcentury, the great European powers are limited democracies and are heavily involved in disputes.In Asia, Japan and Thailand are involved in conicts requently. In Arica and the Middle East,countries like Kenya and Jordan are limited democracies or a short period o time but engage

in disputes relatively requently during that period. Dyads o limited democracies also includeLatin American countries which experienced many conicts with European powers as well aswith each other. For example, Ecuador and Peru repeatedly ought over the Condor Mountainrange (Simmons, 1999). The late 1930s and early 1940s marked a violent turning point in thisconict and, during this period, both countries are classifed as limited democracies. BoliviaParaguay and ArgentinaChile are other conict-ridden dyads o limited democracies. Japan has

10. Merged COW and Polity data in dyadic orm, along with other controls considered in the democratic peaceliterature, are available rom Scott Bennetts EUGene Web site: http://eugenesotware.org.
http://-/?-http://-/?-


15/29


participated in many conicts with various opponents.11 Many o these disputes occurred duringthe nineteenth and the frst hal o the twentieth century. We return to this issue below, when weanalyse the estimates o the model over shorter time subsample.

4.2. Empirical modelOur empirical strategy has two steps. We frst utilize the Polity net democracy score to con-struct a set o dummy variables that characterize the regime types o each dyad in the sample.We then estimate the probability that a MID occurs within each possible dyad. The explana-tory variables include the regime-type dummies and controls typically considered in the demo-cratic peace literature. Our preerred estimation procedure is a panel logit regression modelwith fxed eects at the dyadic level. This simple methodology allows us to study the eectso democracy on conict using a exible unctional specifcation that links conict to politicalregime types. It is particularly well suited or investigating a possible non-monotonic relationbetween democracy and conict. The dyadic fxed-eects control or time-invariant sourceso unobserved heterogeneity, such as historical enmities between countries, or geographical

distance.The main restriction imposed by our methodology is the initial classifcation o regimes into

dictatorships (net democracy scores below 3), limited democracies (scores between 3 and 3)and ull democracies (scores above 3). As a robustness check, we consider a broader defnitiono a limited democracy, corresponding to net democracy scores between 6 and 6 (as well asother defnitions). Some countries which never meet the narrower defnition, like Spain, meetthe broader defnition in some years. Other countries, like France, meet the narrower defnitionduring certain years but meet the broader defnition more oten.

There are six possible confgurations o regime types or a dyad. As shown in Table 1, wedefne a corresponding set o six dummy variables where J {DiDi, DiLi, DiDe, LiLi, LiDe,DeDe}, representing dyad types ranging rom a pair o dictatorships (DiDi) to a pair o ull

democracies (DeDe) as well as all other possible combinations o regimes. The dummy variableD j equals one (D j,dt = 1) i at time t dyad d is composed o a pair o regime types j J (andzero otherwise). For our baseline defnition o a limited democracy, the composition o dyadtypes varies rom a maximum o 31% or a democracydictatorship pair, to 6% or a limiteddemocracy pair (Table 2).

To maximize the amount o data, we consider MIDs rather than wars. Even so, MIDs are rareevents. For instance, in our baseline model, which only considers dyads or which at least onedispute occurred in the data (see below), a total o 40 786 observations are included but only 5%o these involve MIDs (see Table 2).

Let M I Ddt = 1 i dyad d experienced a military dispute at time t, and M I Ddt = 0 otherwise.Our baseline empirical specifcation is a logit model that identifes the conditional probability oa MID or dyad d at time t+ 1 by

Prob{MIDdt+1 = 1|{D j,dt} j =LiLi, Xdt, cd} = G

cd + Xdt +

j =LiLi

j D j,dt

, (13)

11. The online appendix o this paper (available on the journal Web site) provides additional data detail. Table 1lists the countries that are classifed as limited democracies or the longest period o time in our sample. Table 2 reportsthe limited democracies that were engaged most requently in disputes. Table 3 shows the dyads o limited democraciesthat were most conict ridden.


16/29


TABLE 1Defnition o the regime-type dummy variables

Dictatorship Limited democracy DemocracyNetDem2 [10, 4] NetDem2 [3,3] NetDem2 [4,10]

Dictatorship DDiDi DLiDi DDeDiNetDem1 [10, 4]

Limited democracy DLiDi DLiLi DDeLiNetDem1 [3, 3]

Democracy DDeDi DDeLi DDeDeNetDem1 [4,10]

Notes: The table shows the set o regime-type dummy variables included in the regression mod-els or the baseline cut-o points in Polity scores. Each dummy variable D j is equal to one,when NetDem1 and NetDem2 assume values in the relevant intervals, and are equal to zerootherwise.

TABLE 2Sample description or the baseline model

Variable Mean Standard deviation Minimum Maximum

MID onset 0.05 0.21 0 1DDiDi 0.21 0.40 0 1DLiDi 0.13 0.34 0 1DLiLi 0.06 0.24 0 1DDeDi 0.31 0.46 0 1DDeLi 0.12 0.32 0 1DDeDe 0.17 0.38 0 1MajPower 0.54 0.50 0 1

LogCapRatio 2.40 1.67 0 8.44Allianced 0.20 0.40 0 1

Observations: 40 786

Notes: Summary measures or the dependent (MID onset) and explanatory variables in-cluded in the baseline regression Model 1 in Table 3.

where Xdt is a vector o controls, cd is a fxed eect defned at the dyadic level and {D j,dt} j =LiLiare all regime-type dummy variables with the exception o the limited democracies pair.12 Thefxed eects account or unobserved heterogeneity arising rom actors such as geography and

persistence o culture and institutions in the cross section o dyads.13A logit regression models parameters are not identifed i a subset o regressors perectly

predicts the outcome o the dependent variable. Since we include both dyadic and year fxedeects in model (13), only country pairs or which at least one MID occurs in the sample pe-riod can be included in the estimation; in addition, any year in which no MID occurred is also

12. G() is the logistic unction. For a review o qualitative response models, and their panel specifcations, seeWooldridge (2002).

13. For example, the colonial origin o countries in Arica and South America has played a large role in theirsubsequent development. See Acemoglu and Robinson (2006) or a review o much o this work.


17/29


excluded.14 To reduce issues o reverse causality in situations where regime transitions and dis-putes occur during the same year, we lag explanatory variables by one period, as noted by thetime subscripts in (13).

To estimate the parameters in (13), we use Chamberlains (1980) conditional maximum like-lihood procedure, which yields consistent estimates o all parameters except the fxed eects (thec

ds). Because o the exclusion oD

LiLi,dt in (13), the coefcient j on each regime dummyD

jmeasures the partial eect o regime type j relative to a pair o limited democracies. However,because the magnitude o the partial eects depend on the fxed eects which are not estimated,we only obtain an ordinalrather than a cardinalranking o the conict propensity o eachdyadic type. For example, a negative value or DiDi implies that a dyad o dictatorships is lesslikely to experience a conict than a dyad o limited democracies, but the parameter estimatedoes not reveal the magnitude o the eect.15

Returning to the theoretical model, Proposition 1 implies that a dyad consisting o two lim-ited democracies is the most conict ridden, so all coefcients j should be negative. Proposition2 implies that a dyad consisting o two ull democracies is the most peaceul, so we should haveDeDe < j or all j = DeDe. Finally, Proposition 3 implies that i a country ceases to be ullydemocratic, the probability o conict increases most in those dyads where it is matched with a

ull democracy. Proposition 3 cannot be tested in our baseline model because we do not estimatethe fxed eects. Instead, we estimate the parameters o (13) using two alternative procedures: apooled logit model, which excludes fxed eects, and a linear probability model, where G() isan identity map.

The set o controls Xdt includes year fxed eects to account or time varying actors thatare common to all dyadic pairs (e.g. the number o countries in the system, worldwide economicshocks, worldwide conicts.) Furthermore, we include cubic spline terms in the number o yearssince a country pair was last involved in a conict in order to capture the temporal dependence omilitarized disputes on past occurrence o disputes within a pair o countries with a exible unc-tional specifcation.16 We ollow earlier literature on democratic peace (e.g. Oneal and Russett,1997) in selecting the remaining controls. First, major powers may have an increased incentive

to engage in a MID i they think they can escape retaliation, but they may be less aggressivei they can achieve their objectives without conict. These eects are captured by the dummyvariable MajPower, which is equal to one i at least one o the two countries is a major power attime t. Second, an imbalance o military power may create conict (Bremer, 1992). The COWdata set contains an index o military capabilities, constructed rom measures o urban and totalpopulation, energy consumption, iron and steel production, military manpower and expenditure.The variable LogCapRatio that we include in the regressions is the logarithm o the maximumto the minimum level o military capabilities within each country pair taken rom the COW dataset. Third, i two countries in a dyad are ormally allied by a non-aggression or neutrality treaty,

14. The maximized value unction o the likelihood would be unbounded i these observations were instead in-

cluded in the estimation (see, e.g. Albert and Anderson, 1984). The years excluded rom the sample are 1818, 1819,1827, 1841, 1843, 1866 and 1891. The corresponding number o observations are less than 0.5% o the total.

15. Indeed, the partial eect o a dyad o type j relative to a pair o limited democracies is G(cd + Xdt + j )

G(cd+ Xdt), where hatted variables denote estimates o the corresponding parameters. Although absolute magnitudes

o these partial eects cannot be estimated when the cds are unknown, values or j s are sufcient to order the partialeects.

16. Formal tests support the use o the year fxed eects and the spline terms. The estimated coefcient on theour spline terms included indicate that the probability o a MID is higher when another MID occurred in the recentpast within the same dyad. The spline term specifcation allows to account or the temporal dependence with a exiblebut parsimonious specifcation (our parameters in our specifcation, see Beck, Katz and Tucker (1998) or an earlierspecifcation o cubic spline regressions in the democratic peace literature).


18/29


TABLE 3Regression modelsbaseline

Dependent variable: onset o a MID

Model (1) Baseline (2) (3) (4)

Panel a

DDiDi 0.58 0.0027 0.90 0.35[0.21]*** (


19/29


TABLE 4Partial eects or the pooled logit Model 3 in Table 3

Variable Partial eect % Change

DDiDi 0.0010*** 59

DLiDi 0.0006** 38DDeDi 0.0005 29DDeLi 0.0006** 36DDeDe 0.0012*** 74Pr(MID|{Li,Li}) = 0.0017

Notes: Pr(MID|{Li,Li}) denotes the probability or a pair o limited democraciesto engage in a militarized dispute, predicted by Model 3 in Table 3 when all D j sare set to zero, and the other controls are equal to the respective sample means.The reported partial eects indicate the discrete change in conict probabilitywhen the value o the corresponding dummy variable D j goes rom zero to one.The signifcance o a Wald test or the null that each eect is zero is reportednext to the corresponding estimate as in the ollowing ootnote. , *, ** and ***Signifcant at 11%, 10%, 5% and 1%, respectively.

TABLE 5Test o Proposition 3

Regime types Value o (DeT DeDe) (T T TDe)

(T =Di,T =Di) 0.0034*** 0.0013***(T =Li,T =Di) 0.0026*** 0.0008***(T =Li,T =Li) 0.0016 0.00003

Specifcation FE-LPM LOGIT

Notes: The table reports estimates o the conditional probabilities or each pair oregime types T and T that orm (12), when the inequalities are expressed with all

terms appearing on the let-hand side. A positive number indicates that the corre-sponding inequality is satisfed. The parameters are estimated using the fxed eectlinear probability Model 2 in Table 3, and the logit Model 3 in Table 3. The signi-icance o a Wald test or the null that each value is equal to zero is reported next tothe point estimates as in the ollowing ootnote. Reer to ootnote 24 or additionaldetail.*, ** and ***Signifcant at 10%, 5% and 1%, respectively.

or all additional controls included in the regression models with the exclusion o the year fxedeects and the cubic spline terms.

4.3.1. Baseline model. Parameter estimates o our baseline empirical specifcationthe conditional logit model with fxed eectsare shown in the frst column under Model

1 o Table 3. As can be seen rom the table, all estimates j are negative and signifcantat the 1% level. In other words, a dyad o limited democracies is more likely to experiencea militarized dispute than any other dyad type. Also, the estimated coefcient DeDe on theregime-type dummy variable DDeDe is the smallest among the dyad-type dummies. The p-valuesreported in the second column o Model 1 show that these dierences are signifcant at the1% level. We thus confrm previous fndings o a dyadic democratic peace (Babst, 1972; Levy,1988; Maoz and Russett, 1993).18 However, the parameter estimates show a non-monotonic

18. Only the dummy or alliance treaties is statistically dierent rom zero among the additional controls includedin the model. The existence o a treaty reduces the likelihood o a MID within the dyad.


20/29


TABLE 6Regression modelscomparison with Mansfeld and Snyder(2005)

Dependent variable: onset o a MID

Model (1) (2) (3) (4)

Panel a

DDiDi 0.69 0.22 [0.26]*** (0.04)** [0.18] (


21/29


TABLE 7Count o negative D j s in baseline Model 1 in Table 3 or alternative

cut-o points in Polity scores

Maximum

Minimum 2 3 4 5 6

6 (5,5) (5,5) (5,5) (5,5) (5,5)5 (5,4) (5,4) (5,3) (5,2) (5,2)4 (5,5) (5,4) (5,3) (5,3) (5,2)3 (5,5) (5,5) (5,5) (5,5) (5,5)2 (5,5) (5,2) (5,2) (5,3) (5,2)

Notes: The table reports the number o point estimates o the regime-type dummies D j that are negative as well as negative and signif-cant at the 10% levelrespectively, the frst and second entry in eachparenthesesunder alternative defnitions o the regime-type dummyvariable in the baseline regression Model 1 in Table 3. Limited democ-racies are defned or values o the Polity IV net democracy index be-tween a minimum value reported in the frst column and a maximum

value reported in the frst row. The baseline model results, which arereported in ull in column 1 o Table 3, are shown in bold.

these variables are not available or the ull sample, especially prior to 1945, which, as discussedin the next subsection, is a central historical period in our analysis. 20 Moreover, the two sets ofxed eects included in our baseline specifcation help capturing in part the impact o thesevariables. First, the year fxed eects can account or worldwide economic shocks and businesscycles. Hence, economic uctuations that are common to both members o a dyad and aecttheir incentives to be aggressive are captured by the year fxed eects. Furthermore, the dyadiclevel fxed eects account or some important actors, such as the relative disparity o naturalresource endowments and geographic distance, which are largely time invariant but aect the

likelihood o conict through their impact on variables such as dyadic trade ows and the degreeo disparities in national income.

4.3.2. Alternative empirical specifcations. We now consider two alternative empiricalmodels: a linear probability model with fxed eects and a pooled logit model. These modelsallow us to assess the robustness o our baseline results. Moreover, they allow us to compare themagnitudes o the eect o dierent regime types on the probability o conict. This allows usto test the prediction made in Proposition 3.

Robustness and urther results: The estimates o the linear probability model with dyadicfxed eects are also shown in Table 3. From the frst column under Model 2 o Panel a, allestimated coefcients on the regime-type dummies are negative and signifcantly dierent romzero. Furthermore, the coefcient on DDeDe is the smallest and signifcantly so, as shown bythe p values in the second column. As in the baseline model, a pair o limited democracies is

20. We augmented the baseline model specifcation to include a measure o dyadic trade among the set o controls(using historical data rom Barbieri, 1996). The estimated coefcient on dyadic trade was never statistically dierentrom zero. Moreover, due to missing observations, the sample size dropped by more than two thirds relative to thebaseline model, and due to the missing data, only the years 18711992 are included in the estimation. In addition, wealso included measures o gross domestic product per capita in PPP, restricting our sample to the post-1950 sample. Allregime-type dummies remained negative, although the point estimates were much noisier than in the ull sample. Theseresults are similar to the ones or the post-World War II subsample, discussed in the next section.


22/29


TABLE 8Regression modelssubsamples

Dependent variable: Onset o a MID

Model (1) (2) (3) (4)

Panel a

DDiDi 0.84 0.24 1.63 0.14[0.26]*** (0.14) [0.38] (0.03)** [0.65]** (0.65) [0.99] (0.11)

DLiDi 0.64 0.47 1.46 0.30[0.22]*** (0.03)** [0.40] (0.35) [0.71]** (0.50) [0.91] (0.22)

DDeDi 0.62 0.25 1.20 0.04[0.26]** (


23/29


pair o ull democracies. However, we should point out that, due to the very low requency oMIDs in the sample and the linearity o the model, a signifcant portion o ftted probabilitiesimplied by the parameter estimates (in act about 30%) turn out to be negative.

The pooled logit regression model (Model 3 in Table 3) always predicts probabilities betweenzero and one, but it excludes dyadic fxed eects. Instead, we enlarge the set o controls to in-

clude other standard measures used in the democratic peace literature. In particular, we includetwo measures o distance: the logarithm o the distance between the two countries capitals,LogDist, and a dummy variable, Contig, which indicates whether the country pair shares con-tiguous borders. Unlike the other two models, both within- and between-dyadic variation in thedata are used to estimate the model parameters. Furthermore, as or the linear probability model,all parameters are estimated and it is thus possible to quantiy the magnitudes o the partial e-ects associated with the dierent regime types. As shown in Model 3, Table 3, a dyad o limiteddemocracies is again the most conict ridden, and a dyad o ull democracies is the most peaceul(all results are signifcant at conventional levels).23 Table 4 displays the magnitudes o the partialeects o each regime type relative to the limited democracy pair. As shown, all partial eectsare signifcant at conventional levels, with the exception o the one corresponding to DDeLi,which is only barely signifcant. The magnitudes o the partial eects relative to the baseline

probability are again sizable. The likelihood that a dyad engages in a MID alls about 59% whenthe dyad changes rom a pair o limited democracies to a pair o dictatorships, and it alls about74% when the dyad changes rom a pair o limited democracies to a pair o ull democracies.

To summarize, the estimates o the alternative empirical models confrm the baseline resultso a non-monotonic empirical relation between the probability o engaging in conict and politi-cal regime types: a dyad olimited democracies is the mostlikely to engage in a MID and a dyadoull democracies is the leastlikely. Even our more conservative estimates suggest that the lim-ited democratization o a dyad o dictatorships raises the likelihood o conict by more than hal.

Democracies turn hawkish: The three empirical specifcations considered so ar providesupport or the dyadic democratic peace hypothesis. Here, we test an additional implication o

the model. In our theory, whatever the regime type o country i , conict becomes more likelywhen the opposing country j becomes more aggressive. Proposition 3 implies that this eectis maximal when country i is itsel a ull democracy. This prediction can be tested using theestimates o the linear probability model with fxed eects and the pooled logit model. Table 5displays the point estimates, and signifcance or the null that these estimates are zero, or thecombinations o conict probabilities i j that occur in inequality (12), or three dierent pairso regime types T and T.24 The table does not report results or the ourth case T = Di, T = Lisince by the symmetry o the model it is equivalent to the case T = Li, T = Di.

23. The addition o more controls, the elimination o the fxed eects and the dierent estimation sample changessome o the results or the remaining controls. The Alliance dummy is no longer signifcant, while MajPower and Log-CapRatio are now statistically signifcant. The parameter estimates confrm fndings in the democratic peace literature:They imply that country pairs with similar military capabilities and or which at least one country is a major power aremore likely to engage in a MID.

24. In the pooled and fxed-eect logit model specifcations, the dierence in the predicted probabilities condi-tional on the regime pairs, which orm (12), depend on all parameters including the fxed eects. Thus, we cannot testthe proposition using the conditional maximum likelihood estimates o the fxed-eect logit model since the fxed e-ects are not being estimated. In the linear probability model, instead, the dierence in the conditional probabilities in(12) simplifes to the dierence o the corresponding regime-type dummies. In the pooled logit model, the dierenceis calculated by setting the value o all regressors (excluding the regime-type dummies) at their sample means, as inthe previous calculations o the marginal eects. The variance covariance matrix or the Wald test in the logit model isobtained by the delta method.


24/29


The estimates o the linear probability model are reported in the frst column o the table.Point estimates or two o the three pairs o regime types satisy inequality (12) and are sig-nifcantly dierent rom zero. For the case T = Li, T = Li, the parameter estimates are tooimprecise to assess whether the inequality holds. The estimates o the pooled logit model arequalitatively similar and reported in the second column o Table 5. All three point estimates

satisy the inequality (12) and two are statistically dierent rom zero. On net, we thus fnd evi-dence that ull democracies turn hawkish in the sense o responding most aggressively to adversechanges in the environment. Conversely, using the estimates o the linear probability model, ia country changes rom a dictatorship to a ull democracy, then the probability o conict witha ull democracy decreases by about 90%, but the probability o conict with a dictatorship de-creases by only about 12%.25 Thus, even ull democratization does not signifcantly lower theprobability o conict with non-democratic states.

4.3.3. Other empirical results.

Alternative regime defnitions and transitions: In our baseline analysis, we divided the

range o possible net democracy scores, which goes rom 10 to +10, into three intervals oequal length, with limited democracies in the interval [3,3]. In Mansfeld and Snyder (2005),limited democracy instead corresponds to net democracy scores between 6 and 6. As shown inModel 4 o Table 3, the main fndings o our baseline model are confrmed i we use their classif-cation: dyads o limited democracies are most likely to engage in a MID, while ully democraticdyads are the most peaceul. However, the main issue raised by Mansfeld and Snyder (2005) isthat states that have recently been partially democratized are likely to be aggressively national-istic during a transitional period. Accordingly, they defne a transitional dummy that indicateswhether a transition rom a dictatorship to a limited democracy occurred in the previous 5 years.To study the eect o regime transitions, we include Mansfeld and Snyders transitional dummyin our regression, using both the [3,3] defnition o limited democracy rom our baseline em-pirical model and the broader classifcation [6,6].

As shown in Model 1 o Table 6, our baseline regression results are not aected when weinclude the transitional dummy: all regime-type dummies are still negative and signifcant atconventional levels. In other words, limited democracies are not only prone to conict duringa period o transition. They are more likely to be involved in conicts even long ater the tran-sition. Our baseline results appear robust to the inclusion o the additional control: dyads olimited democracies (new and old) are most conict ridden, while dyads o ull democracies aremost peaceul. In our baseline regression, Mansfeld and Snyders (2005) transitional dummyhas the wrong sign (signifcant at the 10% level): a recent democratization actually decreasesthe likelihood o conict. This result also holds when excluding our regime-type dummies(Model 2).

I instead we use the [6,6] classifcation, there is more support or Mansfeld and Snyders

(2005) theory. The transitional dummy is positive and barely signifcant at the 10% level. 26 Allthe coefcients on the regime-type dummies or dyads o regime types are still negative, butonly two are signifcant at conventional levels. We conclude that the data can support both theidea that transitional periods are dangerous, and the idea that limited democracies are inherently

25. That is (DeDe DeDi)/DeDi versus (DeDi DiDi)/DiDi.26. The statistical signifcance o the transitional dummy increases when excluding our regime dummies

(Model 4).


25/29


hawkish, but one or the other theory looks better depending on the exact classifcation o regimetypes.

Polity scores are based on qualitative assessments o governing institutions. Various defni-tions o limited democracies in terms o Polity scores appear in the literature, e.g. Fearon andLaitin (2003) use the range [5,5]. In our theoretical model, the leader o a limited democracy

can stay in power with the support o a hawkish minority, and it is not clear which Polity scorescorrespond more closely to this. Indeed, there is no reason why the defnition should necessar-ily be symmetric around zero. In general, limited democracies can be defned as states with netdemocracy scores in some set [x, y]. As a urther robustness check, we reestimate the baselineconditional logit model or dierent values o x and y.27 Specifcally, x takes values between6 and 2, and y takes values between 2 and 6, or a total o 24 alternative defnitions. Table7 summarizes the results o this analysis. Each entry enclosed in parenthesis reports the num-ber o point estimates or the regime-type dummies D j that are negative as well as negativeand signifcant at the 10% levelrespectively, the frst and second entryor the correspondingdefnition o a limited democracy. For each alternative defnition [x, y] o limited democracy,the minimum net democracy score x is given in the frst column, and the maximum y in the frstrow.

As indicated by the frst entries enclosed in parenthesis, under all 24 alternative defnitionso limited democracy, point estimates or the coefcients o all fve regime-type dummies arenegative or all regression models. As shown by the second entries in parenthesis, the largemajority o these negative point estimates are also statistically dierent rom zero across thealternative defnitions. Thus, although the signifcance somewhat depends on the classifcations,the point estimates indicate that limited democracies are the most conict prone, or a broadrange o possible defnitions o limited democracy.

Subsamples: An inspection o the data indicates that a large number o MIDs or limiteddemocracy pairs occurred beore World War II.28 Thereore, we reestimate the baseline modelover two subsamples o pre- and post-World War II data. As shown in Table 8, the results in the

pre-World War II subsample are analogous to the ones obtained over the entire sample. All coe-fcients on the regime dummies are negative and signifcantly dierent rom zero; the coefcienton DDeDe is again the smallest, and the dierence with respect to DDeDe is signifcant or our othe fve dummies. In the post-World War II sample, we fnd weaker evidence. Although the co-efcients on all regime-type dummies are still negative, only one diers signifcantly rom zero.Furthermore, only three o the dummies now have coefcients that are signifcantly larger thanDDeDe. Overall, the dyadic democratic peace hypothesis seems to be supported both pre- andpost-World War II.29 The hypothesis that limited democracies are inherently aggressive fndsless support ater World War II.

During the Cold War, there is a decline both in the number o limited democracies in oursample and in the number o MIDs in the western hemisphere. During this special period, coun-

tries within the Soviet bloc could not act independently, and the ear o nuclear war preventedminor disputes rom escalating (Gaddis, 2005, pp. 261263). To study i the old patterns might

27. For example, or x = 5 and y = 3, limited democracies are in the interval [5, 3], dictatorships are in[10, 6] and ull democracies in [4,10].

28. See Table 2 in the online appendix to this paper.29. However, i we consider only pre-World War I data, then we reproduce the result oFarber and Gowa (1997),

who ound no evidence or a democratic peace in this period. Thus, it is more correct to say that there is evidence inavour o a democratic peace only ater World War I.


26/29


return ater the erosion o the Soviet bloc, we consider the subsample o militarized disputesater 1984.30

The details o the empirical specifcation are the same as in our baseline model. The estima-tion results are reported as Model 3 in Table 8. A dyad o limited democracies is again moreconict prone than any other dyad type at conventional statistical signifcance levels. The result

holds also or the broader [

6,6] defnition o a limited democracy (not shown in the table). Adyad o ull democracies is again the most peaceul, but the dierence is statistically signifcantonly or two regime-type dummies. Model 4 o Table 8 reports the weaker support or our the-ory when we ocus on only the post-1989 subsample.31 These fndings are not surprising sincethe parameters o our model are only identifed by within-country variation o data (in terms ointerstate conict and regime change), and this is very limited in such a short sample (about halthe size o the post-1984 estimation).

Countries arising rom the disintegration o Yugoslavia and the Soviet Union have recentlybrought war back to Europe. Participants in these conicts, such as Armenia, Croatia, Georgia,Russia and Yugoslavia, are classifed as limited democracies (even using a narrow defnition).The results suggest that the nineteenth and early twentieth centuries constitute a better model orthe contemporary pattern o conict than the Cold War period.

5. CONCLUSION

According to Paine and Kant, democracy is good or peace because wars are disadvantageous tothe population at large. But i wars are caused by ear and distrust, then the link between democ-racy and peace is not obvious. In our simple model o Schellings dilemma, the average citizen isa coordination type. He wants peace, but he supports aggressive actions against potentially hos-tile enemies. The behaviour o coordination types is key to Schellings dilemma. I both leaderswere commonly known to be coordination types (just like their representative citizens), then orany pair o regime types, there would be two pure-strategy Nash equilibria: A A and P P . In ourmodel, this indeterminacy vanishes because each leader is not quite sure that the opponent is a

coordination type. Incomplete inormation and the possibility o dominant strategy types createsa ear spiral that leads to a unique interior equilibrium. In this equilibrium, the marginal incen-tive to be aggressive depends on the regime type. The combination o incomplete inormationand incentives or domestic political survival thus generates the relationship between politicalinstitutions and conict.

The model suggests a possibly non-monotonic relationship between democracy and peace,which is in act ound in the data. Our empirical analysis o militarized disputes in the nineteenthand twentieth centuries reveals that dyads consisting o two limited democracies are the mostconict ridden o all dyads (including mixed dyads). Dyads consisting o two ull democra-cies are peaceul, but as the environment becomes more hostile, ull democracies become more

30. As noted above, the inclusion o the fxed eects in the baseline model requires the estimation to only includedyads or which at least one MID occurred in any given subsample. Given this restriction, the amount o data are quitelimited i we start the subsample ater the all o the Berlin Wall in 1989, as discussed below. Arguably, the erosion othe Soviet bloc began beore 1989. In 1980, Solidarity was ounded in Poland, and in 1985 Mikhail Gorbachev came topower. The cut-o point 1984 is the midpoint between the arrival o Solidarity and the all o the Berlin Wall. In termso robustness o the results discussed below to other cut-os points, we always fnd the point estimates o the regimedummies to be negative through the 1980s. In terms o signifcance, the results would be identical choosing 1985, ratherthan 1984, but would be somewhat weaker in terms o signifcance i we choose 1982, 1983 or 1986. The signifcancedrops considerably post-1987 as the data available to estimate the parameters is signifcantly more limited.

31. Consistent with these results, Gowa (2008) estimates a pooled logit model ater the Cold War and fnds weaksupport or the democratic peace hypothesis.


27/29


aggressive aster than other regime types. These empirical fndings are consistent with the simplemodel.

A non-monotonic relationship between democracy and peace has important policy implica-tions. Many countries in the Middle East are classifed as dictatorships, or vacillate betweendictatorship and limited democracy. For example, between 1981 and 2000, countries such as

Aghanistan, Iraq, Kuwait, Libya, Saudi Arabia and Syria are classifed as dictatorships inour baseline model. Algeria, Egypt, Jordan, Iran and Tunisia are either dictatorships or limiteddemocracies at dierent times. According to President Bush (2003), the advance o reedomleads to peace. Unortunately, the data suggest that this may not be true or a limited advanceo reedom.

Acknowledgment. We thank the Editor and three reerees or many suggestions that improved the paper. We alsothank Igal Hendel, Massimo Morelli, Banu Olcay, Gerard Padro-i-Miquel, Ben Polak, Bruce Russett, Paola Sapienza andFrancesco Trebbi and many seminar participants. Sandeep Baliga thanks the Zell Center or Risk Research or fnancialsupport. The views expressed in this paper are solely the responsibility o the authors and should not be interpreted asreecting the views o the Board o Governors o the Federal Reserve System or o anyone else associated with theFederal Reserve System.

REFERENCES

ACEMOGLU, D. and ROBINSON, J. (2006), Economic Origins o Dictatorship and Democracy (Cambridge:Cambridge University Press).

ALBERT, A. and ANDERSON, J. (1984), On the Existence o Maximum Likelihood Estimates in Logistic RegressionModels, Biometrika, 71, 110.

ASPREY, R. B. (1991), The German High Command at War: Hindenburg and Ludendor conduct World War I(New York: Wiley).

AXELROD, R. (1984), The Evolution o Cooperation (New York: Basic Books).BABST, D. (1972), A Force or Peace, Industrial Research, 14, 5558.BALIGA, S. and SJSTRM, T. (2004), Arms Races and Negotiations, Review o Economic Studies, 71, 351369.BARBIERI, K. (1996), Economic Interdependence and Militarized Interstate Conict, 18701985 (Ph.D. Dissertation,

Binghamton University).BARRO, R. (1973), The Control o Politicians: An Economic Model, Public Choice, 14, 1942.

BECK, N., KATZ, J. and TUCKER, R.(1998), Taking Time Seriously: Time-Series-Cross-Section Analysis withA Binary Dependent Variable, American Journal o Political Science, 42, 12601288.BENNETT, S. (2006), Towards a Continuous Specifcation o the Democracy-Autocracy Connection, International

Studies Quarterly, 50, 313338.BOBBITT, P. (2008), Terror and Consent(New York: Knop).BREMER, S. (1992), Dangerous Dyads, Journal o Conict Resolution, 36, 309341.BUENO DE MESQUITA, B., MORROW, J., SILVERSON, R. and SMITH, A. (1999), An Institutional Explanation o

the Democratic Peace, American Political Science Review, 93, 791807.BUENO DE MESQUITA, B. and SILVERSON, R. (1995), War and the Survival o Political Leaders: A Comparative

Study o Regime Types and Political Accountability, American Political Science Review, 89, 841855.BUSH, G. W. (2003), Remarks by the President at the 20th Anniversary o the National Endowment or Democ-

racy (United States Chamber o Commerce, Washington, DC; http://www.whitehouse.gov/news/releases/2003/11/20031106-2.html).

CARLSSON, H. and VAN DAMME, E. (1993), Global Games and Equilibrium Selection, Econometrica, 61,9891018.

CHAMBERLAIN, G. (1980), Analysis o Covariance with Qualitative Data, Review o Economic Studies, 47,225238.CHASSANG, S. and PADRO I MIQUEL, G. (2009), Economic Shocks and Civil War, Quarterly Journal o Political

Science, 4, 211228.CHASSANG, S. and PADRO I MIQUEL, G. (2010), Conict and Deterrence under Strategic Risk, Quarterly Journal

o Economics, orthcoming.CONCONI, P., SAHUGUET N. and ZANARDI M. (2008), Democratic Peace and Electoral Accountability (Mimeo,

ECARES, Brussels).CRAIG, G. (1978), Germany 18661945 (Oxord: Oxord University Press).DEBS, A. and GOEMANS, H. (2008), War. Who is it Good For? The Relationship between Regime Type, the Fate o

Leaders and War (Mimeo, Yale University).DOWNS, G. and ROCKE, D. (1994), Conict, Agency and Gambling or Resurrection: The Principal-Agent Model

Goes to War, American Journal o Political Science, 38, 36280.


28/29


EPSTEIN, D., BATES, R., GOLDSTONE, J., KRISTENSON, I. and OHALLORAN S. (2006), Democratic Transi-tions, American Journal o Political Science, 50, 551569.

FARBER, H., and GOWA, J. (1997), Common Interests or Common Polities? Reinterpreting the Democratic Peace,Journal o Politics, 59, 393417.

FEARON, J. (1994), Domestic Political Audiences and the Escalation o International Disputes, American PoliticalScience Review, 88, 577592.

FEARON, J. (1995), Rationalist Explanations or War, International Organization, 49, 379414.

FEARON, J. (1996), Bargaining over Objects that Inuence Future Bargaining (Mimeo, Stanord University).FEARON, J. and LAITIN, D. (2003), Ethnicity, insurgency, and civ

Documents

Domestic Political Survival and International Conflict