Leader Survival and Government Spending

in Democracies and Dictatorships∗

Jeff CarterDepartment of Political ScienceThe University of Mississippi

jcarter3@olemiss.edu

March 20, 2014

Word Count: 11,301

Abstract

How do government spending and leader survival affect one another when leaders are motivatedby survival and policy concerns? A formal model developed to address this question indicatesvariation across regime type in winning coalitions’ spending preferences and the cost of leaderreplacement results in democrats and dictators securing their survival with different spendingdistributions and varying in their responsiveness to their constituents’ preferences. Consistentwith expectations, I find that proportionately higher military spending increases the probabilitya dictator remains in power but decreases the probability a democratic leader retains office,mean spending patterns vary across regime type, and greater variance in spending exists indictatorships than in democracies.

∗An earlier version of this manuscript received the Stuart A. Bremer Award for Best Graduate Paper presented atthe 2009 annual meeting of the Peace Science Society (International). Previous drafts were also presented at the 2010annual meetings of the International Studies Association and Midwest Political Science Association, and the 2010Empirical Implications of Theoretical Models Summer Institute. Susan Allen, Phil Arena, Scott Bennett, MichaelBernhard, David Carter, Matt DiGiuseppe, Ben Fordham, Ben Jones, Douglas Lemke, Amanda Licht, Burt Monroe,Tim Nordstrom, Glenn Palmer, Jim Shortle, Laron Williams, Chris Zorn, and, in particular, Heather Ondercin arethanked for helpful comments and discussions at various stages of this project. All remaining errors are my own.

Beginning in the spring of 2010, daily life in Greece frequently ground to a halt as workers

went on strike and citizens protested the austerity programs enacted by the government of

Prime Minister George Papandreou, head of the Panhellenic Socialist Movement party (BBC

2010). Papandreou’s government had substantially reduced government spending on public-

sector salaries, farm subsidies, pensions, and other programs in exchange for bailouts and debt

relief from the European Union, European Central Bank, International Monetary Fund, and

private investors. Not wanting the political blame for further, more severe austerity policies,

Papandreou announced on October 31, 2011 that a national referendum would be held regarding

the latest bailout. This plan was scrapped on November 3rd when he realized the referendum

would likely fail, which would prevent Greece from receiving the funds it needed to avoid default

(BBC 2011). Papandreou agreed to step down on November 6th and officially resigned as Prime

Minister on November 11, 2011, only two years into his government’s four year term (Smith and

Klington 2011, Telegraph 2011). Reflecting on his tenure during the spring of 2012, Papandreou

stood by his decision to pursue austerity despite the political consequences: “I knew that these

(policies) would not be popular. And I knew, sooner or later, I would be very much the target”

(Kakissis 2012).

A fundamental task of government is to finance state policies and programs through the allo-

cation of scarce economic resources. Considerable variation exists in how governments distribute

the resources they possess, particularly across regime type. Explanations for why democracies

and non-democracies allocate their resources differently often argue leaders are motivated by

their own political survival and democratic incumbents are politically accountable and respon-

sive to larger constituencies than are autocratic leaders (e.g., Bueno de Mesquita et al. 2003).1

Rather than being motivated exclusively by their political survival, though, most political in-

cumbents also have preferences over policy outcomes (for example, Fenno 1978, Strøm 1990).

As the example of former Prime Minister Papandreou suggests, some leaders are willing to pur-

sue spending policies that go against their constituents’ preferences, even at the cost of their

political survival.

This article examines the relationship between leader survival and government spending

when both an incumbent and her constituents hold preferences over how a government allocates

its economic resources and leaders can be motivated by survival and policy concerns. I develop

1The terms non-democracy, autocracy, dictatorship, and their respective derivatives are used interchangeablythroughout this article.

1

a game-theoretic model to analyze this situation. In equilibrium, a leader’s optimal play is a

function of each player’s spending preferences, an incumbent’s motivation for implementing pol-

icy, and the structural costs of leader replacement to a winning coalition. The model indicates

that a leader will rationally allocate spending distributions other than those preferred by her

key constituents when she is not exclusively motivated by her political survival. However, a

leader that deviates from the spending distribution preferred by her winning coalition faces a

higher probability of losing office than a responsive leader. Variation in the spending preferences

of democratic and non-democratic winning coalitions implies different optimal spending distri-

butions for survival-motivated leaders across regime type. The model also suggests variation

in the cost of replacement results in dictators allocating spending distributions closer to their

ideal points than democrats. Statistical analyses of the period from 1950 to 2001 are consistent

with the model’s expectations. I find the spending distribution that maximizes the probability

a dictator remains in office contains more military spending and less social spending than the

distribution that maximizes a democratic incumbent’s survival. I also find that dictators allo-

cate more of their state’s total spending to the military, on average, and have greater variance

in their spending distributions than do democrats.

This article makes three significant contributions to our understanding of the relationship

between political leaders and government policy. First, the formal and empirical results il-

lustrate the utility of allowing variation in leaders’ motivation for enacting policy and actors’

policy preferences. Doing so allows for more specific predictions about political behavior than

approaches that assume citizens do not vary in their policy preferences. Further, allowing a

leader to vary in the extent to which she is motivated by policy or survival concerns provides

a useful framework for considering how policies vary with dynamic constraints on a leader’s

behavior (Clark and Nordstrom 2005). Second, my findings indicate that the degree to which

an incumbent is responsive to her constituents’ preferences is a function of her motivation for

implementing policy. This qualifies the common claim that policy responsiveness is a defining

characteristic of democracy (e.g., Dahl 1971, Putnam 1993). Third, the statistical analyses are

the first to empirically demonstrate government spending influences leader survival and have

important implications for a number of topics. For example, my finding that a democratic

leader’s probability of survival is maximized with proportionately less military spending than a

dictator’s probability of survival provides micro-foundations for the claim that democracies are

less likely than non-democracies to fight relatively strong opponents (Reiter and Stam 2002).

2

The idea that leader survival and government spending are related is not novel. Most

notably, selectorate theory argues that leaders spend the resources available to them on pub-

lic goods and private benefits in the way that most efficiently ensures their political survival

(Bueno de Mesquita et al. 1999, 2003). The theoretical argument, formal model, and empirics

presented here differ from selectorate theory in a number of important ways. Theoretically, in

contrast to selectorate theory I do not assume that all leaders seek to maximize their probability

of survival or all citizens have the same preferences. With respect to the formal analysis, the

model developed here indicates that in equilibrium leaders will pursue policies that lower their

chances of remaining in power and, accordingly, that they will occasionally be removed from

office by their winning coalitions. Because selectorate theory assumes leaders are exclusively

survival-motivated, equilibrium behavior in the selectorate model of politics sees incumbents

always providing a mix of public goods and private benefits at least as good as the package

of benefits offered by their domestic challengers and, therefore, never being replaced by their

winning coalitions (Bueno de Mesquita et al. 2003, pg. 120). Finally, the empirical analyses pre-

sented here offer insights into the relationship between leader survival and government spending

not explored or predicted by the research associated with selectorate theory. I demonstrate that

the distribution of government spending influences patterns of leader survival and exhibits sig-

nificantly more variance in non-democracies than is the case in democracies. Selectorate theory

argues that spending on public and private benefits will influence a leader’s political survival

but never tests this expectation.2 Further, while Bueno de Mesquita et al. (2003) offer evidence

that mean levels of government spending differ as a function of regime type, they have no ex-

pectations about and therefore do not examine the variance of spending patterns in democracies

and dictatorships. Thus, the theoretical and empirical results presented here contribute to what

we know about the relationship between leaders and government policy and differ in important

ways from the most prominent research on the topic.

The remainder of the article proceeds as follows. The first section outlines the basic moti-

vations all leaders have for enacting policy: securing their survival and pursuing their personal

preferences. I then develop a simple model of the relationship between leader survival and

government spending that allows an incumbent and her winning coalition to have preferences

over the spending distribution and leaders to be motivated by policy outcomes and remain-

2Bueno de Mesquita et al. (2003) test the influence of public goods and private benefits on leader survival byestimating the effects of GDP growth (public good) and black market exchange rate premiums (private benefit) onthe probability an incumbent loses office (pgs. 302-310).

3

ing in power. The model identifies the general dynamics underlying the relationship between

government spending and leader survival that hold for all incumbents and their winning coali-

tions. The third section incorporates into the model systematic variation across regime type in

the spending preferences of incumbents’ winning coalitions and the cost of leader replacement.

This allows us to see how two institutional characteristics of democracies and dictatorships lead

to variation across regime type in the equilibrium behavior of leaders and winning coalitions.

The fourth empirically assesses a set of the model’s expectations. The article concludes with a

discussion of the larger implications of my findings for existing scholarship and future research.

1 Leader Motivations for Policy Decisions

Political incumbents have two basic motivations for enacting policies. Their first is to secure

their political survival, which is done by maintaining the support of their winning coalition

(Riker 1962, Bueno de Mesquita et al. 2003). An incumbent retains the support of her winning

coalition by enacting their preferred policies, a process known as policy responsiveness (inter

alia, Erikson, MacKuen and Stimson 2002, Burstein 2003). One of the most important ways

in which an incumbent is responsive to her key constituents is by allocating scarce resources to

their preferred policies and programs (Smith and Bueno de Mesquita 2011, Arena and Nicoletti

Forthcoming). Consistent with this claim, scholars have demonstrated that patterns of military

spending and social spending in democracies vary with public preferences across countries and

over time (among others, Wlezien 1996, Soroka and Wlezien 2005, Brooks and Manza 2007).

Much of the explicit theorizing and empirical work on policy responsiveness has focused on

democracies. This is unsurprising as it reflects the intuition that democratic politicians are

accountable for their actions while dictators rarely have to answer for their behavior.3 How-

ever, autocrats also secure their political survival through policy responsiveness. Key notes that

even “the least democratic regime ... needs the ungrudging support of substantial numbers of its

people. If that support does not arise spontaneously, measures will be taken to stimulate by tac-

tical concessions to public opinion” (1961, pg. 3). More recently, Gandhi and Przeworski (2006,

2007) argue that autocratic leaders maintain their hold on political power partially through

policy concessions. Similarly, a dictator keeps his ruling coalition satisfied and retains office

through power sharing agreements and policy concessions in Svolik’s (2009) model of authori-

3For example, Dahl argues “a key characteristic of a democracy is the continuing responsiveness of the governmentto the preferences of its citizens” (1971, pg. 1; see also Putnam 1993).

4

tarian politics. All leaders interested in remaining in power therefore have an incentive to enact

their constituents’ preferred policies.

Most political incumbents, though, are not solely motivated by retaining office. Politi-

cians also have personal preferences that influence the policies they pursue (e.g., Fenno 1978,

Wittman 1977, Strøm 1990).4 As both leaders and their constituents hold policy preferences,

incumbents can face a trade-off between enacting their preferred policy and the policy that best

ensures they remain in power (Strøm 1990, Bender and Lott 1996).5 The concept of “shirk-

ing” nicely captures this dynamic. Shirking occurs when an incumbent’s policy preferences

differ from her constituents’ preferences and she pursues her preferred policy (e.g., Kalt and

Zupan 1984, Bender and Lott 1996, Rothenberg and Sanders 2000). As the logic underlying

policy responsiveness would predict, incumbents are more likely to lose office when they deviate

from their constituents’ preferences (Canes-Wrone, Brady and Cogan 2002).

Two factors significantly increase a leader’s incentive to ignore her constituents’ preferences.

The first is when an incumbent no longer cares about her political survival. Consistent with

this idea, scholars have found that shirking is greatest when an incumbent cannot be re-elected

due to term limits, lame-duck status, or retirement (Persson and Svensson 1989, Carey, Niemi

and Powell 1998, Rothenberg and Sanders 2000). Second, a leader is more likely to deviate

from her constituents’ preferred policy as the institutional or structural cost to her winning

coalition of replacing her increases. For example, McGillivray and Smith (2008) demonstrate

that rulers whom are difficult to remove are more likely to defect from interstate cooperation

agreements that benefit the public than leaders whom are easily replaced by their winning

coalitions. Institutional explanations of the democratic peace often rely on a similar logic.

Claims that democratic incumbents are constrained by electoral accountability implicitly argue

that dictators are able to deviate from the public’s preferences due to the relative difficulty of

removing a dictator from power (e.g., Huth and Allee 2002). Incumbents therefore are more

likely to ignore constituent preferences and pursue their preferred policies as they become less

motivated by survival concerns or the structural cost of leader replacement increases.6

4A leader’s policy preferences are related to his or her political ideology or partisanship, but are also inherentlyidiosyncratic (Clark and Nordstrom 2005, Clark, Nordstrom and Reed 2008, Arena and Palmer 2009).

5This trade-off only obtains if a leader and her constituents prefer different policy outcomes. When a leader andher constituents have the same ideal point, there is no tension between an incumbent pursuing her preferred policyversus her constituents’ preferred policy.

6The political accountability literature offers a related explanation for why an incumbent would pursue policiesdifferent from those preferred by her constituents. This approach argues politicians have greater information aboutthe state of the world and the likely effect of policies than do their constituents, whom possess preferences over policy

5

The next section develops a general model of leader survival and government spending that

incorporates insights from the above discussion. Before moving on, though, it should be noted

that the idea leader survival and government spending are related is not novel. Most promi-

nently, selectorate theory argues the mix of spending on public goods and private benefits that

most efficiently secures an incumbent’s political survival is a function of the size of her win-

ning coalition (Bueno de Mesquita et al. 2003). Thus, it is worth highlighting three differences

between the argument and model developed here and selectorate theory. First, the model pre-

sented below allows citizens to vary in their preferences over government spending. In contrast,

selectorate theory assumes all citizens hold the same four policy preferences: more public goods,

more private benefits, more leisure time, and lower taxes (Bueno de Mesquita et al. 2003, pgs.

78-79, 108).7 As everyone wants the same things from their government, selectorate theory

implies that it does not matter who in society gets into a leader’s winning coalition. Second, the

decision to focus on the abstract concepts of public goods and private benefits allows selectorate

theory to speak to a wider range of state behaviors than the argument developed here. However,

the cost of this abstraction is that it prevents selectorate theory from making precise predictions

concerning government spending on particular policies. In the words of Bueno de Mesquita et al.

(2003), selectorate theory cannot “predict which specific bundle of public or private goods will

be emphasized in different societies” (pg. 104). This is particularly relevant for policies that

do not strictly fit the definition of public or private goods, such as government spending (see

Lake and Baum (2001) on the difficulties of treating spending as a public good). By allowing

citizens to vary in their preferences over how a leader distributes resources, the model devel-

oped here can make predictions about how specific “bundles” of government spending influence

leader survival while avoiding the conceptual and empirical issues of treating spending as either

a public or private good. Third, the model developed here allows leaders to be motivated by

their personal preferences over policy outcomes. This means that, in equilibrium, a leader might

choose a policy unpopular with her winning coalition that results in her removal from power. In

contrast, incumbent leaders always retain office in the selectorate model of politics (Bueno de

options and a politician’s competence (e.g., Canes-Wrone, Herron and Shotts 2001, Woon 2012). An incumbent politi-cian can be either a delegate or a trustee to her constituents (Fox and Shotts 2009). A delegate will ignore her ownexpertise and pursue the policies preferred by her constituents. A trustee is willing to ignore her constituents’ prefer-ences and enact welfare maximizing policies. Within the framework discussed above, then, delegates are responsiveto their constituents’ preferences while trustees enact their personally preferred policy.

7Selectorate theory does allow variation in actors’ personal preferences through the concept of “affinity.” Affinity,however, does not refer to policy preferences but “is simply a preference for one individual over another, independentof the policies of the individuals” (Bueno de Mesquita et al. 2003, pg. 61, emphasis added).

6

Mesquita et al. 2003, pg. 120). This result follows from selectorate theory’s assumption that all

leaders are motivated by securing their political survival. Accordingly, in equilibrium leaders

propose a mix of public and private goods that ensures members of their winning coalitions

will not defect to a domestic challenger. I view it as a strength of the model developed here

that it can account for cases in which an incumbent’s tenure is shortened because she made

a rational decision to pursue a policy counter to the preferences of her constituents; as in the

case of former Greek Prime Minister Papandreou. In sum, while selectorate theory speaks to

similar issues, the model developed here differs in important ways and allows for more specific

predictions regarding the relationship between leader survival and government spending.

2 A Simple Model of Leader Survival and Spending

I model the relationship between leader survival and government spending as a game between

country i’s incumbent leader (Li) and her winning coalition (Wi). The model is informally

outlined here and fully characterized in the appendix. The incumbent controls the distribution

of government spending (xi). Let xi ∈ [0, 1] be a one-dimensional policy space that represents

the proportion of total government spending dedicated to the military. I assume the incumbent

and members of her winning coalition hold symmetric, single-peaked preferences over xi. The

functions z(xi) and v(xi) represent the incumbent’s and winning coalition’s respective valuations

of xi. These functions are maximized, respectively, at the incumbent’s preferred distribution

of spending (xL) and the distribution of spending preferred by the median member of the

incumbent’s winning coalition (xW ). I assume incumbents prefer to remain in office (Ii) over

being ousted (Oi) by their winning coalitions. The term αi ∈ [0, 1] identifies the degree to

which a leader obtains utility from the spending distribution directly relative to whether she

remains in office. I define αi as a function of a leader’s motivation for implementing policy

(mi ∈ [0, 1], where zero represents a purely survival-motivated incumbent and one represents a

purely policy-motivated incumbent) and the structural cost of removing her from power to her

winning coalition (si ∈ (0, 1]). The influence of the cost of leader replacement on α, though,

is a function of the degree to which an incumbent cares about retaining office. That is, the

cost of removal does not enter into the decision calculus of a leader motivated exclusively by

securing her personally preferred policy or remaining in power, but will influence an incumbent

motivated by policy and survival concerns. To capture this relationship, αi(mi, si) is defined as

7

follows:

αi(mi, si) =

0 if mi = 0

mi + (1−mi)(si) if 0 < mi < 1

1 if mi = 1

(1)

The model assumes replacing an incumbent imposes a cost (ci > 0) on a winning coalition.

This cost is a function of two factors. First, the cost of leader replacement is assumed to be

an increasing function of how pleased the winning coalition is with the distribution of spending

(v(xi)). This captures the idea that it is costlier to remove a leader when the winning coalition

is relatively happy with the incumbent than when it is comparatively displeased. Second, and as

discussed above, there are structural costs (si ∈ (0, 1]) to removing any ruler (e.g., McGillivray

and Smith 2008, Debs and Goemans 2010). We can think of these structural costs as the a priori

costs of replacing a leader to a winning coalition regardless of how they value the distribution of

spending (e.g., voting in an election is costly regardless of an incumbent’s popularity). Replacing

a leader brings a benefit (bi) to a winning coalition. In substantive terms, bi can be thought of

as the winning coalition’s relative preference for a domestic political challenger compared to the

incumbent leader. More formally, bi is an exogenous shock to the model revealed after a leader

has distributed xi. Upon receiving xi and the uncertainty over bi being revealed, a winning

coalition decides whether to replace (ri) or keep (ki) the incumbent leader.8 The variable p

takes on a value of one if the winning coalition replaces the incumbent leader and zero if it

keeps the incumbent. The utility functions for a leader and her winning coalition are written as

follows:

ULi = αi(mi, si)z(xi) + (1− αi(mi, si))[p(bi, ci(v(xi), si))Oi + (1− p(bi, ci(v(xi), si)))Ii] (2)

UWi =

v(xi) if ki

v(xi) + (bi − ci(v(xi), si)) if ri

(3)

8See Debs and Goemans (2010) for an analogous approach to modeling a domestic population’s decision to removean incumbent.

8

The game is solved for pure strategy subgame perfect equilibria using backwards induction.

From Mas-Colell, Whinston and Green (1995, pg. 276), there exists a unique equilibrium in

pure strategies for any distribution of a model’s parameters in dynamic games of complete and

perfect information. Beginning with the winning coalition’s decision, it keeps the incumbent if

and only if the cost of replacing her is greater than the benefit of installing a new leader:

Wi plays ki iff bi < ci(v(xi), si) and ri otherwise. (4)

Equation 4 indicates that whether an incumbent is removed from office depends on the

relative cost and benefit of replacement to her winning coalition. An incumbent therefore

influences the probability she retains office through the distribution of government spending.

As the cost of leader replacement is increasing in how much her winning coalition values the

spending distribution, a leader minimizes the probability she is removed from office by allocating

her winning coalition’s preferred mix of spending (xW ). Her personal valuation of xi, though,

also influences her utility. We can see this by examining the partial derivative of Equation 2

with respect to αi:

∂UL

∂α(mi, si)= z(xi)− [p(bi, ci(v(xi), si))Oi + (1− p(bi, ci(v(xi), si)))Ii] (5)

Equation 5 implies that as a leader’s utility is increasingly drawn from the value she places

on the distribution of spending, she obtains more utility from a distribution closer to her ideal

allocation (the first term is positive) and less from whether or not she remains in power (the

second term is negative). The optimal xi for an incumbent therefore is a function of her preferred

distribution of spending (xL), the distribution of spending that maximizes her chances of survival

(xW ), and the extent to which she derives utility from the spending distribution directly and/or

her political survival (αi(mi, si)). Specifically,

x∗i = αi(mi, si)(xL) + (1− αi(mi, si))(xW ) (6)

It is worth providing the intuition behind Equation 6. An incumbent can derive utility from

the distribution of social and military spending directly as a function of her personal valuation

and/or indirectly based on how the spending distribution influences whether she stays in power.

If she derives utility exclusively from the distribution of government spending (α(mi, si) = 1),

9

then she will allocate her preferred spending mix (xL). If a leader derives utility exclusively

from retaining office (α(mi, si) = 0), then she will allocate the distribution of spending that

maximizes her likelihood of staying in power (xW ). If an incumbent derives utility from both

the enacted spending distribution and whether she remains in power (0 < α(mi, si) < 1), then

her optimal allocation is a function of her preferred spending distribution and the mix of guns

and butter that maximizes her probability of retaining office, weighted by α(mi, si).

To demonstrate how x∗i varies with αi(mi, si), xL, and xW , Figure 1 presents x∗

i across all

possible combinations of mi and si given that the leader’s personal preference is to allocate

eighty percent of government spending to the military (xL = 0.8) and the median member

of her winning coalition prefers that twenty percent of spending be dedicated to the military

(xW = 0.2). The optimal distribution of spending is increasing in the surface plot’s location

on the z-axis and is represented by progressively darker colors. The structural costs of leader

replacement to a winning coalition and the incumbent’s motivation for enacting a particular

spending distribution are located on the x-axis and y-axis, respectively.

Figure 1 nicely illustrates three characteristics of equilibrium spending distributions. First,

increasing the degree to which an incumbent is policy-motivated moves the optimal spending

distribution closer to her ideal policy. We can see this in Figure 1 by holding the structural

cost of removal constant at any value and noting the height and color of the surface plot as

the extent to which she is policy-motivated increases. Second, unless an incumbent is purely

motivated by survival or policy concerns, x∗ gets closer to her ideal spending distribution as

the structural costs of leader replacement increases.9 Holding her motivation for implementing

policy constant, increasing the structural cost of removal is associated with x∗ moving away

from the distribution preferred by the winning coalition and closer to the incumbent’s preferred

allocation of government spending. Substantively, this implies that the harder it is for a winning

coalition to remove a leader, the more likely she is to deviate from her constituents’ policy

preferences. Third, Figure 1 nicely highlights that an incumbent not exclusively motivated by

her political survival will pursue policies in equilibrium that increase the probability she will be

removed from office (i.e., all x∗ = xW ). Thus, the model captures the behavior of leaders who

knowingly pursue policies that damage their prospects for remaining in power.

The model developed here captures the general relationship between leader survival and

9Recall that the costs of removal do not influence αi(mi, si), and thus x∗, when an incumbent is exclusivelysurvival-motivated or policy-motivated (Equations 1 and 6).

10

Policy Motivated (m

)

0.0

0.5

1.0

Structural Costs (s) 0.0

0.5

1.0

0.0

0.5

1.0

x*

Figure 1: x∗ as a function of xL, xW , and α(m, s). Calculations assume xL = 0.80 and xW = 0.20.Darker colors represent higher values of x∗.

government spending when a leader and her winning coalition hold policy preferences and an

incumbent can be motivated by her political survival and/or personal preferences. One cost of

the model’s generality is that it abstracts away from the influence of regime type. The next

section addresses this issue.

3 The Influence of Regime Type

Two differences between democracies and dictatorships are of particular importance to the

relationship between leader survival and government spending: the spending preferences of the

members of democratic and non-democratic winning coalitions and the relative cost of removing

11

democratic and non-democratic incumbents. These respective differences and their implications

for the formal model are discussed in turn.

3.1 Spending Preferences of Winning Coalitions

The winning coalitions of autocratic and democratic incumbents systematically differ in

their size and socio-economic composition. The relationship between the size of a leader’s win-

ning coalition and regime type is well understood: democratic incumbents have larger winning

coalitions, in both absolute terms and relative to the size of a country’s selectorate, than au-

tocratic incumbents (Bueno de Mesquita et al. 1999, 2003).10 Less appreciated, though, is

the relationship between regime type and who gets into a leader’s winning coalition. The win-

ning coalitions of democratic leaders consist of proportionately more members of the general

public and fewer civilian elites and members of the military than do the winning coalitions

of autocratic incumbents. The comparatively small winning coalitions of autocratic leaders are

made-up almost exclusively by members of the civilian elite and/or military (Bueno de Mesquita

et al. 2003), with the relative influence of the civilian elite and military varying across types of

non-democracies (e.g., civilian dictatorship vs. military junta).11 The political institutions of

contemporary democracies make it impossible for a democratic incumbent to retain office with

only the support of her country’s civilian elite and/or military. The relatively high levels of

political participation and contestation associated with democracy result in democratic leaders

requiring the support of a large portion of the general public to remain in power (among others,

Dahl 1971, Boix 2003). It therefore follows that democratic winning coalitions are composed of

proportionately more members of the general public and fewer members of a society’s civilian

and military elite than are autocratic winning coalitions. This implies that the median voter

in a democratic leader’s winning coalition is a member of the general public while the median

“voter” in a dictator’s winning coalition is a member of the civilian elite or military.

This variation in the composition of autocratic and democratic winning coalitions has im-

plications for the model developed above because, in general, a society’s elite and public hold

different preferences over patterns of government spending. Compared to the elite, members

10Bueno de Mesquita et al. (2003) are clear that “a large selectorate and a large winning coalition do not inthemselves define a democracy” (pg. 72). However, a large winning coalition is viewed as a necessary condition fordemocratic government: “democracies require larger coalitions than autocracies or monarchies” (pg. 72).

11Space considerations prevent me from adequately investigating variation in the relationship between governmentspending and leader survival among types of non-democracies. I briefly return to this issue in the concluding section.

12

of the public should prefer relatively more resources be allocated to social spending and fewer

resources be allocated to military spending. 12 This claim follows from four observations. The

first two concern the preferences of the public and civilian elite over social spending. First, the

general public derives more direct benefits from social spending than do the wealthy civilian

elite, whom can provide themselves with the services that the public receives via the welfare

state.13 It should be highlighted that this is the case with all means-tested social welfare pro-

grams by definition. Second, spending on social programs typically is financed through taxes

on the wealth of the civilian elite (Przeworski et al. 2000, Boix 2003). Thus, the civilian elite

bear the brunt of the costs of the social welfare state while deriving relatively fewer benefits

than members of the public. It then follows that the general public should prefer its government

allocate more of its resources to social spending than the civilian elite. This claim is consistent

with the overall negative relationship between income and support for the welfare state in the

United States and Europe (Cook and Barrett 1992, Jæger 2006) and research on the respective

preferences of the general public and the wealthy in the United States. Gilens (2009, 2012) finds

that the U.S. public is more supportive of social welfare programs than the affluent, defined as

the top 20% in terms of income. Focusing on the preferences of the very wealthy, defined as

the top 1% of income earners, Page, Bartels and Seawright (2013) find that the general public

prefers increases in spending on health care, food stamps, and social security while the very

wealthy prefer reductions in social spending.

The third and fourth observations concern the preferences of the public and military over

military spending. The third is that military training socializes members of a state’s armed forces

to value a stronger military and favor higher military spending than the civilian population

(Huntington 1957, Nordlinger 1977, Geddes 2003). Fourth, the public and members of the

military have reasons to assess military spending beyond what is required for national security

differently. All citizens benefit from the military expenditures needed to provide the public good

of national security (Samuelson 1954, Olson 1965). Military spending beyond this level can crowd

out consumption spending popular among the public (Sprout and Sprout 1968, Fordham and

12To be clear, I am not claiming that all members of the general public prefer higher social spending and lowermilitary spending than all members of the civilian elite and military. Instead, drawing on existing research, I arguethat, in general, members of the public prefer more social spending and less military spending than the wealthy eliteand members of the armed forces.

13Members of the elite do derive some benefits from social spending. In particular, the elite would benefit fromsome of the long-term consequences of increased social spending, such as an educated and healthier workforce. Thatsaid, it is members of the public, and not the elite, that would be better educated, healthier and have a longerlife-expectancy due to government spending on social programs.

13

Walker 2005). However, military expenditures over and above what is necessary for national

security finances private benefits and club goods for members of the military.14 These private

and club goods include, but are not limited to, higher salaries for members of the military, the

ability to purchase goods and services at reduced prices at base exchanges, and free or subsidized

housing.

Public opinion research supports the claim that the public and members of the military have

different preferences over military spending. Bachman, Blair and Segal (1977) find that members

of the military prefer higher military spending than do members of the public. Addressing a

guns-versus-butter trade-off, Holsti (1998, 2001) and Szayna et al. (2007) find that members of

the public are more likely than members of the military to think that military spending should

be decreased in order to increase education spending. The variation in the spending preferences

of the military and public is consistent with the greater weight members of the military place

on military superiority (Szayna et al. 2007) and is driven by a combination of self-selection

and socialization (Bachman et al. 2000). Considered jointly, these observations imply that, in

general, members of the public should prefer lower military spending and higher social spending

than members of the military and/or the wealthy civilian elite.

Due to variation in the political representation of the elite and public across regime type and

their spending preferences, the median member of a democratic winning coalition prefers his

government allocate fewer resources to military spending and more resources to social spending

than the median member of a dictator’s winning coalition. It is straightforward to incorporate

this insight into the model. Define a state’s form of government as gi ∈ {A,D}, where A

represents an autocracy andD represents a democracy. As the distribution of spending preferred

by the median member of an incumbent’s winning coalition is defined as xW , variation in the

preferred spending distribution of the median members of autocratic and democratic leaders’

winning coalitions is represented by the inequality xDW < xA

W .

3.2 The Cost of Leader Removal

The relative cost of removing an incumbent leader from power varies as a function of regime

type. Namely, it is much harder for a domestic population to replace a dictator than a democrat

(among others, Lake and Baum 2001, McGillivray and Smith 2008, Debs and Goemans 2010).

14See Buchanan (1965) on club goods generally and Smith and Bueno de Mesquita (2011) and Arena and Nicoletti(Forthcoming) on the relationship between club goods and leader survival.

14

Replacing a democratic leader with a domestic challenger requires that the incumbent be de-

feated in her re-election bid. Voting for the opposition in an election is a relatively costless act

of political participation in a democracy (Lake and Baum 2001). Non-democratic incumbents

generally are removed from power through irregular events like a coup or revolution (Chiozza

and Goemans 2011). Participation in such events is much riskier and, especially if they fail,

costlier than voting for a domestic challenger in a democracy (McGillivray and Smith 2008).

Thus, the structural cost of leader replacement is lower for democratic winning coalitions than

it is for the winning coalitions of non-democratic leaders.

The importance of this variation in the cost of removing a leader is hard to overstate. Debs

and Goemans (2010) argue that the “costliness of replacing rulers is a significant, if not the

most significant, difference between dictatorship and democracy” (pg. 435). This fundamental

difference in autocratic and democratic politics is overlooked in the model presented in the

previous section. However, it can be incorporated into the model by constraining the structural

cost of leader replacement to a winning coalition for a given valuation of xi to be higher in

non-democracies than in democracies: sDi < sAi ∀ v(xi).

3.3 Formal Results

Accounting for variation in the spending preferences of winning coalitions and the relative

cost of removing an incumbent across regime type yields a number of insights concerning the

relationship between leader survival and government spending. For expositional purposes, I

present the intuition behind the proofs here and provide formal details in the appendix. Fol-

lowing the logic of backwards induction, I begin with the winning coalition’s decision to remove

a leader.

Proposition 1 argminx

F (ci(v(xi), si)) | gi = D < argminx

F (ci(v(xi), si)) | gi = A

Proof. See appendix.

Proposition 1 indicates the distribution of spending that minimizes the probability a demo-

cratic leader is removed from power contains relatively more social spending and relatively less

military spending than the distribution of spending that minimizes the probability a dictator is

removed from office. While she might rationally allocate a different mix of military and social

spending (see Equation 6 and Figure 1), a leader minimizes her probability of being replaced by

15

providing her winning coalition with their preferred distribution of government spending. Mem-

bers of the public generally prefer more social spending and less military spending than do the

civilian elite or military and the median member of democratic winning coalitions is drawn from

the general public while the median member in an autocratic winning coalition hails from the

military or civilian elite. Accordingly, the spending distribution that maximizes the probability

a democratic leader remains in power contains more social spending and less military spending

than the analogous combination of spending for a dictator. The optimal spending distribution

for survival-motivated democrats and dictators follows from this result.

Corollary 1 If mi = 0, then (x∗i | gi = D) < (x∗

i | gi = A).

Proof. See appendix.

Corollary 1 claims that a survival-motivated democratic incumbent will allocate more re-

sources to social spending and fewer resources to military spending than a survival-motivated

dictator in equilibrium. This follows directly from the spending distributions that minimize the

probability of removal for democratic and non-democratic leaders (Proposition 1). Substan-

tively, Corollary 1 provides a rational explanation for the empirical observations that, compared

to autocracies, democracies generally spend less on the military and more on social spending

(among others, Brown and Hunter (1999), Fordham and Walker (2005), and Huber, Mustillo

and Stephens (2008); although see Mulligan, Gil and Sala-i-Martin (2004)).

Proposition 1 and Corollary 1 follow closely from the assumption that the median member of

a democratic leader’s winning coalition prefers less military spending and more social spending

than the median member of an autocrat’s winning coalition. It is instructive to consider what

happens if this assumption is changed. If we assume the median members of autocratic and

democratic winning coalitions do not vary in their spending preferences, then the model implies

no variation across regime type in the mix of spending that maximizes an incumbent’s probability

of survival or the spending distribution allocated by survival-motivated leaders. Alternatively,

we could assume the public prefers lower social spending and higher military spending than the

civilian elite and military. The model implies two things given this assumption. One, the mix

of spending that maximizes the probability a democratic leader retains office contains relatively

less social spending and more military spending than the corresponding spending distribution

for an autocratic leader. Two, survival-motivated dictators should allocate more resources to

social spending and fewer resources to military spending than survival-motivated democrats,

16

an expectation that runs counter to the empirical research cited above (e.g., Fordham and

Walker 2005, Huber, Mustillo and Stephens 2008). Thus, the model is consistent with a set of

known empirical regularities given the assumption that the public prefers more social and less

military spending than the civilian elite or military, but not when this assumption is changed.

Where Corollary 1 pertains to incumbents driven exclusively by the desire to retain office,

Proposition 2 addresses the behavior of leaders motivated by policy and survival concerns.

Proposition 2 (|x∗ − xL| |gi = A) < (|x∗ − xL| |gi = D) ∀ 0 < mi < 1

Proof. See appendix.

Proposition 2 indicates that a dictator will allocate a distribution of government spending

closer to her ideal spending mix than will a democratic leader when both are motivated by policy

and survival concerns. This result is driven by how the structural costs of leader replacement

and an incumbent’s motivation for enacting policy influence equilibrium behavior. A leader

driven by policy and survival considerations is constrained by her winning coalition’s preferences

to the extent they can easily remove her from power. Accordingly, the equilibrium spending

distribution will be closer to a leader’s ideal point as the structural cost of replacement increases.

The cost of removing an incumbent across regime type therefore implies that dictators will

allocate spending distributions closer to their ideal points than democrats when leaders are

motivated by survival and policy outcomes.

4 Empirical Analyses

The formal model has several implications for the relationship between leader survival and

government spending. I find empirical support for a set of the model’s expectations during the

period from 1950 to 2001. I first present results on how the distribution of government spending

influences the survival of democratic and non-democratic leaders before reporting my findings

regarding patterns of government spending in democracies and dictatorships.15

15The data and code needed to replicate all of the analyses associated with this manuscript will be made availableupon publication.

17

4.1 Leader Survival

The model predicts the spending distribution that best secures a democratic leader’s survival

contains less military spending and more social spending than the distribution that minimizes the

probability a dictator loses office. I test this expectation using data on the universe of political

executives from 1950 to 2001. With the exception of variables measuring the distribution of

government spending, all data were taken from the replication materials associated with Debs

and Goemans (2010).16 The dependent variable is the number of days a leader has been in

power and is drawn from the Archigos project (Goemans, Gleditsch and Chiozza 2009).17

Testing the formal model’s expectation requires five explanatory variables. Spending Distri-

bution is operationalized as the portion of a government’s total expenditures that are allocated

to the military in year t, which precisely mirrors xi from the formal model. Total government

expenditures and military spending data are drawn from the Penn World Tables, version 7.1,

(Heston, Summers and Aten 2012) and Correlates of War (COW) National Material Capabili-

ties, version 3.02 (Singer, Bremer and Stuckey 1972) data sets, respectively. The formal model

assumes that a leader’s winning coalition holds single-peaked, symmetric preferences over the

spending distribution. This implies that spending distributions the same degree less than or

more than xW should increase the probability a winning coalition removes an incumbent to the

same degree. The empirical model captures the formal model’s assumption with the variable

Spending Distribution2.18 The formal model’s dichotomous treatment of regime type is prox-

ied with the variable Democracy, coded one in year t if a leader’s government has a value of

+7 on the 21-point Polity2 index (Marshall and Jaggers 2005) and zero otherwise. The inter-

action terms Spending Distribution*Democracy and Spending Distribution2*Democracy allow

the model to identify the effect of government spending on leader survival in democracies and

dictatorships.

I control for a number of factors thought to influence leader survival. Interstate War takes on

16Available at http://www.rochester.edu/college/faculty/hgoemans/research.htm.17I code leader failure as occurring when an incumbent is removed by a domestic audience. Leader turnover due to

an incumbent being replaced by a foreign power or dying from natural causes reflects fundamentally different datagenerating processes.

18Additionally, following Keele’s (2010) suggestions, diagnostics indicate that Spending Distribution has a non-linear effect on leader tenure in democracies and dictatorships and that including Spending Distribution2 is method-ologically appropriate. More specifically, Spending Distribution and Spending Distribution*Democracy violate thenon-proportional hazards assumption when included alone in the duration models but not when I also includeSpendingDistribution2 and Spending Distribution2*Democracy in the models. These diagnostic results hold whenI use splines to assess the linearity of the effects of Spending Distribution and Spending Distribution*Democracy onleader tenure.

18

a value of one if a leader’s country is involved in a war in year t and zero otherwise. I account for

whether the leader’s country was victorious (War Win), vanquished (War Loss), or obtained a

draw (War Draw) in an interstate war during her tenure. Each war outcome variable is modeled

as a decay function of the number of years since a given outcome was achieved.19 Interstate war

participation and outcome variables are based on data from Brecher and Wilkenfeld (1997). I

control for a state’s involvement in a Civil War in year t with a dichotomous indicator based on

the UCDP/PRIO Armed Conflict Database (Gleditsch et al. 2002). The model also accounts for

four aspects of a state’s economy: its (logged) GDP per capita (Maddison 2008), GDP growth

(Maddison 2008), Trade Openness (International Monetary Fund 2008), and Change in Trade

Openness (International Monetary Fund 2008). A state’s (logged) Population (Singer, Bremer

and Stuckey 1972) also is included as a control variable. The models include three leader-level

characteristics argued to influence an incumbent’s tenure: a leader’s Age when he or she entered

office, the number of previous Times in Office he or she served, and whether he or she came

to power via an Irregular Entry. These three variables are drawn from the Archigos project

(Goemans, Gleditsch and Chiozza 2009). Finally, the models capture region-specific variation

in leader survival with a set of variables that identify whether a state is located in Africa, the

Americas, Asia, or the Middle East (Europe serves as the baseline category).

I estimate leader survival using semi-parametric Cox models. The Cox model is preferable to

parametric event history models when the phenomenon of theoretical interest is the relationship

between a set of covariates and the likelihood of a subject failing and not the distributional

form of subject failure (Box-Steffensmeier and Jones 2004). The models estimated here account

for non-proportional hazards between covariates and leader survival. Analysis of the Schoenfeld

residuals was used to determine whether the influence of an explanatory variable on the hazard

of a subject failing was constant over time (Box-Steffensmeier and Jones 2004). Following

Box-Steffensmeier and Zorn (2001), any offending variable was interacted with the natural log

of a leader’s tenure in office up to time t. I also extend the standard Cox model to capture

shared frailty among leaders of the same state. Shared frailty event history models account

for unobserved heterogeneity across sub-groups that makes subjects within group j more or

less likely to fail than subjects in group j (Box-Steffensmeier and Jones 2004). Formally, the

frailty term ν is a random variable with a mean of one and variance of Θ and is drawn from

19For example, if a leader oversaw a winning war effort in year t, War Win takes on a value of 1 in year t, 0.5 int+ 1, 0.333 in t+ 2, etc.

19

the Gamma distribution. Previously, Chiozza and Goemans (2004), Goemans (2008), and Debs

and Goemans (2010) have used Cox estimators with shared frailty to model leader survival.

Amelia II was used to address concerns about missing data (Honaker, King and Blackwell

2007).20 Multiple imputation of missing observations helps avoid the inefficiency and selection

bias associated with listwise deletion and is more accurate than single imputation (King et al.

2001, Honaker, King and Blackwell 2007). For the data analyzed here, Amelia II is preferable

to other imputation programs as it explicitly takes into account the time-series, cross-sectional

nature of the data (Honaker, King and Blackwell 2007, Honaker and King 2010). The multiple

imputation model was specified with one-year lags, one-year leads, and logical or empirical

upper and lower bounds of variables with missing observations.21 Including the lags, leads,

and bounds improves the predictive performance of the imputation model (Honaker, King and

Blackwell 2007, Honaker and King 2010). Further details about the imputation process are

available in the appendix. The multiple imputation model produced five data sets of 7,935

leader-year observations (1,431 leaders from 162 countries).

The estimates presented in Table 1 reflect the mean coefficients and corrected standard

errors yielded by the estimation of identically specified Cox models on each of the five Amelia

II -generated data sets.22 Three points are worth noting before discussing the theoretically

interesting results. First, positive (negative) coefficients indicate higher values of an explanatory

variable are associated with a leader facing a greater (lower) hazard of losing office.23 Second,

the statistically significant Θ in Table 1 indicates non-trivial variation exists in the likelihood of

a leader being removed from power across states and that accounting for shared frailty among

leaders of the same country is methodologically appropriate. Third, the statistically significant

regional indicators imply the presence of regional variation in leader tenure.

Unfortunately, the use of multiplicative interaction terms greatly limits the possible infer-

ences one can draw from Table 1 for two reasons. First, the coefficient associated with any

20I imputed missing values for the constituent terms of the interaction terms and then created the interactionterms based on variables without missing observations. This ensures that the values the interaction terms take onaccurately reflect the values of the constituent terms.

21Variables that have a lower and upper bound by definition were assumed to fall within those bounds whilevariables without a logical range were bounded by the minimum and maximum values observed in the data set.

22Using code from Goemans (2008), the standard errors are computed by taking the square root of T = U+(1+ 1m)B,

where T is the total variance associate with the mean coefficient estimate, U is the within-imputation variance of theestimated coefficient [U = 1

m

∑mi=1 Ui], B is the between-imputation variance [B = 1

m−1

∑mi=1(Q− Q)2], and 1 + 1

m

is a correction factor to account for simulation error in Q (Rubin 1987, Schafer and Olsen 1998).23An interaction with ln(t) signed in the opposite direction of the constituent term indicates a decay in the original

effect over a leader’s tenure.

20

Table 1: Leader Survival, Government Spending, and Regime Type, 1950-2001

β s.e.

Democracy 0.19 0.47Democracy*ln(t) -0.01 0.07Spending Distribution -0.43 0.31Spending Distribution*Democracy 0.78 0.69Spending Distribution2 -0.03 0.13Spending Distribution2*Democracy 0.26 0.62Interstate War -0.72 0.30∗Win War -1.54 0.85†Draw War -0.51 0.61Lose War 1.69 0.48∗∗Civil War 0.81 0.37∗Civil War*ln(t) -0.06 0.06GDP per capita 2.26 0.19∗∗GDP per capita*ln(t) -0.36 0.03∗∗Growth -2.64 0.41∗∗Trade Openness 0.04 0.16∆ Trade Openness -0.13 0.10Population 0.03 0.04Age 0.01 0.00∗∗Times in Office -0.09 0.06Irregular Entry 4.03 0.38∗∗Irregular Entry*ln(t) -0.61 0.06∗∗Africa 7.35 0.57∗∗Africa*ln(t) -1.20 0.08∗∗Americas 5.27 0.52∗∗Americas*ln(t) -0.78 0.07∗∗Asia 6.52 0.51∗∗Asia*ln(t) -1.04 0.07∗∗Middle East 6.20 0.65∗∗Middle East*ln(t) -0.98 0.09∗∗Observations 7935Leaders 1431Failures 1145Log-likelihood -6648.19Wald-test 25.49 ∗∗θ 0.26 6.70∗∗Two-tailed: †: p ≤ 0.1; ∗: p ≤ 0.05; ∗∗ : p ≤ 0.01

variable tells us the impact of an increase in that variable when all of the other constituent

terms are equal to zero (Braumoeller 2004). Second, and related to the first point, the stan-

dard error associated with a coefficient reflects the uncertainty around the estimated effect of a

21

variable when all of the other constituent terms are equal to zero and does not take into con-

sideration the covariance among that variable, the other constituent terms, and the interaction

terms (Brambor, Clark and Golder 2006). This implies the statistical significance of the inter-

action terms and their constituent terms cannot be assessed using Table 1 alone. I therefore

used a set of simulations to calculate how the distribution of government spending influences

the probability of leader survival in democracies and non-democracies. I took 1,000 draws from

a multivariate normal distribution based on the coefficient and variance-covariance matrices of

each model estimated on each of the five imputed data sets using the MSBVAR package in R

(Brandt 2012). I then used the simulated coefficients to calculate the probability of a democratic

leader and a dictator surviving up to time t over a five-year period as Spending Distribution

ranged from 0 to 1 by intervals of 0.1 via the respective cumulative survival functions. Contin-

uous control variables were set to their mean values while nominal and ordinal control variables

were set to their median values for the simulations.

The results of these simulations are reported graphically in Figure 2.24 Panel A presents

the mean predicted probability a democratic leader will remain in office over a five-year period

(x-axis) as a function of Spending Distribution (y-axis). Higher probabilities of survival are

represented by the surface plot’s height on the z-axis and progressively darker colors. Panel A

indicates the probability a democratic leader remains in power is decreasing in the proportion of

government spending dedicated to the military. This result holds throughout a leader’s tenure,

but is most easily seen by examining how the probability a democratic leader will remain in

office for five years changes as the proportion of government spending dedicated to the military

increases. Focusing on the far end of the x-axis, the height of the surface plot decreases along

the z-axis and becomes lighter in color as the spending distribution contains more military

spending (i.e., moves from 0 to 1 on the y-axis). The probability a democratic incumbent will

be in power for five years given no military spending is 0.25 (medium gray), drops to 0.17 if half

of government spending is dedicated to the military (light gray), and reaches a low of 0.09 if

all government spending is allocated to the military (white). Framed differently, a democratic

leader’s probability of serving five years decreases by 32% if she moves from spending no money

on the military to allocating half of the budget to the military and by 64% if she moves to

24The Supplementary Appendix includes a set of graphs of the probability of leader survival in a democracy and anon-democracy over a five-year period at each value of Spending Distribution. These figures include both the meanpredictions and 95% confidence intervals. Figure 2 does not include confidence intervals because their inclusion makesthe graphs overly busy and difficult to interpret.

22

distributing all government spending to the military.

Days in Office

0

1000Spending Distribution

0.0

0.5

1.0

0.0

0.5

1.0

Pro

ba

bili

ty o

f S

urv

iva

l

Panel A: Democratic Leader

Days in Office

0

1000Spending Distribution

0.0

0.5

1.0

0.0

0.5

1.0

Pro

ba

bili

ty o

f S

urv

iva

l

Panel B: Autocratic Leader

Figure 2: The Probability of Leader Survival and the Distribution of Government Spending inDemocracies and Dictatorships, 1950-2001. Darker colors represent higher probabilities of leadersurvival.

A different relationship exists between government spending and leader survival in non-

democracies. Panel B in Figure 2 reports that the probability a dictator remains in power is

increasing in the proportion of government spending dedicated to the military. Again, it is

easiest to see this result by focusing on how the probability an autocrat will remain in power for

five years changes as his government dedicates more resources to the military. Starting at the

far end of the x-axis, the height of the surface plot increases along the z-axis and becomes darker

in color as more of a government’s total spending is dedicated to the military. The probability a

dictator will rule for five years given he spends nothing on the military is 0.29 (white), increases

to 0.38 if half of expenditures are allocated to military spending (light gray), and reaches 0.46

if all government spending is distributed to the military (medium gray). Thus, moving from

allocating no government spending to the military to half of spending to the military increases

the probability a dictator will rule for five years by 27% while the move from no military spending

to all military spending increases his probability of survival by 53%.

23

Figure 2 indicates the effect of government spending on leader survival is conditional on

regime type. As the proportion of spending dedicated to the military increases, the proba-

bility a democratic leader retains office decreases while the probability a dictator remains in

power increases. These results are consistent with the model’s prediction that a democratic

leader’s probability of removal is minimized with a spending distribution that consists of less

military spending and more social spending than the distribution of spending that minimizes

the probability a dictator will lose power.

The results for the control variables generally match expectations. On average, a leader is

less likely to lose office if she is currently involved in an interstate war or wins an interstate

war (at the 0.1 level) and more likely to be removed from power if she oversees a losing war

effort. Obtaining a draw in an interstate war has no significant influence on a leader’s tenure.

Leaders whose states are embroiled in a civil war face a higher hazard of losing office early

in their tenure, but this relationship weakens over time. Higher levels of GDP per capita are

associated with a greater risk of leader removal early in an executive’s tenure, although this

diminishes the longer she has been in office. An incumbent’s hazard of removal is decreasing

in her state’s economic growth. Trade openness, the annual change in trade openness, a state’s

population, and the previous number of times a leader has been in power have no influence on

an incumbent’s tenure. Older leaders and leaders that rose to power in an irregular manner face

a higher hazard of removal, though the latter relationship weakens over time.

I conducted a set of additional analyses of the relationship between government spending

and leader survival. Using the post-estimation simulations from the model reported in Table 1,

I analyzed the differences in the probability of leader survival as Spending Distribution changes

for a democratic leader and a dictator and whether the effects of Spending Distribution on

leader survival are significantly different across regime type. I also estimated models that omit

outliers and with a more parsimonious specification. Consistent with the results presented

here, these analyses imply that increasing military spending and decreasing social spending

increases the probability a dictator remains in power but decreases the probability a democratic

incumbent retains office. These results are reported in the Supplementary Appendix due to

space considerations.

24

4.2 Government Spending

The formal model has several implications regarding empirical patterns of government spend-

ing. Ideal tests of these expectations require systematic data across countries and over time on

leaders’ motivations for implementing policies and relative preferences for military and social

spending. Unfortunately, these data do not exist. However, we can leverage variation in the

spending preferences of winning coalitions and the cost of leader replacement across regime type

and the implications of leaders pursuing their preferred policies to assess the model’s expecta-

tions for government spending.

Equilibrium spending distributions reflect the preferences of a leader’s winning coalition as

long as she is not exclusively driven by the desire to implement her preferred policy. For all

incumbents that are at least marginally concerned with remaining in power, the spending dis-

tributions allocated by democratic leaders will be biased towards lower military spending and

higher social spending than the mix of spending provided by non-democratic leaders. There-

fore, we should observe proportionately higher military spending and lower social spending in

democracies than in dictatorships, on average.

If we consider scholarship on foreign policy substitutability and intra-democratic conflict

variation, the model implies that we should observe greater variance in spending distributions

allocated by dictators than those provided by democratic leaders. Researchers in these tra-

ditions demonstrate that there exists greater variation in policy outcomes as a leader faces

fewer constraints on her behavior and is better able to pursue her personal preferences. Clark

and Nordstrom (2005) argue that idiosyncratic leader preferences explain why increased po-

litical participation in democracies is associated with significant variance in the probability of

democratic conflict initiation. Clark, Nordstrom and Reed (2008) find that leaders of states

with substantial resources pursue cooperative and conflictual foreign policies with greater vari-

ance than those leaders whom are constrained by their states’ lack of resources. Arena and

Palmer (2009) argue that leaders of right governments are less constrained by their constituents

in the decision to use force than are left leaders and therefore should initiate conflicts with

greater variance than left leaders.25 Consistent with this argument, they find greater variance

around the probability democracies with right governments initiate interstate conflicts than is

25As Arena and Palmer (2009) put it, “While leaders who lack constraints preventing them from using force aremore likely to engage in disputes than leaders who are more heavily constrained (as the latter are universally unlikelyto engage in disputes), being free to take risks does not necessarily mean that one desires to do so” (pg. 962).

25

the case with democracies with left governments. In sum, Clark and Nordstrom (2005), Clark,

Nordstrom and Reed (2008), and Arena and Palmer (2009) all find that the variance in policy

outcomes increases as the constraints on a leader’s behavior decrease and she is able to pursue

her personally preferred policies. The formal model indicates that variation in the structural

costs of replacement allows autocratic leaders to enact their personal preferences to a greater

extent than democratic leaders, except for when leaders are motivated exclusively by survival

or policy concerns. This implies we should observe greater variance in non-democratic spending

distributions than in democratic spending distributions.

A quick examination of the means and interquartile ranges of Spending Distribution among

democracies and non-democracies reveals spending patterns that match the model’s expecta-

tions. Democracies have a mean Spending Distribution of 0.24 and an interquartile range that

spans from 0.09 to 0.33 while dictatorships have a mean Spending Distribution of 0.36 and an

interquartile range of 0.09 to 0.49. Framed differently, dictatorships allocate approximately

51% more of their total spending to the military, on average, and have an interquartile range

approximately 63% greater than do democracies. To offer a more rigorous assessment of demo-

cratic and non-democratic spending distributions, I estimated a set of regression models that

account for multiplicative heteroscedasticity. Estimators with multiplicative heteroscedasticity

allow analysts to model the error variance as a function of a vector of explanatory variables

(Harvey 1976, Alvarez and Brehm 1995). The models were estimated on state-year versions of

the five imputed leader-year data sets described above. I converted the data sets to a state-year

unit-of-analysis in order to model temporal dependence in the Spending Distribution series (see

the next paragraph for more details). The state-year data sets have 6,684 observations.

Time-series cross-sectional data present multiple methodological challenges concerning tem-

poral dynamics and unit heterogeneity (Beck 2008). I focus first on issues related to modeling

dynamics. The dependent variable in my analyses of government spending is a state’s Spending

Distribution. Panel unit root tests indicate that Spending Distribution is not integrated; how-

ever, its value at time t is correlated with its value at t−1 at 0.89. Modeling this autocorrelation

is not feasible using the leader-year data sets because the average number of observations per

panel is 6.6. Moving to a state-year unit-of-analysis increases the mean number of observations

per panel to 41.3, allowing me to better model the temporal dependence present in the de-

pendent variable. Diagnostics indicate that statistically significant autocorrelation exists even

after including a one-year lag of the dependent variable in the models. However, after including

26

both a one-year lag and a two-year lag of the dependent variable no significant autocorrelation

remains among the errors.26 The statistical models therefore include SpendingDistributiont−1

and SpendingDistributiont−2 as explanatory variables.

Concerning cross-sectional issues, failure to account for unmodeled unit heterogeneity in the

dependent variable can lead to faulty inferences. Scholars commonly model unit heterogeneity

with fixed-effects or random effects models (among many others, Cameron and Trivedi 2005,

Beck 2008). These estimators are not ideal for testing the model’s implications for two reasons.

First, the model’s expectations concern variation in spending patterns between democracies

and non-democracies. As fixed effects models rely exclusively on within-panel variation (e.g.,

Cameron and Trivedi 2005), they are inappropriate for theoretical questions related to between-

panel variation. Second, random effects models are not particularly well-suited for time-series

cross-sectional analyses because they rely on the assumptions that unit effects are uncorrelated

with the explanatory variables and are independently and identically distributed, both of which

are violated by time series cross-sectional data (e.g., Cameron and Trivedi 2005, Beck 2008). I

attempted to mitigate the threats to inference that follow from unmodeled unit heterogeneity

in two ways. First, all of the models include the dichotomous indicators Africa, Americas, Asia,

and the Middle East. These variables capture regional variation in Spending Distribution while

allowing for inferences about the between-panel effect of regime type on patterns of spending.

Second, the models use robust standard errors clustered by country in order to account for

within-panel correlation among the errors.

The key explanatory variable in the models of government spending is Democracy. In ad-

dition to the explanatory variables discussed above, a state’s mean Spending Distribution is

modeled as a set of factors argued to influence government spending. Involvement in an Inter-

state War or a Civil War has been shown to increase military spending (Goldsmith 2003). As

social spending is a type of luxury good, it is generally increasing in a state’s GDP per capita

(Palmer 1990). Trade Openness is typically associated with expansive social welfare states,

as governments purchase liberal economic policies with large safety nets (Hays, Ehrlich and

Peinhardt 2005). The variance in annual spending distributions is modeled as a function of

Democracy, GDP per capita, and Interstate War. Democracy is included to test the theoreti-

cal expectation that variation in the structural cost of leader removal should result in greater

26More specifically, errors at t − 1 are a statistically significant predictor of the errors at t when onlySpendingDistributiont−1 is included in the models but are a statistically insignificant predictor of the errors att when SpendingDistributiont−1 and SpendingDistributiont−2 are included in the models.

27

variance in the spending patterns of dictatorships than in democracies. Interstate War should

increase the variance in Spending Distribution as leaders are unlikely to uniformly alter how

they allocate resources as a function their state’s involvement in an interstate war. GDP per

capita is included in the variance equation as states with greater resources have been shown

to exhibit greater variance in policy outcomes (e.g., Clark, Nordstrom and Reed 2008).27 The

primary regression results are reported in Table 2.

Table 2: Spending Distribution and Regime Type, 1950-2001

β s.e.

Spending DistributionDemocracy -0.01 0.00∗∗Interstate War 0.05 0.003∗∗Civil War 0.01 0.001∗∗GDP per capita 0.007 0.001∗Trade Openness 0.001 0.001Africa -0.001 0.001Americas -0.01 0.001∗∗Asia -0.002 0.002Middle East 0.03 0.001∗∗Spending Distributiont−1 0.69 0.01∗∗Spending Distributiont−2 0.19 0.01∗∗Constant 0.03 0.001∗∗VarianceDemocracy -1.77 0.23∗∗Interstate War 1.24 0.16∗∗GDP per capita 0.11 0.06Constant -3.43 0.07∗∗Observations 6360Log-likelihood 3079.45

Two-tailed: †: p ≤ 0.1; ∗: p ≤ 0.05; ∗∗ : p ≤ 0.01

Robust standard errors clustered on country.

The estimates presented in Table 2 reflect the mean coefficients and corrected standard

errors yielded by the estimation of identically specified models on each of the five imputed

data sets.28 Unfortunately, we cannot accurately assess the effect of Democracy on Spending

Distribution using Table 2 because the standard error of a variable included in the mean and

27I useGDP per capita to proxy a state’s resources instead of the more common CINC score (Correlates of War 2001)because the numerator of Spending Distribution (annual military expenditures) is part of the CINC index. Using astate’s CINC score to proxy resources therefore could result in a spurious relationship as increases in a state’s militaryspending would necessarily result in increases in both the dependent variable and a state’s CINC score.

28See Footnote 22 for the equation used to correct the standard errors.

28

variance equations of a model with multiplicative heteroscedasticity varies as a function of the

values of that variable (Harvey 1976, Alvarez and Brehm 1995). I therefore conducted a set

of post-estimation simulations from the model’s coefficient and variance-covariance matrices

to test whether the mean and/or variance of Spending Distribution is significantly different in

democracies and non-democracies. Continuous control variables were set to their mean values

while nominal and ordinal control variables were set to their median values for the simulations.

Figure 3 presents these results.

Panel A in Figure 3 reports the mean spending distributions of autocracies (red square) and

democracies (blue diamond) with 95% confidence intervals (solid black lines) yielded by the

simulations. On average, autocracies dedicate 0.35 of their total government spending to the

military while democracies allocate 0.23 of government spending to the military. Framed differ-

ently, the results indicate that autocracies allocate approximately 52% more of total spending

to the military than do democracies. To assess the significance of this finding, Panel C reports

the mean difference in democratic and non-democratic spending distributions (black circle with

the 95% confidence interval again denoted by the solid black line). The difference across regime

type in the average proportion of total government spending dedicated to the military is sta-

tistically significant if the confidence interval does not contain the zero-line. Panel C therefore

indicates that, on average, democracies allocate significantly less of their total spending to the

military than do non-democracies. This result is consistent with the expectation that spending

distributions allocated by democratic leaders will be biased towards lower military spending and

higher social spending than the spending distributions provided by dictators.

The relatively wider confidence intervals associated with autocratic spending distributions

in Panel A of Figure 3 suggest that there is greater variance around non-democratic patterns

of spending than is the case with democratic spending. To more explicitly assess this rela-

tionship, Panel B presents the mean variance of autocratic (red square) and democratic (blue

diamond) spending distributions yielded by the post-estimation simulations. The mean vari-

ance around autocratic spending distributions is 0.01 while the mean variance surrounding

democratic distributions is 0.001. Panel D reports the difference between these two quantities

(black circle). As the 95% confidence interval lies entirely below the zero-line, Panel D indicates

that there is significantly less variance surrounding democratic spending distributions than is

the case with non-democratic distributions. This is consistent with the claim that dictators are

less constrained by the preferences of their winning coalitions and, thus, vary in the spending

29

Pro

port

ion

to M

ilita

ry S

pend

ing

Autocracy Democracy

0.0

0.5

1.0

Panel A: Mean Spending Distributions

Autocracy Democracy

0.00

0.01

0.02

Panel B: Variance in Spending Distributions

Var

ianc

e

−0.

2−

0.1

0.0

Panel C: Difference in Mean

Dem

ocra

cy−

Aut

ocra

cy

−0.

2−

0.1

0.0

Panel D: Difference in Variance

Dem

ocra

cy−

Aut

ocra

cy

Autocracy

Democracy

Democracy−Autocracy

95% Confidence Interval

Figure 3: Spending Distributions of Democracies and Dictatorships, 1950-2001.

distributions they allocate to a greater degree than relatively constrained democratic leaders.

Turning to the control variables, involvement in an interstate war increases the proportion

of spending dedicated to the military and the variance around a state’s annual spending dis-

30

tribution. Civil war participation is associated with proportionately higher military spending.

Contrary to expectations, increases in GDP per capita are associated with proportionately more

military spending but do not influence the variance in states’ spending distributions. I find no

relationship between spending and a state’s trade openness. Compared to European states,

Middle Eastern countries spend proportionately more on the military while American states

spend proportionately less. The spending distributions of African and Asian states do not differ

significantly from those of European countries.

I conducted a set of robustness checks to ensure the findings reported here hold given alterna-

tive specifications. I estimated models on data sets without outliers and models with different

specifications of the variance equation. I also transformed the data to explicitly analyze the

between panel effect of regime type on patterns of government spending. These models all indi-

cate that, on average, democracies spend less on the military and more on social spending than

dictatorships and that there is more variance in patterns of non-democratic spending. These

results are available in the Supplementary Appendix.

5 Conclusion

Arguably the most salient trend in the contemporary study of international relations and

comparative politics has been the increased theoretical and empirical focus on political leaders

and the choices they make. Some of the most important decisions facing a leader concern

choosing which government programs and policies to finance with the scarce economic resources

available to her. I developed a model of the relationship between leader survival and government

spending when both the incumbent and her winning coalition hold preferences over how a

government allocates resources and a leader can be motivated by securing her political survival

and/or the distribution of government spending. The model indicates that whether a leader

will rationally deviate from her constituents’ policy preferences is a function of her motivation

for enacting policy and the structural costs her winning coalition must pay to remove her from

power. Variation across regime type in the preferences of winning coalitions and the structural

costs of leader replacement result in different equilibrium behavior for democratic and non-

democratic leaders and winning coalitions. I found statistical support for a set of the model’s

empirical implications during the period from 1950 to 2001. I close with a brief discussion of

three of the biggest implications of the findings presented here and avenues for future research.

31

First, the article demonstrates the fruitfulness of allowing variation in citizens’ policy pref-

erences and leaders’ motivation for enacting policy. Selectorate theory assumes that all citi-

zens hold the same policy preferences and winning coalitions vary only in their size (Bueno de

Mesquita et al. 2003). These assumptions imply that the citizens whom can determine whether

an incumbent remains in power are perfectly substitutable. In contrast, I argue socio-economic

status influences policy preferences, which implies the policies that best secure a leader’s politi-

cal survival vary as a function of the composition of her winning coalition. If incumbents are at

least partially motivated by their political survival and retain office through policy responsive-

ness, allowing for variation in their constituents’ policy preferences provides substantial leverage

in predicting political behavior.

One natural extension to the argument presented here is to consider the relationship between

government spending and leader survival among types of non-democracies. It is straightforward

to differentiate non-democracies by whether the civilian elite or military have greater influence

in determining whether a dictator remains in power. Weeks (2012) argues that current and

former members of the armed forces place a higher value on having a strong military than the

civilian elite. Given these respective spending preferences and the equilibrium results reported

here, we would expect military dictatorships to allocate proportionately more resources to the

military than civilian-led dictatorships, on average. We also would expect that lower levels of

military spending should be more likely to threaten the survival of military dictatorships than

civilian-led dictatorships. Further analysis of the relationship between government spending and

leader survival among non-democracies must be left to future research.

The idea that equilibrium policy outcomes are a function of a leader’s preferences and her mo-

tivation for implementing policy has implications beyond patterns of government spending. For

example, a growing literature is beginning to investigate how lame ducks differ from democratic

leaders constrained by the prospects of re-election (e.g., Zeigler, Pierskalla and Mazumder 2013).

Electoral accountability therefore represents a dynamic constraint on democratic leaders’ be-

havior (Clark and Nordstrom 2005). The approach taken here suggests that removing electoral

accountability should reduce the extent to which a leader is concerned with survival and allow

her to pursue her preferred policies. This implies that the effect of term limits on policy should

be conditional on a leader’s preferences. Accordingly, we might expect conservative and liberals

to behave differently with respect to patterns of taxation and spending, protectionists and free

traders to behave differently with respect to trade behavior, and isolationists and international-

32

ists to behave differently with respect to honoring interstate agreements when they are no longer

constrained by the desire to be re-elected. More generally, the results presented here suggest

considering variation in preferences and a leader’s motivation for enacting policy is useful for

understanding political behavior within and across leaders, governments, and regime type.

Second, and related to the previous point, the formal model indicates that policy respon-

siveness is a function of an incumbent’s motivation for implementing policies and the structural

cost of replacement. Leaders motivated by retaining office and those leaders whose winning

coalitions can easily remove them from power will be more responsive to their constituents’

preferences. Given variation in the cost of leader replacement across regime type, this implies

that democratic leaders should be more responsive than dictators, a finding consistent with

most people’s intuition. However, this result only holds when a leader cares about retaining

office. A leader unconcerned with her political survival will ignore her constituents and imple-

ment her preferred policies regardless of the cost of removal. This implies that a dictator who

wants to remain in power will be more responsive to her constituents’ preferences than a demo-

cratic incumbent unconcerned with her survival. While this claim might seem unreasonable

at first blush, I would argue that a lame duck democrat has less of an incentive to enact her

constituents’ preferred policies than a dictator who likely will be killed or jailed if his winning

coalition decides to remove him from power (see Debs and Goemans (2010) on the perils of

being an ex-dictator). Thus, the model’s equilibrium results call into question scholarship that

argues policy responsiveness is a defining feature of democracy (e.g., Dahl 1971, Putnam 1993).

Last, despite the growing focus on the relationship between leaders and policy decisions, this

article is the first to empirically analyze the effect of patterns of government spending on leader

survival in democracies and non-democracies. Duration analysis of all leaders from 1950 to 2001

is consistent with the formal model’s prediction that the distribution of government spending

that best secures a democratic leader’s tenure contains less military spending and more social

spending than the spending distribution that best ensures a dictator remains in power.

The relationship between government spending and leader survival identified here has a

number of implications for existing scholarship and future research. For example, one conse-

quence of interstate war is that it leads governments to allocate relatively more resources to

military spending (Goldsmith 2003). My results indicate that a democratic incumbent would

face a relatively higher hazard of losing office than a dictator given this shift in the spending

distribution. This implies we should observe survival-motivated democrats behaving differently

33

than survival-motivated dictators in at least two ways. First, it suggests that dictators should

be willing to increase the proportion of total spending dedicated to the military during a war

to a greater degree than democrats, especially given the findings that interstate war outcomes

have a larger influence on the survival of dictators (Chiozza and Goemans 2004, Debs and

Goemans 2010, Chiozza and Goemans 2011). This implication is consistent with the findings of

Carter and Palmer (Forthcoming). Second, a survival-motivated democrat has a greater incen-

tive than a survival-motivated autocrat to avoid interstate wars that require substantial increases

in military spending. Consequently, the analyses reported above provide micro-foundations for

the observations that democracies are significantly less likely than non-democracies to initi-

ate interstate conflicts against stronger opponents and the “selection effects” explanation of

democratic success in interstate wars (Reiter and Stam 2002).

The relationship between political survival and government spending is complex. This article

demonstrates how patterns of spending and leader survival affect one another when an incumbent

and her constituents have preferences over how a government spends its resources and leaders

are not exclusively motivated by remaining in power. The theoretical and empirical findings

presented here have implications for how we interpret existing scholarship and suggest several

avenues for future research.

References

Alvarez, R Michael and John Brehm. 1995. “American Ambivalence towards Abortion Policy:

Development of a Heteroskedastic Probit Model of Competing Values.” American Journal

of Political Science 39(4):1055–1082.

Arena, Philip and Glenn Palmer. 2009. “Is it Politics or the Economy? Domestic Corre-

lates of Dispute Involvement in Parliamentary Systems.” International Studies Quarterly

53(4):955–975.

Arena, Philip and Nicholas Nicoletti. Forthcoming. “Selectorate Theory, the Democratic Peace,

and Public Goods Provision.” International Theory .

Bachman, Jerald G, John David Blair and David R Segal. 1977. The All-Volunteer Force: A

Study of Ideology in the Military. University of Michigan Press.

34

Bachman, Jerald G, Peter Freedman-Doan, David R Segal and Patrick M O’Malley. 2000. “Dis-

tinctive Military Attitudes among US Enlistees, 1976-1997: Self-Selection versus Socializa-

tion.” Armed Forces & Society 26(4):561–585.

BBC. 2010. “Greece Hit by Nationwide Strike over Austerity Measures.” February 10,

2010:http://news.bbc.co.uk/2/hi/europe/8507551.stm. Accessed February 1, 2014.

BBC. 2011. “Greek PM Papandreou ‘ready to drop’ Bailout Referendum.” November 3,

2011:http://www.bbc.co.uk/news/world--15575198. Accessed February 1, 2014.

Beck, Nathaniel L. 2008. “Time-Series-Cross-Section Methods”. In Oxford Handbook of Political

Methodology, ed. Janet Box-Steffensmeier, Henry Brady and David Collier. NewYork:

Oxford University Press.

Bender, Bruce and John R. Lott. 1996. “Legislator Voting and Shirking: A Critical Review of

the Literature.” Public Choice 87(1/2):67–100.

Boix, Carles. 2003. Democracy and Redistribution. New York: Cambridge University Press.

Box-Steffensmeier, Janet M and Bradford S. Jones. 2004. Event History Modeling. A Guide for

Social Scientists. Cambridge: Cambridge University Press.

Box-Steffensmeier, Janet M. and Christopher J.W. Zorn. 2001. “Duration Models and Propor-

tional Hazards in Political Science.” American Journal of Political Science 45(4):972–988.

Brambor, Thomas, William Clark and Matt Golder. 2006. “Understanding Interaction Models:

Improving Empirical Analyses.” Political Analysis 14(1):63–82.

Brandt, Patrick. 2012. MSBVAR: Markov-Switching, Bayesian, Vector Autoregression Models.

R package version 0.7-2.

Braumoeller, Bear F. 2004. “Hypothesis Testing and Multiplicative Interaction Terms.” Inter-

national Organization 58(4):807–820.

Brecher, Michael and Jonathan Wilkenfeld. 1997. A Study of Crisis. Ann Arbor, MI: University

of Michigan Press.

Brooks, Clem and Jeff Manza. 2007. Why Welfare States Persist: The Importance of Public

Opinion in Democracies. The University of Chicago Press.

Brown, David S. and Wendy Hunter. 1999. “Democracy and Social Spending in Latin America,

1980-1992.” American Political Science Review 93(4):779–790.

Buchanan, James M. 1965. “An economic theory of clubs.” Economica 32(125):1–14.

35

Bueno de Mesquita, Bruce, Alastair Smith, Randolph M. Siverson and James Morrow. 2003.

The Logic of Political Survival. Cambridge: MIT Press.

Bueno de Mesquita, Bruce, James Morrow, Randolph M. Siverson and Alastair Smith. 1999. “An

Institutional Explanation of the Democratic Peace.” American Political Science Review

93(4):791–808.

Burstein, Paul. 2003. “The Impact of Public Opinion on Public Policy: A Review and an

Agenda.” Political Research Quarterly 56(1):29–40.

Cameron, Colin A. and Pravin K. Trivedi. 2005. Microeconometrics: Methods and Applications.

New York: Cambridge University Press.

Canes-Wrone, Brandice, David W. Brady and John F. Cogan. 2002. “Out of Step, Out of Office:

Electoral Accountability and House Members’ Voting.” American Political Science Review

96(1):127–140.

Canes-Wrone, Brandice, Michael C. Herron and Kenneth W. Shotts. 2001. “Leadership and

Pandering: A Theory of Executive Policymaking.” American Journal of Political Science

45(3):532–550.

Carey, John M., Richard G. Niemi and Lynda W. Powell. 1998. “The Effects of Term Limits

on State Legislatures.” Legislative Studies Quarterly 23(2):271–300.

Carter, Jeff and Glenn Palmer. Forthcoming. “Keeping the Schools Open While the Troops are

Away: Regime Type, Interstate War and Government Spending.” International Studies

Quarterly .

Chiozza, Giacomo and H.E. Goemans. 2004. “International Conflict and the Tenure of Leaders:

Is War Still Ex Post Inefficient?” American Journal of Political Science 48(3):604–619.

Chiozza, Giacomo and H.E. Goemans. 2011. Leaders and International Conflict. Cambridge

University Press.

Clark, David H and Timothy Nordstrom. 2005. “Democratic Variants and Democratic Variance:

How Domestic Constraints Shape Interstate Conflict.” Journal of Politics 67(1):250–270.

Clark, David H, Timothy Nordstrom and William Reed. 2008. “Substitution Is in the Variance:

Resources and Foreign Policy Choice.” American Journal of Political Science 52(4):763–

773.

Cook, Fay Lomax and Edith J. Barrett. 1992. Support for the American Welfare State. New

York: Columbia University Press.

36

Correlates of War. 2001. National Material Capababilities (v.3.02).

Dahl, Robert A. 1971. Polyarchy: Participation and Opposition. New Haven: Yale University

Press.

Debs, Alexandre and H.E. Goemans. 2010. “Regime Type, the Fate of Leaders and War.”

American Political Science Review 104(3):430–445.

Erikson, Robert S., Michael B. MacKuen and James A. Stimson. 2002. The Macro Polity. New

York: Cambridge University Press.

Fenno, Richard F. 1978. Homestyle: House Members in Their Districts. Boston: Little, Brown.

Fordham, Benjamin O. and Thomas C. Walker. 2005. “Kantian Liberalism, Regime Type

and Military Resource Allocation: Do Democracies Spend Less?” International Studies

Quarterly 49(1):141–157.

Fox, Justin and Kenneth W. Shotts. 2009. “Delegates or Trustees? A Theory of Political

Accountability.” Journal of Politics 71(4):1225–1237.

Gandhi, Jennifer and Adam Przeworski. 2006. “Cooperation, Cooptation, and Rebellion under

Dictatorship.” Economics and Politics 18(1):1–26.

Gandhi, Jennifer and Adam Przeworski. 2007. “Authoritarian Institutions and the Survival of

Autocrats.” Comparative Political Studies 40(11):1279–1301.

Geddes, Barbara. 2003. Paradigms and Sand Castles: Theory Building and Research Design in

Comparative Politics. University of Michigan Press.

Gilens, Martin. 2009. “Preference Gaps and Inequality in Representation.” PS: Political Science

& Politics 42(02):335–341.

Gilens, Martin. 2012. Affluence and Influence: Economic Inequality and Political Power in

America. Princeton University Press.

Gleditsch, Nils Petter, Peter Wallensteen, Mikael Erikson, Margareta Sollenberg and Haavard

Strand. 2002. “Armed Conflict, 1945-99: A New Dataset.” Journal of Peace Research

39(5):615–637.

Goemans, H.E. 2008. “Which Way Out? The Manner and Consequences of Losing Office.”

Journal of Conflict Resolution 53(6):771–794.

Goemans, H.E., Kristian Skrede Gleditsch and Giacomo Chiozza. 2009. “Introducing Archigos:

A Data Set of Political Leaders, 1875-2003.” Journal of Peace Research 46(2):269–283.

37

Goldsmith, Benjamin E. 2003. “Bearing the Defense Burden, 1886-1989: Why Spend More?”

Journal of Conflict Resolution 47(5):551–573.

Harvey, Andrew C. 1976. “Estimating Regression Models with Multiplicative Heteroscedastic-

ity.” Econometrica: Journal of the Econometric Society 44(3):461–465.

Hays, Jude C., Sean D. Ehrlich and Clint Peinhardt. 2005. “Government Spending and Public

Support for Trade in the OECD: An Empirical Test of the Embedded Liberalism Thesis.”

International Organization 59(2):473–494.

Heston, Alan, Robert Summers and Bettina Aten. 2012. Penn World Table Version 7.1. Cen-

ter for International Comparisons of Production, Income and Prices at the University of

Pennsylvania.

Holsti, Ole R. 1998. “A Widening Gap between the US Military and Civilian Society?: Some

Evidence, 1976-96.” International Security 23(3):5–42.

Holsti, Ole R. 2001. Of Chasms and Convergences: Attitudes and Beliefs of Civilians and

Military Elites at the Start of a New Millennium. In Soldiers and Civilians: The Civil-

Military Gap and American National Security, ed. Peter D. Feaver and Richard H Kohn.

Cambridge, MA: MIT Press pp. 15–100.

Honaker, James and Gary King. 2010. “What To Do About Missing Values in Time-Series

Cross-Section Data.” American Journal of Political Science 54(2):561–581.

Honaker, James, Gary King and Matthew Blackwell. 2007. “Amelia II: A Program for Missing

Data.” http://gking.harvard.edu/amelia/.

Huber, Evelyne, Tom Mustillo and John D. Stephens. 2008. “Politics and Social Spending in

Latin America.” Journal of Politics 70(2):420–436.

Huntington, Samuel P. 1957. The Soldier and The State: The Theory and Politics of Civil-

Military Relations. Harvard University Press.

Huth, Paul K. and Todd L. Allee. 2002. The Democratic peace and Territorial Conflict in the

Twentieth Century. Vol. 82 Cambridge University Press.

International Monetary Fund. 2008. “International Financial Statistics.”.

Jæger, Mads Meier. 2006. “What Makes People Support Public Responsibility for Welfare

Provision: Self-Interest or Political Ideology? A Longitudinal Approach.” Acta Sociologica

49(3):321–338.

38

Kakissis, Joanna. 2012. “George Papandreou: Greece Had To Make Changes.”

National Public Radio pp. http://www.npr.org/2012/05/01/150804420/

george--papandreou--greece--had--to--make--changes.

Kalt, Joseph P. and Mark A. Zupan. 1984. “Capture and Ideology in the Economic Theory of

Politics.” American Economic Review 74(3):279–300.

Keele, Luke. 2010. “Proportionally Difficult: Testing for Nonproportional Hazards in Cox

Models.” Political Analysis 18(2):189–205.

Key, V.O. 1961. Public Opinion and American Democracy. New York: Knopf.

King, Gary, James Honaker, Anne Joseph and Kenneth Scheve. 2001. “Analyzing Incomplete

Political Science Data: An Alternative Algorithm for Multiple Imputation.” American

Political Science Review 95(1):49–69.

Lake, David A. and Matthew A. Baum. 2001. “The Invisible Hand of Democracy: Political

Control and the Provision of Public Services.” Comparative Political Studies 34(6):587–

621.

Maddison, Angus. 2008. “Statistics on World Population, GDP and Per Capita GDP, 1-2008

AD.”. http://www.ggdc.net/maddison/oriindex.htm.

Marshall, Monty and Keith Jaggers. 2005. “Polity IV Dataset Users’ Manual.”

www.cidcm.umd.edu//polity.

Mas-Colell, Andreu, Michael Whinston and Jerry Green. 1995. Microeconomic Theory. New

York: Oxford University Press.

McGillivray, Fiona and Alastair Smith. 2008. Punishing the Prince. A Theory of Interstate

Relations, Political Insitutions, and Leader Change. Princeton University Press.

Mulligan, Casey B., Ricard Gil and Xavier Sala-i-Martin. 2004. “Do Democracies Have Different

Public Policies than Nondemocracies?” Journal of Economic Perspectives 18(1):51–74.

Nordlinger, Eric A. 1977. Soldiers in Politics: Military Coups and Governments. Englewood

Cliffs, NJ: Prentice Hall.

Olson, Mancur. 1965. The Logic of Collective Action: Public Goods and The Theory of Groups.

Cambridge, MA: Harvard University Press.

Osborne, Martin J. 2004. An Introduction to Game Theory. Oxford University Press.

39

Page, Benjamin I, Larry M Bartels and Jason Seawright. 2013. “Democracy and the Policy

Preferences of Wealthy Americans.” Perspectives on Politics 11(1):51–73.

Palmer, Glenn. 1990. “Alliance Politics and Issue-Areas: Determinants of Defense Spending.”

American Journal of Political Science 34(1):190–211.

Persson, Torsten and Lars Svensson. 1989. “Why a Stubborn Conservative Would Run a Deficit:

Policy with Time-Consistent Preferences.” Quarterly Journal of Economics 104(2):325–345.

Przeworski, Adam, Michael E. Alvarez, Jose Antonio Cheibub and Fernando Limongi. 2000.

Democracy and Development: Political Institutions and Well-being in the World, 1950-

1990. New York: Cambridge University Press.

Putnam, Robert. 1993. Making Democracy Work: Civic Traditions in Modern Italy. Princeton,

NJ: Princeton University Press.

Reiter, Dan and Allan C. Stam. 2002. Democracies at War. Princeton, NJ: Princeton University

Press.

Riker, William H. 1962. The Theory of Political Coalitions. Vol. 578 Yale University Press New

Haven.

Rothenberg, Lawrence S. and Mitchell S. Sanders. 2000. “Severing the Electoral Connection:

Shirking in the Contemporary Congress.” American Journal of Political Science 44(2):316–

325.

Rubin, Donald B. 1987. Multiple Imputation for Nonresponse in Surveys. New York: Wiley.

Samuelson, Paul A. 1954. “The Pure Theory of Public Expenditure.” Review of Economics and

Statistics 36(4):387–389.

Schafer, Joseph L. and Maren K. Olsen. 1998. “Multiple Imputation for Multivariate

Missing-Data Problems: A Data Analysts Perspective.” Multivariate Behavioral Research

33(4):545–571.

Singer, J. David, Stuart A. Bremer and John Stuckey. 1972. “Capability Distribution, Uncer-

tainty, and Major Power War, 1820-1965”. In Peace, War, and Numbers, ed. Bruce Russett.

Beverly Hills, CA: Sage pp. 19–48.

Smith, Alastair and Bruce Bueno de Mesquita. 2011. “Contingent Prize Allocation and Pivotal

Voting.” British Journal of Political Science 42(2):371–392.

40

Smith, Helena and Tom Klington. 2011. “Papandreou Out as Greek Leaders Agree on Unity Gov-

ernment Deal.” The Guardian November 6, 2011:http://www.theguardian.com/world/

2011/nov/06/papandreou--greek--leaders--unity--deal.

Soroka, Stuart and Chrisopher Wlezien. 2005. “Opinion-Policy Dynamics: Public Preferences

and Public Expenditure in the United Kingdom.” British Journal of Political Science

35(4):665–689.

Sprout, Harold and Margaret Sprout. 1968. “The Dilemma of Rising Demands and Insufficient

Resources.” World Politics 20(4):660–693.

Strøm, Kaare. 1990. “A Behavioral Theory of Competitive Political Parties.” American Journal

of Political Science 34(2):565–598.

Svolik, Milan W. 2009. “Power Sharing and Leadership Dynamics in Authoritarian Regimes.”

American Journal of Political Science 53(2):477–494.

Szayna, Thomas S., Kevin F. McCarthy, Jerry M. Sollinger, Linda J. Demaine, Jefferson P.

Marquis and Brett Steele. 2007. The Civil-Military Gap in the United States: Does it

Exist, Why, and Does it Matter? Vol. 379 Rand Corporation.

Telegraph. 2011. “George Papandreou Resigns as Greece’s Prime Minister.” The Telegraph

November 11, 2011(1):http://www.telegraph.co.uk/finance/financialcrisis/

8879647/George--Papandreou--resigns--as--Greeces--prime--minister.html.

Accessed February 1, 2014.

Weeks, Jessica L. 2012. “Strongmen and Straw Men: Authoritarian Regimes and the Initiation

of International Conflict.” American Political Science Review 106(2):326–347.

Wittman, Donald. 1977. “Candidates with Policy Preferences: A Dynamic Model.” Journal of

Economic Theory 14(1):180–189.

Wlezien, Christopher. 1996. “Dynamics of representation: The case of US spending on defence.”

British Journal of Political Science 26(1):81–104.

Woon, Jonathan. 2012. “Democratic Accountability and Retrospective Voting: A Laboratory

Experiment.” American Journal of Political Science 56(4):913–930.

Zeigler, Sean, Jan H Pierskalla and Sandeep Mazumder. 2013. “War and the Reelection Motive

Examining the Effect of Term Limits.” Journal of Conflict Resolution .

41

Appendix

This appendix presents the formal model. The robustness checks discussed in the main text

are available in the Supplementary Appendix.

A Model

The model is a one-shot game between a state’s incumbent leader (Li) and her winning

coalition (Wi). Li’s strategy set is xi ∈ [0, 1], where xi represents the proportion of government

spending dedicated to military spending. Wi’s strategy set is di ∈ {r, k}, where ri replaces Li

and ki keeps Li. The indicator variable pi takes on a value of 1 if Wi plays ri and 0 if Wi plays

ki. Li prefers to remain an incumbent in office (Ii) over being ousted (Oi); Ii > Oi.

Li and Wi hold single-peaked, symmetric preference profiles over xi. The valuations of xi

for Li and Wi are defined as the strictly concave functions z(xi) and v(xi), respectively. I define

xL as Li’s ideal xi ⇒ argmaxx

z(xi) = xL ∀ Li. The ideal policy of the member of Wi holding

the median preference over xi is defined as xw ⇒ argmaxx

v(xi) = xw ∀ Wi.

αi(mi, si) ∈ [0, 1], where mi ∈ [0, 1] represents Li’s motivation for implementing policy and

si ∈ (0, 1] represents the structural cost of leader replacement to Wi. αi(mi, si) is defined as

follows:

α(mi, si) =

0 if mi = 0

mi + (1−mi)(si) if 0 < mi < 1

1 if mi = 1

(A-1)

Replacing Li brings a benefit (bi) to Wi, where bi is an exogenous preference shock to the

model unknown at the beginning of the game that follows the uniform distribution U [l, h], where

l ≤ 0 and h > 0 and arbitrarily large. The cumulative distribution function of bi is written as

F. Removing Li imposes cost ci(v(xi), si)) > 0 upon Wi, where ci is increasing in v(xi) and si:

∂ci(v(xi), si))

∂v(xi)≥ 0 ∀ si (A-2)

∂ci(v(xi), si))

∂si≥ 0 ∀ v(xi) (A-3)

The timing of the game is as follows:

42

1. Incumbent leader Li distributes xi.

2. Uncertainty over bi is lifted.

3. Winning coalition Wi decides to replace (ri) or keep (ki) Li.

As noted in the main text, utility functions for Li and Wi are written as follows:

ULi = αi(mi, si)z(xi) + (1− αi(mi, si))[p(bi, ci(v(xi), si))Oi + (1− p(bi, ci(v(xi), si)))Ii] (A-4)

UWi =

v(xi) if ki

v(xi) + (bi − ci(v(xi), si)) if ri

(A-5)

B Solution

The model is a dynamic game of perfect information with exogenous uncertainty and there-

fore is solved for pure strategy subgame perfect equilibria using backwards induction (Osborne

2004, pg. 226). From Mas-Colell, Whinston and Green (1995), in dynamic games of complete

and perfect information there exists a unique equilibrium in pure strategies for any distribution

of the model’s parameters (pg. 276). Upon receiving xi and observing bi, Wi’s decision rule is

as follows:

Wi plays ki iff bi < ci(v(xi), si) and ri otherwise. (A-6)

As F is the cumulative distribution function of bi, F (ci(v(xi), si)) identifies the probability

with which Wi plays ri and Li loses office.

Li then distributes xi. The first partial derivative of Equation A-4 with respect to αi(mi, si)

identifies how UL changes with αi(mi, si).

∂UL

∂αi(mi, si)= z(xi)− [p(bi, ci(v(xi), si))Oi + (1− p(bi, ci(v(xi), si)))Ii] (A-7)

Note that the first term of Equation A-7 is positive and the second term is negative. The

xi that maximizes Li’s utility then is a function of the values of αi(mi, si), xL, and xw. As

argmaxx

z(xi) = xL ∀ Li and argmaxx

v(xi) = xw ∀ Wi, x∗i ,

43

x∗i = αi(mi, si)(xL) + (1− αi(mi, si))(xW ) (A-8)

C Propositions

Define a state’s type of government as gi ∈ {D,A}. The following inequalities define the

conditions placed on winning coalitions’ preferred spending distributions and structural costs of

leader replacement:

xDW < xA

W (A-9)

sDi < sAi ∀ v(xi) (A-10)

Proposition 1 argminx

F (ci(v(xi), si)) | gi = D < argminx

F (ci(v(xi), si)) | gi = A.

Proof. Li minimizes F (ci(v(xi), si)) with the xi that maximizes ci(v(xi), si). From Equation

A-2, ci(v(xi), si) is increasing in v(xi) ∀ si. As argmaxx

v(xi) = xw, then argmaxx

ci(v(xi), si) =

xw ∀ si ⇒ argminx

F (ci(v(xi), si)) ∀ si. Given Equation A-9, argmaxx

v(xi)| gi = D < argmaxx

v(xi)| gi = A ⇒ argminx

F (ci(v(xi), si)) | gi = D < argminx

F (ci(v(xi), si)) | gi = A .

Corollary 1 If mi = 0, then (x∗i | gi = D) < (x∗

i | gi = A).

Proof. From Equation A-8, x∗i = αi(mi, si)(xL) + (1 − αi(mi, si))(xW ) ⇒ x∗

i = xw ∀ Li if

αi(mi, si) = 0. Equation A-1 defines αi(mi, si) = 0 if mi = 0. The proof of Proposition 1

therefore implies that if mi = 0, then (x∗i | gi = D) < (x∗

i | gi = A).

Proposition 2 (|x∗i − xL| |gi = A) < (|x∗

i − xL| |gi = D) ∀ 0 < mi < 1

Proof. From Equation A-1, αi(mi, si) = mi+(1−mi)si if 0 < mi < 1 ⇒ ∂αi(mi,si)∂si

≥ 0 ∀ 0 <

mi < 1. From Equation A-10, sDi < sAi ⇒ αi(mi, sDi ) < αi(mi, s

Di ) ∀ 0 < mi < 1. Given

Equation A-8, it follows that (|x∗i − xL| |gi = A) < (|x∗

i − xL| |gi = D) ∀ 0 < mi < 1.

44