Defence and Peace Economics, 2008

Vol. 19(2), April, pp. 107–125

PREDATORY BEHAVIOR OF GOVERNMENTS: THE

CASE OF MASS KILLING
SANG HOO BAE* AND ATTIAT F. OTT†

Department of Economics, Clark University, 950 Main Street, Worcester, MA 01610, USA

(Received 10 December 2006; in final form 14 June 2007)

Taylor and Francis
GDPE_A_251550.sgm

In this paper we seek to answer the question: why do governments engage in mass killing? Tullock (1974) gives gain
Defence
10.1080/10242690701516846
1024-2694
Original
0Taylor
00
Assistant
2007
000
[email protected]
00 &and Article
Professor
Francis
(print)/1476-8267
Peace Economics
Sang HooBae
(online)

or avoidance of loss as the motive. We construct a three-stage theoretic framework to explain the choice of a ruler of
a country. The conditions that must be met for a mass killing regime to win over alternative regimes are derived. Using
the COW project data over the period 1816–1997, we estimate two models: negative binomial regression of number
of battle-related deaths and a probit model for the choice of mass killing. The paper concludes with suggestions for
data collections and further research.

Keywords: Mass killing; Vertical differentiation

JEL CODE: C7

In a predatory regime, nothing is done for public reasons. Indeed, the men in charge do not recognize that
‘public purposes’ exist. (James K. Galbraith, The Predator State, Mother Jones, May/June 2006, p.31)

INTRODUCTION

The 20th century is replete with incidents of civilian killings by the state. Conflict-related
deaths in that century were put at 109.7 million or 4.35% of the world population. This
contrasts with 19.4 million deaths in the 19th century, accounting for 1.65% of the world
population. For armed conflicts, although they declined in the late years of the 20th century,
the geographical pattern of such conflicts seems to have shifted towards the poorest countries.
Whereas in the period 1946–1989, 30% of all conflicts took place in low income developing
countries, this percentage rose to 50% in the 1990–2003 period. Africa seems to have born the
brunt of violent conflicts, accounting for 38% of the world conflicts in that period (World
Bank 2004: 153).
Violent conflicts fall into three categories: interstate (military conflicts between states), extr-
astate (between states and non-state players) and intrastate (between fractions within a state).
Although all violent conflicts have similar outcomes, death and destruction, the motivation for

*Corresponding author. Email: [email protected]

†Email: [email protected]

ISSN 1024-2694 print: ISSN 1476-8267 online © 2008 Taylor & Francis
DOI: 10.1080/10242690701516846
108 S.H. BAE AND A.F. OTT

initiating a violent conflict may or may not be the same in all types of conflicts. In some
conflicts such as civil wars, war provides the state or the instigating group with the opportunity
to loot the opposing faction’s resources (Azam and Hoeffler, 2002), to undermine political
support for the opposition parties and to deter future challenges. In the case of interstate war,
its purpose is to liberate the country from colonial power or alien culture.1
In this paper we deal with intrastate conflict resulting in a civil war. Examples include the
Angolan civil war, which lasted 27 years with more than one million deaths; repeated wars in
Sri Lanka, which caused over 500,000 deaths; and the Rwandan civil war, which resulted in
over one million deaths, to name a few (Lacina and Gleditsch, 2005). In a civil war, the
government or the ruler participates either actively in the violence or passively through
complicity with the instigating group.
The study of conflicts and their resolution have been for long a subject of inquiry by both
political scientists and economists. The conflict literature, both theoretical and empirical,
addresses issues ranging from an analysis of onset of war, negotiations of settlement, war and
alliances, duration of conflicts and wars as well as the economic consequences of wars. A
compilation of more than 100 articles dealing with the economics of conflicts are embodied
in three volumes edited by Sandler and Hartley (2003). In the first volume, seven out of
36 papers deal with civil wars. Looting is given as a motivation of civil war (Azam 2002).
Factors that motivate the onset of civil war such as ethnic diversity, poverty and natural
resources are empirically evaluated by Sambanis (2004). The influences of civil wars on
steady state income at home and in neighboring countries were empirically estimated by
Murdoch and Sandler (2002).
With conflicts taking on a ‘genocide-type’ dimension, mass killing of civilians by the state
or by a dominant ethnic or a religious group leads to an inquiry into those factors that would
explain the onset of and duration of these episodes (Easterly et al., 2006; Reynal-Querol, 2002;
Montalvo and Reynal-Querol, 2005; Harff, 2003; Fearon and Laitin, 2003; Sambanis, 2004;
Collier et al., 2006; Eck and Hultman, 2007).
The argument advanced in this paper is that killing is motivated by expectation of gain. This
view has been expressed earlier by Gordon Tullock. In The Social Dilemma: The Economics
of War and Revolution, Tullock (1974) raises the question: ‘when can war be profitable?’ We
answer this question by characterizing the decision process of a ruler with heterogeneous
types. A simple theoretic framework with three stages is constructed to explain the optimal
behavior of a ruler of a country in which there are two distinct groups, A and B. We investigate
the options open to the ruler of the country, assumed to belong to group A, as to whether to
engage in mass killing (attack group B), to form a coalition government with group B, or to
do nothing. In such a framework, we rather focus on the ex ante optimal decision of a ruler
based on benefits and costs under different regimes departing from the previous literature
analyzing the ex post relationship between the degree of mass killing and economic, ethnic,
and institutional factors.

THE THEORETICAL MODEL

We construct a simple theoretic framework to explain the optimal behavior of a ruler of a coun-
try in which there exist two distinct groups [A, B]. Each group has a leader supported by his
own group. Without loss of generality, the leader of group A is assumed to be the ruler of the
nation (e.g. President). Let Li be the leader of group i (i=A, B) and vi the probability distribution

1
For a list of mass killing episodes see Easterly et al. (2006, Appendix 2). For a listing of civil wars 1816–2002
see Gleditsch (2004, Table A-2).
 MASS KILLINGS 109

of Li to be the ruler. Another interpretation of vi is the ruler’s political power, which is similar
to the indicator of the theoretical democracy-totalitarian continuum in Rummel (1995). Similar
to Rummel’s analysis, the ruler’s political power being close to zero means more democratic
decision-making in terms of debate, toleration, and negotiation of differences. On the other
hand, the ruler has absolute power as vi approaches one, which means the ruler has absolute
power over all aspects of society. The distribution of the ruler’s power is given by the cumu-
lative distribution function F(v) with continuous density F ′(v) ≥ 0. For simplicity, we assume
that the distribution of ruler’s power in group A is uniformly distributed over the unit interval
as vi ∈ U[0, 1]. For example if vA = 0 it means that LA, the current ruler and leader of group A,
has probability of 0 of remaining in office. On the other hand, if vA = 1 he has probability of 1
of remaining in office. The objective of LA is to maximize his expected utility, which depends
on his chance to remain in office and the wealth level of his own group. Three options are open
to him: attack group B, form a coalition government with the leader of group B [LB], or do
nothing.
When the ruler of group A chooses to attack group B his expected utility is determined by
three components: his political power, the expected wealth level of his own group and the cost
of attack. θ AA denotes the expected wealth level of group A, which is determined by the sum
of the group’s own production activities and the appropriation of group B’s wealth. To engage
in hostile activities, the ruler of group A needs to allocate his group between productive activ-
ities and predatory activities. This model captures the essential trade-off that the ruler faces in
that an increase in the number of his people allocated to predatory activities, rather than to
productive activities, decreases the level of output of his own group but increases the proba-
bility of a successful attack and hence the appropriation of group B’s wealth.
Violent conflict, henceforth referred to as mass killing, is costly. In this paper we distinguish
between two different types of cost: a fixed explicit cost CAA , which is assumed to be constant
across all political power (vi) of the leader of group A. This constant fixed cost measures the
additional military expenditure associated with mass killing. As a result, CAA is assumed to be
the same across all different political powers of the ruler, or at least independently distributed
with the political power of the ruler.
The second type of cost associated with attack is the opportunity cost of group A (the
attacker) in diverting their members from economic production to war activities. It is the
forgone marginal product of one unit of labor diverted from economic production. This oppor-
tunity cost, therefore, determines the expected wealth level, θ AA . The expected utility of the
ruler of group A under a mass killing regime, VLAA , can be written as:

VLAA = θ AA ⋅v A − CAA (1)

Equation (1) is a trivial form of vertical-differentiation model. Each type of ruler, vA, chooses
his optimal choice of regime based on the expected wealth and the cost indexed by θ AA and CAA ,
respectively. All types of rulers prefer higher expected wealth for a given cost. However, a
ruler with high vA is more willing to pay to obtain a given expected wealth level.
Under a coalition government regime, the expected utility of the ruler will depend on three
components: his political power, the expected wealth level under the coalition regime and the
cost of forming a coalition government. θ AC denotes the expected wealth level of group A
obtained with full allocation of their members to economic production. An additional
assumption is made here that group A’s expected wealth level under the coalition regime is
proportional to the ruler’s political power, vA.
110 S.H. BAE AND A.F. OTT

Forming coalition government is not without cost. It entails two types of costs: a fixed
explicit cost of CAC , which is assumed to be constant across all political powers of the leader
of group A. This may be associated with the additional public spending in order to regain
support from the ruler’s own group for not initiating mass killing of the other group. Again,
CAC is assumed the same across all different political powers of the ruler or at least indepen-
dently distributed from the political power of the ruler. The second type of cost associated
with forming a coalition government is the reduction of the ruler’s political power because of
power sharing with group B leader. The loss of political power is assumed to be proportional
to the ruler of group A’s political power, which can be written as (1−τ)vA where τ measures
the degree of sharing the political power with group B. Thus, the ruler’s expected utility under
the coalition government regime, VLCA , can be written as:

VLCA = θ AC (1 − τ )v A − CAC (2)

Now we turn to the ruler’s [LA] choice over different regimes where his payoff from no activity
is normalized to zero. For a given set of {θ AA , (1− τ )θ AC , CAA , CAC }, the expected net utility for
the ruler [LA] with his political power vA is:

ìVLA = q AA×n A - CAA if he decides to attack group B.

ï CA
VL A = íVL A = q A × (1 - t )n A - CA if he decides to form coalition with group B.
C C

ï0
î if he decides to do nothing.

When the ruler makes his decision over different regimes, he chooses the one that yields the
highest expected net utility. Assume for the moment that θ AA > (1 − τ )θ AC , that is the wealth of
group A after attack (first option) exceeds the wealth acquired by group A under a coalition
with group B. Under this assumption, the ruler’s optimal choice will be determined by the net
gain (or loss) associated with the two options. In order to have a meaningful discussion of the
optimal choice over different regimes, we restrict our analysis to the parameter regions in
which the coalition constraint is binding, that is,2

θ AA (1 − τ )θ AC
< (3)
CAA CAC

which means that the expected wealth per dollar spending is higher for the coalition regime.
When the coalition constraint (3) is binding, we have the following lemma.

Lemma 1
When the expected wealth under the mass killing regime is higher than that under the coalition
regime (θ A > (1 − τ )θ A ), there exists some type of ruler who would choose forming a coalition
A C

government as his optimal strategy if and only if the coalition constraint given by equation (3)
is binding.

2
This is trivial assumption in the vertical differentiation model in industrial organization where consumers have
unanimous ranking of vertically differentiated products, which means everyone prefers the high quality product.
Equation (3), therefore, rules out the case where a choice gives inferior outcome at a higher cost than the high value,
which induces no demand for the inferior outcome.
 MASS KILLINGS 111

Proof
θ AA (1 − τ )θ AC
If the coalition constraint (3) is not binding, we have ≥ , then we have
CAA CAC

(θ AAν A − CAA ) − (θ AC (1 − τ )ν A − CAC ) =

 θ Aν   (1 − τ )θ ACν A  C  (1 − τ )θ Aν A
C 
CAA  A AA − 1 − CAC  C − 1  ≥ ( C A
A − C A 
) C − 1 ≥ 0
 CA   CA   CA 

if θ AC (1− τ )ν A ≥ CAC , which proves VLA ≥ VLA for any V ∈ [0. 1]. As a result, no ruler chooses
A C

the coalition regime as his optimal choice since it yields lower expected wealth to the ruler at
higher costs than under the mass killing regime. This completes the proof.
When the condition in Lemma 1 is satisfied, we have the more meaningful case where the
coalition government regime is not ‘dominated.’ In this set-up, the ruler whose political power
CAA − CAC
exceeds a ‘mass killing threshold’, νˆ , where νˆ ≡ , chooses to attack group B. ν̂
− (1 − τ )θ AC θ AA
denotes the type of the ruler whose net utility makes him indifferent between the mass killing
and the coalition regimes:

θ AA vˆ − CAA = θ AC (1 − τ ) vˆ − CAC ( 4)

Rulers with political power lower than ν̂ but exceeding ‘coalition threshold’, ν̂ˆ , where
CAC
νˆˆ ≡ , choose to form a coalition government. ν̂ˆ denotes the type of the ruler whose net
(1 − τ )θ AC
utility makes him indifferent between forming a coalition government and doing nothing:

θ AC (1 − τ ) vˆˆ − CAC = 0 (5)

We denote the likelihood of mass killing as a probability of the ruler’s type exceeds the mass
killing threshold, ν̂ , as P(vA > ν̂ ) = 1 − F(ν̂ ) = 1 − ν̂ . Similarly, the probability of forming a
coalition government is P(νˆˆ < ν < νˆ ) = F(νˆ ) − F(νˆˆ ) = νˆ − νˆˆ . A
For the ruler of group A, LA, the optimal choice among different regimes is (see Figure 1):

CAA − CAC
≤ν Attack group B,
θ AA − (1 − τ )θ AC
CAC CAA − CAC
≤ ν < Form coalition with group B,
(1 − τ )θ AC θ AA − (1 − τ )θ AC
C
CA
ν< Do nothing.
(1 − τ )θ AC

There are three stages to this framework. In the first stage the political power of each leader vi
FIGURE 1 Leader A’s choice. vA: ruler’s political power; : mass killing threshold; : coalition threshold; : the group A ruler’s expected utility under a mass killing regime; : the ruler’s expected utility under a coalition government regime.

will be realized. Leader A’s choice among attack, coalition or do nothing will take place in the
second stage. If leader A decides to attack group B, he optimizes his decision over the alloca-
tion of his group members between economic production activities and military activities in
the third stage. We proceed with backward induction.
112 S.H. BAE AND A.F. OTT

A
A C
VLA
VLA , VLA

C'
VLA

0 vöö vö 1 vA

Do Coalition Mass Killing

nothing
ˆ
FIGURE 1 Leader A’s choice. vA: ruler’s political power; ν̂ : mass killing threshold; νˆ : coalition threshold; VLAA : the
group A ruler’s expected utility under a mass killing regime; VLCA : the ruler’s expected utility under a coalition
government regime.

Stage Three with Mass Killing Regime

To analyze the optimal behavior of the ruler in this model, we begin by considering the third-
stage choice of the ruler under the mass killing regime. Again we assume that the country
consists of two distinct groups A and B with population size NA and NB, respectively. The
overall size of the population is N = NA + NB. When leader A chooses to attack the other group
in the third stage, he allocates his people among two types of activities: economic production
and military activities. To be more precise, he can channel his people into productive labor,
which is denoted by EA or into soldiering, which is denoted by FA. Moreover, the ruler fully
utilizes his population so that:

N A = E A + FA (6)

For group B, FB replaces FA – where FB is the level of resources devoted to ward off A’s attack.
Therefore, group B’s economic production is constrained by the loss of its members to the war
effort so that group B’s economic effort is:

EB = N B − FB (7)

The reduction in the economic product of group B in equation (7) associated with mass killing,
means that group A will only acquire a fraction of the total wealth of group B (EB < NB). Let
each group’s production level, HA and HB be given by:

H A = βE A and H B = βEB (8)

 MASS KILLINGS 113

where the production level of each group depends on the number of members devoted to
production and a parameter, β, denoting production technology, which is assumed to be the
same for both groups.3
The rewards of mass killing (A attacking B) is measured by θ AA which has been defined
earlier as the value of its own wealth plus the expected wealth resulting from A’s attack on B
and the appropriation of group B’s wealth. θ AA then consists of two wealth components, group
A’s wealth and the addition to group A’s wealth acquired from group B. Since the acquisition
of B’s wealth is uncertain, depending on the probability of success, θ AA may be written as:

θ AA = PAA H B + H A ( 9)

where PAA is defined as a contest success function as specified in Hirshleifer (1988, 1995). The
contest success function (CSF) summarizes the technology of conflict. Pi, each group’s CSF,
is a function of the difference between the two groups’ resource commitments. Using the
F 
Hirshleifer CSF (ratio form), we have PA PB =  A  . Since PA + PB = 1, we have:
 FB 
α A FA
PAA = (10)
α A FA + α B FB

where αA and αB denote the efficiency of conflict effort of the two groups (0 < αi < 1).4
We now consider the optimal decision of leader LA in the third stage when he decides to
initiate the attack. Leader LA maximizes his expected utility by choosing how many people to
allocate to attack and how many to economic production, subject to the constraint NA = EA +
FA:

[ ]
Max VLAA = θ AA v A − CAA = PAA β ( N B − FB ) + β ( N A − FA ) v A − CAA
FA
(11)

where PAA is given by equation (10).

After solving the utility maximization problem of LA, we have either an interior maximum
which satisfies the following condition:

dVLAA é dP A ù
= ê A b ( N B - FB ) - b ú v A = 0 (12a )
dFA ë dFA û

or a corner solution which satisfies the following condition:

dVLAA é dP A ù
= ê A b ( N B - FB ) - b ú v A £ 0 (12 b)
dFA ë dFA û
The first term in the parenthesis in equations (12a) and (12b) represents the marginal benefit
under the mass killing regime. The second term in the parenthesis shows the marginal cost
associated with mass killing, which is measured by the reduction in economic output due to

3
Further, we make the assumption that HA and HB are fixed. This is clearly a simplifying assumption, given that
war affects the size of the economy (Bellany 1999), but is needed for the mathematical derivation.
4
Hirshleifer’s formula does not contain the efficiency parameter αi, although he has suggested that it should be
included.
114 S.H. BAE AND A.F. OTT

an additional increase in the war activities. Therefore, whenever the marginal benefit equals
the marginal cost, the ruler chooses the positive value for FA as in equation (12a). The ruler
chooses FA* = 0 as in equation (12b) if the marginal cost equals or exceeds the marginal benefit.
From equations (12a) and (12b), the optimal level of FA* is as follows:

[
ìï 1 -a F + a a ( N - F ) F > 0
FA* = í a A B B A B B B B ] for 0 £ FB < F
(13)
ïî 0 for FB ³ F

 αA 
where F̃ =  NB .
 αA + αB 
F̃ is the minimum level of group B’s army needed to deter leader A from attack. At the
same time, it could be viewed as the maximum scale of mass killing which is still beneficial
to leader A (F̃ < HB).
Substituting the optimal level of FA* in equation (13) into VLAA in equation (9), we obtain the
expected wealth of group A under the mass killing regime as follows:

ì b
ï
q AA = í a A A [
a ( N - FB ) + a B FB - 2 a Aa B FB EB ] for 0 £ FB < F
(14)
ïî bN A for F³ FB

The equilibrium expected utility level of leader A is obtained by substituting the equilibrium
level of θ AA in equation (14) into equation (11):

ì b
ï
VLAA = í a A
[a A(N ]
- FB ) + a B FB - 2 a Aa B FB EB n A - CAA for 0 £ FB < F
(15)
ïî bN An A for F³ FB

From equation (14), when FB > F̃ the attack will cease, mass killing is deterred.

Stage Three with Coalition Government Regime

When leader A chooses to form a coalition government rather than attack group B, his
expected utility is given by VLCA = θ AC (1 − τ )ν A − CAC . Under the coalition government regime,
the leader of group A allocates his entire group members to economic production, so that A’s
wealth under the coalition regime, θ AC , equals βNA. Although all members of group A are
engaged in economic production (no attack forces), there are two types of costs associated
with forming a coalition government: a reduction in the ruler’s political power (1−τ)vA, and a
fixed cost CAC representing additional spending to regain support from his own group (assum-
ing some opposition to the coalition). The payoff of leader A under the coalition regime is
given by:

VLCA = βE A (1 − τ )ν A − CAC = βN A (1 − τ )ν A − CAC (16)

Stage Two: A Choice Between Attack and Coalition

We consider next the second stage choice of leader A. At the second stage, leader A chooses
between attack (mass killing) and forming a coalition government on the basis of the expected
wealth level associated with the two options. Recall now that leader A engages his group in
 MASS KILLINGS 115

mass killing if the expected utility under the mass killing regime exceeds that under the coalition
government regime (i.e. VLAA > VLCA). To verify the optimal choice of leader A, we need to ensure
that the assumption, θ AA > (1 − τ )θ AC , actually holds. For a given level of political power of leader
A, we show that a comparison of two expected wealth levels under two regimes is as follows.
For 0 ≤ FB < F̃ we have:
b
q AA - (1 - t )q AC =
aA B B
[
a F + a A (tN A + N B - FB ) - 2 a Aa B ( N B - FB ) FB ]
b é (17a )
=
a A ëê
(
a B FB - a A ( N B - FB ) + a AtN A ùú ³ 0
2
û )
For FB ≥ F̃ we have:

θ AA − (1 − τ )θ AC = βN A − β (1 − τ ) N A = βτN A ≥ 0. (17b)

Stage One: Realization of the Ruler’s Political Power

We now turn to the optimal choice of LA according to his political power after verifying
θ AA > (1 − τ )θ AC with equations (17a) and (17b). Substituting from equations (4) and (5) into ν̂
and ν̂ˆ , we calculate the equilibrium configuration of regimes as follows:

ì CAA - CAC
ï b if 0£ FB < F˜
nˆ =
CAA - CAC ï
= í aA
a B FB - a A ( N B - FB ) ( ) 2
+ btN A
(18)
q AA - (1 - t )q AC ï CAA - CAC
ï if FB ³ F˜
ïî btN A

CAC
νˆˆ = (19)
β (1 − τ ) N A
From equations (18) and (19) respectively, we can construct the equilibrium configuration of
the regime in the space of (vA, FB) as in Figure 2. In addition, we can demonstrate the relation-
ship between the scale of defense forces and the likelihood of mass killing.
FIGURE 2 Equilibrium configurations of regimes. vA: ruler’s political power; : mass killing threshold; : coalition threshold; : the minimum level of group B army, needed to deter leader A from attack; FA: size of group A’s military forces.

Proposition 1
Under the mass killing regime, an increase in the scale of defense forces (FB) lowers the like-
lihood of mass killing (1 − ν̂ ).

Proof
To address the effect of an increase in the scale of defense forces on the likelihood of mass
∂ (1 − νˆ )
killing we need to determine the sign of under the mass killing regime:
∂ FB
 αB αA 
α A (CAA − CAC ) +  α B FB − α A ( N B − FB )
α A ( N B − FB ) 
( )
∂ (1 − νˆ )  α B FB
= (20)
∂ FB
( )
2
β  α B FB − α A ( N B − FB ) + βτN A 
2
 
116 S.H. BAE AND A.F. OTT

FB
Deterred
mass
killing
F÷ =
a A NB Do Coalition
a A - aB Nothing
Mass
killing
vA
0 vö vö 1
ˆ
FIGURE 2 Equilibrium configurations of regimes. vA: ruler’s political power; ν̂ : mass killing threshold; νˆ :
coalition threshold; F̃ : the minimum level of group B army, needed to deter leader A from attack; FA: size of group
A’s military forces.

for 0 ≤ F̃ B < .
 ∂ (1- νˆ ) 
Therefore, we obtain sign   = sign ( )
α B FB − α A ( N B − FB ) in equation (20).
 ∂ FB 
Since we have 0 ≤ FB < F̃ , we assume that:

α N
FB = λF˜ = λ A B (21)
αA + αB

where λ ∈ [0, 1).

By substituting equation (21) into ( )
α B FB − α A ( N B − FB ) , we have:

α A NB
( α B FB − α A ( N B − FB ) = ) αA + αB
( )
λα B − α B + (1 − λ )α A < 0 (22)

∂ (1 − νˆ )
As a result, we have < 0 . In addition, since ν̂ˆ in equation (19) is not affected by an
∂ FB
∂ (νˆ − νˆˆ ) ∂ νˆ
increase in FB, we have = > 0.
∂ FB ∂ FB

COMPARATIVE STATICS

We now examine the effect of marginal increases in parameter values on the optimal behavior
of the leader of group A. The focus is on those parameters that affect the two critical values ν̂
and ν̂ˆ , which determine the likelihood of engaging in mass killing or forming a coalition
government. We divide the parameters into three subgroups according to their attributes. The
first subgroup of parameters is {CAA , CAC }, which shift the expected utility curve in a parallel
 MASS KILLINGS 117

fashion. For example, with higher fixed cost, CAA, leaders with different political power face
the same increase in fixed cost, which is equivalent to an inward parallel shift in the expected
utility curve in the mass killing case. With a decreased utility, fewer rulers would choose the
mass killing regime as their optimal choice. Therefore, we observe less likelihood of mass kill-
ing in equilibrium. The second subgroup of parameters {αA, NB, τ} affects the slope of the util-
ity curve. The first two parameters only affect the expected wealth level of group A in the mass
killing case, whereas τ only affects the expected wealth level in the coalition government case.
In the comparative static exercise with the second group of parameters, we observe a pivot
change in one of the expected utility curves that affects the slope of the specific utility curve.
The last subgroup of parameters {NA} affects the slopes of both utility curves. Therefore, the
effect of variation in {NA} depends on the relative magnitude of the slope of both utility curves.

Subgroup 1: Variation in Parameters that Shift the Expected Utility Curve in a Parallel
Fashion

Variation in CAA and CAC

We conduct a comparative static exercise by calculating the following equations:

∂ (1 − νˆ ) 1 ∂ (νˆ − νˆˆ ) 1
A =− <0 and = >0 (23a )
∂ CA ∆ ∂ CAA ∆

∂ (1 − νˆ ) 1 ∂ (νˆ − νˆˆ ) 1 1 
= >0 and = − + <0 (23b)
∂ CAC ∆ ∂ CAC  ∆ β (1 − τ ) N A 

β
where ∆≡
αA
( α B FB − α A ( N B − FB ) )2 + βτN A . Increases in CAA have the same effect as
declines in CAC in the mass killing case since the ruler’s optimal choice rests on the net utility
under different regimes. However, in terms of the likelihood of forming a coalition government,
the effects are substantially different. Due to the constant increases in the fixed cost of forming
a coalition government, a leader with lower probability is more adversely affected by an
increase in CAC . Hence, the probability for forming a coalition government falls further with an
increase in CAC . The effect of an increase in the fixed cost is shown in Figure 3 for the two cases.
FIGURE 3 The effect of an increase in fixed cost. vA: ruler’s political power; : mass killing threshold; : coalition threshold; : the group A ruler’s expected utility under a mass killing regime; : the ruler’s expected utility under a coalition government regime; : new ruler’s expected utility under a mass killing regime; : new ruler’s expected utility under a coalition government regime.

Subgroup 2: Variation in Parameters that Rotate Only One of the Expected Utility
Curves
Variation in αA and NB
The comparative statics effect of conflict efficiency and the size of group B is given by:

∂ (1 − νˆ )
=−
(CAA − CAC ) α B FB α B FB − α A ( N B − FB ) ( ) (24)
∂αA (
β −α B FB + 2 α B FB α A ( N B − FB ) + α A ( N B + τN A ) )2

∂ (1 − νˆ )
=−
β (CAA − CAC ) α B FB − α A ( N B − FB ) ( ) (25)
∂ NB α A ( N B − FB )∆2
118 S.H. BAE AND A.F. OTT

VLAA , VLCA VLAA , VLCA

VLAA
VLAA
VLAA : Old utility under Mass killing VLAA VLAA : Utility under Mass killing
VLAA : New utility under Mass killing VLAC : Old utility under coalition
VLAC
VLAC : utility under coalition VLAC VLAC : New utility under coalition

VLAC

0 vöö vö 1 vA 0 vöö vö 1 vA
Do Coalition Mass Do Coalition Mass
nothing killing nothing killing
(a) Variation in C AA (b) Variation in C AC
ˆ
FIGURE 3 The effect of an increase in fixed cost. vA: ruler’s political power; ν̂ : mass killing threshold; νˆ : coalition
A C
threshold; VLA : the group A ruler’s expected utility under a mass killing regime; VLA: the ruler’s expected utility under
′
a coalition government regime; VLAA : new ruler’s expected utility under a mass killing regime; VLCA′: new ruler’s
expected utility under a coalition government regime.

 ∂ (1- νˆ ) 
Equations (24) and (25) imply that sign  (
 = sign α B FB − α A ( N B − FB ) , and )
 ∂αA 
∂ (1- νˆ ) ∂ (1- νˆ )
with the aid of equation (22), we have > 0 and > 0 . In addition, since an
∂αA ∂ NB
increase in αA and NB only affects the utility curve in the mass killing case, we have
∂ (νˆ - νˆˆ ) ∂ νˆ ∂ (νˆ - νˆˆ ) ∂ νˆ
= < 0 and = < 0 . Therefore, αA and NB have very similar
∂αA ∂αA ∂ NB ∂ NB
effects on the wealth level of group A in the mass killing case so that the leader of group
A would be indifferent between a higher chance to win due to higher αA and the
increased wealth of group B due to bigger NB with given FB.

Variation in τ
An increase in τ has two effects: it reduces the mass killing threshold as the coalition regime
becomes less attractive and, at the same time, it increases the coalition threshold. Thus, it
entails a moderate increase in the likelihood of mass killing but a substantial decrease in the
likelihood of forming a coalition government, as follows:

∂ (1 − νˆ ) 1
∂τ ∆
[
= 2 β (CAA − CAC ) N A > 0 ] (25a )

¶ (nˆ - nˆˆ ) é CAC ù

¶t
= -ê
1
(
+ 2 b (CAA - CAC ) N A ú < 0 ) (26)
ë b (1 - t ) N A D
2
û
 MASS KILLINGS 119

VLAA , VLCA VLAA , VLCA

VLAA' VLAA
V A
LA : Old utility under Mass killing
VLAA V A
LA : Utility under Mass killing

VLAA : New utility under coalition V C

LA : Old utility under coalition VLAC
VLAC : Utility under coalition V C
LA V C
LA : New utility under coalition
VLAC

0 vöö vö 1 vA 0 vöö vö 1 vA



 


 

   
  
(a) Variation in =A (b) Variation in J

ˆ
FIGURE 4 The effect of an increase in αA and τ. vA: ruler’s political power; ν̂ : mass killing threshold; νˆ : coalition
A C
threshold; VLA : the group A ruler’s expected utility under a mass killing regime; VLA: the ruler’s expected utility under
′
a coalition government regime; VLAA : new ruler’s expected utility under a mass killing regime; VLCA′: new ruler’s
expected utility under a coalition government regime.

The effects of variations in αA and τ are depicted in Figure 4.

FIGURE 4 The effect of an increase in αA and τ. vA: ruler’s political power; : mass killing threshold; : coalition threshold; : the group A ruler’s expected utility under a mass killing regime; : the ruler’s expected utility under a coalition government regime; : new ruler’s expected utility under a mass killing regime; : new ruler’s expected utility under a coalition government regime.

Subgroup 3: Variation in Parameters that Rotate Both the Expected Utility Curves
Variation in NA
The response of the two types of rulers to increases in the size of NA is not clear. From
Figure 5, the intersection of the two utility curves produces lower value of ν̂ . Thus, we observe
a definite increase in the likelihood of mass killing. As for the effect on the likelihood of
forming a coalition government, an increase in NA gives incentives for rulers to adopt either
regime. Therefore, the effect on the likelihood of choosing a coalition government regime is
ambiguous and depends on the relative magnitude of the two countervailing effects. These
effects are shown as:

∂ (1 − νˆ ) 1
∂ NA
= 2 β (CAA − CAC )τ > 0
∆
[ ] (27)

∂ (νˆ − νˆˆ ) CAC

∂ NA ∆
1
= − 2 β (CAA − CAC )τ +
β (1 − τ ) N A2
>0
< ( ) (28)

Table I gives a summary of comparative statics.

FIGURE 5 The effect of an increase in NA. vA: ruler’s political power; : mass killing threshold; : coalition threshold; : the group A ruler’s expected utility under a mass killing regime; : the ruler’s expected utility under a coalition government regime; : new ruler’s expected utility under a mass killing regime; : new ruler’s expected utility under a coalition government regime.

To summarize: the three-stage framework developed above sets the conditions for the
ruler’s optimal choice in a country where there are two distinct groups of the population (A
and B). It is an ex-ante model used to explain the behavior of the ruler. The choices before the
ruler, assumed to be from group A, are: to attack group B (mass killing regime), form a coali-
tion government sharing power with the leader of group B, or do nothing. The optimization
equation identifies the conditions that have to be met for the mass killing regime to be selected.
120 S.H. BAE AND A.F. OTT

VLAA
VLAA : Old utility under Mass killing
VLAA : New utility under mass killing
VLAA
VLAA , VLCA VLAC : Old utility under coalition
VLAC
VLAC : New utility under coalition

VLAC

0 vöö vö 1 vA



 

  
ˆ
FIGURE 5 The effect of an increase in NA. vA: ruler’s political power; ν̂ : mass killing threshold; νˆ : coalition
A C
threshold; VLA : the group A ruler’s expected utility under a mass killing regime; VLA: the ruler’s expected utility under
′
a coalition government regime; VLAA : new ruler’s expected utility under a mass killing regime; VLCA′: new ruler’s
expected utility under a coalition government regime.

EMPIRICAL SPECIFICATION AND DATA

Two regression models are estimated. Two dependent variables are used: one count variable
for the number of battle-related deaths (1000 or more per year), the second a binary variable
for the probability of mass killing (10,000 or more battle deaths per year). For the count vari-
able, Model 1, we use a negative binomial regression model,5 and for the binary variable,
Model 2, we use the Probit. The independent variables are those highlighted in the theoretical
model with additional variables as control variables. Before presenting the findings, it is useful
to provide a brief discussion of civil war data.

Civil War Data

The empirical literature on civil wars relies heavily on one data source: The Correlates of War
(COW) project.6 Other data sets are available to users, some only by subscription. Almost all
the data sets suffer from systematic problems (Gleditsch 2004), definitional problems
(Gleditsch et al., 2002; Sambanis, 2001; Doyle and Sambanis, 2000), and coding of onset of
war problems (Sambanis, 2004). A more serious problem is the lack of data on variables such
as the size of the military of the two (or more) war factions, the battle deaths of each of the

5
The negative binomial model as compared to other count model (i.e. Poisson) is assured to be the correct model
– the dependent variable is over-dispersed and does not have an excessive number of zeros.
6
For a complete listing of data sets, see Eck (2005).
 MASS KILLINGS 121

TABLE I Comparative Statics Results

The likelihood of mass The likelihood of forming

Parameters killing coalition government

CAA ∂ (1 − νˆ ) ˆ
∂ (νˆ − νˆ )
<0 >0
∂ CAA ∂ CA A

CAC ∂ (1 − νˆ ) ˆ
∂ (νˆ − νˆ )
>0 =0
∂ CAC ∂ CA C

αA ∂ (1 − νˆ ) ˆ
∂ (νˆ − νˆ )
>0 <0
∂αA ∂αA
τ ∂ (1 − νˆ ) ˆ
∂ (νˆ − νˆ )
>0 =0
∂τ ∂τ A
NA ∂ (1 − νˆ ) ˆ
∂ (νˆ − νˆ ) <
>0
∂ NA >
∂ NA 0

NB ∂ (1 − νˆ ) ˆ
∂ (νˆ − νˆ )
>0 <0
∂ NB ∂ NB

groups separately and, in cases where this information is provided, most of the data points are
usually missing.
Drawing on insights offered in the literature regarding databases, definition of civil war and
coding rules, the selection favored the COW data set (1816–1997) for the following reasons:
it meets the requirement of a relatively large sample (213 observations); it defines a major
armed conflict in a civil war as resulting in at least 1000 annual battle-related deaths; it
identifies the war initiator (government or another group) and the onset and termination of the
war. COW defines a civil war by the internality of the war to the territory of a sovereign state
and the participation of the government as a combatant. A listing of the variables used in the
estimation and the sources are given in Table II.

Findings
The estimation results are presented in Tables III and IV. In both models, the independent
variables are those suggested by the theoretical model: gross domestic product, military
expenditure, military personnel, power of the ruler (measured by the length of executive

TABLE II List of Variables and Sources

Variable name Source

lgdp: Log GDP Angus Maddison (2007), from year 1 to 2003.

lmilexp: Log Military Expenditure Correlates of War (COW).
lpop: Log Population Angus Maddison (2007).
lmilper: Log military personnel Correlates of War (COW).
lmindur: Log war duration Correlates of War (COW).
ethnic: Ethnic fractionalization Alesina et al. (2003).
lengexec: Length executive (in years) Cross-National Time-Series (CNTS) Data Archive
(launched in 1968 by Arthur S. Banks), 1815–1999.
masskill: Dependent variable in Probit model: Dummy Correlates of War (COW).
variable, =1 if number of deaths >=10,000; 0, otherwise
battledeath: Dependent variable in Negative binomial Correlates of War (COW).
model: Number of battle deaths
122 S.H. BAE AND A.F. OTT

TABLE III Negative Binomial Regression of Civil Wars Battle-related Deaths, 1816–1997.

Negative binomial model

Variable Coef/t-stat IRR

lgdp −0.256 0.7740

(−1.44)
lmilexp 0.041 1.0423
(0.54)
lpop 0.546** 1.7266
(2.84)
lmilper 0.071 1.0740
(0.53)
lmindur 0.457*** 1.5799
(9.45)
ethnic 0.085 1.0881
(0.14)
lengexec −0.045*** 0.9560
(−3.48)
constant 1.904
(1.33)
constant 0.773***
(9.38)
No. of obs. 198 89

* p<0.05, ** p<0.01, *** p<0.001

tenure) and ethnicity (measured by the fractionalization index). Two other variables,
population and duration of the war are included as control variables.
In Model 1 the dependent variable is the number of battle related deaths (≥ 1000 annually).
Except for ethnicity and the length of the executive, the independent variables are in logs.
From the regression we find the length of executive tenure and duration to be significant

TABLE IV Probit Model Estimates, Civil Wars 1816–1997

Probit model Marginal change

Variable coef/t-stat coef/t-stat

lgdp −0.399* −0.155*

(−2.11) (−2.11)
lmilexp 0.215* 0.084*
(2.49) (2.47)
lpop 0.367* 0.143*
(2.44) (2.44)
lmindur 0.491*** 0.191***
(7.11) (7.25)
ethnic −1.130* −0.440*
(−2.18) (−2.17)
lengexec −0.025 −0.010
(−1.45) (−1.45)
constant −5.647***
(−4.30)
No. of obs. 198 198

* p<0.05, ** p<0.01, *** p<0.001

 MASS KILLINGS 123

factors for civil war killing. Population is positive and significant; GDP negative but not
significant; ethnicity was positive but not significant. The latter findings are similar to those
reported in the literature (Gleditsch et al., 2002; Fearon and Laitin, 2003; Sambanis, 2004).
From the estimated incidence rate ratio, a one year increase in the length of executive tenure,
the number of battle-related deaths would be expected to decrease by a factor of 0.956, holding
all other variables in the model constant. As for war duration, a one percentage increase in
duration leads to an increase of the rate ratio for battle-related deaths by a factor of 1.58,
ceteris paribus. Population also increases the death rate – a 1% increase in population leads to
an increase of the rate ratio by a factor of 1.72, ceteris paribus.7
In Model 2, the main independent variables: log of GDP, log of military expenditure, log of
war duration, log of population, and ethnicity, have the correct sign and are all significant. The
likelihood of mass killing (≥ 10,000 battle-related deaths annually) increases the greater
the resources the government devotes to the war, the larger the duration of the conflict and the
larger the population. The negative sign for ethnicity, measured by the fractionalization index,
suggests that the more fragmented the population (many ethnic groups), the less likely is mass
killing. With respect to GDP, again the finding is consistent with the literature – the higher the
income of the country, the less likely the ruler would engage in predatory behavior.
Considering the marginal effects on the probability of mass killing from a change in an
independent variable, we note that one unit increase in military expenditure increases the
likelihood of mass killing by 8.4%; a one unit increase of population by 14.3%; a one unit
increase in duration by 19%. The strongest effect is for ethnicity, where a one unit increase in
the index reduces the probability by 44%, for GDP the reduction is 15.5%.
A further insight into the ruler’s choice of the mass killing regime may be gained from data
reported in Table V. In the table the contest success functions (CSF) are compared with the
actual outcomes for 24 civil war conflicts.
In the theoretical model it is argued that the ruler of group A in initiating the attack on
group B expects to enhance his group’s wealth (θ AA ). Since θ AA depends on his group’s wealth
and the expected wealth appropriated from group B, the likelihood of this appropriation can
be inferred from the scores of the CSF. A value between 0.5 and 1.0 predicts a winning
outcome for the party initiating the conflict (the government). This turned out to be the case
for 11 cases. In four cases, the opposition won even though the CES scores for the govern-
ment were above 0.5. For the remaining nine cases, the CSF scores did not predict the
outcome. Of these cases the government won four, the opposition won two and for three wars
the outcome for each war was a stalemate or no win.
Several factors may explain the ‘limited success’ of the CSF. Because of missing informa-
tion on the relative size of the two armies in the COW data set, the small sample size (24 out
of 213 observations) may have biased the result. In addition, factors not captured in the esti-
mation, such as the efficiency of the conflict technology, information that could have been
gained from prior war episodes and/or support from outside sources by either party to the
conflict, may have played a more significant role in determining the outcome.

CONCLUSIONS

The statistical analysis provides insights into the factors that induce a government, or a leader
of a country, to engage in mass killing of its own people.

7
We estimated another variant of the model where the number of battle-related deaths was set at ≥ 10,000 annually.
For this sample, none of the coefficients changed sign, unlike the significance of some of the independent variables.
124 S.H. BAE AND A.F. OTT

TABLE V A Comparison of CSF and Civil War Outcome

Name of war FA/FB CSF War Outcome (winner)

Two Sicilies vs. Anti-Monarchists 6.5 0.866 Opposition

Sardinia vs. Sardinian Rebels 6.3 0.864 Government
Portugal vs. Conservatives 3.6 0.781 Government
Ottoman Empire vs. Mehmet Ali 1.3 0.566 Government
Austria-Hungary vs. Magyars 0.6 0.369 Government
China vs. Taipings 0.3 0.234 Government
Morocco vs. Fez Caids of 1907 103.2 0.990 Government
Morocco vs. Fez Caids of 1911 106.3 0.990 Government
China vs. Kuomintang 0.2 0.139 Opposition
Greece vs. Communists 326.6 0.997 Government
Lebanon vs. Leftists of 1958 267.8 0.996 Government
Republic of Vietnam vs. NLF 14.4 0.935 Opposition
Zaire vs. Katanga & Leftists 4.4 0.815 Government
Yemen Arab Republic vs. Royalists 55.0 0.982 Government
Dominican Republic vs. Leftists 129.6 0.992 Government
Jordan vs. Palestinians 0.9 0.462 Government
Ethiopia vs. Eritrean Rebels 2.7 0.735 Opposition
Iraq vs. Kurds of 1974 0.5 0.314 Government
Afghanistan vs. Mujahedin 33.3 0.971 Opposition
Cambodia vs. Khmer Rouge of 1978 102.0 0.990 Stalemate
Sri Lanka vs. Tamils 70.1 0.986 No winner
Iraq vs. Kurds & Shiites 1.7 0.630 Government
Liberia vs. Anti-Doe Rebels 0.05 0.05 Opposition
Liberia vs. NPFL & ULIMO 15.2 0.938 Stalemate

Source: Correlates of War Project, 1816–1997

The findings presented in the paper differ from those reported in the literature in two major
respects: the threshold for mass killing is defined at ≥ 10,000 battle-related deaths annually,
whereas the literature uses either ≥ 1000 per year or some number between 25 and 1000.8
Secondly, two of the independent variables – the power of the leader measured by the length
of executive tenure, and military expenditure – turned out to be significant, but were not
investigated in the civil war studies.
As is the case in most empirical research, the testing of hypotheses is held hostage to data
availability and data quality. Here is an avenue for future data collection and research.
Researchers undoubtedly will continue to use COW data sets. There is room for improvement
there. For one thing, some effort should be made to fill in missing values, especially for two
critical variables, i.e. pre-war armies of the two factions in the war and battle-related deaths
separately for the state and the group challenging the state. The suggested improvements in
data collection will enable researchers to shed light on the motivation as well as the capabili-
ties of the opposing groups. Moreover, factors that would induce the government and the
opposition parties to resolve the conflict through the formation of a coalition government
rather than engaging in mass killing need to be filtered out of the data. Another avenue of

8
The choice of ≥ 10,000 battle-related deaths per year, although arbitrary, conveys the scale of mass killing in a
civil war more than the practice of using ≥ 25 or ≥ 1000 battle-related deaths. The higher number used in this paper
was in response to an anonymous referee’s suggestion.
 MASS KILLINGS 125

research is to examine the role played by outsiders in initiating or concluding a violent

intrastate conflict.

ACKNOWLEDGEMENT

The authors are grateful for comments and suggestions from two anonymous referees and the
editor. They thank participants to the African Outreach Program of The Institute for Economic
Policy Studies for useful comments. Chyanda Querido provided valuable research assistance.

References
Alesina, A., Devleeschauwer, A., Easterly, W., Kurlat, S. and Wacziarg, R. (2003) Fractionalization. Journal of
Economic Growth 8(2) 155–194.
Azam, J. (2002) Looting and conflict between ethnoregional groups: lessons from state formation in Africa. Journal
of Conflict Resolution 46(1) 131–153.
Azam, J. and Hoeffler, A. (2002) Violence against civilians in civil wars: looting or terror? Journal of Peace Research
39(4) 461–485.
Bellany, I. (1999) Modelling war. Journal of Peace Research 36(6) 729–739.
Collier, P., Hoeffler, A. and Rohner, D. (2006) Beyond greed and grievance: feasibility and civil war. CSAE Working
Paper 2006-10, Department of economics, Oxford University.
Doyle, M. and Sambanis, N. (2000) International peacebuilding: a theoretical and quantitative analysis. The American
Political Science Review 94(4) 779–801
Easterly, W., Gatti, R. and Kurlat, S. (2006) Development, democracy, and mass killings. Journal of Economic
Growth 11(2) 129–156.
Eck, K. (2005) A Beginner’s Guide to Conflict Data: Finding and Using the Right Dataset. UCDP Paper 1, Depart-
ment of Peace and Conflict Research, Uppsala University, Sweden.
Eck, K. and Hultman, L. (2007) One-sided violence against civilians in war: insights from new fatality data. Journal
of Peace Research 44(2) 233–246.
Fearon, J. and Laitin, D. (2003) Ethnicity, insurgency and civil war. American Political Science Review 97(1) 75–90.
Gleditsch, K. (2004) A revised list of wars between and within independent states, 1816–2002. International
Interactions 30 231–262.
Gleditsch, N., Wallensteen, P., Eriksson, M., Sollenberg, M. and Strand, H. (2002) Armed conflict 1946–2001: a new
dataset. Journal of Peace Research 39(5) 615–637.
Harff, B. (2003) No lessons learned from the holocaust? Assessing risks of genocide and political mass murder since
1955. American Political Science Review 97(1) 57–73.
Hirshleifer, J. (1988) The analytics of continuing conflict. Synthese 76(December) 201–233.
Hirshleifer, J. (1995) Anarchy and its breakdown. The Journal of Political Economy 103(1) 26–52.
Lacina, B. and Gleditsch, N. (2005) Monitoring trends in global combat: a new dataset of battle deaths. European
Journal of Population 21(2–3) 145–166.
Maddison, A. (2007) World Population, GDP and Per Capita GDP, 1-2003 AD. At http://www.ggdc.net/maddison.
Montalvo, J. and Reynal-Querol, M. (2005) Ethnic polarization, potential conflict, and. civil wars. American
Economic Review 95(3) 796–813.
Murdoch, J. and Sandler, T. (2002) Civil wars and economic growth: a regional comparison. Defence and Peace
Economics 13(6) 451–464.
Reynal-Querol, M. (2002) Ethnicity, political systems and civil wars. Journal of Conflict Resolution 46(1), 29–54.
Rummel, R. (1995) Democracy, Power, Genocide, and Mass Murder. Journal of Conflict Resolution 39 3–26.
Sambanis, N. (2001) A note on the death threshold in coding civil war events. Conflict Processes Newsletter online,
at http://mailer.fsu.edu/∼whmoore/cps/newsletter/june2001.
Sambanis, N. (2004) What is civil war? Conceptual and empirical complexities of an operational definition. Journal
of Conflict Resolution 48(6) 814–858.
Sandler, T. and Hartley, K. (2003) Economics of Conflict. Aldershot, UK: Edward Elgar.
Tullock, G. (1974) The Social Dilemma: The Economics of War and Revolution. Blacksburg, VA: University
Publications.
World Bank (2004) World Development Report 2005: A Better Investment Climate for Everyone. Washington, DC:
World Bank Publications.