Greed and Grievance in Civil War
Paul Collier
The World Bank, CEPR and CSAE
Anke Hoeffler
CSAE, Oxford
The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors.
They do not necessarily represent the views of the World Bank, its Executive Directors, or the countries
they represent.
2
1. Introduction1
Civil war is both a tragedy and an impediment to development, affecting most of the world’s
poorest countries. It is now far more common than international conflict: of the 27 major armed
conflicts listed in the Stockholm International Peace Research Institute Yearbook for 1999
(SIPRI 1999), all but two were internal.
In Section 2 we compare two contrasting motivations for rebellion: greed and grievance. Most
rebellions are ostensibly in pursuit of a cause, supported by a narrative of grievance. However,
since grievance-assuagement through rebellion is a public good which a government will not
supply, economists would predict that such rebellions would be rare. Empirically, many
rebellions appear to be linked to the capture of resources: diamonds in Angola and Sierra
Leone, drugs in Colombia, and timber in Cambodia. We set up a simple rational choice model
of greed-rebellion and contrast its predictions with those of a simple grievance model. A second
empirical regularity is that some countries are prone to repeat conflict. This may be either
because their underlying characteristics make them highly conflict-prone, or because of a
feedback effect whereby conflict generates grievance which in turn generates further conflict.
We show why such a feedback effect might be present in greed-motivated rebellions as well as
in grievance-rebellions. Finally, we consider an integrated model in which the motivation for
rebellion is both greed and grievance.
In Section 3 we discuss the construction of a comprehensive data set of 161 countries for each
of the eight, five-year periods between 1960 and 1999, giving a total of 1288 potential
observations. For 73 of these observations the society was at peace at the start of the period
but experienced civil war during it. In Section 4 we use logit regressions to explain these
collapses into civil war in terms of characteristics at the start of the period. We perform nonnested
tests on the greed and grievance models. Although the greed model substantially
outperforms the grievance model, we cannot reject the hypothesis that the grievance model
adds to explanatory power. An integrated model, which incorporates some features of
grievance into the greed model, outperforms all other models. An economic calculus of the costs
and opportunities for the control of primary commodity exports appears to be the main
systematic initial impetus to rebellion, with an additional effect from fear of domination by an
ethnic majority. After peace has been restored, the legacy of conflict-induced grievance enables
rebel movements to restart conflict by drawing on the support of their diasporas. We show that
the results are robust to the inclusion of a wide range of alternative variables, and to tests for
random and fixed effects and to correction for rare events bias. The last section concludes.
2. Theories of Rebellion
1
Previous versions of this paper have benefited from presentations at CERDI, NYU, LSE, Oxford, Princeton,
the World Bank and a CEPR conference at Lisbon. We would like to thank participants for comments, and
especially Todd Sandler for helpful written suggestions.
3
Civil war occurs as a result of rebellion. Hence, the phenomenon to be explained is the
emergence of a rebel organization. A rebellion is somewhat analogous to three other types of
organization: protest movements, armies, and organized crime. In popular and political analysis,
the most common analogue to rebellion is protest. The theory of protest movements, (Kuran,
1989), focuses on the coordination problem. In effect, by joining each other on the streets,
people can create an election. However, the analogy is not very close. Protest is not a sustained
economic activity: the participants are not full-time employees of the protest organization, and
they risk little. Indeed, this is the essence of Kuran’s `prairie fire’ model: protest only takes off if
the risks fall fast enough (through increasing participation) to make it safe. By contrast,
participants in a rebellion must be prepared to fight over a prolonged period against organized
force which aims to kill them. In this, a rebel organization is more analogous to a regular army
than to a protest movement. Like an army it must solve immense problems of hierarchy and
cohesion in order to get people to risk their own death in order to further the military objective
of the group. However, unlike both an army and a protest organization, a rebellion must
generate revenue in order to feed and pay its workers. The payroll of an army is financed out of
taxation which the army itself does not raise. A protest movement does not have a significant
payroll. The rebel organization must generate income despite not being directly productive, and
in this respect rebellion is like crime. Indeed the analogy with crime is standard in the present
economic theory of rebellion For example, Grossman (1999) states `in such insurrections the
insurgents are indistinguishable from bandits or pirates’ (p.269). However, the very scale of
rebellion makes it distinctive from other crime. Rebel organizations usually have between 500
and 5,000 workers, whereas most criminals are self-employed or work in small groups. Hence,
within crime, the closest analogy is that with organized crime. Even organized crime is usually a
rather small scale activity, but the largest groups are approximately comparable in size to the
smaller rebel organizations. The recent theory of organized crime (e.g. Konrad and Skaperdas,
1998), explains its larger scale than other crime by scale economies in extortion, eventually
counterbalanced by policing.
The analogies with the economic theories of organized crime and protest movements form the
basis for the `greed-rebellion’ theory of Section 2.1. In this model rebellion is a distinctive type
of organized crime, although subject to constraints similar to those facing an incipient protest
movement. In Section 2.2 we present a contrasting political model of rebellion in which we put
aside the economic considerations and focus instead upon various forms of grievance. Section
2.3 compares the two models and considers various ways in which rebellion might combine
both greed and grievance.
2.1. A Model of Greed-Rebellion
Consider a rebellion which, like organized crime, generates its income from extortion. The
rebels menace legitimate economic activities and exact a tribute. Empirically, rebellion is
nevertheless distinctive from other types of organized crime, if only in terms of scale and
casualties. We are concerned only with those rebellions which are sufficiently large and
4
sustained, and come into sufficient conflict with government forces, to generate at least 1,000
battle-related deaths. This is the conventional empirical definition of civil war, and this is the
phenomenon which we will be seeking to explain. Hence, rebellion is distinctive from other
crime in the scale of organized violence. We suggest that rebellion is distinctive because the
object of extortion is different. The typical object of criminal extortion is a street of
shopkeepers. The criminal organization is small because the scale economies in menacing
shopkeepers are modest, and because the rate of predation cannot get very high before
businesses relocate. Rebellion is at the apex of organized crime because the object of extortion
is not a street of shopkeepers but the export of primary produce. As a result, both the scale
economies of menace and the sustainable rate of extortion, are atypically high. Primary
commodity exports are sustainable targets for predation because their production is intensive in
irreversible and immobile assets, and because produce must be transported to a port. The
owners of irreversible and immobile assets such as land, trees and mines, receive rents which
can be expropriated without curtailing the activity, whereas similar predation of the incomes
generated by mobile factors would simply produce relocation. As primary commodities are
transported to a port they are exposed to predation at many geographic `choke points’. The
government itself typically levies taxes at the tightest choke point, namely the port, but rebels
may attack at any point along the transport route.
We begin with a model which, for a country of given population, analyzes how the risk of such a
predatory rebellion is affected by variations in the level of income and its structure.
Corresponding to our subsequent empirical analysis, we will be seeking to explain only the
initiation of conflict, rather than its duration.
We first specify a function for the revenue which a rebellion can generate from predation upon
primary commodity exports, while being opposed by government forces which are protecting
those exports. The gross revenue function has two components, the tribute which can be
exacted conditional upon a successful threat of military force, and a military contest function.
The tribute is assumed to be a function of the value of primary commodity exports: the tribute
increases in this base for predation, but at a diminishing rate. It is convenient to specify primary
commodity exports as the product of income, y, and the share of income constituted by such
exports, n. For the military contest function, we follow the specification of Konrad and
Skaperdas (1998), in which military outcomes reflect the balance of opposing forces, rebel, r,
and government, g. Hence:
Rr = (ny)α
[r/(r+g)]. (1)
We next introduce an insight from the economic theory of protest movements (Kuran, 1989).
Government forces are not simply trying to impede rebels from predation, they are trying to kill
them. If the rebel force is too small, when it attacks the choke points, the government forces
which it encounters will turn from defense to attack. In Kuran’s model, there is a tipping point of
participation above which the protest becomes viable. Here, we assume that there is a threshold
size of rebel forces relative to government forces, below which predation of primary commodity
5
exports is too dangerous to be viable. We introduce this as a survival constraint: rebel forces
must exceed some fraction of government forces in order to engage in predation without
suffering punishment. Thus,
r ≥ βg. (2)
The rebel cost function, Cr, is its wage bill:
Cr = rwr. (3)
Where wr = rebel wage rate.
The rebel wage is assumed to be linear in the level of income, y:
wr = ϕy. (4)
The rebel leader thus chooses the size of the rebel laborforce so as to maximize net revenue,
(1)-(3), subject to (2). Rebellion will occur when net revenue is non-negative, which we refer to
as the financial viability constraint:
Rr – Cr ≥ 0. (5)
We next endogenize the initial government defense effort, g, which precedes the initiation of
rebellion. The government cost function, Cg, mirrors that of the rebels:
Cg = gwg (6)
wg = λy. (7)
We assume that the government sets its precautionary military expenditure as fraction of its
revenue, Rg. Reflecting observed fiscal patterns, government revenue is assumed to be elastic in
income, and in the share of income from primary commodity exports. As discussed, the same
factors make natural resource exports atypically taxable and lootable.2
Rg = δyθ
+ γny. θ>1 (8)
Prior to rebellion, the government devotes a given share of its expenditure, ν, to military
expenditure. Hence, prior to rebellion, government forces, g, are:
2
Government military expenditure may also be increasing in the proportion of natural resource exports
because the latter are recognized as being vulnerable to predation. This would simply reinforce the results
below on the non-monotonic effect of natural resource exports on the risk of conflict.
6
g = ν(δyθ-1
+ γn)/λ. (9)
The minimum size of a viable rebellion is that at which the survival constraint is binding. If the
rebellion is financially viable at this size, the rebel leader may choose to expand it further.
However, since there are diminishing marginal returns to rebel labor, the rebellion will only be
worth expanding beyond the size imposed by the survival constraint if it is also financially viable
at that size. Hence, financial viability at the size which just satisfies the survival constraint is the
condition for the initiation of a rebellion.
Substituting, and rearranging, the financial viability condition is:
λnα
/(1+β) ≥ ν(δϕyθ-α
+ nϕγy1-α
) (10)
Equation (10) is thus the key condition for the initiation of greed-rebellion.
Now consider whether the rebellions so initiated are likely to become sufficiently large to be
recorded as civil wars: that is, whether they generate at least 1,000 combat-related deaths. For
this a key consideration is whether the incipient rebellion generates a phase of `arms race’ in
which both government and rebel forces grow in response to each other. If there is such a
phase, then we will assume that the rebellion is large enough to generate a civil war.
The sequence of rebellion is as follows. In the pre-history of the rebel movement it builds its
forces until they reach the level which just satisfies the survival condition. At this stage rebel
forces can start operations, and so combat deaths commence. We assume that the initial
response of the government to rebellion is to increase its expenditure upon its military forces and
focus on how rebel forces respond to this increase. If rebel forces respond by contracting, then
it is possible that the rebellion gets snuffed out before combat-related mortalities reach the
critical level of 1,000 at which the conflict is classified as a civil war. By contrast, if the rebel
group initially responds by increasing its own forces, then there is a phase of mutual escalation,
although this phase may end with the bankruptcy of the rebellion. The Konrad-Skaperdas (KS)
military combat function predicts how the optimal size of the rebel organization changes in
response to an increase in government forces: rebel forces will increase if r>g, and decrease if
r<g. However, at the point at which the rebel organization commences predation its forces are
below their optimal size: the rebel organization builds its forces from zero, and once they satisfy
the survival constraint, operations commence. At this point the government expands its own
forces. Hence, regardless of whether the optimal size of the rebel organization increases or
reduces in response to the expansion of government forces, the actual size must increase in line
with the rise in the minimum size necessary to satisfy the survival constraint. The rebel
organization expands its forces at least in reaction to this increase and perhaps additionally in a
continuing adjustment towards the optimum. In response, the government further expands its
own forces.3
Hence, once (10) is satisfied, not only does the rebellion come into existence, but
3 Collier (2000) further analyzes this escalation.
7
it enters an arms race phase with government forces. While this race may end with the
bankruptcy of the rebellion, we assume that in the interim it will have caused sufficient combat
deaths to meet the criterion used in our empirical analysis. Condition (10) is thus both a
necessary condition for rebellion and a sufficient condition for civil war. We now derive from it
specific predictions as to how the risk of rebellion will be affected by the level of income and its
structure.
First, consider the level of income. The LHS of (10) is not a function of y. Since θ>1>α, the
derivative of the RHS w.r.t. y is strictly positive:
dRHS/dy = (θ-α)νδϕyθ-α-1
+ (1-α)νnϕγy-α
> 0. (11)
Hence, the conditions under which rebellion is profitable become more restrictive the higher is
income. Restated, higher income is predicted to reduce the risk of conflict.
Next, consider the structure of income. Differentiating (10) w.r.t. n yields:
d(10)/dn = [αλnα-1
/(1+β)] - νϕγy1-α
. (12)
In the neighborhood of n=0, (10) is negative and (12) is positive, so that as n increases from
zero the risk of conflict increases. However, (12) is not monotonic in n. Setting (12) to zero and
solving for n yields a critical value, n*, below which the conditions for rebellion become easier
as n increases, and above which they become tighter. Hence, n* denotes that intermediate level
of primary commodity dependence at which the risk of rebellion is at its peak:
n* = {[(1+β)νϕγy1-α
]/αλ}1-α
. (13)
The above analysis applies to a country of given population. Now consider how the risk will be
affected by variations in population size. If the risk in each of two identical territories is p, the
risk that there will be at least one conflict in the two territories is p + p(1-p). The elasticity of
risk with respect to population, controlling for n and y, will therefore be (1-p). Globally, over
the period 1960-1999 the mean risk of conflict per period was around 0.06, so that the
expected elasticity of conflict with respect to population is around 0.94.
Thus far, the model has four testable predictions. The risk of conflict should be decreasing in per
capita income, increasing in the share of natural resource exports in GDP at low levels,
decreasing at high levels, and be slightly less than unit elastic in population. We now introduce
three refinements which introduce further testable propositions: relative military advantage, rebel
costs of recruitment, and start-up finance.
Relative military advantage
8
The military combat function included in (1) is a convenient simple special case of the KS
function. The more general form of the KS combat function is:
r/(r+g+k), (14)
where k denotes the relative military advantage of the two forces. For k=0, the special case we
have adopted above, neither force has technological superiority. For k>0 government forces
have the advantage, and conversely for k<0. Relative military advantage affects the necessary
conditions for financial viability. Trivially, the greater is the relative military advantage of the
government, (the higher is the value of k), the lower is the value of the combat function, (14),
and hence the lower is gross and net revenue from rebellion. Rebellion is less likely the greater is
the relative military advantage of the government. We endogenize relative military advantage
with respect to geography, cohesion and motivation.
Rebels need a refuge, and two geographic features, mountains and forests, are commonly
supposed to make counter-insurgency more difficult. We measure these as the proportion of a
country’s terrain which is mountainous, m, and the proportion which is forested, Λ.
Additionally, Herbst (2000) has suggested that states such as Zaire are prone to rebellion
because the population is dispersed around the edges of the country, making government
military control more difficult. We measure this by the geographic dispersion of the population,
Ψ. We predict that the risk of conflict is increasing in these three measures.
In war, the size of a fighting force is probably less important than its willingness to fight. Military
history offers many instances in which cohesion and motivation offset numerical inferiority. The
government army has the advantage that it predates the rebellion and so has had time to build
cohesion and motivation. The rebel organization must create these attributes as it recruits. The
need for cohesion places a simple constraint upon the composition of rebel recruits: the
organization cannot afford diversity. If recruitment spans ethnic and religious divides it will be
more difficult to forge the resulting laborforce into a cohesive fighting force. The strategy of
homogenous recruitment will sometimes come at a cost, since in socially fractionalized societies
it will reduce the size of the recruitment pool and so raise labor costs for any given size of
recruitment. Thus, the more fractionalized the society, the more costly is rebel recruitment (or
the less cohesive is the rebel force). We thus predict that social fractionalization should reduce
the incidence of civil war. We measure social fractionalization, f, by the extent to which a
society is divided by ethnicity and religion.
In addition to cohesion, an army needs motivation: rebel forces must be persuaded to want to
kill the enemy. The Leninist theory of rebel organization, which many rebel groups follow even if
they are not Marxist, is that initially the population from which forces are recruited does not
realize that it is oppressed, so that an awareness of grievance has to be built by the rebellion.
Many rebel organizations invest substantial resources in this process of indoctrination. For
example, the highly successful Eritrean People’s Liberation Front routinely withdrew troops
from the front line for periods of six months for ideological training. Thus, subjective grievance
9
is consciously generated by conflict to enhance efficiency, rather than being either an accidental
by-product of conflict, or a prior cause of conflict. The greed model departs from Leninist
theory in assuming that the ability of a rebel organization to inculcate subjective grievance is
unrelated to objective grounds for grievance. This assumption is partly a convenience to
distinguish the model as sharply as possible from the subsequent grievance model, but also has
some a priori claim to credibility. Many rebel groups target their recruitment on children and
drug addicts, and such selectivity, in conjunction with the phenomenon of relative deprivation,
whereby a majority of the population considers itself to be unfortunate, may enable
indoctrination to be effective in generating grievance regardless of objective circumstances. This
theory of military motivation has two testable predictions. First, objective causes of grievance
will not explain the risk of conflict. We return to this in Section 2.2 where we measure objective
grounds for grievance. Secondly, since conflict produces military cohesion, rebel military
advantage will be greater in societies which have recently experienced conflict. We measure this
by the time, τ, which has elapsed since the end of any previous civil war.
The model implies that in ethnically diverse societies, rebellion will appear to be caused by
ethnic differences even though diversity actually reduces the risk of conflict. Where rebellion
occurs in such societies, the need for homogeneity in recruitment will make it ethnically specific,
and the need for the rebel organization to generate a subjective sense of grievance may focus
the discourse on ethnicity.
The hypothesized effects of geography, cohesion and motivation on relative military advantage
are summarized in (15):
k = k(m, Λ, Ψ, f, τ). (15)
+ + + - Rebel
costs of recruitment
In the basic model set out above, rebel and government labor costs are treated as being
symmetrical: for both the wage is simply linear in per capita income. However, this ignores an
important difference between the rebel force and the government army. Our focus is on the
initiation of rebellion, and in this phase the rebel force must grow, and probably needs to grow
rapidly to reach the survival constraint at minimum cost. By contrast, the government army is
already in steady state. We would expect this to make the labor costs of the rebel organization
more sensitive to the state of the labor market than those of the government. The tighter is the
labor market, controlling for the level of per capita income, the more costly would be rebel
recruitment, and so the lower would be the risk of conflict. We measure the state of the labor
market through three proxies: education, population growth, and economic growth. Our
education variable is the proportion of young males enrolled in secondary education, s. This
both directly measures an important alternative activity to rebellion for that part of the population
most likely to be recruited, and proxies income-earning opportunities for the group. The
10
population growth rate (
•
p ), proxies the change in labor supply, and will consequently proxy the
change in pressure on fixed assets, notably land. The growth rate of income proxies the demand
for labor. Income growth decomposes into
•
p and the growth in per capita income,
•
y , and
since the former is already directly included, we measure income growth net of population
growth. The difficulty of rebel recruitment is assumed to be increasing in s and,
•
y , while being
decreasing in
•
p . To reduce problems of endogeneity, these variables are all measured prior to
the period of prediction. In (16) we introduce this as a modification to the rebel wage function,
(4):
wr = ϕy.ƒ(s,
•
y ,
•
p ) (16)
+ + Since
this change in the wage function affects the financial viability of rebellion, we predict that
the risk of conflict decreases with s and
•
y and increases with
•
y .
Start-up finance
We now consider how rebel organizations might acquire start-up finance. The survival condition
imposes a minimum size on rebel forces below which they cannot be operational in resource
predation. This implies that there are threshold start-up costs. Since rebellions may not be able
to raise funding from conventional sources, they must look elsewhere. We consider two
potential sources of finance, foreign governments and diasporas.
Foreign governments may provide pump-priming finance for rebellions which subsequently
become self-sustaining. An unusually clear example was the financing of Renamo in
Mozambique by the government of Southern Rhodesia. The latter collapsed shortly after
Renamo was established, but Renamo sustained itself as a rebel force for many years before
evolving into a political party. To test the proposition empirically, we need some proxy for the
willingness of foreign governments to finance military opposition to the incumbent government.
Among the possible proxies, we regard the end of the Cold War as providing the clearest test
of the proposition. Undoubtedly, during the Cold War the two major powers tried to enlist
other governments as allies, and each power had an incentive to destabilize those governments
which allied with the opposing power. Hence, we predict that with the end of the Cold War, the
supply of foreign power finance for rebellions declined and that this should have reduced the
risk of conflict. We test this by introducing a dummy variable, D1, for the post-Cold War
period. We might note that this is a controversial prediction. For example, Kaplan, (2000)
argues that the end of the Cold War `lifted the lid’ off previously suppressed conflict.
11
A further potentially important source of start-up finance for rebellion is a diaspora living in
OECD countries. Such diasporas are usually much richer than the population in their country of
origin. They are better-placed for collective action: emigrants have a cultural incentive to create
diaspora organizations which can then discipline free-riding. They do not suffer the
consequences of the conflicts they finance. As with grievance among the local population, in the
greed-model grievance among the diaspora is assumed to be manufactured by the rebel
organization rather than being an original cause of conflict. Hence, the diaspora increases the
risks of conflict renewal, but not the initial risk of conflict. We measure the size of diasporas in
the USA relative to the population in their country of origin. We test whether this variable, d, is
significant as an initial cause of conflict, and by interacting it with the time since any previous
conflict, d/τ, whether it is a significant risk factor post-conflict. The theory of greed-conflict
predicts that only the latter will be significant.
2.2. A model of grievance-rebellion
In a grievance-rebellion the objective is not predation, but rather to assuage grievance. This is
indeed the standard characterization of rebellion in the popular literature. The most proximate
economic theory is, as noted above, that of protest movements. Within this framework, a
rebellion might be thought of as a protest movement which has failed to escalate into mass
participation. Recall that rebellions seldom have more than 5,000 participants, whereas
successful protest movements are generally many times this size. In Kuran’s model, the early
phase of a protest movement is generated by a few ardent adherents who are insensitive to the
risks involved. It escalates if this induces a cascade of participation by those who are initially
deterred by fear of government reprisals. A theory of rebellion as a failed protest movement
must thus explain two stages: the initial impetus for protest, and the failure of escalation.
Consider, first, the circumstances in which some people feel sufficiently strongly to protest, to
the extent of being insensitive to the risks of government reprisal. Evidently, the impetus for such
protest should be rooted in objective grievances, some of which are potentially measurable. We
distinguish three of them: inter-group hatred, political exclusion, and vengeance.
Inter-ethnic or inter-religious hatreds, h, are probably the most common popular explanation for
civil conflict. A possible example of such a conflict might be Bosnia. There is indeed evidence
that such hatreds exist, and since many conflicts are inter-ethnic or inter-religious, it is widely
presumed that the hatreds are the cause of the conflict. Although such hatreds are usually not
directly observed, they can evidently only occur in societies which are multi-ethnic or multireligious.
Inter-group hatred may be monotonic in the extent of social fractionalization, or the
relationship may be a quadratic: for example, societies with two groups may have a higher
incidence of inter-group hatred than societies with many groups. Hence,
h = h(f), h(0) = 0; h(f>0) > 0 (17)
12
A second presumed cause of grievance is political exclusion, j. The quantitative political science
literature has already explored the relationship between conflict and the political rights of a
society, q, as the latter range from dictatorial repression to full representative democracy.
Econometric studies have found that other than when repression is very severe, it tends to
increase the risk of conflict (Gleditsch and Hegre, 1997). However, even democracies may
generate grievance if one voting block is able to forge a persistent majority and uses its power
to disadvantage a minority. Normally, such persistent majorities are illusive because excluded
groups are able to make offers which split the majority. However, one circumstance in which a
stable winning coalition can form is where political allegiance is pre-determined by ethnic identity
and one ethnic group constitutes a majority of the population. Whether such a group uses its
power to extract transfers from the minority depends in part upon the size of the majority.
Conditional upon power, large majorities have less incentive to be exploitative than small
majorities. For example, if there are fixed costs to inter-group transfers, there is a critical size of
minority below which the minority is not worth exploiting. The circumstances in which one or
more ethnic groups are permanently politically exploited by a dominant group thus depend upon
the largest group having a sufficiently large share of the population to control the political
process, but not being so large that exploitation is not worthwhile. The precise range of
population shares over which the largest ethnic group exercises such dominance cannot be
determined a priori, although simple models would bound the minimum at 50% and the
maximum at strictly less than 100%. The grievance model postulates that a dummy variable for
societies characterized by such ethnic dominance, D2, will be significant and positive, increasing
the risk of conflict. A possible example of grievance-rebellion driven by an excluded ethnic
minority might be the conflict in Sri Lanka, where the Tamil Tigers claim to defend the interests
of the 12% of the population who are Tamil.
A second circumstance of political exclusion is where the poor are marginalized from the
political process. As Hirshleifer (1991) shows, normally the poor will succeed in using the
political contest to ameliorate their economic position. A high degree of economic inequality, i,
is therefore some indication that the poor are atypically politically marginalized. The `rage of the
poor’ at high inequality is indeed probably the single most popular explanation for conflict after
that of inter-ethnic hatred, and may be exemplified by the Castro rebellion in Cuba. A final
circumstance of persistent political exclusion is where a rich minority is heavily taxed by the
majority. The rich may contest the government or, following Buchanan and Faith (1987), they
might attempt to secede from the fiscal jurisdiction of the state. A possible example of the
former might be the Contra rebellion in Nicaragua, and of the latter, the eventually successful
secession of the Eritrean region from the much poorer state of Ethiopia. Since there is more
incentive for the majority to tax the rich the larger is the share of income or wealth accruing to
them, the `rage of the rich’ might also be expected to increase with economic inequality.
Hence, grievance due to political repression is postulated to be a function of the general level of
political rights, the ethnic composition of the society, and the degree of inequality:
j = j(q, D2, i) (18)
13
Finally, much of the case study literature dwells upon history: current rebellion is motivated by
the desire to revenge atrocities committed during a previous conflict. Corresponding to the
greed theory of induced conflict, we assume that the longer the period since a previous conflict,
τ, the less strong may be the demand for grievance assuagement for such atrocities. Thus, the
intensity of grievance, G, is assumed to be a function of these three components:
G = G(h, j, τ). (19)
Now consider the second stage in protest-rebellion, the failure of the initial protest to escalate
into successful political revolution. In the Kuran model the protest movement succeeds in
generating political change if, as people join the protest, the resulting reduction in the risk of
punishment attracts enough new entrants further to reduce the risk, yielding a cascade. The key
parameter is the elasticity of participation with respect to the risk of punishment. Kuran suggests
that this elasticity will be greater the more homogenous the society. The participation cascade
breaks if there are gaps in the distribution of preferences such as might occur if the society is
fractionalized. Hence, social fractionalization enters twice in a model of rebellion as aborted
protest. It is both an impetus for the initial cadre of protestors, and an impediment to political
revolution. If rebellions are aborted protests they should therefore be significantly more common
in fractionalized societies.
However, unlike a protest, a rebellion which generates a civil war is a sustained, full-time effort.
Adherents must eat, and so, as with a greed-rebellion, the rebel organization is constrained to
be financially viable. Here the grievance-rebellion faces a severe test. Grievance-assuagement is
predominantly a public good and evidently it is one which will not be provided by the
government. Its financing will therefore face acute difficulties of free-riding. The willingness of
non-participants to support the rebel organization financially may be affected by their own level
of grievance, so our grievance proxies might affect the risk of conflict both through generating
hard core adherents, and by increasing the scope for funding. A further possible source of
grievance-related funding is diasporas. There is some evidence that diasporas are more
grievance-conscious than the populations from which they originated. Hence, it is possible that
diasporas are willing to finance the initiation of rebellion even when local populations are not
sufficiently concerned to do so themselves. We have already discussed how this proposition can
be tested. However, the collective action problem is so severe in the temporary provision of
public goods by non-governmental means, that their supply will usually be negligible. Thus, the
principle prediction of an economic theory of grievance rebellion is that such a rebellion will not
occur, regardless of the intensity of the grievance. We test this prediction using the grievance
proxies in (19).
2.3 Greed-Grievance Interactions
We now compare the greed and grievance models of conflict and consider a synthesis.
14
The greed model postulates that the cause of initial conflict is an economic calculus of relative
military advantage, the government’s ability to finance defense expenditure, the scale of primary
commodity exports, and the costs of rebel recruitment. If a conflict occurs, the rebel
organization will generate subjective grievance. Post-conflict, until this gradually decays, it will
increase the risk of subsequent conflict. The extent of this post-conflict risk depends upon the
size of diasporas, since they are able to finance rebellion. By contrast, the grievance model
postulates that the cause of initial conflict is not an economic calculus but rather is a protest
generated by objective grievances: ethnic or religious hatreds, inequality, oppression, or
historical vengeance. Rebellions are protests which fail to cascade into non-violent revolution.
The scope for rebels to find a refuge, proxied by the three geography variables, can reasonably
be added to the grievance model without upsetting its essentially non-economic spirit.
Greed and grievance can co-habit. Where the conditions for greed-rebellion exist but those for
grievance-rebellion do not, a group initially motivated by grievance may become dependent
upon primary commodity predation for survival, thus transforming itself into a greed-rebellion.
Conversely, greed-rebellions need to manufacture subjective grievance for military cohesion and
may find an objective grievance an effective basis for generating it. Hence, the presence of
primary commodity exports may sustain rebellions which are motivated by objective grievance,
while the presence of objective grievance may sustain rebellions motivated by predation. Such
interdependence may make case study evidence difficult to interpret.
The two models are evidently not nested, relying largely upon distinct variables, and in Section 4
we compare their predictive power using standard tests. However, one variable, ethnic and
religious fractionalization, is included in both models with an opposite predicted sign and so is of
particular interest. The grievance model naturally sees ethnic and religious difference as a
potential source of hatreds as compared with a homogenous society, and as an impediment to
non-violent revolution. It thus predicts that the risk of conflict will be lower in homogenous
societies. The greed model sees fractionalization as tightening the constraint on rebellion: the
need for organizational cohesion imposes homogeneity and so if the society is diverse the costs
of rebel recruitment are raised.
An integrated model of conflict would allow both greed and grievance to initiate conflict.
Objective grievance could make recruits more willing to join, and so lower the costs of
rebellion. Potentially, diasporas could provide funding for initial rebellions, motivated by
objective grievance, rather than just financing subsequent conflicts motivated by induced
grievance, as proposed by the greed model. The greed and grievance models thus potentially
nest into an integrated model which combines both sets of variables as causes of initial, as well
as subsequent conflicts. Testing these models is the agenda for the rest of the paper.
3. Data
Our empirical analysis incorporates several new data sets constructed for the study. In this data
15
section we present a brief description of the war variable and the various proxies for greed and
grievance. The data source for the war variable and all other variables is Hoeffler and Sambanis
(2000) where a more detailed data description can be found.
3.1. War data
We analyze the risk of civil war using a panel data set for 161 countries. For each country we
have potentially one observation for each of the eight sub-periods 1960-64, 1965-69, ...,
1995-99. For 73 of these observations the country was at peace at the beginning of the period
but a civil war started during it. We predict only war-starts, not their continuation.4
Hence, once
civil war has commenced, the possibility for a further initiation of conflict only occurs once
peace has been re-established. In fact, most of the countries which had one conflict went on to
have further conflicts. Indeed, 47 of the 73 wars were in countries which had already had at
least one previous civil war since 1945. We define a civil war as an internal conflict in which at
least 1000 battle related deaths (civilian and military) occurred per year. This definition has
become standard in the literature following the seminal data collection exercise and analysis by
Singer and Small (1982, 1994). Table 1 lists the 73 observations for which a war started during
a sub-period. Our new war data set primarily updates that of Singer and Small from 1992 to the
end of the decade.
3.2. Rebel Revenue Sources
The three main sources of rebel revenue are primary commodity exports, diasporas and, during
the Cold War, foreign powers.
Primary commodity exports are measured as a ratio of GDP at the beginning of each subperiod.
This data was obtained from various World Bank sources.
In order to proxy the size of the diaspora we used US immigration data. The US Bureau of the
Census provides detailed figures on the size of the foreign born population.5
Inter-census figures
were interpolated. To capture the relative size of the population we divided the US immigration
figures by the total population of the country of origin.
The end of the Cold War was proxied by a dummy variable which took the value of unity for
the periods 1990-94 and 1995-99.
3.3 Relative Military Advantage
Military advantage reflects geography, rebel cohesion and rebel motivation.
4
In our earlier work (Collier and Hoeffler, 1998), we used a much smaller sample of wars and conflated war
starts with war duration, using a tobit procedure. We now regard this approach as flawed since duration
appears to be determined by rather different factors from starts.
5
Source: http://www.census.gov/population
16
Using geographic data we calculated a Gini coefficient measuring the dispersion of the
population in a country. Analogous to the income Gini coefficient, the Gini coefficient of
population concentration will be high if a large proportion of the country’s population is
concentrated in a relatively small area of the country. The more evenly the population is
dispersed across the country the lower is the Gini coefficient of population concentration. For
the calculation of this Gini coefficient we used population data per 400km2
cells.6
We also
experimented with two further measures of population dispersion, population density and the
proportion of the population living in urban areas, neither of which were significant. The data
source for both series is WDI 1998.
Data on forest coverage was obtained from the Food and Agriculture Organization (FAO)7
. For
each period this provided an estimate of the percentage of a country’s land area covered in
forests and woods. We could find no satisfactory existing data set on mountainous terrain. The
study commissioned a new data set from Dr. Gerrard, a specialist on the subject. The measure
allows for the ruggedness of the terrain, rather than simply relying upon altitude and is available
from the World Bank.8
We measure social fractionalization by ethno-linguistic and religious fractionalization. The ethnolinguistic
fractionalization data was obtained directly from the Atlas Naradov Mira (1964) which
lists the ethno-linguistic groups for each country. Based on the Altas Naradov Mira data an
ethno-linguistic fractionalization index can be calculated.9
For each country it measures the
probability that two randomly selected people do not speak the same language. An index of 0
means that the entire population speaks the same language while a higher index indicates a
higher degree of linguistic heterogeneity. The maximum value of the index is 100. Using the same
concept we also constructed a religious fractionalization index which measures the probability
that two randomly selected people do not share the same religious affiliation.10
The ethnolinguistic
data was measured at about 1960 and the religious data was measured in 1970 and
1980. Since these measures only change very slowly over time we use the linguistic data for all
sub-periods and the 1970 data on religion for the observations 1960-70 and the 1980 data for
1980-99. For 1975 we use the average of the 1970 and 1980 data. Social fractionalization can
potentially be equal to, additive or multiplicative in religious and ethnic fractionalization. For
example, if all the religious divisions coincide with some or all of the ethnic divisions, then the
overall fractionalization of the society may be no greater than that measured by ethnic
fractionalization. With such coincident divisions, social diversity would simply by the maximum
of the underlying ethnic and religious diversity measures. A second possibility is that the religious
6
We would like to thank Uwe Deichman (World Bank) for extracting the original data from the Geographic
Information System (GIS).
7
Source: http://www.fao.org/forestry
8
We would like to thank John Gerrard, University of Birmingham, for the compilation of this variable.
9
We would like to thank Tomila Lankina, University of Oxford, for translating the data entries from the Atlas
Naradov Mira. The first study to use this measure in the economics literature is Mauro (1995).
10
Most of the data was kindly made available by Robert Barro (Barro 1997) and for some countries we used
the data from the original source (Barrett 1982).
17
divisions occur within ethnic groups (or conversely). In this case, social diversity would be
approximated by the sum of the underlying measures. A third possibility is that cleavages are
cross-cutting, so that ethnic groups are divided by religion and religious groups by ethnicity. In
this case social diversity would be approximated by the product of the underlying measures.
More precisely, since the index for each measure ranges 0-100, cross-cutting cleavages would
be proxied by the product of the two indices plus the maximum of the two underlying indices. It
is necessary to add the maximum of the underlying indices to the product, since otherwise, a
society which is (say) ethnically diverse but homogenous in religion would appear to be as
homogenous as a society that was homogenous in both ethnicity and religion. In the analysis of
Section 4 we use only this third measure of social fractionalization, demonstrating in Table 8 that
it dominates both the other approaches to aggregating ethnic and religious diversity.
The time since the previous civil conflict is measured in months. If a country experienced a civil
war we measure the peace period as the number of months between the end of the civil war and
the beginning of the sub-period, 1960, 1965, ..., 1995. If a country never experienced a civil
war we measure the peace period since the end of World War II, i.e. in 1960 these countries
have 172 months of peace and they accumulate 60 additional peace months in every sub-
period.
3.4 Opportunity Cost
We use four measures of the opportunity cost of recruitment: secondary school enrolment rates
for men, average income, population growth and income growth.
Male schooling is measured by the gross male secondary school enrollment ratio. This was
obtained from the World Bank World Development Indicators (WDI) 1998. Gross male
enrollment ratios are defined as the ratio of total male enrollment, regardless of age, to the male
population of the age group that officially corresponds to the level of secondary schooling. Thus,
the gross enrollment ratio can exceed 100 percent.
Income is measured as the per capita GDP at the beginning of each period. In order to be able
to compare income data over time and across countries we mainly use the real PPP adjusted
figures as provided in the Penn World Tables Mark 5.6 (PWT). For the 1995 data we used the
growth rates from the real PPP adjusted GDP per capita data from the WDI 1998 in order to
update the PWT data which only provide data up to 1992.
Population growth is measured as the average annual growth rate of the population in the
previous five years. The main source of the population data is WDI 1998. Income growth is
measured as the average annual growth rate of real per capita GDP in the previous five years.
The income data used to calculate the growth rates is as described above. We combine
population growth and per capita income growth into a single variable, which is the weighted
difference between them. We experimented with weights, choosing that which gave the variable
the highest t-statistic, this being when the population growth rate was given three times the
18
weight of per capita income growth.
3.5 Sources of Grievances
Here we concentrate on two sources of grievances, political exclusion and inequality, since the
variables which proxy sources of hatred (social fractionalization) and vengeance (previous
conflicts) have already been discussed above.
Data on political exclusion is available in the Polity III data set (Jaggers and Gurr 1995). We
concentrate on the democracy variable which characterizes the general openness of political
institutions. The democracy score ranges from 0 to 10 where 10 denotes a highly open regime.
In addition we tried the autocracy score from the Polity III data set and a measure of political
openness published by Freedom House (“Gastil Index”).11
We constructed a measure of ethnic dominance based on the ethno-linguistic data from the
Atlas Naradov Mira. This dummy takes a value of one if the largest linguistic group makes up
between 45 and 90 percent of the total population and zero otherwise. We constructed a
number of different dummies to proxy ethnic domination. As reported below, we found that
when ethnic dominance was defined by this range the economic and statistical significance of the
variable was at a maximum.
Income inequality is measured as in Deininger and Squire (1996) either as the Gini coefficient of
income distribution or as the ratio of top quintile’s share of income to bottom quintile’s share.
We proxy land inequality by a Gini coefficient of land distribution, for a more detailed
description please refer to Deininger and Squire (1998).
In Table 2 we present some descriptive statistics for the main variables of interest, distinguishing
between the peace observations, the war observations and the entire sample. The war
observations are on average characterized by a higher proportion of previous civil wars, lower
opportunity costs of war (lower secondary school enrolment rates, lower per capita income,
lower economic growth and higher population growth) larger populations, higher ethnic
fractionalization, more mountainous terrain, a lower democracy score and higher income
inequality. However, war observations have a similar average share of primary commodity
exports in GDP, similar religious fractionalization, a similar incidence of ethnic dominance and
have a smaller diaspora than peace observations.
We now turn to the regression analysis to examine the relationship between these possible
causes of conflict and the risk of civil war.
4. Empirical Results
11
Source: http://www.freedomhouse.org/ratings
19
Our empirical analysis attempts to predict the risk that a civil war will start during a five-year
sub-period, through a logit regression in which the explanatory variables are characteristics at
the start of the sub-period.
We start with the greed model (see Table 3). Because per capita income and enrolment in
secondary schooling are highly correlated, they cannot be used in the same regression. The first
four columns include secondary schooling but not per capita income, which permits a sample of
688 episodes of which 43 are war observations. The final column replicates the core regression
using per capita income instead of secondary schooling, which permits a sample of 747
episodes of which 47 are war observations.
The first column omits the variables which proxy the effect of previous civil war. The two
opportunity cost proxies are significant with the expected signs. A higher gross secondary
school enrolment rate for males reduces the risk of war. The difference between income growth
per capita and population growth, both measured for the previous five year period, decreases
the risk of war. As predicted in (12), the effect of primary commodity exports on the risk of war
is significant but non-linear. The highest risk of war is at a share in GDP of about 25 percent.
The positive linear effect on risk suggests that there is a predatory element in rebellion. The
negative quadratic effect suggests that the increased tax revenue eventually augments the
capacity of the government to defend itself sufficiently to offset the attraction of enhanced loot.
The elasticity of the risk of conflict with respect to population is approximately unity, as
predicted. A higher degree of fractionalization makes societies significantly safer, high crosscutting
ethno-linguistic and religious fractionalization decrease the risk of civil war. Two of the
technology parameters, the dispersion of the population and mountainous terrain, are also
significant. More mountainous countries are more likely to experience a civil war, while a higher
concentration of the population decreases the risk of conflict.
We then introduce the effect of previous conflicts. We hypothesize that a previous civil war may
increase the risk of experiencing a civil war. Initially, we test this hypothesis by including a
dummy which equals one if the country experienced a civil war in the past. Here we include any
civil wars which occurred between the end of World War II and 1995. The coefficient on this
previous war dummy is highly significant, previous wars increase the risk of new civil war.
However, such a dummy could equally be picking up unobserved country-specific effects. We
then introduce the length of the peace period measured in months since the end of the last civil
war. In column 3 we add this peace duration variable to the model. The coefficient is negative
and significant, i.e. the longer the peace period the lower the risk of conflict. The coefficient on
the previous war dummy is now insignificant. Thus, the peace period is a more precise measure
of the effect than is the previous war dummy. This indicates that the risk decays after conflict, as
might grievance, rather than being a proxy for unobserved country-specific effects. In column 4
we therefore drop the insignificant previous war dummy, leaving a model in which greed causes
initial conflict; conflict causes some effect such as grievance; and this grievance causes further
20
conflict.12
In the last column we present an alternative specification in which we include the
average income per capita as a measure of opportunity costs instead of schooling. The results
are very similar, although the overall performance of the regression deteriorates.
In Table 4 we investigate whether the effect of post-conflict grievance on the risk of conflict is
reinforced by the diaspora, as suggested by the analytic model. Since the data set on the size of
diasporas in the USA reduces the number of countries on which we have data, the sample size
is consequentially radically reduced from the 747 observations and 47 wars which is our
maximum sample in the previous analysis. In order to preserve sample size we therefore retreat
to a more parsimonious version of the model, dropping four sample-constraining peripheral
explanatory variables: ethnic and religious fractionalization, geographic concentration of the
population, the extent to which the terrain is mountainous, and the rate of growth in the previous
five year period. The remaining explanatory variables are thus per capita GDP, primary
commodity exports, population, and the number of months since the previous conflict. Even with
these data-restoring deletions, the sample size is reduced to 595 observations (containing 29
war observations). However, all the included explanatory variables remain significant.
On this sample, we then test for the effect of a diaspora. We measure the size of the diaspora
relative to the resident population in the country of origin. To allow for a fading post-conflict
effect, we interact this measure with the number of months since the previous conflict, dividing
the former by the latter. This variable, dias/peace in Table 4, is added to the regression, the
result being shown in the first column. The variable is positive and significant, indeed, slightly
more significant than the number of months of peace by itself, although the latter also remains
significant. A large diaspora considerably increases the risk of further conflict. If we compare
the post-conflict society with the largest diaspora against that with the smallest, with other
variables at their mean values, after five years of peace the risk of renewed conflict is around six
times greater.
However, while this result is consistent with the analytic model, it is also open to a more
anodyne interpretation. Evidently, diasporas are to an extent endogenous to the intensity of
conflict: when civil war occurs, many people leave and settle in the USA. Hence, the size of the
diaspora might simply be proxying the intensity of conflict. The result may therefore simply show
that more severe conflicts have a higher risk of renewed conflict. To test for this we
decomposed observed diasporas into a component which is exogenous to the intensity of
conflict and a residual component which is therefore endogenous to its intensity. For this
decomposition we estimated a simple migration model, reported in the Appendix, Table A3.
The size of the diaspora in a census year is predicted to be a function of its size in the previous
12
We examined this time effect in more detail by including the natural logarithm of the peace variable or its
square, however, a linear decay term provides a better fit. Note that the measure of peace since the end of
the civil war is somewhat imprecise since we only measure it from the end of the war to the initial year of
each sub-period. A duration model of post-war peace would allow a more detailed analysis of this peace
effect, however, the duration model results in Collier, Hoeffler and Söderbom (1998) support the results
presented in this paper.
21
census, time, per capita income in the country of origin, and whether there was a war in the
intervening period. This model predicts the size of the diaspora with reasonable accuracy. We
then replace the diaspora data used in the model with estimated diaspora size in all cases where
the observed diaspora is for a year subsequent to a conflict. Thus, all post-conflict observations
of diasporas are estimates which are purged of any effect from the intensity of conflict. The
difference between these estimates and the actual figures are then used as an additional variable,
measuring that part of the diaspora which is potentially endogenous to the intensity of conflict.
Both of these measures are then introduced into the regression in place of the previous single
measure of the diaspora. The results are reported in the second column of Table 4. The purged
measure of the diaspora remains significant, and the size of the coefficient is only slightly altered.
Further, its coefficient is not significantly different from that on the endogenous diaspora
measure. Had the effect of the diaspora been simply a proxy for the intensity of conflict, neither
of these would have been the case. The purged variable would have been insignificant, and the
coefficient on the endogenous measure would have been larger. This suggests that the
substantial effect of the diaspora on the risk of conflict renewal is indeed due to its financial
contribution to war start-up.
In Table 5 we turn to the examination of a rebellion which is motivated only by grievance. In the
first column we examine the relationship between ethnic dominance, social fractionalization,
democracy and the duration of peace. As in the greed model, we control for geographic military
advantage by including the dispersion of the population and mountainous terrain. Since we are
not including any lagged variables we can use 884 observations of which 57 observations
experienced an outbreak of civil war. The results suggest that societies which are dominated by
one large ethnic group (45-90 percent of the population belonging to the same ethnic group) are
more likely to experience conflict. Greater openness of political institutions reduces the risk of
conflict. In the second column we add the gini coefficient of income inequality as an explanatory
variable and in column 3 we add the gini coefficient of land inequality as an alternative measure
of inequality. Neither measure is significant. Note that the sample size is reduced when we
include the income inequality measures. However, we are still analyzing a substantial number of
war occurrences (40 in column 2 and 38 in column 3).13
All three grievance models have very
low explanatory power, the regressions only have an R2
of 0.14 or lower.
We now turn to the question of which model, greed or grievance, provides a better explanation
of the risk of civil war. Since the two models are non-nested, i.e. one model is not a special
case of the other, we use the J-test as suggested by Davidson and MacKinnon (1981). This
non-nested test is based on the following artificial nesting procedure. First we explain the risk of
civil war, p, in terms of the two different models, greed and grievance.
(1) p=f(greed)
(2) p=f(grievance)
13
We also tried the ratio of the income shares of the top to the bottom quintiles. This was also insignificant.
22
Based on these logit regressions we calculate the predicted probabilities and add these
predicted values, pˆ greed
and pˆ grievance
to our alternative models.
(1) p=f(greed, pˆ grievance
)
(2) p=f(grievance, pˆ greed
)
According to the J-test the significance of the coefficients of these added variables enables us to
choose between the two different models. If pˆ grievance
is significant in the greed model we reject
the greed model in favor of the grievance model. If pˆ greed
is significant in the grievance model
we reject the grievance model in favor of the greed model. As can be seen in columns 2 and 3
of Table 6 pˆ grievance
is significant in the greed model and pˆ greed
is significant in the grievance
model. Thus, we conclude that we cannot choose between the two models, grievance adds
explanatory power to the greed model and vice versa.14
Since we find neither model dominates
the other, we next investigate the combination of the two models as presented in column 3 of
Table 6.
In this combined model all variables apart from democracy are significant and the overall fit is
reasonable (pseudo R2
of about 0.3). Since both the grievance and greed models are nested in
the combined model, we can use a likelihood ratio test to determine whether the combined
model is superior. We can reject the validity of the restrictions proposed by both the greed and
the grievance models.15
Although the combined model is superior to the greed and grievance models, one variable,
democracy, is completely insignificant.16
Dropping it yields the model in column 4. Without the
democracy variable the coefficients and standard errors of all of the other variables remain
virtually unchanged. Based on a likelihood ratio test we cannot reject the hypothesis of a zero
coefficient on the democracy variable.17
While all variables in the model of column 4 are
significant, that for mountainous terrain is only significant at the ten percent level and in column 5
we investigate a more parsimonious model in which it is dropped. Since this model is nested
within the two previously presented models, we again use likelihood ratio tests to compare it to
those of columns 3 and 4. The tests narrowly favor this model over a model which includes
14
When we tested an alternative specification of the greed model which includes income rather than
schooling, we found that grievance did not seem to add explanatory power to the greed model, i.e. the
coefficient pˆ grievance
was insignificant.
15
Using the same sample as for the combined model (n=665) we obtain the following results: Greed model
versus combined model, 2 degrees of freedom, Likelihood Ratio Test statistic 6.34, p-value 0.042; grievance
model versus combined model, 5 degrees of freedom, LRT statistic 41.76, p-value 0.000.
16
We tried different specifications to test for the effect of political repression by including the autocracy
score instead of the democracy score and by using the difference between the two variables as suggested
by Londregan and Poole (1996). We also tried the Freedom House measure of political freedom, but neither
of these alternative political repression measures were found to be significant.
17
Using the same sample (n=665) we obtained a log likelihood of –111.25 for the restricted model. Thus, the
LRT statistic is 2.01, 1 degree of freedom, p-value 0.16.
23
mountainous terrain.18
No further reduction in the model is accepted, and no additions of
variables included in our previous models are accepted. We refer to the model in column 5 as
our baseline model.
Our baseline model allows us to calculate the change in the probability of war-starts for different
values of the explanatory variables. We present these calculations in the Appendix, Table A5.
At the median of all variables a country would have a 1.4 percent probability of experiencing a
civil war, while at the mean of all variables the risk of a war-start is higher, about 14 percent.
Our model predicts that a country with the worst characteristics (lowest per capita income,
lowest GDP growth and highest population growth, a primary commodity export share of 0.26,
the largest population, the lowest fractionalization, ethnic dominance, a geography Gini
coefficient of zero and only one month of peace) would have a near-certain risk of war while a
country with the best characteristics would be a very safe society. We now calculate how the
change in one variable (while the others are assumed to take their mean values) affects the
probability of experiencing a civil war. Whether a country is characterized by ethnic dominance
has a considerable impact on the risk of experiencing a civil war. Countries without a dominant
ethnic group have a risk of war of about 10 percent, however, the risk is doubled if countries
have a dominant ethnic group. If the male secondary school enrollment rate is 10 percentage
points higher than the average, the risk of war is reduced by about four percentage points (a
decline in the risk from 14% to 10%). Higher income growth reduces the risk of conflict: an
additional percentage point on the growth rate reduces the risk of war by one percentage point
(a decline from 14% to 13%). A reduction in population growth has a larger effect on the risk of
war, a one percentage point decrease implies a 2.5 percentage point reduction in the risk of
conflict (from 14% to 11.5%). The effect of primary commodity exports is very considerable: at
the risk maximizing value of the primary commodity export share (0.26), the risk of civil war is
about 23 percent, while a country with no natural resource exports only has a probability of a
war-start of 0.5 percent. A society characterized by the maximum social fractionalization has a
very small risk of experiencing a civil war (about three percent), however, the risk is about eight
times higher for a country which is socially homogenous. Directly after a civil war there is a high
probability of a re-start, the risk is about 42 percent. This risk declines over time, however,
sustained peace only slowly decreases the risk of experiencing a new conflict. Five years after
the end of the civil war the risk is still about 37 percent. The elasticity of the risk of conflict with
respect to population is 0.95, close to the analytic expectation of 0.94 discussed in Section 2.
Lastly, we examine the impact of the dispersion of the population on the risk of conflict.
Countries with a highly concentrated population are very safe from conflict while countries
which are characterized by a homogenously dispersed population have a very high risk of civil
war (about 60 percent).
In column 6 of Table 6 we run the baseline model with average per capita income instead of
18
First, using the same sample (n=665) we test the combined model (column 3) versus the baseline model
(column 5). We obtain a LRT statistic of 3.24, 2 degrees of freedom, p-value 0.198. Second, we test the
reduced model (column 4) against the baseline model (column 5). We obtain a LRT statistic of 2.03, 1 degree
of freedom, p-value 0.16.
24
schooling. Recall that since we have more data on income than on schooling we are able to use
a larger sample which includes four additional war observations. The variables are all significant
at conventional levels, although the overall explanatory power is slightly lower. Using this
maximum sample we analyze the effect of external assistance as causes of war. Adding a post
cold-war dummy (which takes a value of one for the periods 1990-94 and 1995-99) in column
7 we find that that the risk of civil war was significantly higher during the cold war period. This is
consistent with the prediction that the super-powers eased the financing constraint on
rebellion.19
Thus, we find some weak evidence that external assistance increased the risk of civil
war during the cold war.
In Table 7 we investigate a number of different estimation issues. We concentrate on the
analysis of random effects, fixed effects, time effects and a correction for rare events.
We re-estimated our models using random effects. For the baseline model, and indeed, all those
models which include a feedback effect from previous conflict, we accept the hypothesis that
we can pool across the observations.20
The estimation of fixed effects logits was only possible
on a very small sub-sample of the observations. The countries for which the dependent variable
does not vary over time (the majority of countries experienced only peace) cannot be included
in the analysis. Although the fixed effects test is very severe, the non-monotonic effect of
primary commodity exports remains significant. Were the effect of primary commodity exports
dependent only upon cross-section data, it might suggest that the variable was proxying some
other characteristic such as geography. However, the fixed effects regression uses only changes
in primary commodity dependence, and so reduces the scope for alternative interpretations.
We analyzed whether time effects matter by including time dummies in the model. Based on a
log likelihood ratio test we cannot reject the hypothesis that the time dummies are zero.21
However, as discussed above, the dummies for 1990-94 and 1995-99 are jointly significant
and support consolidation into a single dummy for the post Cold War period in the variant of the
baseline model.
Finally, in the last column of Table 7 we use a recently developed correction method for rare
events data (King and Zeng, 2000). Our data is characterized by a relatively small number of
events (wars), only about six percent of the observations are characterized by a civil war
outbreak. The results presented in King and Zeng (2000) suggest that standard logit estimation
underestimates the probability of an event occurring when the events are rare. We used their
correction procedure but found the differences between the standard logit results and the rare
19
This result only holds for the augmented sample which includes 47 civil wars. Due to data constraints the
baseline model excludes four of these wars all of which started during the cold war period. The wars we have
to exclude are Chad (1980-84), China (1965-69), Jordan (1970-74) and Liberia (1985-89).
20
The null-hypothesis (rho=0) is not rejected: the chi-squared statistic is 0 (p = 0.998). When the feedback
effect from previous conflict is excluded, pooling is marginally rejected: the chi-squared statistic is 2.02
(p=0.155).
21
The LRT statistic is 2.1, 6 restrictions, p=0.91.
25
events corrected results to be negligible. Using the rare events logit procedure, all coefficients on
the variables have the same signs and are significant at the same levels. The mean of the
predicted probabilities obtained from the standard logit regression is 0.063 and the predicted
probabilities from the rare events logit regression have a mean of 0.068. The standard
deviations and the extreme values of the predicted probabilities obtained from the standard logit
and the rare events logit were also similar. Thus, contrary to King and Zeng (2000), we do not
find that the logit substantially underestimates the probability of the event occurring.
We examined a number of different model specifications. We found that none of the following
geographic and demographic characteristics were significant: forest coverage, population
density, the proportion of the population living in urban areas and the proportion of young men
aged 15 to 29. In Tables 8 and 9 we present a more detailed analysis of ethnic fractionalization.
In Table 8 we demonstrate that the measure of cross-cutting cleavages (the product of religious
and ethnic fractionalization plus the maximum of each component separately), dominates the
other possible aggregation procedures for ethnic and religious diversity. When this measure of
fractionalization is included with the ethnic and religious diversity indices either together or
individually, it is significant whereas the underlying indices are not significant. In Table 9 we
investigate alternative definitions of ethnic dominance. In the first column we include the ratio of
the largest ethnic group to the total population in the model and find its coefficient to be
insignificant. We then construct a number of different ethnic dominance dummies, taking the
value one if 45-90, 45-85, 45-95, 30-90, 40-90, or 50-90 percent of the total population
belong to the same ethnic group. As can be seen, the relationship between ethnic dominance is
most significant when the largest ethnic group makes up 45-90 percent of the total population.22
5. Conclusion
We started from two simple models of rebellion. Greed-rebellion was motivated by predation of
the rents from primary commodity exports, subject to an economic calculus of costs and a
military survival constraint. Grievance-rebellion was motivated by hatreds which might be
intrinsic to ethnic and religious differences, or reflected objective resentments such as
domination by an ethnic majority, political repression, or economic inequality. Both of these
models could allow for technological constraints upon rebellion such as differences in
geography. We then introduced the possibility of a feedback effect, whereby conflict increased
the risk of subsequent conflict through generating grievance. In grievance-rebellion this added a
simple destabilizing loop: grievance causes conflict which then causes further grievance. In
greed-rebellion the feedback was more complex because induced grievance only raised the risk
of further conflict if it augmented the potential financial rewards to rebellion. We postulated that
such a loop might arise if grievance enabled rebellions to tap into the resources of diasporas.
Finally, we allowed for interdependence between greed and grievance. Greed-rebellions need
to generate grievance for military cohesion, grievance rebellions might be driven to predation to
raise finance.
22
We also experimented with various other definitions of ethnic dominance, such as the dummy taking a
value of one if the largest ethnic group made up 50-75, 50-80, 50-85, 50-90 percent of the population. Neither
of these definitions provided a better fit than our preferred definition based on 45-90 percent.
26
Using a comprehensive panel data set of conflict over the period 1960-99 we examine the risk
of civil war using logit regressions. Starting with the distinct greed and grievance models, we find
that most of the proxies for objective grievance are insignificant and that the best-performing
grievance model has very low explanatory power. By contrast, the simple greed model
performs well. The extent of primary commodity exports is the largest single influence on the
risk of conflict. Strikingly, ethnic and religious fractionalization makes a society safer, as
predicted by the greed model, rather than more dangerous as predicted by the grievance model.
We find that a war substantially increases the risk of subsequent war, although this risk fades
over time. Thus, although objective grievance is not a powerful primary cause of conflict,
conflict may generate grievances which become powerful additional risk factors. However, we
can show that by far the strongest effect of war on the risk of subsequent war works through
diasporas. After five years of post-conflict peace, the risk of renewed conflict is around six
times higher in the societies with the largest diasporas in America than in those without American
diasporas. Presumably this effect works through the financial contributions of diasporas to rebel
organisations.
Finally, we test for an integrated greed-grievance model. We find that only one of the potential
sources of objective grievance significantly adds to the explanatory power of the greed model,
namely ethnic dominance. Societies in which the largest ethnic group has between 45% and
90% of the population have around double the risk of conflict of other societies. Presumably,
this is because such societies have both the power and the incentive to exploit their minorities.
Our results thus contrast with conventional beliefs about the causes of conflict. A stylized
version of these beliefs would be that grievance begets conflict which begets grievance which
begets further conflict. With such an analysis, the only point of intervention is to reduce the level
of objective grievance. Our model suggests that what is actually happening is that opportunities
for primary commodity predation cause conflict, and that the grievances which this generates
induce diasporas to finance further conflict. The policy intervention points here are reducing the
absolute and relative attraction of primary commodity predation, and reducing the ability of
diasporas to fund rebel movements.
27
References
Barro, R. J. 1997. Determinants of Economic Growth. Cambridge, MA: MIT Press.
Barrett, D. B., ed. 1982. World Christian Encyclopedia. Oxford: Oxford University Press.
Buchanan, J.M. and R.L. Faith. 1987. Secession and the Limits of Taxation: Towards a Theory
of Internal Exit, American Economic Review, 77:1023-31.
Collier, P. 2000. Rebellion as a Quasi-Criminal Activity, Journal of Conflict Resolution 44,
(forthcoming).
Collier, P. and A. E. Hoeffler. 1998. On the Economic Causes of Civil War, Oxford
Economic Papers, 50:563-73.
Collier, P., A. E. Hoeffler, and M. Söderbom. 1998. On the Duration of Civil War and
Postwar Peace. Centre for the Study of African Economies, Discussion Paper 98/8.
Davidson R. and J.G. MacKinnon. 1981. Several Tests for model specification in the presence
of alternative hypotheses. Econometrica 49:781-793.
Deininger, K., and L. Squire. 1996. A New Data Set Measuring Income Inequality. World
Bank Economic Review 10:565-91.
Deininger, K. and Squire, L. 1998. New Ways of Looking at Old Issues: Inequality and
Growth. Journal of Development Economics 57:249-287.
Department of Geodesy and Cartography of the State Geological Committee of the USSR.
1964. Atlas Naradov Mira. Moscow.
Gerrard, A.J.W. 2000. ‘What is a Mountain?’ Mimeo. DECRG, World Bank.
Gleditsch, Nils Petter, and Havard Hegre. 1997. Peace and Democracy: Three Levels of
Analysis, Journal of Conflict Resolution, 41:283-310.
Grossman, H. I. 1999. Kleptrocracy and Revolutions. Oxford Economic Papers 51:267-283.
Herbst, Jeffrey. 2000. States and Power in Africa, Princeton, Princeton University Press.
Hirshleifer, J. 1991. ‘The Paradox of Power’. Economics and Politics 3:171-200.
Hoeffler, A. E. and N. Sambanis. 2000. Data Notes - World Bank Project on ‘The Economics
of Civil Wars’. Mimeo.
28
Kaplan, Robert, D. 2000. The Coming Anarchy, New York, Random House.
Keith Jaggers and Ted Robert Gurr. 1995. ‘Tracking Democracy's Third Wave with the Polity
III Data’, Journal of Peace Research 32:469-482.
King, G., and L. Zeng. 2000. Logistic Regression in Rare Events Data.
http://gking.harvard.edu/preprints.shtml
Konrad, Kai, A. and Stergios Skaperdas. 1998. ‘Extortion’, Economica, 65:461-77.
Kuran, T. 1989. ‘Sparks and Prairie Fires: A Theory of Unanticipated Political Revolution’.
Public Choice 61:41-74.
Londregan, J.B. and K. T. Poole. 1996. ‘Does High Income Promote Democracy?’ World
Politics 49:1-30.
Mauro, P. 1995. ‘Corruption and Growth’. The Quarterly Journal of Economics 110:681-
712.
Sachs, J.-D. A.M. Warner. 1995. ‘Natural Resource Abundance and Economic Growth’.
National Bureau of Economic Research Working Paper: 5398.
The Stockholm International Peace Research Institute. 1993. Yearbook of World Armaments
and Disarmaments. Oxford: Oxford University Press.
29
Table 1: Outbreaks of War
country year war prev. country year war prev.
startedwar startedwar
Afghanistan 1975-79 1 0 Mozambique 1960-64 1 0
Afghanistan 1990-94 1 1 Mozambique 1975-79 1 1
Algeria 1960-64 1 1 Myanmar/Burma 1965-69 1 1
Algeria 1990-94 1 1 Myanmar/Burma 1980-84 1 1
Angola 1960-64 1 1 Nicaragua 1975-79 1 0
Angola 1975-79 1 1 Nicaragua 1980-84 1 1
Azerbaijan 1990-94 1 0 Nigeria 1965-69 1 0
Bosnia 1990-94 1 0 Nigeria 1980-84 1 1
Burundi 1970-74 1 0 Pakistan 1970-74 1 0
Burundi 1984-89 1 1 Peru 1980-84 1 0
Burundi 1990-94 1 1 Philippines 1970-74 1 1
Cambodia 1970-74 1 1 Romania 1984-89 1 0
Chad 1980-84 1 0 Russia 1990-94 1 0
China 1965-69 1 1 Russia 1995-99 1 1
Colombia 1980-84 1 1 Rwanda 1960-64 1 1
Dominican Republic 1965-69 1 0 Rwanda 1990-94 1 1
El Salvador 1975-79 1 0 Somalia 1980-84 1 0
Ethiopia 1970-74 1 1 Somalia 1984-89 1 1
Georgia 1990-94 1 0 Sri Lanka 1970-74 1 0
Guatemala 1965-69 1 1 Sri Lanka 1980-84 1 1
Guatemala 1970-74 1 1 Sudan 1960-64 1 0
Guatemala 1975-79 1 1 Sudan 1980-84 1 1
Guinea-Bissau 1960-64 1 0 Tajikistan 1990-94 1 0
India 1980-84 1 1 Turkey 1990-94 1 0
Indonesia 1975-79 1 1 Uganda 1965-69 1 1
Iran 1970-74 1 1 Uganda 1980-84 1 1
Iran 1975-79 1 1 Vietnam 1960-64 1 1
Iran 1980-84 1 1 Yemen 1990-94 1 0
Iraq 1970-74 1 1 Yemen AR 1960-64 1 1
Iraq 1984-89 1 1 Yemen PR 1984-89 1 0
Iraq 1990-94 1 1 Yugoslavia 1990-94 1 0
Jordan 1970-74 1 0 Yugoslavia 1995-99 1 1
Laos 1960-64 1 1 Zaire 1960-64 1 1
Lebanon 1975-79 1 1 Zaire 1990-94 1 1
Liberia 1984-89 1 0 Zaire 1995-99 1 1
Liberia 1990-94 1 1 Zimbabwe 1970-74 1 0
Morocco 1975-79 1 1
30
Table 2: Descriptive Statistics
sample (n=1174) no civil war (n=1101) civil war (n=73)
mean st.dev. min max n mean st.dev. min max n mean st.dev. min max n
war starts 0.06 1174 0 1101 1 73
previous war 0.34 1174 0.32 1101 0.64 73
male secondary
schooling
43.3 31 0.3 147 965 44.3 31 0.3 147 902 30.2 26.6 1 102 63
GDP per capita
(const. US$)
4056 4317 222 33946 1014 4200 4388 257 33946 958 1644 1353 222 5832 56
GDP per capita
growth t-1
1.62 3.83 -
22.08
14.41 864 1.72 3.78 -22.1 14.41 815 -0.05 4.28 -10.66 7.09 49
population growth
t-1
2.15 1.54 -6.38 16.62 991 2.12 1.57 -6.38 16.62 929 2.50 1.01 -0.013 5.81 62
primary commodity
exports/GDP
0.17 0.19 0 2.14 1090 0.17 0.19 0 2.14 1023 0.14 0.1 0.01 0.48 67
population
(millions)
25.4 94.5 0.014 1200 1161 24.5 92.9 0.014 1200 1089 40.0 116.4 0.54 715 72
ethnic
fractionalization
40 28 0 93 1107 39 28 0 93 1039 50 29 4 90 68
religious
fractionalization
36 24 0 79 1147 36 24 0 79 1078 36 25 0 69 69
geographic
dispersion
0.57 0.23 0 0.97 1028 0.57 0.23 0 0.97 961 0.60 0.15 0.24 0.92 67
mountainous
terrain
15.8 20.73 0 94.3 1174 15.12 20.28 0 94.3 1101 26.73 24.2 0 81 73
ethnic dominance
(45-90%)
0.47 0 1 1107 0.47 1039 0.47 68
democracy 3.90 4.21 0 10 987 4.21 4.21 0 10 987 1.67 2.58 0 9 67
diaspora 0.02 0.03 0 0.19 654 0.02 0.03 0 0.19 654 0.004 0.005 0 0.16 37
Income Inequality
(Gini)
41.35 10.27 17.83 63.2 721 40.65 10.27 17.83 63.2 721 41.97 10.30 23.38 62.3 44
31
Table 3: Greed Model
1 2 3 4 5
male secondary schooling -0.036
(0.011)***
-0.037
(0.011)***
-0.028
(0.012)**
-0.027
(0.012)**
ln GDP per capita -0.822
(0.288)***
(GDP growth - 3*population
growth) t-1
-0.080
(0.036)**
-0.074
(0.036)**
-0.084
(0.036)**
-0.086
(0.036)**
-0.101
(0.035)***
primary commodity
exports/GDP
34.088
(8.186)***
34.109
(8.423)***
32.147
(8.497)***
32.058
(8.465)***
22.923
(6.903)***
(primary commodity
exports/GDP)2
-67.792
(18.683)***
-67.172
(18.679)***
-62.307
(18.954)***
-62.091
(18.962)***
039.192
(14.710)***
ln population 1.026
(0.191)***
0.960
(0.194)***
0.832
(0.198)***
0.832
(0.197)***
0.574
(0.146)***
social fractionalization -0.0002
(0.0001)**
-0.0003
(0.0001)***
-0.0003
(0.0001)***
-0.0003
(0.0001)***
-0.0003
(0.0001)***
geographic dispersion -3.517
(1.142)***
-3.888
(1.181)***
-3.384
(1.200)***
-3.289
(1.180)***
-1.568
(1.051)
mountainous terrain 0.019
(0.009)**
0.013
(0.009)
0.014
(0.010)
0.015
(0.010)
0.012
(0.009)
previous war dummy 1.016
(0.405)***
0.252
(0.527)
Peace duration -0.004
(0.002)**
-0.004
(0.001)***
-0.004
(0.001)***
n 691 691 688 688 747
no of wars 43 43 43 43 47
pseudo R2
0.25 0.27 0.29 0.29 0.26
log likelihood -120.98 -117.63 -114.32 -114.43 -129.44
Notes: All regressions include a constant. Standard errors in parentheses.
***, **, * indicate significance at the 1, 5 and 10 percent level, respectively.
32
Table 4: Greed – Diaspora Model
1 2
ln GDP per capita -1.032
(0.285)***
-1.037
(0.286)***
primary commodity
exports/GDP
23.98
(8.542)***
23.78
(8.56)***
(primary commodity
exports/GDP)2
-48.163
(22.1)***
-47.75
(22.12)**
ln population 0.318
(1.156)**
0.319
(1.155)**
peace duration -0.002
(0.001)***
-0.002
(0.001)*
diaspora/peace duration 648.77
(367.57)*
predicted diaspora/peace 687.09
(393.12)*
(diaspora-predicted
diaspora)/peace duration
768.9
(562.42)
n 595 595
no of wars 29 29
pseudo R2
0.23 0.23
log likelihood -89.05 -89.01
Notes: All regressions include a constant. Standard errors in parentheses.
***, **, * indicate significance at the 1, 5 and 10 percent level, respectively.
33
Table 5: Grievance Model
1 2 3
social fractionalization 0.00003
(0.00007)
0.00000
(0.00009)
0.00008
(0.00009)
ethnic
dominance (45-90%)
0.263
(0.290)
0.534
(0.342)
0.567
(0.374)
democracy -0.117
(0.046)***
-0.091
(0.051)**
-0.138
(0.052)***
geographic dispersion 0.095
(0.787)
-0.276
(0.958)
-0.152
(1.039)
mountainous terrain 0.015
(0.006)**
0.011
(0.008)
0.015
(0.006)**
income inequality 0.014
(0.016)
land inequality -0.150
(1.159)
Peace duration -0.005
(0.001)***
-0.004
(0.001)***
-0.005
(0.001)***
n 884 614 620
no of wars 57 40 38
pseudo R2
0.13 0.11 0.14
log likelihood -184.17 -132.11 -122.61
Notes: All regressions include a constant. Standard errors in parentheses. ***, **, * indicate significance
at the 1, 5 and 10 percent level, respectively.
34
Table 6: Combined Greed and Grievance Model
1 2 3 4 5 6 7
male secondary schooling -0.026
(0.012)**
-0.030
(0.012)***
-0.031
(0.012)***
-0.036
(0.011)***
ln GDP per capita -1.007
(0.281)***
-1.006
(0.283)***
(GDP growth - 3*population
growth) t-1
-0.079
(0.036)**
-0.085
(0.037)**
-0.090
(0.037)**
-0.085
(0.036)**
-0.103
(0.035)***
-0.113
(0.035)***
primary commodity
exports/GDP
31.902
(8.333)***
33.576
(8.514)***
34.220
(8.610)***
32.089
(8.375)***
22.983
(6.806)***
23.025
(6.795)***
(primary commodity
exports/GDP)2
-61.491
(18.489)***
-64.234
(18.873)***
-65.676
(19.186)***
-62.511
(18.964)***
-39.293
(14.505)***
-39.026
(14.394)***
ln population 0.840
(0.203)***
0.912
(0.211)***
0.934
(0.210)***
0.946
(0.204)***
0.625
(0.148)***
0.678
(0.153)***
social fractionalization -0.0003
(0.0001)**
0.00003
(0.0001)
-0.0003
(0.0001)**
-0.0003
(0.0001)**
-0.0003
(0.0001)***
-0.0004
(0.0001)***
-0.0003
(0.0001)***
democracy -0.042
(0.058)
-0.043
(0.058)
ethnic
dominance (45-90%)
0.895
(0.373)**
0.909
(0.477)***
0.882
(0.374)***
0.847
(0.370)**
0.623
(0.348)*
0.676
(0.351)**
geographic dispersion -2.443
(1.270)**
0.023
(1.016)
-3.670
(1.208)***
-3.600
(1.200)***
-3.891
(1.172)***
-1.851
(1.006)*
-2.020
(1.023)*
mountainous terrain 0.003
(0.011)
0.0004
(0.009)
0.016
(0.010)*
0.016
(0.010)*
peace duration 0.001
(0.003)
0.0004
(0.0015)
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
post-cold war dummy -0.980
(0.499)**
grievance residual 0.897
(0.481)*
greed residual 1.066
(0.200)***
n 665 665 665 688 688 747 747
no of wars 43 43 43 43 43 47 47
pseudo R2
0.30 0.30 0.30 0.31 0.30 0.27 0.28
log likelihood -112.27 -111.02 -110.97 -111.61 -113.03 -128.71 -126.47
Notes: All regressions include a constant. Standard errors in parentheses. ***, **, * indicate significance at the 1, 5 and 10 percent level, respectively.
35
Table 7: Robustness Tests of the Baseline Model
1 2 3 4
Random Effects Fixed Effects Pooled Logit plus
Time dummies
Rare Events Logit
male secondary schooling -0.036
(0.011)***
0.0001
(0.033)
-0.034
(0.012)***
-0.034
(0.011)***
(GDP growth -
3*population growth) t-1
-0.085
(0.036)**
-0.063
(0.067)
-0.105
(0.041)***
-0.080
(0.036)**
primary commodity
exports/GDP
32.089
(8.375)***
34.885
(14.945)**
32.978
(8.648)***
29.399
(8.667)***
(primary commodity
exports/GDP)2
-62.511
(18.964)***
-63.363
(28.824)**
-65.483
(19.822)***
-55.982
(22.781)***
ln population 0.946
(0.204)***
0.373
(1.374)
0.960
(0.210)***
0.891
(0.155)***
social fractionalization -0.0003
(0.0001)***
-0.007
(0.005)
-0.0003
(0.0001)***
-0.0003
(0.0001)***
ethnic
dominance (45-90%)
0.847
(0.370)**
0.913
(0.381)**
0.807
(0.350)**
geographic dispersion -3.891
(1.172)***
78.555
(80.216)
-3.936
(1.196)***
-3.722
(1.041)***
Peace duration -0.004
(0.001)***
0.01
(0.003)***
-0.004
(0.001)***
-0.004
(0.001)***
T70-74 -0.960
(0.841)
T75-79 0.145
(0.776)
T80-84 -0.002
(0.777)
T85-89 0.477
(0.749)
T90-94 -0.820
(0.855)
T95-99 -0.229
(0.168)
n 688 135 668 688
no of wars 43 43 43 43
pseudo R2
0.33
log likelihood -113.03 -112.27 -107.44 -113.03
Notes: All regressions include a constant. Standard errors in parentheses. ***, **, * indicate significance
at the 1, 5 and 10 percent level, respectively.
36
Table 8: Social Fractionalization
1 2 3 4 5 6 7 8 9 10
Baseline
Model
Baseline Model
plus ethnic
fractionalizatio
n
ethnic
fractionalizatio
n only
Baseline Model
plus religious
fractionalizatio
n
religious
fractionalizatio
n only
ethnic and
religious
fractionalizatio
n
Baseline Model
plus ethnic and
religious
fractionalizatio
n
social frac.=
ethnic plus
religious
fractionalizatio
n
social frac.=
ethnic minus
religious
fractionalizatio
n
social frac.=
ethnic times
religious
fractionalizatio
n
male secondary
schooling
-0.036
(0.011)***
0.037
(0.012)***
-0.025
(0.001)***
-0.037
(0.012)***
-0.034
(0.011)***
-0.036
(0.113)***
-0.037
(0.116)***
-0.034
(0.011)***
-0.021
(0.010)**
-0.036
(0.011)***
(GDP growth -
3*population
growth) t-1
-0.085
(0.036)**
-0.083
(0.037)**
-0.093
(0.035)***
-0.084
(0.037)**
-0.078
(0.036)**
-0.081
(0.036)**
-0.084
(0.037)**
-0.086
(0.036)**
-0.082
(0.036)**
-0.085
(0.036)**
primary
commodity
exports/GDP
32.089
(8.375)***
32.700
(8.425)***
23.442
(7.213)***
32.019
(8.390)***
28.368
(7.872)***
29.447
(8.060)***
32.862
(8.454)***
28.809
(8.003)***
21.120
(6.822)***
32.147
(8.381)***
(primary
commodity
exports/GDP)2
-62.511
(18.964)***
-64.283
(19.044)***
-44.962
(16.501)***
-62.791
(19.035)***
-56.891
(18.260)***
-58.464
(18.633)***
-64.450
(19.020)***
-56.049
(18.491)***
-42.019
(15.923)***
-62.621
(18.970)***
ln population 0.946
(0.204)***
0.925
(0.206)***
0.749
(0.177)***
0.941
(0.205)***
0.828
(0.186)***
0.892
(0.201)***
0.924
(0.206)***
0.904
(0.199)***
0.643
(0.164)***
0.946
(0.204)***
social
fractionalizatio
n
-0.0003
(0.0001)***
-0.004
(0.001)***
-0.0003
(0.0002)*
-0.0005
(0.0003)*
-0.014
(0.005)***
0.004
(0.007)
-0.0003
(0.0001)***
ethnic
fractionalizatio
n
0.009
(0.010)
-0.0138
(0.007)**
-0.007
(0.007)
0.011
(0.013)
religious
fractionalizatio
n
-0.006
(0.014)
-0.026
(0.009)***
-0.023
(0.009)**
0.004
(0.018)
ethnic
dominance
(45-90%)
0.847
(0.370)**
0.754
(0.381)**
0.923
(0.369)***
0.816
(0.375)**
0.773
(0.372)**
0.833
(0.379)**
0.756
(0.381)**
0.940
(0.369)***
0.923
(0.373)***
0.847
(0.370)**
geographic
dispersion
-3.891
(1.172)***
-4.099
(1.201)***
-2.530
(1.042)**
-4.070
(1.242)***
-4.129
(1.222)***
-4.129
(1.230)***
-4.027
(1.245)***
-3.711
(1.163)***
-2.581
(1.053)***
-3.894
(1.172)***
peace duration -0.004
(0.001)***
-0.0004
(0.001)***
-0.005
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
N 688 688 668 688 688 688 688 688 688 688
no of wars 43 43 43 43 43 43 43 43 43 43
pseudo R2
0.30 0.30 0.27 0.30 0.29 0.29 0.30 0.29 0.26 0.30
log likelihood -113.03 -112.60 -118.56 -112.93 -114.42 -113.96 -112.57 -114.62 -118.98 -113.00
Notes: All regressions include a constant. Standard errors in parentheses. ***, **, * indicate significance at the 1, 5 and 10 percent level, respectively.
37
Table 9: Ethnic Dominance
1 2 3 4 5 6 7
Largest group as a
proportion of the total
population
45-85% 45-90% 45-95% 30-90% 40-90% 50-90%
male secondary
schooling
-0.034
(0.011)***
-0.036
(0.011)***
-0.036
(0.011)***
-0.032
(0.011)***
-0.039
(0.011)***
-0.038
(0.012)***
-0.035
(0.011)***
(GDP growth -
3*population
growth) t-1
-0.079
(0.036)**
-0.084
(0.036)**
-0.085
(0.036)**
-0.083
(0.036)**
-0.085
(0.036)**
-0.089
(0.037)**
-0.084
(0.036)**
primary
commodity
exports/GDP
31.444
(8.422)***
31.351
(8.289)***
32.089
(8.375)***
30.546
(8.296)***
31.918
(8.285)***
30.802
(8.218)***
31.292
(8.265)***
(primary
commodity
exports/GDP)2
-62.283
(19.037)***
-61.182
(18.712)***
-62.511
(18.964)***
-59.257
(18.837)***
-63.073
(18.656)***
-60.491
(18.433)***
-61.363
(18.776)***
ln population 0.842
(0.196)***
0.909
(0.201)***
0.946
(0.204)***
0.870
(0.196)***
0.951
(0.204)***
0.915
(0.200)***
0.907
(0.198)***
social
fractionalization
-0.0004
(0.0001)***
-0.0003
(0.0001)***
-0.0003
(0.0001)***
-0.0003
(0.0001)***
-0.0004
(0.0001)***
-0.0004
(0.0001)***
-0.0003
(0.0001)***
ethnic
dominance
-1.223
(1.076)
0.695
(0.373)**
0.847
(0.370)**
0.371
(0.554)
0.752
(0.385)**
0.736
(0.374)**
0.659
(0.373)*
geographic
dispersion
-3.907
(1.186)***
-4.042
(1.190)***
-3.891
(1.172)***
-3.891
(1.172)***
-4.155
(1.182)***
-3.931
(1.170)***
-3.686
(1.150)***
peace duration -0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
-0.004
(0.001)***
N 688 688 688 688 688 688 688
No of wars 43 43 43 43 43 43 43
Pseudo R2
0.28 0.29 0.30 0.28 0.29 0.29 0.29
Log likelihood -115.05 -113.98 -113.03 -115.45 -113.69 -113.67 -114.14
Notes: All regressions include a constant. Standard errors in parentheses. ***, **, * indicate significance at the
1, 5 and 10 percent level, respectively.
39
Table 9A: Ethnic Dominance - Descriptive Statistics
N=688 mean std. dev. min max
Largest group as a
proportion of the total
population
0.70 0.25 0.18 1
45-85% 0.33 0.47 0 1
45-90% 0.44 0.50 0 1
45-95% 0.77 0.42 0 1
30-90% 0.60 0.49 0 1
40-90% 0.55 0.50 0 1
50-90% 0.39 0.49 0 1
40
Country year war started previous war
Dominican Rep. 1965 1 0
Nigeria 1965 1 0
Burundi 1970 1 0
Pakistan 1970 1 0
Sri Lanka 1970 1 0
Zimbabwe 1970 1 0
El Salvador 1975 1 0
Nicaragua 1975 1 0
Somalia 1980 1 0
Peru 1980 1 0
Romania 1985 1 0
Turkey 1990 1 0
Guatemala 1965 1 1
Myanmar/Burma 1965 1 1
Uganda 1965 1 1
Iraq 1970 1 1
Iran 1970 1 1
Guatemala 1970 1 1
Philippines 1970 1 1
Ethiopia 1970 1 1
Morocco 1975 1 1
Iran 1975 1 1
Guatemala 1975 1 1
Mozambique 1975 1 1
Angola 1975 1 1
Indonesia 1975 1 1
Nicaragua 1980 1 1
Colombia 1980 1 1
Iran 1980 1 1
Myanmar/Burma 1980 1 1
India 1980 1 1
Sri Lanka 1980 1 1
Sudan 1980 1 1
Uganda 1980 1 1
Nigeria 1980 1 1
Burundi 1985 1 1
Iraq 1985 1 1
Somalia 1985 1 1
Burundi 1990 1 1
Algeria 1990 1 1
Rwanda 1990 1 1
Zaire 1990 1 1
Zaire 1995 1 1
43 31
41
Table A2: Descriptive Statistics
sample (n=688) no civil war (n=645) civil war (n=43)
mean st.dev. min max mean st.dev. min max mean st.dev. min max
War starts 0.06
Previous war 0.36 0.33 0.72
Male secondary schooling 44.5 31 1 147 45.6 31 1 147 27.3 21.7 3 87
GDP per capita (const. US$) 4025 4065 222 18993 4182 4137 290 18993 1673 1360 222 5832
GDP per capita Growth t-1 1.63 3.51 -13.1 14.41 1.72 3.43 -13.1 14.41 0.25 4.46 -10.66 7.09
Population growth t-1 2.09 1.19 -6.38 6.61 2.05 1.2 -6.38 6.61 2.73 0.64 0.47 4.09
Primary commodity
exports/GDP
0.16 0.14 0 0.94 0.16 0.14 0 0.94 0.16 0.1 0.02 0.48
Population (millions) 30 30 40
Ethnic fractionalization 41 29 0 93 40 29 0 93 50 30 4 90
Religious fractionalization 36 25 0 79 36 25 0 79 34 25 2 69
Geographic dispersion 0.6 0.21 0 0.97 0.6 0.21 0 0.97 0.57 0.15 0.24 0.91
Mountainous terrain 15.8 20.19 0 82.2 15.06 19.73 0 82.2 26.81 23.77 2.4 74.5
Ethnic dominance (45-90%) 0.28 0.27 0.4
Democracy 4.26 4.31 0 10 4.43 4.34 0 10 1.92 2.9 0 9
Diaspora 0.01 0.02 0 0.16 0.01 0.02 0 0.16 0.003 0.02 0 0.16
Income Inequality (Gini) 41.35 10.04 20.97 62.3 41.14 10.0 20.97 62.3 44.33 10.29 23.38 62.3
42
Table A3: A simple Migration model
Dependent Variable: Diasporait
Diasporai, t-1 1.163
(0.045)***
Ln GDP per capita i, t-1 -0.002
(0.001)**
War dummyi, t-1 0.003
(0.003)
N 216
R2
0.76
The regression is based on the following model:
diasit=f(diasi, t-1, lnGDPi, t-1, wari, t-1, Tt)
Where dias denotes diaspora which is measured as the ratio of emigrants in the USA to the total
population of the country of origin. The variable “war” is a war dummy, measured at t-1 it takes
a value of one if the country experienced a civil war in the previous period. The method of
estimation is OLS. The data is measured at the beginning of each decade, i.e. 1960, 1970,
1980 and 1990. The regression includes time dummies, T, which are jointly significant.
Based on this simple migration model we estimated the size of the diaspora at time t.
Dias estit=xit*^ß
For countries which experienced a previous civil war23
we used these estimated values to
correct for a possible endogeneity problem. We replaced a total of 64 observations. For
countries which did not experience a civil war we use the actual diaspora data. We took the
averages of this corrected diaspora data measured in 1960 and 1970 (1970 and 1980, 1980
and 1990) in order to obtain values for 1965, 1975 and 1985. For 1995 we use the
observations measured in 1990.
In Table A4 we present some descriptive statistics of the original diaspora data, the corrected
diaspora data and the difference between the two variables.
23
Here we only consider wars after 1960.
43
Table A4: Diaspora – Descriptive Statistics
N=704 mean st.dev. min max
diaspora 0.0161 0.0286 0.00003 0.1948
predicted diaspora 0.0162 0.0283 0.00003 0.1948
diaspora-predicted
diaspora
-0.0001 0.0043 -0.0200 0.0625
44
Table A5: Marginal Probabilities
The probability of a war-start was calculated in the following way:
probability=exp(X^ ß )/(1+exp(X^ ß))
Variabl
e
Coeff
.
Mean of X median mean worst best no ethnic ethnic plus 10
%
plus
2%
less 2%
dom
.
dom
.
education gy1 popg1
education -0.037 44.485 -1.460 -1.624 -0.011 -5.366 -1.624 -1.624 -1.989 -1.624 -1.624
gy1-
3*gpop1
-0.085 -4.696 0.433 0.397 2.804 -2.857 0.397 0.397 0.397 0.226 -0.114
exports 32.089 0.158 3.594 5.054 8.022 0.000 5.054 5.054 5.054 5.054 5.054
exports2 -62.511 0.025 -0.784 -1.551 -3.907 0.000 -1.551 -1.551 -1.551 -1.551 -1.551
lnpop 0.946 17.233 14.909 16.301 19.777 11.256 16.301 16.301 16.301 16.301 16.301
fractionaliz. -0.0003 1785.878 -0.607 -0.607 -0.004 -2.372 -0.607 -0.607 -0.607 -0.607 -0.607
ethnic dom. 0.847 0.440 0.000 0.373 0.847 0.000 0.000 0.847 0.373 0.373 0.373
geog dispersion -3.891 0.575 -2.366 -2.238 0.000 -3.778 -2.238 -2.238 -2.238 -2.238 -2.238
peace -0.004 347.670 -1.507 -1.488 -0.004 -2.534 -1.488 -1.488 -1.488 -1.488 -1.488
constant -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443
X*betahat -4.231 -1.826 11.081 -22.093 -2.199 -1.351 -2.191 -1.996 -2.337
probability 0.014 0.139 1.000 0.000 0.100 0.206 0.101 0.120 0.088
Variable Coeff
.
Mean of X Exports= Exports= min max min 5 yrs 10yrs geog dis- geog
dis-0.26 0 fraction. fraction. peace peace peace persion=1 persion=0
education -0.037 44.485 -1.624 -1.624 -1.624 -1.624 -1.624 -1.624 -1.624 -1.624 -1.624
gy1-
3*gpop1
-0.085 -4.696 0.397 0.397 0.397 0.397 0.397 0.397 0.397 0.397 0.397
exports 32.089 0.158 8.343 5.054 5.054 5.054 5.054 5.054 5.054 5.054
exports2 -62.511 0.025 -4.226 -1.551 -1.551 -1.551 -1.551 -1.551 -1.551 -1.551
lnpo
p
0.946 17.233 16.301 16.301 16.301 16.301 16.301 16.301 16.301 16.301 16.301
fractionaliz. -0.0003 1785.878 -0.607 -0.607 -0.004 -2.372 -0.607 -0.607 -0.607 -0.607 -0.607
ethnic dom. 0.847 0.440 0.373 0.373 0.373 0.373 0.373 0.373 0.373 0.373 0.373
geog dispersion -3.891 0.575 -2.238 -2.238 -2.238 -2.238 -2.238 -2.238 -2.238 0.000 -3.891
peace -0.004 347.670 -1.488 -1.488 -1.488 -1.488 -0.004 -0.257 -0.514 -1.488 -1.488
constant -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443 -16.443
X*betahat -1.212 -5.329 -1.223 -3.590 -0.342 -0.595 -0.851 0.412 -3.479
probability 0.229 0.005 0.227 0.027 0.415 0.356 0.299 0.602 0.030