© AP Photo/David Guttenfelder
About 140 civil wars around the world since 1945 have
killed approximately 20 million people and displaced 67
million.1
These figures are based on Doyle
and Sambanis 2003. Other data sets list
different numbers of civil wars; data on deaths vary widely.
Despite this massive scale of human misery, the academic community did
not pay much attention to the problem of civil war until very recently.
Two important papers—one by Paul Collier and Anke Hoeffler
(henceforth CH), and the other by James Fearon and David Laitin
(henceforth FL)—have generated much interest in the question of
why civil wars occur.
2 Collier and
Hoeffler 2001.
These papers present
the counterintuitive finding that civil wars are not caused by ethnic
division or political grievances, but by the opportunity structure for
the organization of rebellion or insurgency.
3 Rebellion is CH's term; insurgency
is FL's term.
Both papers use macro-level data to test
hypotheses about civil war that are based on ideas about micro-level
behavior (e.g., why individuals would join an insurgency).
4 In this article, the terms micro and
macro are not used to distinguish microeconomic from
macroeconomic models, but rather to set individual-level incentives and
actions (the “micro” level) apart from social movements and
opportunity structures (the “macro” level).
They
identify a set of statistically significant correlations between civil
war onset and a number of explanatory variables.
The already significant gap between the micro-level theories and
their macro-level implications is magnified when the micro-macro
relationships are studied solely through cross-national statistical
analyses. Such studies often overlook information about the causal
pathways that link individual or group behavior with the outbreak of
civil war. In this article, I argue that by combining statistical and
case study work we can better understand the political processes that
lead societies to civil war. I also emphasize the importance of
focusing on the macro level in order to explain how individual episodes
of violence are organized in the form of a civil war.
Given the significance of the CH and FL models, I begin with them to
demonstrate how a comparative case study design can be combined with a
formal-quantitative approach to build a better-specified theory of
civil war.5
See Ballentine and Sherman
2003, a book of case studies that makes this
same point.
I use cases to develop—not test—theory
and to qualify the causal inferences that we can draw from a
quantitative model of civil war. I draw on the Case Study Project on
Civil Wars, a set of 21 case studies on civil war onset and war
avoidance
6 Throughout this article, I cite
revised drafts of case studies originally presented as part of the Case
Study Project on Civil Wars in New Haven, Connecticut, in April 2002.
In each instance, I include a reference listing for the relevant
version of the paper, not necessarily the original. For the Caucasus
study, I include both the 2002 and the 2003 versions, because revisions
were extensive. See Zürcher et al. 2002
and Baev et al. 2003. (Please note that these
are two versions of the same study; the authors changed the order of
their names.)
that systematically apply the CH model.
7 The Case Study Project on Civil Wars began in
the spring of 2000. Teams of country experts wrote the case studies; in
most instances, an author from the country under consideration worked
with an author from a U.S.-based institution. The following countries
were included: Algeria, Azerbaijan, Bosnia, Burundi, Colombia,
Democratic Republic of Congo, Georgia, Indonesia, Ivory Coast, Jamaica,
Kenya, Lebanon, Macedonia, Mali, Mozambique, Nigeria, Russia, Senegal,
Sierra Leone, Sudan, and United Kingdom (Northern Ireland). Some of
these countries had more than one civil war. The following case studies
were commissioned but never completed: Afghanistan, El Salvador,
Moldova, Somalia, Sri Lanka, and Uganda.
(I was the primary
investigator from the summer of 2001 until the project's
completion in the spring of 2004.) Two online supplements to this
article include details about the project's research design and
summaries of the cases.
8 The first
guidelines given to authors are posted at
www.yale.edu/unsy/civilwars/guidelines.htm. (More detailed
instructions were given at two conferences—in Oslo, Norway, in
June 2001, and in New Haven, Connecticut, in April 2002.) Go to
pantheon.yale.edu/∼ns237/index/research.html#Cases
for a summary assessment of the models used in each case. The cases
will be published in an edited volume (date to be
determined).
The case studies make several contributions to our understanding of
civil war:
- They help us identify a number of causal mechanisms through which
independent variables in the CH and FL models influence the dependent
variable—i.e., the risk of civil war onset. It quickly becomes
clear that the CH model's distinction between “greed”
and “grievance” as competing motives for civil war is
illusory, because greed and grievance are usually shades of the same
problem.9
The distinction was proposed in early versions of the CH
model. It corresponds to the distinction between grievances and
opportunity structures in the FL model.
- They question assumptions and premises of the quantitative studies
and make clear that CH and FL are often right for the wrong reasons yet
also wrong for the wrong reasons. (In other words, the cases identify
mechanisms that are different from those underlying their theories,
both where the statistical models make good predictions and where they
make bad predictions).
- They sometimes point to a poor fit between the empirical proxies
and the theoretically significant variables—i.e., they identify
measurement problems in the statistical studies. Later in this article,
I offer some examples and suggest ways to improve the connection
between theory and data.
- They help us identify new variables that might explain civil war
but are omitted from CH and FL (e.g., external intervention,10
External intervention is likely partially determined by
other variables in the model. Thus, if it were added to the model, the
estimation method would probably have to be adjusted and an
instrumental-variables approach used. Such an approach might also be
more appropriate for the current versions of the CH and FL models,
since they include variables that may not be purely exogenous.
This is
not simply an exercise intended to maximize the R2 in
quantitative models, though who would mind a higher R2 if it
could be achieved by adding significant, exogenous variables to the
model?
There is not room in
this article to present new quantitative tests of an expanded civil war
model or even to present a complete new theory of civil war. This is
the subject of a book-length manuscript that I am writing.
- They highlight interactive effects between variables in the
statistical models and help us identify exogenous and endogenous
variables by presenting narratives of the series of events and the
processes that led to civil war.
- They suggest substantial unit heterogeneity in the data, as the
mechanisms that lead to civil war seem to differ substantially across
different sets of countries and types of civil war.
This last observation informs a new theoretical contribution made in
this article. I argue that it is not as useful to view civil war
outcomes as the result of deep-seated and hardly changing structural
conditions as it is to observe the links among different forms of
political violence and to analyze the dynamics of conflict escalation
and the transition from one form of violence to another.
It is accepted practice in the literature to pool events of civil war
without exploring whether, in fact, they all result from the same
causal process.13
Exceptions are Sambanis
2001, Sambanis 2002, and Sambanis 2003b,
in which I explore differences between ethnic and revolutionary civil
wars and between civil war and politicide or genocide.
Although civil wars are thought to be different from other forms of
political violence, current models of civil war do not allow us to make
that determination, and it is important to establish whether there are
such differences. If civil wars have different causes than do other
forms of violence, then the policy interventions designed to prevent
them should differ from those meant to ward off terrorism or other
violence. For example, perhaps we will eventually find that riots and
minor armed conflict occur more frequently than civil wars when there
are sharp drops in a country's economic growth or when intense
price shocks reduce real incomes in primary commodity–dependent
countries.
14 Horowitz 2001 casts doubt on the hypothesis that price
shocks are associated with ethnic riots and refers to an inconclusive
literature on the presumed association between price shocks and food
riots. I have no evidence to add to the debate; I use this only as an
example of a possible association that might differentiate forms of
violence.
Perhaps these conditions also lead to military
uprisings, or coups, if soldiers are not well paid and if military
elites are not controlled by political elites. Prevention or
intervention policies should then be designed to stabilize prices,
ensure that the troops are paid, and gradually replace the leadership
of the army to establish better military-civilian relations. By
contrast, these factors may not be as important in preventing
separatist civil war, which may be a greater concern for some countries
(e.g., countries with large and territorially concentrated ethnic
groups).
15 For a theory and evidence on
the relationship between ethnic violence and the territorial
concentration of ethnic groups, see Toft 2003.
Separatist violence may be more
strongly associated with regional inequalities in the distribution of
income and may be more likely to occur in peripheral regions bordering
countries that are governed by ethnic kin of the majority population in
the region.
16 See Sambanis and Milanovic
2004 for a theory of separatist conflict that
focuses on regional differences in income and ethnic composition as
determinants of the demand for self-determination.
Thus,
policy interventions should involve an increase in the state's
police presence in those regions to prevent the organization of
insurgency,
17 See Herbst 2000 for a relevant argument. Herbst explains that
state formation in Africa has evolved into a concept of a
“core” territory that the state must control to remain
viable. Peripheral regions are often excluded; and the military,
weakened because of its competitive relationship with the state, allows
political conflicts to turn into insurgencies much more easily in
peripheral regions than in the “core” territories.
or they may involve redistribution to reduce regional disparities.
However, if a country faces a higher risk of a popular revolution
because of widespread class inequities, then redistribution should take
a different form (it should not focus on regions, but rather on
reducing interpersonal inequality). If our research shows that
genocides are likely to take place in the context of civil wars in
countries that have autocratic systems and ethnic elites with
exclusionary ideology,
18 Harff 2003 identifies the significance of the elite
ideology argument. Public statements should make identification of
these elites relatively straightforward. All genocides also seem to
occur in autocracies. Out of 244 country-years of politicide from 1960
through 1999 in Harff's data set, only Sudan from 1965 through
1968 and 1986 through 1988 and Guatemala in 1996 can be labeled deep
democracies with a net democracy-autocracy score of 7/10 or higher
in the Polity IV database (see Marshall and Jaggers 2000). And these are cases of civil war, not
genocide, which is usually understood as intentional targeting of an
ethno-religious group with the intent to destroy that group.
then the international community could develop a strategy for early
intervention in those civil wars to prevent genocide.
19 A concrete discussion of policy
interventions that better target civil war as compared to other forms
of political violence must await the results of empirical studies that
identify differences in the causal paths leading to different forms of
political violence.
If civil wars are not unique, however, then by analyzing them in
isolation we may be getting inefficient (or even biased) statistical
estimates, because we are arbitrarily restricting our analysis to a
subsample of the data. Either way, to explain civil war, we must
construct a theory that tells us why conflict escalation leads to civil
war as opposed to other kinds of violence.
It is not my intention here to pit CH against FL and pronounce one
the winner. Both models are useful starting points as I try to outline
a new way of analyzing the causes of civil war. By focusing on the
distinction between process and outcome, I offer a way to reconcile
quantitative and qualitative research designs, which political
scientists often (mistakenly) consider as mutually exclusive rather
than complementary approaches.
A Primer on the State of Civil War Theory
Civil war in both CH and FL is defined as armed conflict between the
government of a sovereign state and one or more organized groups that
are able to mount effective resistance against the state. More details
go into these definitions (e.g., a threshold of the number of deaths)
and the definitions used in the two studies differ slightly, but these
differences do not influence their key empirical results much.20
On the definition and measurement of civil
war and empirical differences that result from different definitions,
see Sambanis 2003a.
The CH model of war onset is based on the idea of an
individual's tradeoff between production and appropriation.21
The theoretical underpinnings of this idea
are modeled by Grossman 1991; Hirshleifer
1989; Hirshleifer 1995; Konrad and Skaperdas 1999.
According to CH, people decide to
become rebels after weighing the economic opportunity cost of violence
against its expected utility. War is an inefficient way to settle
disputes because it is costly and reduces the net value of rents
available to the state. Yet war occurs, because of “three
interacting determinants:
preferences,
opportunities,
and
perceptions,” and because it is hard to negotiate
credible peaceful settlements without the use of force.
22 Hirshleifer 1995,
172. The credibility argument is developed for interstate war (Fearon
1995) and civil war (Skaperdas 2001).
Preferences for private gain
(“greed,” in the CH model) lead to political violence if
there are opportunities to rebel. Defined in this way, rebels are
indistinguishable from bandits or greedy war entrepreneurs.
23 See Collier 2000.
Collier and Hoeffler write, “On the literal greed interpretation
the extortion of primary commodity exports will occur where it is
profitable, and the organizations which perpetrate this extortion will
need to take the form of a rebellion.” Collier and Hoeffler 2001, 3.
Grievance, in CH, is simply
rhetoric used to legitimize a person's decision to engage in
appropriation rather than production. The instrumental cause of civil
war is the availability of loot, combined with the opportunity to
organize an insurgency.
According to the CH model, rebellion is sustained through looting of
natural resources, extortion of local populations, and support from
ethnic diasporas. Insurgency is less likely when the state is strong or
when economic opportunity costs of rebellion are high. State strength,
not well theorized in CH, is approximated by the country's
economic strength (gross domestic product per capita). The expectation
is that richer states will be bureaucratically more efficient and will
have the resources needed to defend themselves against a rebellion.
Additionally, strong states can reduce the available rebel labor supply
by decreasing the net expected gains of rebellion. Rebel labor supply
should also decline as the economic opportunity costs of rebellion
increase (time devoted to production rather than appropriation pays
more in rich states than in poor ones). In sum, a country's
economic opportunity structure—its income level and growth, as
well as its economic system—determines the “supply”
of insurgency for a given level of insurgency “demand.”
The FL model shares much of the same logic. FL rejects primordialism,
nationalism, modernization, and other grievance-based explanations in
favor of an opportunity-structure argument that explains civil war as a
function of state strength (also measured by GDP per capita) and the
technology of insurgency (rough terrain in which to hide and navigate,
diaspora funding, et cetera). FL dismisses CH's resource-predation
argument—that is, FL finds no support for linking primary
commodity exports to war onset. Although FL finds that oil-dependent
states are more prone than others to civil war, the study interprets
this as a weak-state effect, since oil dependence frequently causes
weak state structures. The FL model is state-centric; it includes a
host of variables that measure state capacity, such as political
instability, new state formation, and regime type. However, many of the
differences between CH and FL are mainly interpretive, since the models
often use the same (or closely related) proxy variables to test
different hypotheses about civil war onset.
Regarding opportunity structure for rebellion, CH and FL resemble
earlier important works.24
For example,
Gurr 2000, chapter 3.
But what is
new is the quantitative empirical testing of their models, which
suggests several important and robust results. Both CH and FL find that
GDP lowers the risk of civil war: CH takes this to mean that poverty
exacerbates risk, while FL interprets it as support for the
state-weakness theory. While FL sees no significant association between
high economic growth and civil war risk, for CH, high economic growth
and high levels of secondary-school enrollment (particularly among
males) lower risk by increasing opportunity cost. Both CH and FL find
that countries with mountainous terrain and large amounts of external
financial support have higher risk of civil war; but they disagree on
other ways of financing rebellion.
25 For
instance, CH focuses on the looting of natural resources as both a
motive for war and a means to sustain it. FL agrees with the hypothesis
but contradicts CH's empirical findings. CH finds that diaspora
funding only influences war duration and recurrence, not the risk of
initial war onset. Both models use a dummy variable denoting the Cold
War period as a proxy for external support from superpowers, though who
receives such support (the government or the rebels) is not
specified—and this factor should influence civil war
risk.
According to both models, civil war is more likely in
populous countries. However, CH claims this is because large
populations are more likely to include aggrieved groups, whereas FL
attributes it to states' having a harder time controlling large
populations.
26 China and the Soviet Union
(under Stalin) are obvious exceptions to FLs' argument, though
there are also many examples of unruly small countries (Cyprus,
Georgia, Azerbaijan, and others).
Finally, both CH and FL find
that democracy does not significantly reduce the risk of civil war and
that ethnic fractionalization does not increase it—although,
according to CH, ethnic
dominance increases risk. CH and FL
seem to agree that countries in the middle of the democracy-autocracy
spectrum and those with political instability are more prone to civil
war.
27 These results were first reported
by Hegre et al. 2001 and can also be found in
the literature on social movements.
CH and FL get their results from pooled logit analysis of time-series
cross-sectional data covering roughly the same set of countries and
much of the same time period.28
FL
analyzes data per year from 1945 through 1999; CH analyzes five-year
panel data from 1960 through 1999.
Both models analyze war
onset, but CH drops observations of ongoing war, while FL codes such
periods as
0's so as not to leave out war onsets that
occur in countries with an ongoing civil war.
29 CH finds a highly significant negative relationship
between time at peace and the risk of civil war onset, but FL does
not.
With this telegraphic review in mind, let us now turn to the case
studies.
Case Study Methodology
Most researchers who work with quantitative methods are averse to
doing case studies. They presume that case studies cannot test or
inspire theory because they suffer from selection bias, omitted
variable bias, endogeneity, and measurement error. However, these
problems are also frequently present in quantitative studies. In what
follows, I address these concerns by describing the methodology of the
Case Study Project on Civil Wars, arguing that well-designed case
studies can actually help reduce the presence of such problems in
quantitative analyses of civil war.
The project's main purpose was to improve causal inferences
drawn on the basis of the CH model. CH was tested statistically, so we
did not need to retest it using case studies. In fact, such retesting
was not even desirable, because statistical identification associated
with case study designs is widely known to be problematic.30
Thus, each case study tried to
complement our theory, refine our empirical measures, explore
interactions among explanatory variables, identify endogenous and
exogenous variables in the model, and identify causal mechanisms
underlying the theory.
To achieve these goals, we used a structured-focused comparison
design, exploring the fit of the CH theory to each case. Country
experts wrote narratives that addressed a common set of questions about
the onset and organization of rebellion. Since all narratives responded
to the same questions, our research design allowed us to draw
generalizations.
Unit of observation
As in the CH and FL statistical tests, the unit of observation for
the case studies is the country—or, more accurately, several
periods of peace and war within each country. Some countries have had
recurrent wars. The authors discuss all periods of war, exploring the
connections among these events. In effect, each case study provides us
with several observations. For example, the Indonesia study focuses on
patterns of war and peace in Aceh over eight five-year periods; it
therefore yields eight observations (two of war and six of no war). The
Nigeria study analyzes several periods that included two wars, while
the Ivory Coast study examines three periods, each with no war but with
a different structure of war risk.31
On
the Ivory Coast, see Azam and Koidou 2003. On
Indonesia, see Ross 2003. On Nigeria, see
Zinn 2003.
Case selection
Given that we knew the values of the dependent variable for each case
and we had a sense of which variables were on average significant
predictors of outcomes, we did not pick cases with a view to predicting
outcomes. Rather, we selected them with two goals in mind: first, to
maintain sufficient variation among the independent variables,32
On this selection rule for qualitative
research designs, see King et al. 1994.
and second, to include cases that
the model predicted well
and ones that it predicted
poorly.
33 Including cases of no war
resembles John Stuart Mill's indirect method of
difference. See Ragin 1987. Each case
study effectively includes no-war comparisons—i.e., periods of
war and no war in each country—and the authors explain the
transition across these two states.
Among cases predicted
poorly by the model, we considered both periods of war onset not
predicted by CH and periods of peace in countries for which CH
predicted high risk of civil war.
34 Even
where it generates false negative or false positive predictions, the
model is still technically correct because, strictly speaking, it does
not predict war, but rather the probability of war. Thus, war may not
have happened in high-risk countries for reasons that remain outside
the model. But the same is true for “accurate” statistical
predictions: war might occur in high-risk countries for reasons that
the model does not consider.
The case studies helped us sort
out which of these poor predictions failed for idiosyncratic reasons
and which ones resulted from possible flaws in the theoretical model,
the operational measures used in quantitative analysis, or the
estimation method.
35 Of course, all of
this assumes that the case studies themselves are relatively free of
measurement error and other biases. This is one of the reasons why
commissioning experts to write the cases may be superior to simply
summarizing existing case studies. Ultimately, we can check the
usefulness of the case studies when hypotheses and model refinements
based on the cases are taken back to the data for further
testing.
Where CH generated accurate predictions, the case
studies primarily aimed to trace the mechanisms linking the independent
variables to the dependent variable, so we could assess whether the
theory underlying the CH (and related FL) tests of hypotheses
corresponded to the empirical measures used. Our selection criteria and
the large number of cases in our project allowed us to build theory
inductively; the new theory could later be taken back to the data for
further hypothesis testing.
We sought to ensure some variation in the independent variables,
though our case-selection mechanism was not sophisticated. We picked
democracies and autocracies, countries with long and short histories of
violence, high and low levels of ethnic fragmentation, and high and low
natural-resource dependence.36
All
selection criteria were controlled for in the CH model, and the fact
that the case studies analyzed long periods in each country's
history ensures over-time variation in the independent
variables.
Choosing a random sample of cases was not
necessary, since we did not intend to use them to test the theory.
37 Fearon and Laitin (2004) are currently engaged in a similar case study
project, based on a random selection of cases.
Random
selection from the CH or FL data sets might have resulted in a sample
that predominantly included cases of no war, given that civil war is
relatively rare. It could also have resulted in a sample with no
significant variation in the independent variables. We could have
avoided the first (but not the second) problem by sampling more heavily
on cases of war, but nonlinearities that may have been present in the
data could have complicated the sampling process.
38 Nonlinearities imply that the theorized linear
relationship between the dependent variable and the independent
variable does not apply to the entire sample. If ethnic identity
matters in different ways in developed and less-developed countries
(cf. Horowitz 1985), then adding interaction
terms is one way to properly explore conditional effects. When such
effects are present, a stratified sampling method is correct if cases
are used for hypothesis testing.
For example, civil wars occur
predominantly in poorer countries, so a case-control design that
selected three no-war cases for every case of war
39 A case-control design was used by Esty et al. 1995 in their study of state failure.
would have included far too many middle-to-high-income countries to
allow us to disentangle the relationship between the dependent variable
and other independent variables at different levels of income. Properly
controlling for this problem would have entailed a highly stratified
selection rule and a sample size even larger than the one we had in our
project. So we opted instead for a design that would offer theoretical
insights into the complex pathways and sequence of events that lead to
civil war. These insights, at a later stage, could help us retest the
theory with quantitative methods.
A comparison of Chechnya and Dagestan (both parts of Russia) in our
case studies offers an example of the complicated causal path linking
political instability to civil war. Both regions were harmed by the
political instability that followed the collapse of the Soviet Union.
In Chechnya, political institutions fell at the hands of Chechen
nationalists, led by General Dzhokhar Dudaev, who pursued secession on
behalf of the Chechens. In Dagestan no titular majority group existed,
and the national independence movement acted through preexisting
political institutions that represented 54 regional soviets, or
parliaments. Only 39 soviets favored secession, and political elites
decided that secession was too costly.40
Thus,
the effects of political instability on civil war risk in Chechnya and
Dagestan were filtered through the type of regional political
institutions available in each case and were conditioned by the
regions' ethnic composition as well as by nationalist elites'
actions. The helpful perspective on sequence offered by these cases can
lead us to refine our hypotheses in quantitative models (e.g., some
variables might be endogenized or considered in interaction with
others).
Unit heterogeneity
The CH and FL tests are based on a pooled sample of all civil wars,
reflecting a strong assumption of unit homogeneity.41
“Two units are homogeneous when the expected values
of the dependent variables from each unit are the same when our
explanatory variable takes on a particular value.” King et al.
1994, 91. In other words, in a sample that
includes homogeneous observations, if GDP (or any other explanatory
variable) changed by, say, 10 percent and all other variables in the
model were held constant, the expected risk of war should change by an
equal amount in all cases.
If this assumption is violated, it
can bias causal inferences from the model. Case studies are a good way
to test the validity of the homogeneity assumption—for instance,
by exploring the organization of rebellion.
42 Ragin 1987. This is an
exploratory, not formal, test of the assumption.
If there is
substantial unit heterogeneity in the data, then tests should be
restricted to particular regions or ranges of key variables across
which outcomes differ significantly. For example, we should use
slightly different models to analyze the risk of civil war in Africa or
in the Middle East, or we should analyze the likelihood of civil war
only in low-income or heavily resource-dependent countries.
43 These are sometimes called
“within-systems relationships.” See Przeworski and Teune
1970, 57–9; Ragin 1987.
We did not adopt such an approach in
selecting cases, because both CH and FL put forth propositions that can
be widely generalized, so we sought to have broad representation
roughly corresponding to the geographic incidence of civil war.
44 Collier and Hoeffler (2002a) show that there are no statistically
significant patterns in civil war onset in different regions. There is,
however, a heavier representation of African wars in our
sample.
Identifying causal mechanisms
Case studies can identify interactions between variables and
establish a chronological sequence of events that helps map out the
pathways linking the independent variables to the dependent variable.
This also makes it easier to deal with the problems of endogeneity and
selection that likely affect the results of both CH and FL. Several
plausible causal mechanisms can be identified by the case studies,
though testing the significance of these mechanisms and rank-ordering
them is probably best done by going back to the statistical models.45
This is because of the well-known
“degrees of freedom” problem in small-N studies—i.e.,
the problem of having too few observations of outcomes to estimate the
effects of several independent variables.
Mechanisms are, in effect, variables that operate in sequence.46
Mechanisms have been defined as a
“delimited class of events that alter relations among specified
sets of elements in identical or closely similar ways over a variety of
situations.” McAdam et al. 2001,
24.
They can also be outcomes: depending on the level of
aggregation at which a theory is built, different outcomes become
intervening variables, connecting a set of variables to other outcomes.
Consider the theory that HIV causes AIDS and AIDS causes death. HIV
infection can be an independent variable in a regression on mortality
rate. Contraction of AIDS is the mechanism through which HIV usually
leads to death. But at a finer level of specificity, we could discern a
number of competing mechanisms that link HIV to death (e.g., pneumonia,
cancer, viral infections). Finally, HIV is itself a mechanism through
which actions (e.g., needle sharing among intravenous drug users) might
cause death. It is therefore up to the analyst to determine the level
of aggregation at which he or she formulates a theory; and according to
that level of aggregation, one could proceed to identify the mechanisms
linking independent variables to the dependent variable. There could be
an almost infinite regression toward more micro-level cause-and-effect
relationships, but each step down the ladder changes the real focus of
the theory (in our example, we go from a theory of how HIV causes death
to a theory of how cancer causes death). This infinite regression to
more remote causes ultimately becomes irrelevant, as more distant
causes will explain less and less of the variance in the outcome that
we are trying to explain.
47 Consider
another medical example. Assume that obesity causes clogging of the
arteries, which can lead to a heart attack, which can cause death. One
could therefore say that heart attack is the mechanism through which
obesity causes death. But assume, also, that obesity is the result of
bad eating habits, which are at least partially the result of
socialization at home by one's parents. Therefore, it is
theoretically consistent to argue that obesity is the mechanism through
which bad parenting causes death. That statement, however, has very
little prima facie credibility. Put more formally, bad parenting should
have little power as an explanatory variable in a model that tries to
describe death as a function of a person's physical
characteristics.
Thus, we come back to the distinction between micro- and macro-level
research on civil wars that I made earlier. It is true, as argued by
Stathis Kalyvas in an earlier issue of this journal, that motives for
violence vary wildly at the micro level and that micro-level and
macro-level determinants of civil war often do not directly correspond
to one another.48
Kalyvas 2003 focuses on variation in micro-level actions to
challenge the idea that the ontology of a civil war can be established
on the basis of macro-level identities. In my analysis, ontology is a
broader concept, referring to civil war as an analytically and
empirically meaningful category that is distinct from other forms of
political violence.
But analysis at different levels of social
conflict will necessarily reveal different causal patterns. The
interesting question is how and why disparate private motives for
violence are disciplined into a single organizational form, such as
civil war. The disjuncture between micro-level actions and macro-level
identities that Kalyvas expertly demonstrates implies one of two
things: the variation at the micro-level is irrelevant to the question
of civil war onset, or civil war is such an aggregate concept that it
is not useful as an analytical category. Supposing the latter, if civil
war includes coups, riots, gang violence, crime, and genocide, are we
right in analyzing civil war as distinct from all these other forms of
violence? If what we are trying to explain is the outbreak of civil
war, then the process of interest is that by which divergent incentives
and myriad personal calculations generate civil war rather than another
type of violence. The theories proposed by CH and FL
assume
that civil war is a distinct category of violence and try to explain
when and where civil wars are likely to occur. (This assumption is
clear in their arguments but is also revealed by the fact that they
test their models only on data of civil war, not on data on all forms
of violence.) But these theories can partially explain many forms of
violence—even organized crime—and are not specific to civil
war. We must therefore consider a wider array of both micro- and
macro-level theories, including ones that explain how emotions,
ideology, revenge, or coercion can interact to produce collective
action that culminates in a civil war.
Ultimately, it is the interaction between micromotives and
macrostructures that determines the expression of violent conflict.
Different organizational forms (e.g., civil war) are at the same time
outcomes of such a micro-macro interaction and mechanisms that
can explain violence at higher levels of aggregation. Thus, terrorism,
coups, and riots may be leading indicators of civil war (and may
precipitate civil war), yet we must explain why in some countries we
observe those forms of violence without also observing violence
escalation into civil war. A general theory of political violence must
explain how and why we shift from one form of violence to another, and
it must analyze civil war as part of a dynamic process.49
Developing such a theory is the subject of
my ongoing work.
At the same time, if we do not know why a
civil war instead of some other type of violence occurs, then we do not
understand civil war.
50 A counterargument
might be that once the preconditions for some political violence are
established, the organizational form of violence is entirely
contingent. If that is true, then focusing on civil wars as a distinct
outcome is not meaningful, and we must analyze all political violence
as a whole. Empirical testing of such a theory must be based on a data
set that pools all events of political violence, not just civil
wars.
There is potentially a hierarchy of mechanisms that explain how civil
war starts, but such a hierarchy might be hard to establish through
case studies. However, even without rank-ordering these mechanisms, we
can identify patterns that keep coming up in the cases.51
It is true that the process of using the
case studies to identify variables to build theory is not systematic in
the sense that there is no rule specifying that a variable must appear
x number of times before we should consider adding it to the
statistical model. But the same might be said of any process of
theorizing: how many times must a theorist who is developing a
deductive model of a phenomenon observe the phenomenon before
he/she decides to model it? It is important that case-study authors
who suggested rival explanations of outcomes in their countries did so
while being guided by the CH model. But in selecting variables from the
case studies to develop the theory further, it may be possible for
several, potentially conflicting, theories to be developed, depending
on the reader's interpretation of the case studies. In the end,
these competing theories can be sorted out empirically when we take
them back to the data.
These patterns can be explored further
in statistical analysis. The narratives presented in the case studies
show dynamic interactions over time between explanatory variables in
the CH model, and they help identify mechanisms and variables that may
be more commonly found in civil wars than in other types of violence.
Later in this article, I focus on a few variables and processes that
might explain civil war: I consider the effects of state repression,
the impact of neighborhood effects, and external intervention in the
process of conflict escalation leading to civil war.
Measurement Error and Causal Inference
A discussion of the CH and FL empirical results must start with an
assessment of the consequences of measurement error in these data sets.
Although CH and FL take much care in coding cases and collecting data,
in several instances they use problematic proxy variables. This point,
borne out by several cases, casts some doubt on several causal
inferences that might be drawn from the CH and FL findings.
Economic variables: GDP, growth, education
The key measures of opportunity cost in the CH model are GDP per
capita, secondary education (among males), and economic growth.
Consistent with the CH theory, many countries in the Case Study Project
on Civil Wars had low or declining income in their prewar years. In
Sierra Leone, real per capita income was barely more than $900 before
the war started, down from $1,400 in the 1970s. In Indonesia, the East
Asian financial crisis caused income to fall by 9.8 percent in the
province of Aceh in 1998, right before a war started. The oil and gas
sector, which accounted for 65 percent of Aceh's GDP, shrank by
almost one-fourth during the financial crisis.52
In Nigeria,
recession in the late 1970s caused unemployment to double to more than
20 percent before the onset of the Maitatsine rebellion. In Yugoslavia,
income dropped and unemployment soared after the liberal reforms of
1989, just two years before the first of several wars in the former
Yugoslavia.
But GDP per capita can also be used as a proxy for state strength, as
in the FL model. It is unclear, therefore, how to interpret the
negative correlation between GDP per capita and civil war. CH's
use of GDP adjusted by purchasing power parity is consistent with the
opportunity-cost argument; yet FL's constant-dollar GDP
figures are more in line with the state-strength argument,
since they describe the overall size of the economy. However, both uses
of GDP are questionable. The size of an economy probably correlates
with country size or population size and may not indicate the extent of
a state's control over its territory. For the CH opportunity-cost
argument, unemployment levels might have been a more direct measure,
because they would have indicated potential rebel supply. In prewar
Yugoslavia, income per capita was two or three times that of
war-affected countries in our sample, but unemployment surged and in
some regions reached 40 percent of the adult population. Using
unemployment rates (especially region-specific rates) in the
quantitative analysis might help distinguish the CH opportunity-cost
theory from the FL state-strength theory—hypotheses that are now
conflated by their similar use of GDP as a proxy variable.
The interpretation of GDP as a measure of state strength is also
problematic in a number of cases. Consider the United Kingdom (Northern
Ireland insurgency) and Kenya (Rift Valley ethnic violence). In the
United Kingdom, the FL theory could be used to explain (at least
partially) why the “Troubles” and their aftermath did not
escalate into a larger war. The strength of the British state may have
forced the insurgents to adopt a strategy of low-level urban violence
and terrorism.53
See Woodwell 2003 regarding the deterrent effect of the Royal
Ulster Constabulary's strength of 13,500 members.
Yet
this explanation omits other plausible factors, such as the role of
civil society and public opinion in the United Kingdom and neighboring
Ireland. A more intense campaign by the Irish Republican Army (IRA) and
a more indiscriminate and forceful response from the British army would
have caused negative reactions from civil society on both sides. In an
established democracy like Britain, war-fighting tactics such as the
ones used by the Russian state in the war in Chechnya (for instance,
bombing the capital, Grozny) are not viable—indeed, they are
unthinkable. But it is hard, on the basis of the FL model, to
disentangle the impact of the state's strength (defined in terms
of military and economic capacity) from the consequences of the
state's liberal-democratic characteristics.
The problem lies with the poor analytic value of the empirical proxy
used to measure state strength. The case of Kenya is instructive. Here,
we have a weak economy with a weak civil society, but a strong
authoritarian state. The state's ability to repress its opponents
was unconstrained until the recent liberalization in Kenya. Despite
strong ethnic antagonism, significant electoral violence, and a coup
attempt in August 1982, no civil war has occurred in Kenya, mainly as a
result of the state's strength and authoritarianism.54
CH and FL do not code a war in Kenya. Other
data sets (Doyle and Sambanis 2003) code a
war from 1991 to 1993 because of the state's indirect involvement
in the violence, but this is an ambiguous case and could also be
classified as intercommunal violence.
The mechanism of
exercising state control was corruption.
55 Kimenyi and Ndung'u 2002.
The government used local police
forces to violently repress local opposition groups that could not be
bought off; it rewarded government supporters with gifts of public
land. Our case study on Kenya shows that a low GDP is not necessarily a
good measure of a state's capacity to prevent civil war.
Additionally, we cannot even distinguish between the effects of state
strength and those of absent or weak civil-society institutions.
Clearly, to sort out the relative significance of these explanations,
we must return to large-N data analysis. But case studies help us
identify plausible mechanisms and hypotheses to test.
Consider next the CH argument about education. Several countries seem
to fit the CH model well, especially in Africa (there were virtually no
educated Congolese at the time of their independence and right before
the start of their first civil war). But regional disparities may
weaken the underlying theory linking education to violence. Other civil
war countries—Cyprus, Yugoslavia, Georgia, and Russia—had
high education rates. Lebanon had one of the highest in the Arab world,
with adult literacy at 60 percent (compared with 15 percent in Iraq) in
the 1950s and 1960s, and school enrollment at 76 percent in the
1950s.56
Saudi Arabia, by contrast, had a
secondary schooling rate of 4 percent (males enrolled), but no war.
To make sense of these contradictions, we need an explanation of
how schooling might influence civil war risk. A close-up look
at what is being taught in schools might help. In many countries, the
curriculum is the primary way of inculcating children with nationalist
ideology. It is not surprising that CH does not focus on this
micro-level mechanism, since the model dismisses the significance of
nationalist ideology as a motive for civil war. But others have shown a
causal link between nationalist education in schools and the
persistence of nationalist ideology.57
This might
explain cases such as Lebanon, where sectarian education fueled war by
nurturing ideologies of intolerance. Indeed, a recent study on
terrorism found that Hezbollah recruits in Lebanon have been drawn from
the ranks of the highly educated, and this pattern also seems to hold
cross-nationally.
58 Krueger and
Malečková 2003. The authors
present micro-level results but also cross-national results consistent
with this finding.
If economic opportunity cost is the
mechanism through which education influences the decision to join a
civil war, then why does this mechanism not apply to the decision to
join a terrorist organization? This puzzle suggests that a refinement
of the theory is necessary to distinguish civil war from terrorist
violence.
Turning to economic growth, several of our countries seem to be
perfect examples of the CH argument. In Sierra Leone, economic growth
was negative before the start of the war in 1991.59
Yugoslavia's growth rate declined 15 percent to 20 percent between
1988 and 1992, fueling social unrest.
60 Kalyvas and Sambanis 2003.
Economic growth was also negative in
the five years before civil war broke out in Senegal, Mali, Azerbaijan,
and other countries in our sample. However, the relationship between
economic growth and civil war is complex. While CH posits a linear
relationship, there are undeniable dynamic effects between growth and
civil war.
First, the effects of economic growth may be channeled through other
variables. Even rapid growth may (indirectly) cause civil wars. In
Indonesia, it indirectly reinvigorated GAM—the rebel Free Aceh
Movement—because it resulted in the expansion of the extractive
resource industries in the region and an increase in the number of
migrants, leading to land seizures in Aceh.61
Thus,
government policies implemented in high-growth periods exacerbated the
risk of war. This example seems to suggest that internal migration is a
mechanism for war outbreak; but migration was the result of a
deliberate government policy of repression, so the cause of violence
was state repression, not migration (or growth) per se.
Second, quantitative studies of civil war fail to account for the
effects of low-level violence that typically precedes war, reducing
both income and growth by reducing investment and encouraging capital
flight. This is particularly true for studies using data sets that code
civil war onset during the year that deaths cross the 1,000 threshold
(as is common in the literature), even though armed conflict may have
been occurring for several years. Quantitative models should therefore
consider modeling the endogeneity of economic growth to armed conflict.
Political violence in Caucasian states caused massive declines in
income. Georgia's GDP per capita dropped from approximately $3,670
in 1991 to somewhere between $777 and $913 in 1997.62
All former Soviet states had drastically falling growth
rates during the collapse of the Soviet Union. It is unclear how much
of the growth decline in Georgia, Azerbaijan, and Chechnya was a result
of the war and how much was a result of Soviet collapse.
In
Azerbaijan, GDP fell from around $4,400 per capita in 1985 to around
$400 in 1996, and then rose to $510 in 1999. In the Democratic Republic
of Congo (DRC)—one of the most war-ravaged countries, with up to
five civil wars—income per capita in the late 1990s was half of
what it had been at the time of independence in 1960. In Burundi,
another country with recurrent civil wars, GDP per capita fell by half
in the 1990s, from $211 in 1991 to $110 in 1999.
63 Ngaruko and Nkurunziza 2002.
If at least some of these declines
were attributable to escalating violence that led to civil war, then we
have a feedback effect that should be properly modeled in quantitative
studies of war onset.
In response to these drastic changes in economic conditions, the CH
theory predicts that the risk of civil war would increase as income
fell. This is consistent with evidence that the risk of war recurrence
is far greater immediately after the end of a war than several periods
later.64
CH finds that the risk of war is
50 percent greater immediately after the previous war ends than in
other time periods.
Declining income can be the mechanism
through which time at war increases the risk of future wars. But if we
interpreted GDP per capita as a measure of state strength, we would
reach a similar conclusion, as falling GDP would imply loss of
strength, which would increase the risk of a new war. This logic might
suggest a specification change in the CH model: adding an interaction
term between GDP and ongoing war to a model of civil war onset would
measure such an effect. Furthermore, the CH model neglects the effect
that ongoing civil war has on the risk of a new war. The authors drop
ongoing periods of war from their analysis, ignoring the feedback
effects mentioned above. If we used the FL coding of the dependent
variable, we could add such an interaction term to control for the
potentially differential effects of some variables during periods of
war as compared with periods of peace. For instance, as argued in the
case study on Georgia, a war's depletion of national income may
discourage a long war duration because the available “loot”
shrinks, making war unprofitable.
65 In short, a close examination of the cases reveals the need to refine
our empirical measures so we can explain the relationship between
economic variables and war onset. The same is true with respect to the
CH and FL arguments on resource dependence.
Resource predation and primary commodity exports
The resource-predation hypothesis is a cornerstone of the CH
“greed” (or “opportunity”) model of civil war.
DRC is perhaps the best example of this argument. Five Congolese
rebellions originated in the resource-rich regions of Katanga, Kivu,
and Kasai.66
Ndikumana and Emizet 2003.
DRC, formerly Zaire, produces 50
percent of the world's cobalt, 30 percent of its diamonds, 20
percent of its copper, and sizable amounts of gold and tin—and
most of this production is concentrated in the eastern provinces.
Mineral exports equal 25 percent of the country's GDP.
However, testing the resource hypothesis is difficult, because CH
uses the ratio of primary commodity exports to GDP—a very distant
measure of resource dependence. CH finds maximized risk of war onset
when primary commodity exports constitute 25 percent to 32 percent of
GDP (depending on whether GDP or education is included in the model).
While this is a useful result (for one thing, it suggests that
diversifying a country's economy may reduce the risk of civil
war), the proxy includes agricultural commodities that are not easy to
loot. A more targeted test would consider the exports individually and
focus on easily lootable usual suspects, such as diamonds, timber, or
gold.67
In a more recent paper, CH
presents such a disaggregated test. See Collier and Hoeffler 2002b.
Testing the resource-predation argument is also difficult because, as
the case studies show, civil wars in resource-dependent countries often
have nothing to do with natural resources. The Maitatsine rebellion in
Nigeria in the 1980s took place in an oil-dependent country, but the
rebellion was not financed by natural-resource rents. Rather, the
rebels—drawn from the ranks of the homeless, the unemployed, and
the politically exiled—were recruited through ideological
indoctrination and Koranic teaching. They used primitive weapons; their
limited finances came from beggars' hoards, small-scale thefts,
and profits from sales of charms and medicines.68
Zinn 2003. This war is omitted
from many data sets, but it meets the case studies' definitional
criteria (ongoing violence from 1980 to 1984 that caused 5,646
deaths).
For predicting civil war in Senegal in the 1990s, the
FL model seems more appropriate than the CH model. Senegal reexports
oil, so although it does not produce a significant amount, it is coded
as an oil exporter; the country also depends on primary commodity
exports. However, “subscriptions” to the movement, not
resource extraction and looting, initially financed its civil war.
69 Humphreys and ag Mohamed
2003. The Mouvement des Forces Démocratiques de Casamance,
led by veterans from the Senegalese army, staged protests in 1983 for
independence of the Casamance region. Violence increased in the late 1980s
and turned into a civil war in the 1990s.
Exploitation of the
cannabis and cashew nut crops in Casamance, Senegal, helped the rebels pay
for the war once it had started, while the army extorted timber from the
region.
In some countries, resource predation is a motive for civil war even
though the CH data set codes them as having low primary commodity
exports. Oil was a key cause of the 1967 Biafran rebellion in Nigeria.
There had been no demands for self-determination of the eastern regions
before the discovery of oil. But once oil was found, Odumegwu Ojukwu,
governor of the eastern region, mandated “that oil revenue be
paid to the regional treasury,” and demands for independence
grew.70
This may explain the CH
model's false negative prediction of the Biafran war.
Nigeria's primary commodity export share of GDP increased to 38
percent between 1990 and 1994. It was 12.3 percent from 1965 to 1969,
much lower than the highest risk level. Zinn 2003.
Using the country/year as the unit of analysis makes it hard to
avoid spurious correlations in CH and FL. Consider Azerbaijan, which
had a nationalist war between Azeris and Armenians from 1991 to
1994.71
The conflict had actually started
earlier, in 1988, when Azerbaijan was still part of the Soviet
Union.
The country is dependent on exports of oil and natural
gas, amounting to 23 percent of GDP in 1999 and 91 percent of total
exports in 2001.
72 On the surface, this
observation seems consistent with CH and FL, since both models
predicted a high risk of civil war in Azerbaijan partly because of its
oil dependence. But war occurred in Nagorno-Karabakh, a region bereft
of natural resources, with a small economy based on agriculture and
food processing. Because econometric tests are based on
country/year analysis, they could not establish the fact that this
conflict was entirely unrelated to oil.
73 Another important question left unanswered by CH and FL is whether
natural resources create motives for war, opportunities to sustain
wars, or both. Although case studies cannot uncover hidden motives,
they can trace the sequence of events that betrays them, and they can
analyze public rhetoric and actions of rebel leaders. In many countries
with no natural resources—Bosnia, Lebanon, Burundi,74
Burundi has a high ratio of primary commodity
exports to GDP because of coffee exports. No study has argued
that control of coffee production is related to the Burundi
war.
Georgia, Mozambique—we nonetheless observe predatory
behavior by rebels.
75 All of these
countries had lower resource dependence than the population
mean.
Theft from houses and small businesses, car jackings,
extortion, and kidnappings for ransom are all ways rebel groups can
finance rebellion. So it is not natural resources per se, but
opportunities for looting, that are widely used to sustain insurgency.
Some resources cannot be exploited unless the rebels have gained
control of the territory (this is the case with oil). One would
therefore expect resource predation to be especially important for
sustaining rebel organizations once the violence has started. Conflict
escalation is likely to occur, then, in oil-dependent countries with
secessionist ethnic groups. The state cannot afford to lose revenue
from oil exports or to let the rebel groups grow by taking control of
oil fields.
Unit Heterogeneity: Ideology, Ethnicity,
and the Organization of Civil War
As the previous section of this article makes clear, the data from
the Case Study Project on Civil Wars demonstrate that the risk of civil
war is not spread evenly across countries. Rich, industrialized
countries are virtually risk-free. Middle-income countries have low or
declining risk. Poor countries face the most risk. While the case
studies' quantitative models control for differences in
development, the reason that countries have different proclivities to
civil war might have more to do with the way other independent
variables, such as ethnicity and democracy, behave at various levels of
income.
Take the United Kingdom, where a high level of economic development
contributes to the CH model's estimate of a 2 percent probability
of civil war from 1970 to 1974, an estimate that is three times lower
than the population average. However, a war did break out in Northern
Ireland in 1971, despite high levels of income.76
Some data sets do not code a civil war in the United
Kingdom. Even though the 1,000-aggregate threshold of deaths has
certainly been exceeded, fighting has been sporadic, which has led to a
slow accumulation of deaths in violence that some label as
terrorism.
Secondary school attendance there was high (in
1970, Northern Ireland was among the top 13 countries in the world), as
was its income level (not far from Britain's), and there were no
natural resources.
77 At the same time, if
the CH model had accounted for high unemployment among Catholic men,
the probability estimate for a rebellion among Catholics would have
increased because of easier rebel recruitment, consistent with
CH's predictions.
But in a sense, all of that may be
irrelevant if the civil war was motivated by religious difference and
fueled by repressive government policies. The opportunity-cost argument
does not apply well to “volunteer forces” such as the Irish
Republican Army (IRA) that are more concerned with ideology than with
looting.
A key reason that CH does not predict this case well may be the fact
that the model pools all types of civil war (as does FL).78
In Sambanis 2001,
I first made the argument that the trade-off between the economic costs
of rebellion and the gains of political and cultural freedom may be
different in pure ethnic conflicts as compared to nonethnic
conflicts.
Thus, CH assumes that the opportunity-cost argument
applies to all civil wars, even wars that are ethnicized. But
ethnicization of a war and other aspects of the organization of
violence might be related to the causes of violence. For example,
Miriam Lowi asserts that economic decline and demographic pressures led
to the emergence of Islamist protest in Algeria in 1992.
79 But several periods of serious
economic decline in that country did
not produce Islamist
backlash—e.g., under Houari Boumedienne (1965–1978). Rather
than economic failure alone, the country's bankrupt political
system (leading to illegal intervention in the 1991
elections)—
combined with increased economic
woes—may have caused the Islamist protest in Algeria.
Some analysts classify wars fought by ethnic groups as ethnic
wars.80
Others argue that ethnicity is just
a cover for economic motives,
81 Collier
and Hoeffler 2001.
personal
hostilities,
82 criminality,
83 or an
assortment of other objectives that are not truly ethnonationalist at
their core.
84 Brubaker and Laitin 1998.
But even though many conflicts can
become ethnicized after they start, and ethnic mobilization can be used
by political elites to support nonethnic rebellions, it may be
significant that ethnicity is mobilized in some wars but not in others.
After all, many rebel groups are organized within ethnoreligious
parameters (in Burundi, recruitment follows tribal lines;
85 Ngaruko and Nkurunziza 2002.
in Lebanon, recruitment and alliance
patterns follow religious lines
86). A common-sense
definition of
ethnic war is a war fought between ethnic groups
over issues that relate to ethnicity. It does not matter whether ethnic
identity can be manipulated by elites pursuing private goals; the fact
that ethnicity lends itself to manipulation and can be used to motivate
collective action is in itself significant. If there is
anything special about ethnic ties, then wars that are
possible only through the mobilization of ethnic identities should be
viewed as distinct from other wars and may have different causes. Wars
over self-determination, for example, should be usefully distinguished
from popular (class-based) revolutions. (Different variables—for
example, regional inequality or the territorial concentration of ethnic
groups—may explain wars over self-determination while being less
relevant for other types of war or other forms of violence, such as
terrorism or coups.)
To avoid making false inferences about the causes of civil war,
before we pool all civil wars, we must establish that there is no
significant unit heterogeneity (i.e., no differences between ethnic and
nonethnic war).87
Another test of the
homogeneity assumption would be to see whether explanatory variables
(e.g., democracy) have different effects on the risk of civil war
during different periods (e.g., pre- and post-Cold War).
To
better predict civil war, we must also explore another type of
heterogeneity: the differences across forms of political violence.
Assume, for a moment, that coups have different causes from civil wars
but that genocides or politicides have similar causes. Current practice
is to include coups in civil war lists when the total number of deaths
is high enough and when the state loses some lives, and to exclude
genocide unless it takes place during civil war (but not right before
or right after it).
88 See Sambanis 2003b for a more formal analysis of the differences
between genocide and civil war.
If many coups are categorized
as civil wars, then our inferences about the causes of civil war will
be conflated with results about the causes of coups.
89 More formally, the coefficient estimates for explanatory
variables in the civil war model may be biased.
By contrast,
if genocides or politicides are excluded from our data, there will
simply be more uncertainty in the estimates, because of the
observations lost. If we assume, however, that both coups and genocides
are different from civil war, as is commonly argued in the literature,
then we can consider whether coups (both failed
and
successful) are causes of civil war. In that case, we would want our
data sets to distinguish a coup's “coup” phase from
its “civil war” phase, but neither our data nor our
definitions are currently fine-tuned enough to do so for all cases.
Genocides might in fact require that we invert the equation, since they
are often consequences of civil war.
90 In
Harff's (2003) list of politicides and
civil wars, I could find only one case (Chile, 1973–1976) that
took place outside civil war.
If so, only a model or narrative
explaining how these two outcomes fit together could show how we shift
from civil war to genocide. These hypothetical examples demonstrate
that, rather than studying civil war by way of a static model and
ignoring the differences between civil war and other violence, we can
gain leverage by identifying and explaining the differences both within
the category of civil war and between civil wars and other types of
political violence.
Along the same lines, in order to grasp the dynamics of civil war, we
must study the growth of rebel organizations. One pattern identified in
our case studies is that most insurgent groups do not have war-fighting
capital (troops and weapons) from the outset; they start small.91
The exception to this rule is when the
national army becomes split between warring groups, as in
Bosnia.
In Colombia, the National Liberation Army grew from 30
men in 1965 to 270 in 1973, and then to 4,500 in 2000.
92 In Azerbaijan, the NK rebels grew
from under 1,000 in 1988 to 21,000 between 1992 and 1994.
93 In Aceh, GAM started with 24 members in 1976 and,
by 2001, had 2,000 to 3,000, plus a militia of 24,000; it came to
control 80 percent of Aceh's villages.
94 In Georgia,
the South Ossettian force went quickly from 300 to 400 volunteers in
early 1990 to 1,500 full-time fighters and 3,500 reservists.
95 The war in Mali started when a small
group of Libyan-trained fighters belonging to the Mouvement Populaire
de Libération de l'Azawad killed four people and took
control of a dozen rifles, which they gradually used to increase their
ranks and military strength.
96 Based on these
examples, one could argue that to get to civil war (as opposed to
smaller-scale violence) violent groups need time to grow. A factor that
facilitates growth is external assistance as well as the state's
inability to effectively repress the insurgency in its early stages.
The government's actions during the early stages of rebellion
are often a critical determinant of escalation to war. A civil war (but
potentially, also, other violence) is likely if the government uses
repressive tactics that reduce the effectiveness of nonviolent protest
while not being sufficiently strong to end all protest.97
See Sambanis and Zinn 2003 for empirical evidence in wars fought over
self-determination.
In several countries, we have observed
conflict escalation that transformed preexisting nonviolent political
groups, political parties, or ethnic organizations into violent rebels.
In Burundi, the army was “a permanent threat” that caused
much violence.
98 Ngaruko and Nkurunziza
2002, 36.
Hutu groups from the 1960s
and 1970s participated in the violence of 1990, although new ones also
formed. In Lebanon, most militias were associated with a preexisting
political party or religious group.
99 In Northern Ireland, a clearly
defined ethnic base supported insurgency, since the IRA recruited
predominantly from the (mostly working-class) Catholic community. The
violent wing of the IRA splintered off a nonviolent movement. In the
DRC, ethnicity formed the basis of five rebellions. The Lunda, Ndembu,
and Yeke groups, for example, led the Katanga secession and Shaba wars,
while the 1996–1997 Kabila rebellion drew its first recruits from
the Banyamulenge.
100 Ndikumana and
Emizet 2003.
The fact that preexisting ethnic organizations are effectively
mobilized to support violence in some cases, but not others, might
offer the basis for making a distinction between ethnoreligious and
other wars. If, however, ethnoreligious affiliation is widely used to
mobilize support for rebellion in different types of wars, establishing
such a typology may be difficult101
—unless we focus on wars
over self-determination. (Such “ethnic wars” are of the
ideal type discussed earlier: they are fought between ethnic groups
over ethnicity.) If some countries are at greater risk of
ethnic war than of nonethnic war (perhaps as a function of their
demographic characteristics), and if ethnic wars are caused by
different combinations of factors than are nonethnic wars, then there
is both analytical and policy interest in establishing a typology of
ethnic war. Case narratives can help us develop such a typology by
offering details on the patterns of recruitment as well as on the
issues over which the war is fought.
The case studies highlight yet another feature of rebel organization
during civil war that economic models have overlooked: forced
recruitment of soldiers. Coercion was widely practiced by the
guerrillas of the (Communist) Democratic Army of Greece from
1947–1949,102
for instance, and by the Lord's
Resistance Army in Uganda in the 1990s. In Burundi, rebel groups
purchased Kenyan street children at the price of $500 for 150 boys.
103 Ngaruko and Nkurunziza 2002.
In Mozambique, the Front for the
Liberation of Mozambique (FRELIMO) used repression, imprisonment,
reeducation, and indoctrination to increase its forces, while the
Mozambican National Resistance (RENAMO) “used force at every
point for almost every purpose.”
104 Weinstein and Francisco 2002.
This pattern creates problems for
the economic opportunity-cost theory, since many recruits do not have
the luxury of deciding independently to join the rebels.
105 The FL model is less compromised by this
finding, because its focus on state strength allows it to claim that a
strong state should be able to prevent forced rebel
recruitment.
A crucial question is whether coercion and material gain together are
enough to motivate political violence. To answer this question, we need
to go beyond CH and FL. Roger Petersen's research, for example,
associates violence with emotional responses to structural change.106
This perspective appears to negate
the presumption of CH and FL that everyone is a potential rebel, given
the opportunity. Rather, Petersen might say, a rebel (at least one who
is not coerced) should have an emotional makeup that differs from that
of someone committed to nonviolence.
I want to emphasize that emotional and economic theories of civil war
can be complementary, not mutually exclusive (at least inasmuch as
emotions are controllable by thought). The relationship between the two
is especially clear in Robert White's analysis of insurgency in
Northern Ireland. White explains the Catholic Nationalists' switch
from peaceful protest to violence as a reaction against the initiation
of internment and violent acts committed by the state.107
Most of the protesters interviewed
by White were outraged by government repression. But those who turned
to violence tended to be working class, unemployed, or in school; by
contrast, those who continued to believe that violence was not the
answer tended to be older, employed, middle-class people with college
degrees—in other words, people who faced higher opportunity costs
if they fought.
Many emotions (e.g., resentment toward injustice or fear of
repression) are clearly consistent with economic models of war, since
some of the models' variables, such as political instability and
economic decline in an ethnically fractionalized society, could elicit
emotions in favor of insurgency. Moreover, emotions can lead people to
pursue private goals, like greater security or more wealth. However,
other emotions (e.g., rage) are inconsistent with economic models
because they generate violent acts that are not necessarily selective
in their targets. Perhaps rage is more closely associated with
terrorism, as terrorism is an example of usually nonselective
violence;108
Krueger and
Malečková 2003.
civil
war, however, has been shown to be selective.
109 One way to combine the emotional and economic theories is to view
emotion-based explanations as focusing on the demand side of the
equation and economic models as focusing on the supply side. As we
develop more of the demand side, it becomes obvious that ideology and
psychology cannot be ignored as explanations of civil war. It may be
the case, however, that ideology, ethnicity, and emotions play
different roles in different forms of violence. To determine whether
this is true, we must look first at what causes violence and then at
how violence takes specific forms.
Reconceptualizing the Dependent Variable
I have been arguing that we cannot understand the causes of civil war
without looking both below civil war (at individual-level violence) and
around it (at different forms of organized political violence). Case
studies provide useful insights on the similarities and differences
among coups, wars, and politicides, and on the pathways that lead from
one to another.
Recurrent cycles of violence
For many countries caught in a conflict trap, civil war is a phase in
a cycle of violence. By isolating civil war in quantitative studies, we
choose to focus on an event rather than a process, and we discard a lot
of useful information that explains how we end up having a civil
war.
It is common to see violent anticolonial movements give way to civil
war (Algeria in 1962, Mozambique in 1976, West Papua in the 1960s,
Sudan in the late 1950s), or to see civil wars grow out of
international wars and occupations (Greece in the 1940s; Yugoslavia in
the 1990s, with an almost 50-year delay since the intra-Yugoslav
fighting during World War II; and, more recently, Iraq).110
On the Greek civil war, see Kalyvas 2000. On the link between Yugoslavia's civil
wars and memories of atrocities since World War II, see Glenny 2001. Iraq in 2004 can be coded as a civil war,
with the United States occupying the government.
Civil wars
can be very bloody coups (as in Costa Rica in 1948, Bolivia in 1952,
and Argentina in 1955) or international wars fought by proxy in a third
country (Lebanon from 1979 to 1991, the DRC from 1998 to 2002). They
can be born out of riots. The 1947 riots in India had both the
organizational complexity and the level of destruction usually found in
a civil war; bloody spikes of riots and pogroms have marred Indian
history ever since. Sometimes, as in Rwanda, civil wars are
indistinguishable from intercommunal fighting, since all politics seems
to be colored by ethnic divisions. In Burundi, nationalist strife in
the 1950s led to ethnic violence after independence had been obtained;
this continued with massacres in 1972 and 1988, and culminated in a
civil war in the 1990s. Nigeria went from a massive civil war in Biafra
in the late 1960s to relative peace in the 1970s, but then experienced
ethnic rioting and massacres and a second bout of civil war in the
1980s. Though Nigeria has not had a civil war in the 1990s, widespread
rioting has killed 10,000 people. Other countries—such as Cyprus,
which was embroiled in violent conflict from the late 1950s until the
mid-1970s—have gone from anticolonial movements to coups, to
civil war, to interstate war. DRC has seen every imaginable form of
political violence since the 1960s, including a possible genocide in
the Kivu area in the 1990s.
Indeed, the reality of civil war is messy, and our decision to code a
period of violence as a civil war (instead of a riot, a coup, or
politicide) hinges on rather vague criteria.111
On the complexities of defining and measuring civil war,
see Sambanis 2003a.
In many
definitions of civil war, a distinguishing feature is effective
resistance—i.e., the state must also suffer. But no coding rule
to date specifies the time period over which the state must incur its
deaths. If it suffers 100 deaths in the first year of a four-year
conflict that causes 30,000 total deaths (with 29,900 on the side of
the opposition), the entire four years might qualify as civil war. But
when state deaths are concentrated in a short period of time, is it not
more accurate to code one phase of the conflict as civil war and the
other as politicide?
The example of Cambodia illustrates the difficulty of distinguishing
between civil war and politicide. In Cambodia, civil war from March
1970 to April 1975 killed, on average, 122,500 people per year. Most
data sets code a politicide that followed immediately after, killing an
average of 347,500 people per year from April 1975 to January 1979. A
second civil war is coded from 1979 until 1991.112
The killing fields in Cambodia (1975–1979) are thus excluded from
civil war lists, even though this period of extreme violence was
closely linked to the war. But fighting (which included Vietnamese
troops) was ongoing around the border during these years. Because of
the low quality of the data, we cannot establish with certainty that
100 or so troops from the stronger side were not killed during that
period; if we were certain about this, we would code an ongoing civil
war in Cambodia along with politicide in 1975–1979. More
importantly, we cannot understand the onset of the new conflict in 1979
as merely a function of Cambodia's GDP per capita or natural
resource dependence—the killing fields are an obvious missing
link here.
Lack of attention to the relationship between civil war and other
forms of violence hurts models like CH and FL, which code
1975–1979 in Cambodia as a period of “peace” (i.e.,
no war). Such periods of “peace” can hide much violence. By
modeling the proximity of other forms of violence, we can indirectly
model the complex links among past riots, coups, and interstate
wars—all of which add to the risk of a future civil war by
destroying property and human capital, undermining economic incentives,
increasing levels of intergroup hostility, and accumulating
conflict-specific capital. Revising the peacetime measure, for
instance, would allow the CH model to make better predictions.113
The CH model under-predicts the risk of
civil war onset in Burundi, Indonesia, Lebanon, and Sierra Leone,
because of the narrow measurement of the “peacetime”
variable.
Beyond this, we can follow a two-pronged approach:
First, if our data do not allow us to make clean distinctions between
forms of violence, we should not throw away cases for not meeting
arbitrary definitions, but rather analyze all events of political
violence as an aggregate category. (The theory must be stated at a
higher level of aggregation in that case—it will be a theory of
political violence.) Second, we can try to improve our definitions and
data coding, and develop models that explain different forms of
violence as well as the transition from one form to another. This
dynamic approach has considerable appeal, as political violence may
itself be understood as a part of a process of state evolution, and
states at different stages of their development might be at risk of
different forms of violence.
Crime, grievance, and politics: The organization of
violence
Another interesting link identified by the case studies is one
between criminal and political violence. In Colombia, for example, 90
percent of the regions in 1995 with the highest homicide rates also had
active guerrilla groups, whereas these groups were active in only 54
percent of the country's total number of regions. Furthermore, 70
percent of these high-homicide regions had substantially high drug
trafficking (as compared with 23 percent of regions nationally).114
Criminal and political violence are both favored by state weakness.
Mafias are organizations designed for extortion, smuggling, and drug
trade, but they can also provide security and authority in areas where
the state has no monopoly over the means of violence.115
Organized crime thus flourished with
the decline of the Soviet state's strength. In the 1990s,
Russia's inability to maintain the prison population led to mass
releases, which led to more crime.
116 Andrienko and Shelley 2003.
Haphazard privatizations increased
the amount of loot, spurring the formation and growth of criminal
gangs, much in the same way that rebel organizations have grown in
Sierra Leone and other resource-rich failing states.
Looting is also a common way of supporting both political and
criminal violence. The form of violence that emerges—Mafia or
rebel group—may be determined by the type of available loot and
the means necessary to appropriate it. If ordinary crime or corruption
is sufficient to acquire the desired amount of loot, then it could
prevail as the organizational mode of looting. If large-scale looting
is needed and if economies of scale in looting can be exploited, then
organized crime will flourish and can coexist with a weak state. If
appropriation of loot requires control of the state or control of
territory (as in the case of oil deposits), then rebel organizations,
particularly secessionist groups, will likely grow. Rebellion may not
be necessary if control of the state can be achieved indirectly through
the exploitation of ethnic or kinship networks, turning the state into
a source of rents for the group.
In sum, organized violence is the result of four interacting factors:
the demand for loot, the demand for political change, the opportunity
to mobilize criminal or insurgent groups, and the mechanisms
(relational, emotional, cognitive, or environmental) that characterize
claim making and resource extraction. And there are important links
between political and criminal forms of violence: while a strong state
can deter the escalation of a conflict to violence, criminal and
political violence can reinforce each other and thereby undermine a
state's authority and capacity. Consider Sierra Leone, where
criminal activity accumulated violence-specific physical and human
capital, and war diverted the state's attention from fighting
crime.117
Over time, the rebels and criminals
became indistinguishable from one another: the Revolutionary United
Front (RUF) recruited and supported illicit diamond diggers while
fighting against the state. Similarly, in Colombia, guerrillas provided
protection for drug cartels, which in turn financed the rebellion. War
economies create constituencies that benefit from war, and violence is
sustained by the same logic of profiteering that supports criminal
activity. Over time, low levels of crime and violence can take control
of a state.
In cases where resources are insufficient to create much public
support for rebellion or where economic gains do not trump emotional or
ideological motives for violence, terrorism may occur. Terrorism can
also grow where large-scale rebellion is likely to be crushed by a
strong state (terrorism might be considered an “incomplete”
civil war). It can feed off civil war, and vice versa. In Egypt,
terrorism against Westerners was the direct result of government
suppression of the al-Gama'at al-Islamiyya, an insurgent group.
The Israeli-Palestinian conflict (some have called it a civil war since
the first intifada, in 1987) was at the heart of international
terrorism before the Oslo Accords of 1997. Kidnappings in Colombia
provided a means for rebels to finance their insurgency. Today, Chechen
terrorism in Russia is an outgrowth of the Russo-Chechen war. In fact,
many Chechen terrorists are “heroes” from the Abkhaz war
against Georgia (including Shamil Basaev, the Chechen prime minister,
who fought in Georgia in 1992 and 1993). As one form of violence feeds
into another over time and across space, we become less able to study
each form in isolation.
Theory Building and Omitted Variables
Just as various forms of violence in the same country can influence
the risk of civil war onset, civil wars themselves may have such an
effect on neighboring countries. The CH and FL models disagree on this
point. While CH finds that a country's civil war risk increases if
its neighbors have had a civil war in the previous year, FL finds no
such evidence. Both models, however, seriously underestimate the
international dimension of civil war. Several case studies point to
ways a “neighborhood” influences a country's war
risk.118
On neighborhood effects, see
Brown 1996; Lake and Rothchild 1998; Sambanis 2001.
Contagion and diffusion
As our cases show, such influence can be seen in the significant
demonstration (diffusion) effects of civil war. A good example is the
Aceh rebellion in Indonesia, where an independence movement simmered
for decades and a brief civil war was quickly suppressed in 1991. The
war reignited in 1999 when, in a climate of political instability and
economic recession, East Timor's referendum on independence
emboldened Acehnese resistance. The onset of mass protest in favor of
independence in Aceh can be traced to November 1999, soon after the
September 1999 referendum in East Timor.119
In Senegal,
the Casamance independence movement was influenced by the ideology of
the independence struggle in Guinea-Bissau.
120 Another sign of neighborhood influence, direct contagion (the
spillover of war across borders), is abundantly evident in the cases
under review. In the Casamance conflict, Guinea-Bissau was used as a
venue for cross-border raids in Senegal, a market for goods, and a
source of arms. Yugoslavia's wars (Croatia in 1991 and from 1992
to 1995; Bosnia from 1992 to 1995; and Kosovo from 1998 to 1999) all
were shaped by the rival irredentist nationalisms of Greater Serbia and
Greater Croatia, and many of the same groups were active in each war.
In the former Soviet republics, wars clustered around the Caucasus in
the early 1990s, “benefiting” from the region's
accumulation of war-specific physical and human capital.121
Sierra Leone's civil war did
not start until Charles Taylor's Liberia provided sanctuary to the
rebels; it was sustained by international crime networks that engaged
in arms-for-diamonds trade. The civil wars in the African Great Lakes
region are perfect examples of contagion. Burundi's and
Rwanda's recurrent wars influenced each other as well as fighting
in DRC, ultimately pulling in Uganda and Zimbabwe. In both Burundi and
Rwanda, wars have occurred between the same two ethnic groups: the Hutu
and the Tutsi. The Rwandan social revolution of 1959 caused a transfer
of power from a Tutsi monarchy to a Hutu majority, leading to massacres
of Tutsi and massive refugee movements, some to Burundi. Tutsi groups
in Burundi feared a similar development, since the Hutu were also the
majority there and the Tutsi sought to consolidate their power over
state institutions, especially security forces.
122 Ngaruko and Nkurunziza 2002.
This ongoing ethnic conflict was at
the core of seven episodes of civil war in the two countries.
Quantitative studies that point to significant neighborhood effects
of civil war have a hard time distinguishing among many possible
diffusion and contagion mechanisms.123
A prominent instrument of
contagion is cross-border ethnic kin.
124 Recent empirical work at the dyadic level suggests that
the presence of common ethnic groups across national borders increases
the risk that domestic ethnic conflicts will become internationalized.
See Woodwell 2004; see also Gurr 2000.
We see this in some of the cases. In
Macedonia, the risk of civil war in the 1990s increased when ethnic
Albanians in Kosovo actively supported Albanian independence across the
border by organizing armed opposition to the Macedonian government.
125 Indeed, some civil wars are better
understood as regional communal conflicts. The wars in Burundi and
Rwanda are really wars between Hutus and Tutsis in the Great Lakes
region, with significant temporal and spatial dependence connecting
civil war outbreaks in these two countries. Another contagion mechanism
is the flow of refugees from one country to another. A large number of
refugees from Burundi and Rwanda to the Eastern Congo threatened the
ethnic-demographic balance of the Kivu region, contributing to conflict
among natives, migrants, and refugees.
126 Ndikumana and Emizet 2003.
An even more common neighborhood
factor is the cross-border trade in small arms.
127 This was one of the most common observations in most of
our case studies.
The end of a civil war creates a surplus of
small arms and lowers their price in the neighborhood.
The examples above suggest that if we want to predict where and when
a civil war will occur, we can no longer afford to ignore the temporal
and spatial dependence of various forms of political violence.
External intervention and internationalized civil war
External intervention is another form of international influence that
many case studies have identified as a key factor in civil war onset.
In Mozambique, DRC, Burundi, Georgia, and several other countries,
external economic and military assistance was critical in both inciting
and supporting rebellion. Jeremy Weinstein and Laudemiro Francisco make
a powerful argument that Mozambique's civil war was largely the
result of South Africa's intervention. When FRELIMO became the new
government in Mozambique, it offered safe haven to all African
liberation movements and threatened the country's neighbors,
Rhodesia and South Africa. FRELIMO's opposition in Mozambique,
RENAMO, initially had a small base of support and amounted to a proxy
war against Zimbabwe African National Liberation Army (ZANLA)
guerrillas. The level of violence dropped markedly in 1979, when
Rhodesian support for ZANLA stopped after the collapse of the Ian Smith
regime.128
Weinstein and Francisco 2002.
RENAMO became incorporated into the
South African Defense Forces, from which it acquired supplies as well
as logistical and technical support, accounting for its tight,
centralized structure.
129 Intervention
cannot succeed without local support. In Mozambique, FRELIMO's
failed socialist agricultural policies, intense repression, and
southern political dominance combined to create a favorable climate for
external agitation to civil war.
Mozambique's experience is not unique. The third Congolese
war—the Shaba rebellion—was the result of an invasion by
Congolese expatriates from Angola.130
Ndikumana and Emizet 2003.
Yugoslavia's ethnic conflict in
Kosovo rose to the level of civil war only after NATO's military
intervention. And earlier, the Bosnian Serbs and Croats would probably
not have had sufficient military resources to wage war in Bosnia
without the support of Serbia and Croatia, respectively.
131 Kalyvas and Sambanis 2003.
In Georgia, Abkhazian resistance
could not have been organized or sustained without direct Russian
assistance.
132 Similarly, the Lebanese war cannot
be understood as distinct from the multiple external interventions and
counterinterventions by the United States, Syria, and Israel. Most of
the local factions in Lebanon represented a foreign government's
interest. In Sierra Leone, persistently high levels of poverty, slow
economic growth, low levels of education, and high dependence on
natural resources did not cause a civil war until soon after the onset
of Liberia's civil war in 1989. Charles Taylor offered Foday
Sankoh, leader of Sierra Leone's RUF, a base from which to mount a
rebellion. Sankoh received his insurgency “schooling” in
Libya.
133 In this and many other wars in
Africa, but also in countries as far away as Indonesia, Libya's
Muammar Gaddafi proved exceptionally meddlesome.
Despite these observations, we have no quantitative research to date
on the relationship between external intervention and civil war onset.
This suggests that an important variable may be omitted from the CH and
FL models. If intervention occurs with high frequency and ends up being
statistically associated with civil war onset, and if the intervention
is also correlated with any of the independent variables in CH or FL,
then the CH and FL parameter estimates will suffer from omitted
variable bias. It would be useful to test for such bias by adding a
variable measuring intervention (or the expectation of intervention) to
the models and re-estimating them.
Conflict escalation and civil war
External intervention is itself usually the result of an escalating
pattern of conflict. To understand the conditions under which
intervention will lead to war, we need to analyze process and not
simply focus on outcome. Case studies can give us a better sense of the
dynamics of conflict escalation by presenting a sequence of
events—a series of actions and reactions—linking several
independent variables together in a process that culminates in war.
They can help us determine whether civil wars occur with little
warning, like earthquakes, or with much buildup, like volcano
explosions.
Some of the most useful insights from the case studies discussed here
come from analyses of the dynamics of conflict in countries where civil
war did not occur despite a large number of risk factors. Consider
Nigeria since the mid-1980s. The risk of war outbreak there was among
the highest in the world, according to both CH and FL. Nigeria has seen
a lot of ethnic rioting over the past 20 years, but not war. Because
the state's response has not been indiscriminately violent,
antistate violence by ethnic rioters did not escalate. (By contrast,
the state's reaction during the Biafran secession in 1967 was
swift and overwhelming, since Nigeria felt threatened by demands for
secession by an oil-rich region.) In the 1990s, a rebellion by the Ijaw
did not grow into a civil war. Because the Ijaw used violence mostly
against other communities and oil companies, the Nigerian government
did not feel sufficiently threatened to respond forcefully. In fact, it
granted concessions to the Ijaw. Whenever the Ijaw targeted the police
or other government institutions, however, the government responded
with decisive force. It is generally easier to gain concessions from
the state if those concessions do not threaten state security; if they
do, the potential for further violence is heightened.
A strong state can afford to be accommodating or repressive, at low
cost.134
But even accommodating policies may
not effectively curb opposition if the state is weak and therefore
cannot uphold its end of the bargain. In Aceh, for instance, the new
democratically elected government's decentralization laws of 1999
were not credible. Because the government did not control the military
and because Indonesia relied on Aceh for oil and gas exports, there was
a significant risk that the government would renege on its promises of
fiscal autonomy for Aceh.
If accommodation does not work, repression is usually next; but as
our case studies show, government repression typically leads to more
opposition and violence.135
In Burundi, the government excluded
Hutus from elite positions and inflicted violence on their leaders.
Over time, this repression led to fewer educational opportunities and
less economic power for Hutus—but it eventually backfired,
resulting in a large-scale Hutu rebellion that entailed Hutu coup
attempts, Tutsi countercoups, Hutu massacres of Tutsi in 1965 and 1972,
and the involvement of the army and ethnic militias. The lack of
democratic governance and the collapse of political and judicial
institutions meant that there was no source of legitimate authority
that could break the cycle of violence.
136 Ngaruko and Nkurunziza 2002.
A strategy of incomplete repression is likely to do more harm than
good, while complete repression by strong states can eliminate the
threat of war. However, in the case of the Muslim Brotherhood (one of
Nigeria's many rebellious groups), war was avoided through
selective repression. The Brotherhood had backing from Iran, was guided
by an antigovernment ideology, and used violent tactics. It called for
an Islamic state, but the arrest of its leader decapitated the movement
in its early stages. Selective repression worked because it was applied
quickly. Nondemocratic states can use selective repression more easily
than democratic ones can to reduce the risk of conflict escalation.
Thus, a government's likelihood of using repression or
accommodation—and the effects of these approaches—may be
determined by state capacity and regime type combined.
The conflict escalation potential of incomplete repression strategies
may explain why democratization increases civil war risk.137
This result is identified in FL, but not
in CH. See also Snyder 2000.
A
democratic or democratizing regime cannot easily use repression,
because the state's enforcement apparatus becomes weaker as its
activities become more transparent. The state is therefore less able to
root out opposition in its early stages. During Indonesia's
authoritarian period, state repression obliterated the GAM, the main
rebel group in Aceh, until the early 1990s. But then a period of
incomplete democratization caused friction between the state and the
military, leading to incomplete repression strategies in Aceh. Since
the state lacked control of the army, human rights abuses in Aceh went
unpunished; this undermined the government's credibility and
elevated popular grievance, which the GAM capitalized on to mobilize
public support and increase its ranks of fighters.
138 In
Senegal—a democratic country—large-scale expropriations of
indigenous land in Casamance began in 1979, and a systematic
denigration of Casamançais followed through the imposition of
Wolof in education, the media, and the administration. The protests of
December 1982 and 1983 triggered a harsh reaction from the state,
helping to radicalize the movement as some of the protesters sought
refuge in the forests and created the maquis (rebel bases). Protesters
who were imprisoned in Dakar started organizing the political wing of
the party.
139 Humphreys and ag Mohamed
2003.
These examples of growing discontent because of failed
democratization or incomplete repression suggest that most cases of
civil war are better explained by the “volcano” (or
escalation) theory rather than by the “earthquake” theory.
Northern Ireland perhaps best illustrates how civil war can result from
a slow but steady escalation of protest. The pivotal event was the
march in Derry/Londonderry on October 5, 1968, when Royal Ulster
Constabulary forces assailed protesters, leading to efforts at partial
appeasement of Catholics with British intervention, and a package of
political concessions to the Northern Ireland Civil Rights
Association.140
The unsubstantial reforms came late
and caused extreme negative reactions by Unionists. Political
instability and protest led to Terence O'Neill's resignation
in 1969 and a victory for extremists, paving the way to the
“Battle of the Bogside,” which marked the start of the
Troubles on August 12, 1969, and the development in 1970 of the
Provisional IRA (PIRA). The PIRA abandoned the strategy of
“abstentionism” that had been used up to that
point—something akin to the peaceful protest of the U.S. civil
rights movement—in favor of a radically militant stand against
Protestants and the British, transforming a disorganized sectarian
protest into an organized campaign for political violence.
Another important piece of the puzzle is that escalation potential
varies across subnational regions and is greater in regions whose
“special status” privileges have been revoked, as in
Casamance (Senegal), Kosovo (Yugoslavia), Aceh (Indonesia), and
elsewhere. In DRC, the Loi Fondamentale overturned long-standing
legislation on minority rights and was seen as a precipitant to war. A
series of nationality laws designed to “protect” the local
population in the Kivu region led the Transitional Parliament to strip
Banyarwanda and Banyamulenge of their Congolese nationality in April
1995. The Banyamulenge refused to leave and turned to Rwanda for help.
Rwanda's intervention led to the massacre of Hutu refugees in
DRC.141
Ndikumana and Emizet 2003.
Escalation risks are also a function of what goes on in nearby
countries. Conflict was brewing all over the neighborhood for years
before the Lebanese civil war erupted. The power of Palestinian
organizations in the country grew after the 1967 Arab-Israeli War
heightened anti-Israeli emotions and Palestinians forged alliances with
Lebanese groups. The civil war can actually be traced back to 1968,
when armed conflict broke out between rival Lebanese groups and between
the government and Palestinian groups that wanted to use Lebanon as a
stage for action in Israel.
These examples suggest that civil wars do not erupt without warning.
The state and challengers go through a process of conflict escalation,
often involving external influences. The process leads to civil war as
the result of extreme demands by the challengers or repression by the
state. The CH and FL models' logic of opportunity structure
applies here, too, but it is only a part of a more complicated picture.
Both repression and economic opportunity figure prominently as
explanations of insurgency in this comment from John Garang, leader of
the Sudan People's Liberation Army:
The burden of neglect and oppression by successive Khartoum
clique regimes has traditionally fallen more on the south than on other
parts of the country. Under these circumstances, the marginal cost of
rebellion in the south became very small, zero or negative; that is, in
the south it pays to rebel.142
John
Garang de Mabior on the founding of the Sudan People's Liberation
Army (SPLA) and the Sudan People's Liberation Movement (SPLM).
Quoted in Ali et al. 2003,
1.
Conclusion
In this article, I have tried to make a substantive contribution to
the literature on civil war by relying on a methodological approach
that combines quantitative and qualitative research. I have briefly
reviewed two major contributions to the field: the CH and FL models of
civil war. These models make fundamentally sound propositions, but
their scope could be expanded and their fit to the data improved.
Elaborating on the interplay between statistics and case studies has
led me to consider the relationship between micro-level and macro-level
explanations of civil war. Economic models such as those reviewed here
offer one of several possible explanations of violence, but they do not
explain why violence will take the form of civil war. I have argued
that micro-level explanations must be aggregated to make sense of civil
war (though I have not presented sufficient evidence to prove this
claim). That is, micro-level theories explain violence in
general, but it is hard for them to distinguish among the many
different motives behind violence in civil war as opposed to violence
in a riot, a coup, or a genocide. Macro-level theories, by contrast,
are most helpful in explaining why, for a given set of private motives,
violence is organized in the form of civil war or something else.
Process-driven explanations, best explored through historical
narratives that focus on the dynamic interaction among actors and
between actors and opportunity structures, can elucidate particular
outcomes and cases.
One difficulty is that this discussion presumes clear definitions of
the various forms of political violence. If our concept of civil war is
muddy, then it will be hard to explain. In the quantitative literature,
this problem is addressed empirically, by running the same model on a
few different definitions of civil war: if the results are the same,
the model is accepted. But if there is no substantive difference
between, say, coups and civil war (at least with respect to the
variables typically included in quantitative models), then all
competing definitions of civil war can be wrong together. Thus, to
understand the interaction between micro-level and macro-level analysis
with a view to explaining civil war, we must first establish and
measure the differences across forms of political violence and identify
the “ontology” of civil war. Ultimately, civil war cannot
be defined only or even predominantly at the micro level. It refers to
a specific organization of violence and is therefore a macro-level
phenomenon. For civil war to have meaning as a distinct category, there
must exist a combination of micro-level motives and macro-level
structures that is unique to civil war.
Case studies can help map the processes that link individual motives
and actions to broader outcomes. In so doing, they may identify
plausible mechanisms or potentially missing variables that can feed
into theories of civil war onset. Case studies may point to causes of
particular wars by systematically comparing several periods within each
country, holding several variables constant across periods. But if too
much is changing in tandem, then the cases can only be used to
highlight processes. Either way, case studies offer theoretical
inspiration, context, and texture that improve formal and quantitative
models.
This article demonstrates one way to combine quantitative and
qualitative methodologies: cases selected after model estimation can be
used to identify measurement error, explore the exogeneity and
homogeneity assumptions in the quantitative model, identify potentially
omitted variables, and discuss causal mechanisms. This knowledge can
help us to refine a theory and retest it in new quantitative analysis,
improving our causal inferences from the model. As our cases suggest,
the role of external assistance (direct intervention or provision of
cross-border sanctuaries), neighborhood effects, and escalation
processes may be important theoretical additions in explaining how
political violence takes the form of civil war. Adding these variables
and others to existing models may let us distinguish civil war from
other forms of violence.
The Case Study Project on Civil Wars has also pointed to ways in
which concepts and operational definitions of variables used in
statistical analysis should be improved. Quantitative studies of civil
war aim primarily at explanation, not prediction (though the two are
obviously related). But most of the variables included in these models
have few time-varying covariates and can pick up mostly cross-sectional
variation, so their capacity to predict the timing of civil war onset
will be limited. Thus, these models should at least make the correct
predictions for the right reasons. We have seen that use of poorly
measured and poorly conceived empirical data sometimes leads CH and FL
to make the right predictions, but for the wrong reasons. If large-N
studies make incorrect assumptions about causal paths, they will lack
explanatory power. Moreover, the inferences drawn from these models
cannot yet inform policy except in a very general and indirect way. We
know that by increasing GDP per capita, we will somehow reduce
the risk of civil war, but a more targeted policy intervention might be
both more effective and easier to implement.
Some of the cases under review have challenged the unit-homogeneity
assumption that underlies current quantitative work. This should prompt
analysts to test for fixed effects by country, region, or period.
Furthermore, periods with no civil war often hide important social
conflicts and violence—a fact that is not properly modeled in
quantitative studies. The path dependence of violence implies that we
must model the transitions across different forms of political
violence, such as riots, genocide, civil war, and terrorism. To better
measure differences across forms of violence, we need to go back to
theory-building and statistical tests. In such tests, the unit of
analysis cannot always be the country and year. Narrowing down the
analysis to the subnational level (the largest administrative region
below the state) might reduce the risk of finding spurious correlations
between natural-resource dependence proxies and the onset of civil war.
Moreover, various units of analysis might be required to answer
questions about the causes of different forms of violence. When
analyzing the risk of secession, focusing on the subnational region is
more profitable than focusing on the country, and it allows us to
introduce new variables to the model, such as precise measures of
interregional inequality, that might help explain the risk of secession
in particular regions but may be unrelated to other forms of
violence.
So my message in this article is not only that economic models of
civil war must be expanded by bringing politics back in. In addition, I
have outlined the need to combine several theories of violence and use
both qualitative and quantitative methods to cumulatively construct a
theory that establishes the boundaries of the concept of civil war.
Instead of modeling simplistic distinctions between “greed”
and “grievance,” our theory-building efforts should be
redirected toward understanding how different forms of violence are
organized during escalating conflict. Such a theory is both more
intellectually satisfying and more policy-relevant, as different
interventions must be designed to address the risks and consequences of
various forms of violence. To develop models to guide our policies, we
must proceed interactively, complementing statistical inference with
in-depth case knowledge.
