Journal ofhttp://jcr.sagepub.

com/
Conflict Resolution

Sensitivity Analysis of Empirical Results on Civil War Onset

Håvard Hegre and Nicholas Sambanis
Journal of Conflict Resolution 2006 50: 508
DOI: 10.1177/0022002706289303

The online version of this article can be found at:

http://jcr.sagepub.com/content/50/4/508

Published by:

http://www.sagepublications.com

On behalf of:
Peace Science Society (International)

Additional services and information for Journal of Conflict Resolution can be found at:

Email Alerts: http://jcr.sagepub.com/cgi/alerts

Subscriptions: http://jcr.sagepub.com/subscriptions

Reprints: http://www.sagepub.com/journalsReprints.nav

Permissions: http://www.sagepub.com/journalsPermissions.nav

Citations: http://jcr.sagepub.com/content/50/4/508.refs.html

>> Version of Record - Jul 13, 2006

What is This?

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

Sensitivity Analysis of Empirical Results
on Civil War Onset

HÅVARD HEGRE
Centre for the Study of Civil War, PRIO
NICHOLAS SAMBANIS
Political Science Department
Yale University

In the literature on civil war onset, several empirical results are not robust or replicable across
studies. Studies use different definitions of civil war and analyze different time periods, so readers cannot
easily determine if differences in empirical results are due to those factors or if most empirical results are
just not robust. The authors apply a methodology for organized specification tests to check the robustness
of empirical results. They isolate causes of variation in empirical results by using the same definition of
civil war and analyzing the same time period while systematically exploring the sensitivity of eighty-eight
variables used to explain civil war in the literature. Several relationships with the onset of civil wars prove
robust: large population and low income levels, low rates of economic growth, recent political instability
and inconsistent democratic institutions, small military establishments and rough terrain, and war-prone
and undemocratic neighbors. Variables representing ethnic difference in the population are robust only in
relation to lower level armed conflict.

Keywords: civil war; internal armed conflict; sensitivity analysis; ethnic fractionalization

1. INTRODUCTION

“A fragile inference is not worth taking seriously” (Leamer 1985, 308). Most
political scientists would agree with that statement. However, most empirical stud-
ies in political science typically do not subject their inferences to rigorous robustness
or specification tests. Replication studies, to some extent, address this issue: in the
empirical study of civil war onset, there is now consensus that the risk of war

AUTHORS’ NOTE: We thank Scott Bennett, Halvard Buhaug, Scott Gates, Nils Petter Gleditsch, and
Håvard Strand for helpful suggestions and Gudrun Østby and Christin Ormhaug for excellent research
assistance. The research was supported by the Polarization and Conflict Project CIT-2-CT-2004-506084,
funded by the European Commission-DG Research Sixth Framework Programme, and by the World
Bank. It reflects only the authors’ views, and the Community and the World Bank are not liable for any
use made of the information contained therein. An online supplement is available at http://jcr.sagepub.com/
cgi/content/full/50/4/508/DC1/ and at http://folk.uio.no/hahegre/publications.htm.
JOURNAL OF CONFLICT RESOLUTION, Vol. 50 No. 4, August 2006 508-535
DOI: 10.1177/0022002706289303
© 2006 Sage Publications

508
Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014
 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 509

decreases as average income increases and the size of a country’s population

decreases. Beyond these two results, however, there is little agreement. For example,
Fearon and Laitin (2003, 85) and Hegre (2003, 26) find mountainous countries to
have a higher risk of war than other countries (because they provide rebels with nat-
ural sanctuaries), but Collier and Hoeffler (2004) and Sambanis (2004a) do not find
this result. Collier and Hoeffler find that a high degree of dependence on primary
commodity exports increases a country’s civil war risk, but few other empirical stud-
ies reach this conclusion.1 There is thus uncertainty about the inferences that we can
make based on empirical results on civil war.
Even though studies such as Fearon and Laitin’s (2003) and Collier and
Hoeffler’s (2004) are exemplary in that they conduct several robustness checks, the
additional analyses presented in any single article on civil war are typically not
enough to sort out which of the discrepant findings are most tenable. Part of the
reason for this is that there is no single method that is clearly better than others to
test the fragility of empirical results.2 Moreover, most scholars’ robustness tests are
ad hoc: they identify a set of competing explanations and see if their empirical
results hold once they control for some variables that might be consistent with those
explanations. While testing different model specifications allows us to discriminate
between rival theories and increases our knowledge about the relationship between
the dependent variable and a set of independent variables, ad hoc specification tests
do not allow us to clearly establish if variation in empirical results is due to differ-
ences in research design, data sources, measurement and operationalization of IVs,
period coverage, or estimation methods.
To establish if small changes to the model specification influence the robustness
of empirical results and the inferences that we can draw from those results, Leamer
(1985, 308) argued that we need a framework for “‘global sensitivity analysis’ in
which a neighborhood of alternative assumptions is selected and the corresponding
interval of inferences is identified. Conclusions [would be] judged sturdy only if the
neighborhood of assumptions is wide enough to be credible and the corresponding
interval of inferences is narrow enough to be useful.” We follow Leamer’s suggestion
and explore two possible sources of variation in empirical results in the civil war lit-
erature: first, we test the sensitivity of commonly cited, substantively important
results to small changes in the set of variables included in a regression. Second, we
test how fragile our substantive inferences are to small changes in the way we opera-
tionalize theoretically significant variables (which we call “concept variables”). We
focus on changes in the statistical significance of parameter estimates, but we also
discuss changes in substantive effects that are due to small specification changes.
The article is organized in five sections beyond the introduction. Section 2 presents
the methodology for sensitivity analysis. Section 3 provides a brief overview of the

1. See, for example, the special issue of Journal of Conflict Resolution (Ron 2005) exploring the link
between natural resources and civil war.
2. From a purely theoretical standpoint, one might reasonably argue that only a single model spec-
ification should be estimated and that this specification should follow directly from the theory. However,
most theories are not comprehensive, and we often need to test rival theories to determine empirically
which one fits the data better.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

510 JOURNAL OF CONFLICT RESOLUTION

range of variation in empirical results in the literature on civil war onset. The large
number of apparently fragile results motivates our analysis. Section 4 presents the
dependent variable. Section 5 details our results and identifies the most robust and
most fragile results in the literature on civil war onset. Section 6 concludes with a
summary of our findings.

2. METHODOLOGY OF SENSITIVITY ANALYSIS

Leamer (1985) referred to the largest possible set of inferences that can be drawn
from a given data set as “extreme bounds.” The size of these “extreme bounds”
depends on the number of models that can be estimated (i.e., variations in model spec-
ifications) within the limits of the data set. These variations must be theoretically con-
sistent and aim to show how minor changes in the list of variables alter the
conclusions of estimation. In Bayesian terms, the extreme bounds approach (EBA)
suggests that the analyst explores the range of posterior distributions that result from
specification changes to the prior distribution of a variable. To credibly identify the
range of inferences that may be drawn from a model, a “global sensitivity analysis”
should be applied, “large numbers of variables should be included, as should differ-
ent functional forms, different distributions, different serial correlation assumptions,
different measurement error processes, etcetera, etcetera” (Leamer 1985, 311). Given
the severe computational burden of such an approach, a reasonable compromise is to
focus on selected dimensions of the model and regression coefficients in particular.
Levine and Renelt (1992) used a variant of Leamer’s (1985) EBA to check the
sensitivity of cross-country regression estimates on the determinants of economic
growth. They were motivated by what they perceived as frequently contradicting
empirical linkages between long-run growth rates and a wide array of explanatory
variables. The EBA used by Levine and Renelt specifies equations of the following
form: Y = βι I + βµ M + βz Z + u, where Y is the dependent variable, I is a set of vari-
ables always included in the regression, M is the “focus” variable (i.e., the one whose
behavior we are interested in observing as we change the model specification), and
Z is a subset of control variables selected among several potentially significant
explanatory variables. They first estimate a model that includes only the I-variables
and the focus variable and then estimate regressions for “all possible combinations
of up to three Z-variables and identify the highest and lowest values for the coeffi-
cient on the variable of interest, βµ, that cannot be rejected at the 0.05 significance
level” (Levine and Renelt 1992, 944). The design tries to reduce multicollinearity
problems by restricting the total number of explanatory variables to “eight or fewer,”
choosing a “small pool of variables from which the extreme bounds procedure
selected Z-variables,” and “excluding variables that, a priori, might measure the
same phenomenon” (Levine and Renelt 1992, 944–5). This specification design
minimizes the risk of underspecified models while also minimizing the computer
power needed to estimate the models, and it reduces problems associated with
multicollinearity. The extreme bounds on the coefficient βµ denote the confidence that
we can have in the partial correlation between Y and M. The upper extreme bound is

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 511

defined as the maximum value of βµ plus two standard deviations, and the lower
extreme bound is βµ minus two standard deviations of the estimate. Coefficient βµ is
considered robust if it “remains significant and of the same sign at the extreme
bounds” (Levine and Renelt 1992, 944).
Levine and Renelt’s (1992, 959) analysis leads them to conclude that “very few
economic variables are robustly correlated with cross-country growth rates.” But this
extreme result may suggest that their analysis sets too high a hurdle for robustness.
According to Sala-i-Martin (1997, 179), “If the distribution of the estimators of β[µ]
has some positive and some negative support, then one is bound to find one regres-
sion for which the estimated coefficient changes signs if enough regressions are run.
Thus, giving the label of nonrobust to all variables is all but guaranteed.” This argu-
ment is reasonable, so to assess the robustness of empirical results in the literature
on civil war onset, we apply Sala-i-Martin’s less stringent test, which involves look-
ing at the entire distribution of parameter estimates to determine the level of confi-
dence in each of the explanatory variables.
Sala-i-Martin (1997) estimates M models of the following form:

γj = αj + βyj y + βzj zj + βxj xj + ε, (1)

where γ is the dependent variable and the subscript refers to the model, y is a vector of
three variables that always appear in the regressions, z is the variable of interest, and x is
a vector of up to three variables taken from a pool χ of available variables. He then uses
these estimates of the βzj to compute the cumulative distribution function—CDF(0)—
the proportion of estimates that are larger or smaller than zero, selecting the largest of
the two. The distribution function is computed under two different assumptions: the
first aggregation assumes that their distribution across models is normal. The average

M
estimate is computed as β̂z = j =1 ωzj βzj , where the weights ωzj are proportional to the
Lzj
integrated likelihoods ωzj = M (Sala-i-Martin 1997, 179). The weights ensure that
i=1 Lzi

models with better fit contribute more to the estimate, and the fact that the same
number of variables is always included in the regression implies that we do not get
“artificially” better fit by increasing the number of variables. Similarly, Sala-i-
Martin (1997) computes the average variance of the estimates, σ̂z2 =
M ωzj σzj2 . If the
j =1

assumption that the βzj are normally distributed holds, the probabilities that β < 0 and
β > 0 can be computed from βˆz and as σ̂2z functions of the average t-ratio βˆz/σ̂ 2z . We
will refer to the smallest of these two probabilities as the “average p-value.” If the
assumption of normality does not hold, then Sala-i-Martin computes the aggregate
CDF(0) from the individual CDF(0)s,


M
ˆ z (0) =
CDF(0) =
ωzj βzj
zj (0|β̂zj , σ̂zj2 ).
j =1

We make a number of adaptations to Sala-i-Martin’s (1997) procedure to fit our

purpose. First, an obvious difference is that, while Sala-i-Martin applied the method

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

512 JOURNAL OF CONFLICT RESOLUTION

to cross-sectional data with a continuous dependent variable, we are using time-series

cross-sectional data, and is the logit-link function of the probability of an outbreak
of civil war. Nonetheless, most papers in the civil war literature treat the data as
cross-sectional since they use pooled logit or pooled probit estimators. Thus, we can
apply Sala-i-Martin’s method directly. Yet the time series in our data may create tem-
poral dependence, which we account for by using a function of “time at peace” as
one of our core variables (Beck, Katz, and Tucker 1998; Raknerud and Hegre 1997).
We include the same three y variables in all regressions and refer to them as our
“core” variables. In a review of the quantitative literature on civil war, Sambanis
(2002) identified the following three core variables that are almost always included
in models of civil war onset: the natural log of population (lnpop), the length of
peacetime until the outbreak of a war (pt8, which we model as a decay function of
time at peace), and the natural log of per capita gross domestic product (GDP) in
constant dollars (lngdp). From a theoretical perspective, income is perhaps the most
important variable in models of civil war and other forms of political violence.
Keeping these three core variables always in the model, we estimate many other mod-
els for each of the variables that we include in the pool χ of relevant variables (we
explain how we selected variables in the next section).
Sala-i-Martin (1997) estimates each possible combination of three x-variables for
each z-variable. We deviate somewhat from this in our treatment of the x-variables.
Our data contain several different measures for the same concept variable (the theo-
retically significant variable). For instance, we can measure “the level of democracy”
using data from the Polity project, from Przeworski et al. (2000), or from Freedom
House. To avoid including variables that measure the same thing in the same model,
we restrict the combination of x-variables to those that measure three different con-
cept variables—not simply different operationalizations of the same concept. If a
control is an alternative operationalization of the same concept as the focus variable,
this is likely to diminish the effect of the focus variable. This might not be an insur-
mountable problem if all concepts had the same number of alternative operational-
izations since any bias would be roughly equal for each focus variable. However, the
number of alternative operationalizations for each concept varies (see next section).
Allowing more than one operationalization for each concept then would tend to hurt
variables belonging to concepts with many operationalizations.3
Another difference from Sala-i-Martin’s (1997) procedure is that we have to allow
for the fact that several variables in our data set have missing data, and the number of
missing observations is different for different variables. In Sala-i-Martin’s model, para-
meter estimates are weighted by the model’s log-likelihood to ensure that models with
better fit to the data are given greater weight. The log-likelihood, however, is dependent
on the number of observations that are included in the estimation. Hence, we calculate
weights based on the likelihood ratio index (LRI), which is analogous to the R2. It is

3. These could be concept variables of great interest in the literature, such as ethnic heterogeneity,
which explains why we have many different ways of measuring them. For example, ethnic heterogeneity
is measured by Vanhanen’s (1999) racial, linguistic, and religious heterogeneity index, or ehet; Fearon’s
(2003) ef index; and the widely used ethnolinguistic fractionalization index, elfo.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 513

given as 1 – LLm/LL0, where LL0 is the log-likelihood with only the constant term, and
LLm is the log-likelihood of the model having just been estimated. This statistic has the
advantage that it is less dependent on sample size, and this is important for us as some
of our variables are missing observations and sample size varies across models.4
A few important variables in the civil war literature cannot be represented in a
single z term; in particular, we cannot use a single variable to capture nonlinear
effects, such as the hypothesized inverted U-shaped association between civil war
risk and democracy level or the parabolic association between civil war risk and
primary commodity export dependence. We enter each of these as individual terms
(so we control, for example, for primary commodity exports and their square) and
report results for the robustness of the individual terms since we cannot evaluate
their joint significance or joint robustness with the method we employ.5
The fact that we perform extensive specification tests implies that we do not know
the true model.6 We only know that three variables should be in the model: GDP per
capita, population size, and time since the previous war. While most scholars would
agree that such a model is underspecified, they would disagree over which other vari-
ables to add.7 The method that we apply here could help provide some information on
what is a better specified regression equation for civil war onset by testing the fit of
several theoretically relevant variables. But it certainly does not replace the need for
theorizing about civil war, and the main usefulness of our approach is to give us a sense
of the distribution of empirical estimates for all potentially relevant variables.
A concern with our approach as well as with all studies of civil war onset is that
some of the variables included in civil war models may be endogenous. We cannot
deal with this problem while trying to explore empirically the robustness of different
model specifications here, and we assume exogeneity for all variables (as do almost
all the studies that we surveyed in this literature). To deal with endogeneity, we would
have to switch estimators for those models where we suspected endogeneity as a result
of the variable combinations in the model. We could not simply use instrumental

4. Fearon and Laitin (2003) suggest that some of the differences between their results and those
from other models may be due to listwise deletion of cases due to nonrandomly missing data in other stud-
ies. Our approach should be less vulnerable to this problem. The LL0 will also change if the number of
observations changes. Also, 90 percent of likelihood ratio indexes (LRIs) fall between .050 and .11, so
even if differences in sample size influence weights, we do not weigh any models too heavily.
5. For related reasons, we leave out interaction terms to simplify the analysis and reduce statistical
output, but the method can be applied while using interaction terms.
6. In the presence of theoretical ambiguity, others have proposed a very different approach, which
prefers parsimony to empirical exploration. Achen (2002), for example, proposes using no more than
three independent variables in regression analyses in the absence of a formal model that justifies the inclu-
sion of more controls. We do not share this view and agree more with much of the literature in develop-
ment and labor economics, which is now moving away from an emphasis on formal models as a
motivation for empirical analysis. While we certainly see the value of mathematical models in political
science, we do not think they represent the only way to theorize about politics or to identify hypotheses
for empirical testing. For a related, useful methodological and applied discussion of this point, see Oneal
and Russett (2005).
7. Scholars often add controls to avoid the risk of omitted variable bias, even though their theory
does not call for additional variables. For a useful methodological perspective on this issue, see Clarke
(2005).

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

514 JOURNAL OF CONFLICT RESOLUTION

variables estimation for all models since, if exogeneity cannot be rejected, this would
reduce the efficiency of the estimates. And we could not hope to identify valid instru-
mental variables for all model specifications. Since very few of the papers in the lit-
erature on civil war deal with the issue of endogeneity,8 we also ignore it and simply
try to reduce the risk by lagging independent variables.9
Finally, there is some concern with how missing data influence the results of this
exercise. Many of the variables in our data set have missing observations. The missing
variables problem is both more and less problematic for our approach as compared to
the standard regression-based analysis of civil war. It is more problematic because we
include in our model variables with a large number of missing observations. However,
given that there is no consensus on which variables to include in a civil war model and
hence no basis for identifying or excluding the possibility of omitted variable bias, this
problem applies in principle to any empirical study of civil war. A potentially more
serious problem is that we aim to isolate the degree to which individual parameter esti-
mates are sensitive to changes in the model specification. Varying the model specifi-
cation, however, leads to changes in the sample when variables are missing data for
different observations. This makes it difficult to identify whether changes in results are
due to changes in the specification or to changes in the sample. We reduce this prob-
lem by omitting from the analysis variables that lead to large losses of cases. This prob-
lem should not bias any of the parameter estimates as long as the reason for missing
observations is not correlated with the dependent variable.10

3. VARIATION IN EMPIRICAL RESULTS ON CIVIL WAR ONSET

A brief overview of the theory and evidence in the literature on civil war onset
explains how we selected the specification of our core model. We include income per
capita as a measure of the economic opportunity cost of the war (Collier and
Hoeffler 2004), or of some aspect of state capacity (Fearon and Laitin 2003), and
other economic factors influencing the decision to rebel. We control for population
size because, for a constant per capita propensity to initiate armed resistance, the
definition of a civil war, which classifies armed conflict as a civil war only if there
is a high threshold of deaths, implies that civil wars are more likely to occur in

8. Exceptions include Sambanis (2000, 2001), Blomberg and Hess (2002), and Miguel, Satyanath,
and Sergenti (2004).
9. Endogeneity may result in artificially good fit between models; hence, weighting by log-likelihood
or the LRI could result in weighting the wrong models more. This problem is limited in magnitude, how-
ever. Ninety percent of the LRIs fall between .050 and .11, and the largest weight given is .182. In addi-
tion, some combinations of variables may present more problems than others. For example, the inclusion
of both economic growth and trade may violate the exogeneity assumption since both trade and growth
are affected by conflict or the expectation of conflict. Variables that measure the degree of militarization
(e.g., military expenditures; size of the government army) may also be endogenous to the expectation of
political violence or, certainly, to ongoing war.
10. Any systematic measurement error in the independent variables would influence parameter esti-
mates and would therefore make it harder to identify robustly significant coefficients.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 515

populous countries. We include time at peace since the last civil war since peace is
likely to exhibit positive time dependence: the longer a country is at peace, the lower
should be the risk of (another) war as conflict-specific capital remains unused and
peace-specific capital is accumulated (Collier and Hoeffler 2004; Hegre et al. 2001).
These three variables are always part of civil war models.
Both theoretical and empirical support for many other explanatory variables is
mixed. There is no theoretical agreement on what is the “right” set of variables to
include in the model, and there is also mixed empirical support for many variables. Our
approach consists of creating categories of “concept variables” that correspond to dif-
ferent theoretical arguments found in at least one study. For each concept, we have sev-
eral empirical proxies that are supposed to measure the theoretically significant
variables. We list all concept categories in Table 1, along with a description of variables
in each category and the sources for each variable. A full discussion of the various the-
ories and empirical results would take us too far, so we refer readers to cited sources for
more information. We only briefly address some of the key debates to motivate the
empirical focus of our article and provide some context for the use of these variables.
A heavily debated question is the effect of ethnic difference (concept 1) on the
risk of civil war. The most commonly used measure for ethnic difference is the now
well-known index of ethnolinguistic fractionalization (elfo).11 Most studies find this
to be nonsignificant, although studies of economic growth and public goods provision
(see, e.g., Mauro 1995; Alesina, Baqir, and Easterly 1997; Easterly and Levine 1997)
find it to be significant.12 Elfo ranges from 0 (ethnic homogeneity) to 100 (extreme
ethnic heterogeneity) and measures the probability that two randomly selected indi-
viduals belong to different ethnolinguistic groups.13 We also include the square term
to capture arguments about nonlinear effects of fractionalization (Collier and
Hoeffler 2004). We assume that elfo is exogenous to civil war.14 Elfo, like all similar
measures, is time invariant and acts much like a country-specific effect.
We also use Vanhanen’s (1999) index of ethnic heterogeneity (ehet), which ranges
from 0 (homogeneity) to 144 (heterogeneity) based on 1990s data and combining
subindices of racial division, linguistic division, and religious division. Vanhanen
measured the size of the largest group as a percentage of total population and
(text continues on p. 521)

11. We use an expanded version of the index that includes data from the original Soviet atlas that
were not included in Taylor and Hudson (1972).
12. This suggests an obvious indirect link between elfo and civil war—ethnic difference can
increase the risk of civil war through its effects on these other variables, such as economic growth and
income. However, no study has yet considered these indirect effects. The same is true for inequality. Most
studies look for a direct linear relationship between economic inequality and civil war. The relationship,
however, may be indirect: inequality may lead to regime transition, which may lead to civil war (see
Sambanis 2005).
13. We do not explore here important questions such as what constitutes a different language, how
groups are identified, what level of disaggregation is used to distinguish different languages, or how bilin-
gual and trilingual groups are coded.
14. Certainly, ethnic identity, or at least the salience of ethnic identity, is shaped by social processes,
including war. However, these are long-term processes, so we maintain an assumption of weak exogene-
ity for ethnic and religious divisions in the short run. In addition, elfo was coded in 1960, the start of our
data set, which further justifies the exogeneity assumption.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 TABLE 1

516
Concept Categories and Variable Names

Variable Name Description Source

Dependent variable warstns War onset Sambanis (2004a)

Core variables uppac Internal armed conflict onset Gleditsch et al (2002)

ln_popns Population, log-transformed Fearon and Laitin (2003)
ln_gdpen GDP per capita, log-transformed Authors’ coding, based on
pt8 2^(-years in peace/8) Sambanis (2004a)

Concept Label Concept No. Variable Name Description Source

Ethnic fragmentation 1 dlang Linguistic component of ehet Vanhanen (1999)

1 drace Racial component of ehet Vanhanen (1999)
1 drel Religious component of ehet Vanhanen (1999)
1 ef Ethnic fractionalization index Fearon and Laitin (2003)
1 ef2 Ef squared Fearon and Laitin (2003)
1 ehet Ethnic heterogeneity index Vanhanen (1999)
1 elfo Ethnolinguistic diversity Collier and Hoeffler (2004)
1 elfo2 Ethnolinguistic diversity, squared Collier and Hoeffler (2004)
1 numlang Number of languages in Ethnologue Fearon and Laitin (2003)
1 plural Share of largest ethnic group Fearon and Laitin (2003)
1 plurrel Size of largest confession Fearon and Laitin (2003)
1 relfrac Religious fractionalization Fearon and Laitin (2003)
Ethnic dominance/polarization 2 etdo4590 Ethnic dominance measure Collier and Hoeffler (2004)

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

2 second Percent population in second largest group Fearon and Laitin (2003)
Level of democracy 3 auto4 Autocracy index from Polity IV Marshall and Jaggers (2000)
3 dem Dummy: 1 for democracies and 0 for Marshall and Jaggers (2000)
autocracies
3 dem4 Democracy index from Polity IV Marshall and Jaggers (2000)
3 exrec Executive recruitment concept variable; Marshall and Jaggers (2000)
Polity IV
 3 mirps2 Consistent autocracy Gates et al. (2006)
3 mirps3 Consistent democracy Gates et al. (2006)
3 parcomp Competitiveness of participation; nonelites; Marshall and Jaggers (2000)
Polity IV
3 parreg Regulation of participation; Marshall and Jaggers (2000)
Polity IV
3 part ln(share of population voting × opposition’s share Vanhanen (2000)
of votes cast)
3 pol4 Polity index; Polity IV Marshall and Jaggers (2000)
3 pol4m Polity Index; Polity IV; 77 & 88 Marshall and Jaggers (2000)
coded = 0
3 polcomp Political competition: concept variable; Marshall and Jaggers (2000)
Polity IV
3 reg Dummy: 1 for dictatorships and 0 for Przeworski et al. (2000)
democracies; ACLP
3 sip2 Continuous measure of democracy Gates et al. (2006)
3 xconst Executive constraints—operational Marshall and Jaggers (2000)
independence of CE; Polity IV
Inconsistency of 4 anoc Dummy: anocracy = 1 Fearon and Laitin (2003)
political institutions 4 mirps0 Inconsistent polity (semi-democracy) Gates et al. (2006)
4 mirps1 Caesaristic polity Gates et al. (2006)
4 partfree Partially free polity Freedom House (2002)
4 pol4sq Pol4 squared Marshall and Jaggers (2000)
Political instability 5 ager Age in years of the current regime as classified Przeworski et al. (2000)

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

by REG; ACLP

(continued)

517
 TABLE 1
(continued)

518
Variable Name Description Source

Dependent variable warstns War onset Sambanis (2004a)

Core variables uppac Internal armed conflict onset Gleditsch et al (2002)

ln_popns Population, log-transformed Fearon and Laitin (2003)
ln_gdpen GDP per capita, log-transformed Authors’ coding, based on
pt8 2^(-years in peace/8) Sambanis (2004a)

Concept Label Concept No. Variable Name Description Source

5 autch98 Autocracy annual change; Polity 98 Gurr and Jaggers (2000)

5 demch98 Democracy annual change; Polity 98 Gurr and Jaggers (2000)
5 durable Years since last regime transition/since 1949; Marshall and Jaggers (2000)
Polity IV
5 Inst3 Political instability; whether there was a change in Fearon and Laitin (2003)
Polity score in the past three years
5 nwstate New state Fearon and Laitin (2003)
5 p4mchg Annual change in modified polity; Polity IV Marshall and Jaggers (2000)
5 polch98 Polity annual change; Polity 98 Gurr and Jaggers (2000)
5 proxregc 2^(-durable/0.5) Gurr and Jaggers (2000)
Political system 6 incumb Consolidation of incumbent advantage; ACLP Przeworski et al. (2000)
6 inst 0-dict; 1-parliam; 2-mixed dem; 3-pres dem; ACLP Przeworski et al. (2000)
6 major Majoritarian system Przeworski et al. (2000)
6 presi Presidential system Przeworski et al. (2000)
Centralization 7 autonomy Country has de facto autonomous regions Sambanis (2004a)

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

7 centpol3 Centralized state? (Polity III data plus updates Sambanis (2004a)
for post-1994)
7 fedpol3 Federal state? (Polity III data plus Sambanis (2004a)
updates for post-1994)
7 semipol3 Semi-federal state? (Polity III data plus updates Sambanis (2004a)
for post-1994)
 Neighborhood political 8 avgnabo Average SIP score of neighbors Gates et al. (2006)
economy 8 nmdgdp Neighbors’ average ln(GDP per capita) Fearon and Laitin (2003);
Sambanis (2001)
8 nmdp4_alt Neighbors’ median polity (both land and water Marshall and Jaggers (2000)
contiguity; using polity2)
8 regd4_alt Median regional polity (using polity2) Sambanis (2002)
Region 9 geo1 Region: Western Europe and the United States Authors’ coding
9 geo2 Region: Eastern Europe and Central Asia Authors’ coding
9 geo34 Region: Middle East and North Africa Authors’ coding
9 geo57 Region: South and East Asia and Oceania Authors’ coding
9 geo69 Region: Latin America Authors’ coding
9 geo8 Region: Sub-Saharan Africa Authors’ coding
Neighborhood war 10 nat_war Whether a neighbor is at war in a given year Sambanis (2004a)
10 tnatwar Total number of neighbors at war in a given year Sambanis (2004a)
Growth 11 gdpgrowth Annual change in GDP, percent Computed from Fearon and Laitin
(2003) GDP series
Economic policy 12 expgdp Exports of goods and services as percent World Bank (2000)
GDP; WDI data
12 trade Trade as percept of GDP; in 1995 constant dollars World Bank (2000)
Social welfare 13 illiteracy Percent adult population illiterate; WDI World Bank (2000)
13 infant Infant mortality; WDI World Bank (2000)
13 life Life Expectancy at birth; WDI World Bank (2000)
13 pri School enrollment, primary, percent gross; WDI World Bank (2000)
13 seceduc School enrollment, secondary, percent gross; WDI World Bank (2000)
Resources 14 agexp Agricultural raw materials exports as percentage World Bank (2000)
of merchandise exports

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

14 fuelexp Fuel and oil products exports as percentage World Bank (2000)
of merchandise exports
14 fuelexp Fuel and oil products exports as percentage World Bank (2000)
of merchandise exports

(continued)

519
 TABLE 1
(continued)

520
Variable Name Description Source

Dependent variable warstns War onset Sambanis (2004a)

Core variables uppac Internal armed conflict onset Gleditsch et al (2002)

ln_popns Population, log-transformed Fearon and Laitin (2003)
ln_gdpen GDP per capita, log-transformed Authors’ coding, based on
pt8 2^(-years in peace/8) Sambanis (2004a)

Concept Label Concept No. Variable Name Description Source

14 manuexp Manufactures exports as percentage of World Bank (2000)

merchandise exports; WDI
14 oil Oil exports/GDP Fearon and Laitin (2003);
Sambanis (2004a)
14 sxpnew Primary commodity exports/GDP Collier and Hoeffler (2004)
14 sxpsq Primary commodity exports/GDP, squared Collier and Hoeffler (2004)
Terrain, geography, 15 lmtnest Rough terrain Fearon and Laitin (2003)
population distribution 15 ncontig Noncontiguous state Fearon and Laitin (2003)
15 popdense Population density: people per square World Bank (2000)
kilometer; WDI
Militarization 16 army85 Size of government army in 1985 World Bank (2000)
16 milper Military personnel in thousands World Bank (2000)
Time 17 decade1 Dummy: 1960s Authors’ coding
17 decade2 Dummy: 1970s Authors’ coding

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

17 decade3 Dummy: 1980s Authors’ coding
17 decade4 Dummy: 1990s Authors’ coding
17 coldwar Code 1 for cold war year—before 1990 Authors’ coding
Colonial war 18 warhist War in the country since 1945? Authors’ coding

NOTE: GDP = gross domestic product; ACLP = Alvarez, Cheibub, Limongi, and Przeworski’s democracy score; CE = chief executive; REG = regime;
SIP = scalar index of polities; WDI = world development indicators.
 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 521

summed across the inverse of these ratios for all three categories (largest racial,
linguistic, religious group).15 We also include variables, coding the number of ethnic,
linguistic, or religious groups from Fearon and Laitin (2003).
Several authors argue that polarization (a small number of large groups) or dom-
inance (one very large group) more frequently lead to civil war than does mere frac-
tionalization (e.g., Collier and Hoeffler 2004; Reynal-Querol 2002). This is our
concept 2 variable. We include two measures of it: Fearon and Laitin’s (2003)
measure of the share of the population in the second largest group and Collier and
Hoeffler’s (2004) dichotomous “ethnic dominance” measure.
The significance of political institutions (concept 3) is also heavily debated. Some
recent studies point to a significant negative association between civil war and the
level of democracy (Esty et al. 1995; Esty et al. 1998; Gurr 2000), although most
find the level of democracy to be nonsignificant (Collier and Hoeffler 2004; Fearon
and Laitin 2003). We have several measures of democracy.
There is more evidence that concept 4, institutional inconsistency or anocracy
(Hegre et al. 2001; Fearon and Laitin 2003), is significant and that it increases the
risk of war, as does concept 5—political instability (Hegre et al. 2001; Fearon and
Laitin 2003; Elbadawi and Sambanis 2002). We include several measures of these
two concepts. Among others, we include a decay function proxregc of the Polity
durable variable, which measures the number of years since an institutional change
that leads to a minimum of three points’ change on the Polity index.
Moving beyond democracy and considering the type of political institution
(concept 6), Reynal-Querol (2002) finds that proportional representation reduces
the risk of civil war. However, she only analyzes ethnic wars and focuses on war
prevalence, not onset. We include her measures along with two measures from
Przeworski et al. (2000).
The degree of political centralization may also influence civil war risk (concept 7).
We include this concept because many civil wars are fought over self-determination
claims, and several authors have argued that more decentralization could reduce the
incentives for violence by peripheral groups (Lijphart 1977; Hechter 2001).
Another debated question is the regional dimension of civil wars. We group these
variables into three concepts: concept 8 refers to the political economy of the neigh-
borhood, including the average income and democracy level of neighboring
countries; concept 9 includes indicator variables for the regions themselves; and
concept 10 measures the extent of war in the neighborhood. Some authors find
important economic spillover effects of civil war (Murdoch and Sandler 2002), and
others (Sambanis 2001) find that civil war in a neighboring country in the previous
year increases a country’s risk of civil war onset, although Hegre et al. (2001) and
Fearon and Laitin (2003) find that these neighborhood spillovers are not significant.

15. Vanhanen’s (1999) computations may have an error: ehet’s lowest value is 0 (perfect homo-
geneity). But if ehet is the sum of the inverse ratio of the largest group over the population, then its small-
est value should be 3 (a country with a single group comprising 100 percent of the population in all three
cleavages should be coded as 1 + 1 + 1 = 3).

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

522 JOURNAL OF CONFLICT RESOLUTION

Similarly, there is disagreement over the effects of living in “bad” neighborhoods—

that is, in regions with mostly poor, autocratic regimes, which some say will increase
the risk of civil war in the neighborhood (Sambanis 2001; Gleditsch 2002). We
explore the empirical robustness of these concept variables.
The role of economic factors has also been the focus of several studies. Several
authors have shown that economic growth (concept 11) reduces civil war risk (Collier
and Hoeffler 2004; Hegre 2003). Results are ambiguous, maybe since the variable
used in most data sets has many missing observations. We use a variable with a more
complete time series.
Some scholars have considered concept 12—the relationship between economic
policy, particularly trade, and civil war (Esty et al. 1995; Hegre, Gleditsch, and
Gissinger 2003; Elbadawi and Hegre 2004). So, we also control for that.
Another concept variable is social welfare (concept 13), which includes infant
mortality (Esty et al. 1998) and education (Collier and Hoeffler 2004). Many schol-
ars argue these should have a separate effect beyond that of income, but the results
are again inconclusive.
The importance of natural resources (concept 14) has figured prominently in the
debate (Collier and Hoeffler 2004; Collier et al. 2003; Fearon and Laitin 2003; Ross
2004; Ron 2005). We include several measures of how important natural resources
are as compared to manufactures.
Other aspects of the country’s demographic characteristics and geography are
also potentially important (concept 15), as they affect the government’s ability to
stage a counterinsurgency campaign. We include measures of the terrain (from
Fearon and Laitin 2003) and a measure of population density.
The degree of militarization of the country and the government’s military strength
(concept 16) may influence the risk of war as governments with strong militaries can
more effectively deter insurgency or repress it before it rises to the level of civil war
(Sambanis 2000). We include a measure of the size of military expenditures and the
size of the government army.
The risk of civil war may also be a function of systemic factors (concept 17). The
end of the cold war is sometimes thought to have unleashed ethnic conflict, although
the evidence is inconclusive (Sambanis 2001; Fearon and Laitin 2003). We control
for the cold war and also use decade dummies to control for time trends.
Finally, we include information on countries’ colonial war history (concept 18) as
some authors have argued that countries with different colonial histories followed dif-
ferent trajectories of institutional development, and this affects their civil war proneness.
The fact that we have several proxies for each concept means that we can gauge
the extent to which our causal inferences about a theoretical concept are specific to
a proxy variable. This is a topic of debate in the literature since scholars have iden-
tified a disjuncture between the theoretically significant variables and some of the
proxies used in empirical tests (Sambanis 2004b). If results for different proxies of
the same concept are inconsistent, this means that we need to define the concept
better and find a more direct measure of it.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 523

4. THE DEPENDENT VARIABLE

Differences in empirical results may also be due to the fact that there are impor-
tant differences in the coding of civil war across studies (Sambanis 2004a).
Depending on which data set we use, we can have twice as many civil war starts and
country-years at war. These differences across civil war lists have been shown to
affect some empirical estimates significantly. We do not combine our sensitivity
analysis with one that checks robustness to changes in the civil war list because this
would make the analysis very hard to follow.
We primarily present results based on Sambanis’s (2004a) civil war definition. A
civil war is defined as an armed conflict between an internationally recognized state
and (mainly) domestic challengers able to mount an organized military opposition to
the state. The war must have caused more than 1,000 deaths in total and in at least a
three-year period. We refer to this variable as “CW” below. We also ran our analysis
using the Uppsala/PRIO list of internal armed conflict (Gleditsch et al. 2002). The
Uppsala/PRIO data set has an annual twenty-five battle deaths threshold and there-
fore is not limited to civil war. The variable is labeled IAC in what follows.16
Our dependent variable (civil war onset) is coded 0 for all country-years with no
war, 1 for the year that a war started, and missing for periods of ongoing war. This
causes us to lose some observations of wars that start while another war is ongoing in
the same country. The alternative way is to code periods of ongoing war as 0s (except
the year of onset), but countries with ongoing wars may have a systematically differ-
ent risk of a new war, and we would need to control for that as well as to consider the
effects of the ongoing war on the other explanatory variables. We therefore drop ongo-
ing war years in the CW analyses.17 For the IAC conflicts, where multiple conflicts
going on in the same country are more frequent, we code ongoing conflict years as 0s.

5. RESULTS OF SENSITIVITY ANALYSIS OF

MODELS OF CIVIL WAR ONSET

We analyze the sensitivity of the eighty-eight variables included in Table 1, orga-

nized by concept category. We cover the same set of countries over the same period
and draw data from Sambanis (2001, 2004a), Hegre et al. (2001), Fearon and Laitin
(2003), Reynal-Querol (2002), and Collier and Hoeffler (2004). We standardized vari-
ables to make estimates easily comparable, and we also log-transformed many vari-
ables to minimize the effect of extreme values.18
The combinations of the eighty-eight variables, following the method presented
in the previous section, produce 4.7 million logistic regressions. Table 2 reports the
results for the three core variables for the CW and IAC dependent variables.
16. See Sambanis (2004a) for more details on the coding rules and explanations for how each case
was coded. See our supplement for complete results based on the Uppsala/PRIO list.
17. We did run the analysis, coding periods of ongoing war as 0s, and there were no significant dif-
ferences. See our online supplement for a table with results.
18. We standardized by subtracting the mean and dividing by the standard deviation.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

524 JOURNAL OF CONFLICT RESOLUTION

TABLE 2
Results for the Core Variables

Estimations Weighted Weighted p Normal (One- p Nonnormal

Variable Label (Sample) Mean βˆz Mean SD (βˆz) Sided Test) (One-Sided Test)

Sambanis civil wars

ln_popns Ln(population) 4,700,394 .280 .090 .00092 .0055
ln_gdpen Ln(GDP 4,700,394 –.504 .197 .0053 .031
per capita)
pt8 2^(-years in 4,700,394 .730 .492 .069 .105
peace/8)

Uppsala/PRIO internal armed conflicts

ln_popns Ln(population) 4,700,394 .399 .093 < .00001 .000085
ln_gdpen Ln(GDP per 4,700,394 –.438 .171 .0052 .032
capita)
pt8 2^(-years in 4,700,394 .248 .341 .234 .175
peace/8)

The standardized coefficient for the population variable had a mean estimate of
.280 over the CW estimations and a mean standard error of .090 (standard errors are
corrected for clustering by country). On average, this corresponds to a t-value of
3.12 and an average p-value of 0.00092. Recall that this statistic is based on the
assumption that the βzj is normally distributed and corresponds to a one-sided test of
statistical significance. The top panel in Figure 1 shows how the t-values are distrib-
uted over the 4,700,134 estimations. We have added vertical lines at t = –1.65 and
t = 1.65 to represent the boundaries of a fairly lenient definition of statistical signif-
icance. The bars inside these lines indicate the frequency of nonsignificant parame-
ter estimates. The distribution of the population variable estimates is quite close to
a normal distribution. If we relax the assumption of the normal distribution of esti-
mates, the p-value is .0055 (see the right-most column in Table 2).
The GDP per capita variable also is robustly significant, with a mean t-value of
–2.5 and an average p-value of .0053 under the assumption of normal distribution.
The middle panel in Figure 1 shows that the estimates have a much more skewed dis-
tribution than the population variable. The mode of the distribution of t-values is
below –3, but in approximately a quarter of the models, the variable is not signifi-
cant at the 5 percent level. Relaxing the distribution assumptions, the average p-
value is .031. The mean parameter estimate is –.504, so income is substantively more
important than population. The estimates for log-transformed variables may be inter-
preted as elasticities, or roughly as the relative change in the odds of conflict when
the variable is increased by 1 percent. A 1 percent increase in income reduces the
risk by .5 percent. The corresponding figure for population is an increase of .28 percent.
This comparison is possible because we have standardized the variables, allowing us to
also interpret the estimates in terms of how much the odds of war change when the
variable is increased by one standard deviation of its own distribution. Decreasing
income by one standard deviation increases the risk of civil war by 65 percent,

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 525

.75
Density
.5
.25
0

-6 -3 0 3 6
Estimated t-values, ln_popns: Log population
.75
Density
.5
.25
0

-6 -3 0 3 6
Estimated t-values, ln_gdpen: Log gdp per capita
.75
Density
.5
.25
0

-6 -3 0 3 6
Estimated t-values, pt8: Decay function of time since previous conflict

Figure 1: Distribution of Estimated t-Values: Core Variables

whereas increasing population by one standard deviation increases civil war risk by
32 percent.
Time since previous war is less robust than the other two core variables. The aver-
age t-statistic is 1.48, corresponding to an average p-value of .07 for a one-sided test.
The bottom panel in Figure 1 shows the distribution of estimated t-statistics.
The lower half of Table 2 shows that the estimates for population size and GDP
per capita are roughly comparable for the IAC dependent variable. The pt8 variable
is not robustly significant in these models.19
Table 3 lists the most robust variables for the CW estimations. The variables are
ordered by the average p-value (normal distribution assumption). The third column
19. This has to do with the coding of ongoing conflict years as 0s. See Hegre (2003) for a discus-
sion of this problem.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 TABLE 3

526
Civil Wars: The Eighteen Most Robust Variables, Sorted by p Nonnormal

Weighted Weighted p Normal p Nonnormal Rank,

Variable Label Concept Mean βˆz Mean SD (βˆz) (One-Sided) (One-Sided) Mean (βˆz)

inst3 Political instability; 5 .448 .098 .00002 .00029 4

whether Polity coded a
change or 77 or 88 in
previous three years
parreg Regulation of participation; 3 –.406 .140 .0018 .0064 6
Polity IV
geo34 Region: Middle East and 9 .289 .113 .0039 .012 13
North Africa
proxregc 2^(-durable/.5) 5 .287 .115 .0049 .019 16
gdpgrowth Annual change in GDP, percent 11 –.548 .223 .0055 .024 2
anoc Dummy: anocracy = 1 4 .299 .122 .0060 .041 11
partfree Partially free polity 4 .325 .131 .0062 .021 8
nat_war Whether a neighbor 10 .318 .141 .010 .025 9
is at war in a given year
lmtnest Rough terrain 15 .294 .136 .014 .019 12
decade1 Dummy: 1960s 17 –.301 .155 .029 .041 10
pol4sq Pol4 squared 4 –.287 .158 .031 .072 15
nwstate New state 5 .236 .134 .035 .065 20
regd4_alt Median regional polity 8 –.416 .240 .035 .061 5
(using polity2)

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

etdo4590 Ethnic dominance measure 2 .235 .142 .038 .065 21
milper Military manpower in thousands 16 –.541 .330 .044 .015 3
geo1 Region: Western Europe 9 –.587 .351 .047 .052 1
and the United States
tnatwar Total number of neighbors 10 .172 .107 .049 .074 28
at war in a given year
presi Presidential system 6 –.373 .228 .050 .063 7

NOTE: Sambanis civil war list; ongoing war years omitted from data set. Complete results are given in the online supplement. GDP = gross domestic product.
 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 527

lists the concept category corresponding to each variable. The fourth and fifth
columns report the weighted average of parameter estimates and their standard
errors, respectively. We report the average p-values in column 6 and p-values with-
out the normality assumption in column 7. Table 4 presents the same information for
the Uppsala/PRIO list of internal armed conflict. More detailed results for all eighty-
eight variables can be found in our supplement.
We have identified robust operationalizations for up to twelve concept categories
in Table 3, and eighteen variables have average p-values less than .05 under the least
restrictive assumption (these are one-tailed tests since the hypotheses are usually
directional; only nine variables are robustly significant with a two-tailed test). Since
there are more onsets of Uppsala/PRIO armed conflicts (Table 4), p-values are gen-
erally lower. Twenty variables are robustly significant in this table. Two political
instability variables are among the most robust variables in both tables. The inst3 vari-
able reflects whether there was a change in the country’s Polity score within the three
years prior to the year of observation or whether the country was coded as being in
“transition” or “interregnum” by the Polity project. The inclusion of the transition
cases in the definition of this variable may raise concerns about endogeneity, however.
One of the Polity project’s criteria for coding a country as being in a state “interreg-
num” is a complete collapse of central authority, which is likely to be connected with
civil war outbreak. However, these transition cases form only a minority of the Polity
score changes recorded in these variables and an even smaller proportion of the
changes used in the coding of the proxregc variable, which is also significant and
robust and increases the risk of civil war onset.20 The proximity of regime change
(proxregc) is also very robust using the Uppsala/PRIO definition (Table 4). Concept
5 variable nwstate (a dummy variable for the two first years of independence) is also
among the robust variables (it is not robust for internal armed conflict).21
The second most robust variable in Table 3 is one of the “level of democracy”
variables (concept 3). Parreg is one of the subindicators forming the Polity IV
Project’s Democracy index, and it reflects the extent to which political participation
is regulated. However, endogeneity concerns are even more serious here than in the
instability variables since Polity codes countries with extensive political violence as
having nonregulated participation.22 Other “level of democracy” variables are not
robust, but several “inconsistency of political institutions” variables are quite robust
(e.g., partfree, anocracy).

20. We reemphasize that our interest is to map the distribution of parameter estimates in the quan-
titative literature as it now stands; almost all of the literature assumes exogeneity of political variables,
and if we change the estimator, these results will also probably change.
21. New state is a theoretically ambiguous variable. The particularities of new states are likely to be
captured by several of the other variables since new states are more likely to have instability, be anocra-
cies, and be more ethnically fractionalized. See Sambanis (2004a).
22. We reran the analysis, excluding the variables that are affected by this problem: parreg, par-
comp, anoc, and pol4sq. The list of the most robust variables changes only marginally when these vari-
ables are omitted. The most notable changes are that the p (normal) increases from .0041 to .0082 for the
prxregc variable and from .049 to .072 for the tnatwar variable. All other results change only marginally.
This exercise increases our confidence that endogeneity problems do not overly distort the weights we
gave to individual estimations (cf. fn. 9). Complete results are given in the online supplement.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 528
TABLE 4
Internal Armed Confluct: The Twenty Most Robust Variables, Sorted by p Nonnormal

Weighted Weighted p Normal p Nonnormal Rank,

Variable Label Concept Mean βˆz Mean SD (βˆz) (One-Sided) (One-Sided) Mean βz

ehet Ethnic heterogeneity index 1 .388 .108 .000 .001 3

inst3 Political instability; whether 5 .344 .080 .000 .000 8
Polity coded a change or
77 or 88 in previous three years
gdpgrowth Annual change in GDP, percent 11 –.436 .093 .000 .001 2
dlang Linguistic component of ehet 1 .382 .121 .001 .002 4
oil Oil exports/GDP 14 .309 .114 .003 .007 9
decade4 Dummy: 1990s 17 .292 .106 .003 .014 11
ef Ethnic fractionalization index 1 .364 .142 .005 .009 5
parreg Regulation of participation; Polity IV 3 –.261 .108 .007 .020 15
elfo Ethnolinguistic diversity 1 .348 .149 .010 .017 7
proxregc 2^(-durable/.5) 5 .198 .086 .011 .024 23
ef2 Ef squared 1 .264 .129 .019 .028 13
coldwar Code 1 for cold war year—before 1990 17 –.215 .100 .019 .056 21
plural Share of largest ethnic group 1 –.264 .131 .021 .031 14
autonomy Country has de facto autonomous regions 7 .076 .039 .021 .063 60
milper Share of population in military forces 16 –.451 .220 .021 .036 1
drel Religious component of ehet 1 .232 .114 .023 .048 19
nat_war Whether a neighbor is at war in a given year 10 .207 .101 .023 .047 22

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

elfo2 Ethnolinguistic diversity, squared 1 .268 .141 .029 .040 12
geo34 Region: Middle East and North Africa 9 .215 .114 .033 .066 20
sxpsq Primary commodity exports/GDP, squared 14 .299 .175 .040 .063 10

NOTE: Uppsala/PRIO internal armed conflict list. Complete results are given in the online supplement.
 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 529

The “neighborhood at war” (natwar) variable has robustly positive estimates for
both conflict variables, lending support to hypotheses regarding the significance of
diffusion and contagion effects in civil war, although we cannot distinguish among
these different mechanisms through which civil war influences are transmitted across
countries. The tnatwar version also makes the top eighteen in Table 3. Economic
growth (gdpgrowth) is robustly negative in both tables, lending support to Collier and
Hoeffler’s (2004) “opportunity cost” theory of civil war. The regional democracy
level (regd4_alt) is also robustly negative for the CW variable, suggesting that living
in a bad neighborhood may be more important than a low in-country democracy level
(Sambanis 2001). The rough terrain variable (lmtnest) is robustly positive for the CW
variables. This highlights an important difference if we compare our specification
tests to results obtained by using a single model while varying the measure of the
dependent variable (as in Sambanis 2004a), where rough terrain is not significant.
The military personnel (milper) variable is negatively and robustly correlated
with civil war onset. Countries with large militaries may be better able to deter insur-
gency or repress any opposition before it rises to the level of civil war.
The last two variables are less robust for the IAC dependent variable, however.
Conversely, the square of primary commodity exports, the oil and autonomy vari-
ables, and the 1990s dummy are positive and robust for the IAC variable. There are
several important differences in the robust determinants of lower armed conflict as
compared to civil war, and explaining those differences should be an area of further
research in the literature on political violence. The most striking difference between
the two conflict variables concerns measures of ethnic fractionalization.
Only one measure of ethnic composition in the country has an average p-value
smaller than .05 in Table 3: “ethnic dominance,” as measured by Collier and Hoeffler
(2004). This variable is not robust for the IAC variable. On the other hand, eight ethnic
fractionalization variables are robustly associated with the risk of IAC conflict (Table
4). Apart from the “share of largest ethnic group” (plural), all are positive, increasing
the risk of internal armed conflict. One of the fractionalization variables, ehet, even
ranks as the most robust variable for IAC. None of these, however, are robust in the CW
estimations. The weighted mean β̂z for Fearon’s (2003) ethnic fractionalization
measure is many times higher when we look at all internal armed conflict using
Uppsala/PRIO data relative to the civil war list (this is also true for other measures).
Figure 2 shows that in the vast majority of the estimations, the ethnic fractionalization
variables were not statistically significant for civil war, but they were for all internal
armed conflict according to the Uppsala/PRIO definition (Figure 3). Thus, we have
identified an important puzzle: why might ethnic fractionalization be associated with
low-intensity armed conflict but not full-scale civil war?23 And why does ethnic domi-
nance seem to affect the risk of the most serious conflicts but not the lower level ones?
Finally, a result that has not received much attention is that there may be some
geographic and time effects: the Middle East and North Africa region is robustly
positive, and the 1960s dummy is robustly negative. Period and region effects have
not been adequately explored in the civil war literature.
23. See Sambanis (2004a) for more discussion.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

530 JOURNAL OF CONFLICT RESOLUTION

dlang: linguistic ehet drace: racial ehet drel: religious ehet ef2: ef squared
0 .3 .6 .9 1.2

ef:ethnic fract. index ehet: ethnic heterogenity elfo2: elfo squared elfo: ethnolinguistic diversity
0 .3 .6 .9 1.2
Density

numlang: no. of languages plural: largest ethn. gr. (%) plurrel: largest confession (%) relfrac: religious fractionaliz.
0 .3 .6 .9 1.2

-6 -3 0 3 6 -6 -3 0 3 6 -6 -3 0 3 6 -6 -3 0 3 6
t-values, concept 1: Ethnic fragmentation, by variable name

Figure 2: Distribution of Estimated t-Values for Ethnic Fragmentation: Civil Wars

dlang: linguistic ehet drace: racial ehet drel: religious ehet ef2: ef squared
0 .25 .5 .75 11.25

ef: ethnic fract. index ehet: ethnic heterogenity elfo2: elfo squared elfo: ethnolinguistic diversity
0 .25 .5 .75 11.25
Density

numlang: no. of languages plural: largest ethn. gr. (%) plurrel: largest confession (%) relfrac: religious fract.
0 .25 .5 .75 1 1.25

-6 -3 0 3 6 -6 -3 0 3 6 -6 -3 0 3 6 -6 -3 0 3 6
t-values, concept 1: Ethnic fragmentation, by variable name

Figure 3: Distribution of Estimated t-Values for Ethnic Fragmentation: Internal Armed

Conflict

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 531

The estimates for “weighted meanβˆz” (column 4) show that most of the variables
that are significant in most models also are substantively important. The right-most
column reports the variables’ rank in terms of size of the mean estimateβˆz. Most of
the variables in Table 3 are also among the top twenty variables with large effects
(estimates larger than 0.25). There are some exceptions: military personnel (milper)
ranks fifteenth in terms of average p-value but third in terms of average impact
(mean βˆz). The variable with the largest substantive effect is a dummy for Western
Europe and the United States (geo1). This is also statistically significant and rela-
tively robust. Presidential political systems (presi) also have a large effect, but that
variable is less robust. Other variables are also estimated to be substantively impor-
tant but are rarely significant, as is the case with secondary school enrollment (one
of Collier and Hoeffler’s [2004] key variables), life expectancy at birth, and the share
of manufacturing in exports.24
In our supplement, we present figures displaying the distribution of estimated
t-values for the concept variables that are heavily debated in the literature: ethnic
dominance/polarization, level of democracy, institutional inconsistency, political insta-
bility, and natural resource dependence. The graphs allow a direct comparison of the dif-
ferent proxies for each concept variable. The precise measure used to test the effects of
democracy or instability matters, for example. Many variables that have received atten-
tion in policy debates and theoretical models of civil war are not robust. For example,
nearly all measures of natural resource dependence (one of our concept variables) do
not do well. Only dependence on oil exports may be considered marginally robust.

6. CONCLUSION

Using a methodology developed by Leamer (1985) and Sala-i-Martin (1997), we

assess the robustness of a number of empirical results in the burgeoning civil wars
literature. We find some of those results to be highly sensitive to small changes in the
model specification. These specification changes amount to estimating models that
include all possible seven-variable combinations of variables that conform to a set of
rules set by our methodology and that might be theoretically relevant to explaining
the onset of civil war. Some of the empirical results in the civil war literature are frag-
ile, but others are not, and they are worth taking seriously. We confirm that a large
population and low per capita income increase the risk of civil war, and this is con-
sistent with many studies of civil war. We have found other robust relationships: civil
wars are more likely to occur in countries with recent political instability and incon-
sistent democratic institutions; countries with small militaries and rough terrain;
countries located in war-prone, undemocratic regions; and countries with low rates of
economic growth. Most other variables fail our significance tests on average.

24. All three variables are highly correlated with gross domestic product per capita, which is a core
variable in our analysis. This collinearity may lead to either inflated βz estimates or to deflated t-values.
Our analysis may therefore underestimate somewhat the robustness of these variables.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

532 JOURNAL OF CONFLICT RESOLUTION

It was not our purpose in this article to sort out different mechanisms that under-
lie these significant correlations, and we do not claim to prove or disprove causal
relationships. Rather, we only wanted to establish which correlations are more
believable given the data and estimation methods used by the vast majority of
researchers in this field. By establishing which empirical results are more robust, our
article helps researchers focus their theory-building efforts and identifies empirical
patterns that are worth more explanation. Our analysis gives readers a sense of the
range of variation of parameter estimates when we try different model specifications.
This sort of sensitivity analysis could be carried out in all empirical literatures and
subfields in political science. While it might not be feasible for every empirical study
to conduct as systematic an empirical investigation as the one that we have presented
here, our results indicate that analysts should worry about the fragility of their infer-
ences and should substitute closely related but differently measured variables in their
regressions and add control variables to make sure the results on focus variables are
robust before making causal inferences. There are several possible sources of uncer-
tainty in the empirical literature on civil war that we have not considered, including
uncertainty about the meaning and measuring of civil war. Different estimation
methods are likely to produce different results, and adding interactions may do the
same. All these sources of variability in empirical estimates should be considered.
Our sensitivity analysis is therefore limited as it only explores robustness to changes
in the set of control variables. But other researchers could (should) check robustness
to changes in period and country coverage, sampling and estimation techniques,
and so on.
Our analysis does not replace the need for careful theorizing about civil war. Our
approach may be a complement to theoretically focused studies. We were motivated
by the fact that no study to date has produced a clear theoretical justification for the
model used in econometric tests. We do not know the model of civil war. Although
most scholars agree that there are some variables that belong in the model, there is
disagreement about others. Given the theoretical uncertainty about the causes of civil
war, our analysis can help point to questions that need further scrutiny (e.g., the
results on ethnic fractionalization that we mentioned earlier). It can also help by
identifying empirical patterns that are robust, and we must be able to explain. The
fact that all our variables have been considered by at least one other scholar as the-
oretically relevant for civil wars justifies their inclusion in the model and makes it
hard to ignore robust empirical results. Some of our stronger results have clearly
important theoretical implications that are not well understood. We found strong
neighborhood effects of civil war consistent with Sambanis (2001), while other
major studies (Fearon and Laitin 2003; Fearon 1998) have argued against the rele-
vance of spillover effects. Our empirical approach, however, cannot explain why or
how civil wars spread across neighboring countries. We need more theorizing and
different sorts of empirical tests to uncover the mechanisms underlying this correla-
tion. There could be different—often competing—mechanisms that explain how one
country’s civil war risk spreads around the region. There could be demonstration
effects or diffusion effects, as rebel groups learn by observing others, or there could

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 533

be contagion effects, as rebels spill over borders to incite rebellion by their coethnics
in neighboring countries. Alternatively, the geographic clustering of civil wars may
be due to shared attributes in countries with common borders. Our analysis points to
the need to better understand the international dimensions of civil war.
Our results strongly support some, but not all, aspects of economic theories of
civil war. We find oil export and primary commodities dependence to be robustly
associated with lower level armed conflict only and not with civil war, contrary to
arguments made by major studies of civil war. Other commonly used measures of
resource dependence are not significant. The problem here may be more with the
available proxies than with the underlying theory connecting resource dependence to
civil war proneness. The key variables in economic theories—per capita income and,
to a lesser extent, the rate of growth of income—are very robust.
But while our results on income and growth agree with Collier and Hoeffler’s
(2004) theory, we also find that political variables—institutional instability, incom-
plete democracy, and undemocratic neighborhoods—are very important and
increase the risk of civil war, as argued in some recent studies (Sambanis 2001). By
contrast, states with strong militaries are able to effectively deter or preempt civil
war. These results are consistent with political theories of civil war (Hegre et al.
2001). The effects of neighborhood variables and political variables may have to be
considered together: bad neighborhoods—those full of autocracies and civil war—
might make democratic transitions in individual countries harder (cf. Gleditsch
2002). Thus, in undemocratic regions, we might be more likely to see democratiza-
tion efforts resulting in prolonged periods of institutional instability, which we have
argued are associated with increased risk of civil war. This conjecture is worth inves-
tigating further as we expand our theoretical focus to consider the spatial dimensions
of civil war.

REFERENCES

Achen, Christopher. 2002. Toward a new political methodology: Microfoundations and ART. Annual
Review of Political Science 5:423-50.
Alesina, Alberto, Reza Baqir, and William Easterly. 1997. Public goods and ethnic divisions. Working
Paper No. 6009, NBER.
Beck, Nathaniel, Jonathan N. Katz, and Richard Tucker. 1998. Taking time seriously: Time-series-cross-
section analysis with a binary dependent variable. American Journal of Political Science 42:1260–88.
Blomberg, S. Brock, and Gregory D. Hess. 2002. The temporal links between conflict and economic
activity. Journal of Conflict Resolution 46:74–90.
Clarke, Kevin A. 2005. The phantom menace: Omitted variable bias in econometric research. Conflict
Management and Peace Science 22:341-52.
Collier, Paul, Lani Elliott, Håvard Hegre, Anke Hoeffler, Marta Reynal-Querol, and Nicholas Sambanis.
2003. Breaking the conflict trap: Civil War and development policy. World Bank Policy Research
Report. Oxford, UK: Oxford University Press/World Bank.
Collier, Paul, and Anke Hoeffler. 2004. Greed and grievance in civil war. Oxford Economic Papers
56:563–95.
Easterly, William, and R. Levine. 1997. Africa’s growth tragedy: Policies and ethnic divisions. Quarterly
Journal of Economics 112:1203–50.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

534 JOURNAL OF CONFLICT RESOLUTION

Elbadawi, Ibrahim, and Håvard Hegre. 2004. Globalization, economic shocks, and armed conflict. Paper
presented at the “Globalization, Territoriality and Conflict” conference, January 16-18, University of
California, San Diego. http://folk.uio.no/hahegre/papers.
Elbadawi, Ibrahim, and Nicholas Sambanis. 2002. How much war will we see? Explaining the prevalence
of civil war. Journal of Conflict Resolution 46:307–34.
Esty, Daniel C., Jack Goldstone, Ted Robert Gurr, Barbara Harff, Pamela T. Wurko, Alan N. Unger, and
Robert S. Chen. 1998. State failure task force report: Phase II findings. McLean, VA: Science
Applications International Corporation.
Esty, Daniel C., Jack A. Goldstone, Ted Robert Gurr, Pamela T. Surko, and Alan N. Unger. 1995. Working
papers state failure task force report. McLean, VA: State Failure Project.
Fearon, James D. 1998. Commitment problems and the spread of ethnic conflict. In The international
spread of ethnic conflict, edited by David A. Lake and Donald Rothchild, 107-26. Princeton, NJ:
Princeton University Press.
———. 2003. Ethnic and cultural diversity by country. Journal of Economic Growth 8:195-222.
Fearon, James D., and David D. Laitin. 2003. Ethnicity, insurgency, and civil war. American Political
Science Review 97:75–90.
Freedom House. 2002. Annual surveys of freedom ratings 1972–73 to 2000–01. http://freedomhouse
.org/ratings.
Gates, Scott, Håvard Hegre, Mark P. Jones, and Håvard Strand. 2006. Institutional inconsistency and
political instability: The duration of polities. American Journal of Political Science 50 (4).
Gleditsch, Kristian Skrede. 2002. All international politics is local: The diffusion of conflict, integration,
and democratization. Ann Arbor: University of Michigan Press.
Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg, and Håvard Strand.
2002. Armed conflict 1946–2001: A new dataset. Journal of Peace Research 39 (5): 615–37.
Gurr, Ted Robert. 2000. Ethnic warfare on the wane. Foreign Affairs 79:52–64.
Gurr, Ted Robert, and Keith Jaggers. 2000. Polity98 project: Regime characteristics, 1800–1998.
http://www.bsos.umd.edu/cidcm/polity.
Hechter, Michael. 2001. Containing nationalism. Oxford, UK: Oxford University Press.
Hegre, Håvard. 2003. Disentangling democracy and development as determinants of armed conflict.
Paper presented at the 44th Annual Convention of the International Studies Association, February
26–March 1, Portland, OR. http://folk.uio.no/hahegre/Papers/DisentanglingWB.pdf.
Hegre, Håvard, Tanja Ellingsen, Scott Gates, and Nils Petter Gleditsch. 2001. Toward a democratic civil peace?
Democracy, political change, and civil war, 1816–1992. American Political Science Review 95:33–48.
Hegre, Håvard, Nils Petter Gleditsch, and Ranveig Gissinger. 2003. Political institutions, globalization,
and conflict. In Globalization and armed conflict, edited by Gerald Schneider, Katherine Barbieri, and
Nils Petter Gleditsch, 271-5. Boulder, CO: Rowman & Littlefield.
Leamer, Edward E. 1985. Sensitivity analyses would help. American Economic Review 75:308–13.
Levine, Ross, and David Renelt. 1992. A sensitivity analysis of cross-country growth regressions.
American Economic Review 82:942–63.
Lijphart, Arend. 1977. Democracy in plural societies: A comparative exploration. New Haven, CT: Yale
University Press.
Marshall, Monty, and Keith Jaggers. 2000. Polity IV project: Codebook and data files. http://www.
bsos.umd.edu/cidcm/inscr/polity.
Mauro, Paulo. 1995. Corruption and growth. Quarterly Journal of Economics 110:681–712.
Miguel, Edward, Shanker Satyanath, and Ernest Sergenti. 2004. Economic shocks and civil conflict: An
instrumental variables approach. Journal of Political Economy 112:725–53.
Murdoch, James C., and Todd Sandler. 2002. Economic growth, civil wars, and spatial spillovers. Journal
of Conflict Resolution 46 (1): 91–110.
Oneal, John R., and Bruce Russett. 2005. Rule of three, let it be? When more really is better. Conflict
Management and Peace Science 21:293-310.
Przeworski, Adam, Michael E. Alvarez, José Antonio Cheibub, and Fernando Limongi. 2000. Democracy
and development: Political institutions and well-being in the world, 1950–1990. Cambridge, UK:
Cambridge University Press.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014

 Hegre, Sambanis / SENSITIVITY ANALYSIS OF CIVIL WAR ONSET 535

Raknerud, Arvid, and Håvard Hegre. 1997. The hazard of war: Reassessing the evidence for the democratic
peace. Journal of Peace Research 34:385–404.
Reynal-Querol, Marta. 2002. Political systems, stability and civil wars. Defence and Peace Economics
13:465–83.
Ron, James. 2005. Paradigm in distress? Primary commodities and civil war [Introduction to special
issue]. Journal of Conflict Resolution 49 (4): 443–50.
Ross, Michael. 2004. What do we know about natural resources and civil war? Journal of Peace Research
41:337–56.
Sala-i-Martin, Xavier X. 1997. I just ran two million regressions. American Economic Review 87:178–83.
Sambanis, Nicholas. 2000. Partition as a solution to ethnic war: An empirical critique of the theoretical
literature. World Politics 52:437–83.
———. 2001. Do ethnic and non-ethnic civil wars have the same causes? A theoretical and empirical
inquiry (Part 1). Journal of Conflict Resolution 45:259–82.
———. 2002. A review of recent advances and future directions in the quantitative literature on civil
wars. Defence and Peace Economics 13:215–43.
———. 2004a. What is a civil war? Conceptual and empirical complexities of an operational definition.
Journal of Conflict Resolution 48:814–58.
———. 2004b. Expanding economic models of civil war using case studies. Perspectives on Politics
2:259-80.
———. 2005. Poverty and the organization of political violence: A review and some conjectures. In
Brookings Trade Forum 2004, edited by Susan M. Collins and Carol Graham, 165-222. Washington,
D.C.: Brookings Institution.
Taylor, Charles Lewis, and Michael C. Hudson. 1972. World handbook of political and social indicators.
3rd ed., vol. 1. New Haven, CT: Yale University Press.
Vanhanen, Tatu. 1999. Domestic ethnic conflict and ethnic nepotism: A comparative analysis. Journal of
Peace Research 36:55–73.
———. 2000. A dataset on measures of democratization. Journal of Peace Research 37:251–65.
World Bank. 2000. World development indicators, 2000. Washington, D.C.: International Bank for
Reconstruction and Development/World Bank.

Downloaded from jcr.sagepub.com at Swinburne Univ of Technology on August 26, 2014