What Is Civil War? Conceptual and Empirical Complexities of an Operational Definition
Author(s): Nicholas Sambanis
Source: The Journal of Conflict Resolution, Vol. 48, No. 6 (Dec., 2004), pp. 814-858
Published by: Sage Publications, Inc.
Stable URL: https://www.jstor.org/stable/4149797
Accessed: 08-11-2018 15:30 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
Sage Publications, Inc. is collaborating with JSTOR to digitize, preserve and extend access to
The Journal of Conflict Resolution
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
What Is Civil War?
CONCEPTUAL AND EMPIRICAL COMPLEXITIES
OF AN OPERATIONAL DEFINITION
NICHOLAS SAMBANIS
Department of Political Science
Yale University
The empirical literature on civil war has seen tremendous growth because of the compila
tive data sets, but there is no consensus on the measurement of civil war. This increases th
inferences from unstable empirical results. Without ad hoc rules to code its start and end and
from other violence, it is difficult, if not impossible, to define and measure civil war. A wide
tion in parameter estimates makes accurate predictions of war onset difficult, and differe
results are greater with respect to war continuation.
Keywords: civil war; Correlates of War; data sets; coding rules
Advances in the empirical literature on civil war depend critically on
and refinement of data on civil war occurrence. Most civil war lists rely
Correlates of War (COW) project.' Since the first COW list was publish
been little peer review of COW coding rules. Within the walls of the C
there has been substantial debate on how to improve data quality. That d
ever, has not benefited from open scholarly discourse.2 Given the proje
nence, many large-N studies use COW data unquestioningly or limit th
making only minor changes to COW data.3 Currently, about a dozen rese
1. Singer and Small (1972, 1994), Small and Singer (1982), Sarkees and Singer (200
(2000).
2. Personal communication (July 31, 2001) with Stuart Bremer, founding member of
of War (COW). COW recently initiated an online forum for public debate, but the forum
3. Walter (2002) and Collier and Hoeffler (2001) use COW data without making any m
Mason and Fett (1996) make only minor changes.
AUTHOR'S NOTE: For comments, I thank Keith Darden, Michael Doyle, James Fear
Gleditsch, HAvard Hegre, Stathis Kalyvas, David Laitin, John Mueller, Monica Toft, Ben V
beth Wood, participants in the Laboratory in Comparative Ethnic Processes (Dartmouth M
2001 ), and participants in the University of British Columbia's workshop on "Econometric
War." I also thank Annalisa Zinn, Douglas Woodwell, Katherine Glassmyer, Ana Maria A
Rose Armentia Bordon for research assistance. All files necessary to replicate the analysis i
be found at http://pantheon.yale.edu/-ns237/index/research.html#Data and www.yale.edu/unsy
JOURNAL OF CONFLICT RESOLUTION, Vol. 48 No. 6, December 2004 814-858
DOI: 10.1177/0022002704269355
O 2004 Sage Publications
814
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 815
have produced civil war lists based on apparently divergent definitions of civi
there is less pluralism here than one might think. Most projects do not condu
historical research and depend heavily on COW. The result may be replic
errors due to the original COW coding rules and uncertainty about whethe
definitions generate different results.
I take a close look at the operational definition of civil war and review se
ing rules used in the literature and analyze their implications for our under
civil war. To make the discussion concrete, I refer to specific cases. This ex
erates insights into our ability to measure civil war and distinguish it from o
of political violence. Drawing on those insights, I propose a new coding rule
a new civil war list. I then measure the impact of differences across codin
regressing the same civil war model on 12 different versions of war onset
lence. Although it may be impossible (and some would argue undesirable) t
a consensus on a single definition and measurement of civil war, it is im
know if our understanding of civil war is affected by different coding rules.
sis shows that some of our substantive conclusions are sensitive to differences in
coding rules, whereas others are remarkably robust.
Significant differences across civil war lists are mainly due to disagreement on
three questions: What threshold of violence distinguishes civil war from other form
of internal armed conflict? How do we know when a civil war starts and ends? How
can we distinguish between intrastate, interstate, and extrastate wars? Answers to
these questions are not only relevant for the purposes of accurate coding, but they also
reveal the degree to which we share a common understanding of the concept of civil
war.
I discuss these issues as a way to explore what differences across codin
tell us about the meaning of civil war. I do not offer new theory that helps
civil war from other forms of violence.4 Rather, I want to see if availab
definitions (coding rules) allow us to measure and analyze civil war as a d
gory with a "natural history" (cf. McAdam, Tarrow, and Tilly 2001), a
that is different from that of other forms of political violence (cf. Tilly
implicitly accept here the premise that civil wars are different from other
consider if the coding rules we have at our disposal are sufficient to m
empirical distinction between civil war and other violence.
I find that it is not possible to arrive at an operational definition of civil
adopting some ad hoc way of distinguishing it from other forms of arm
Although a core set of "ideal" cases of civil war may exist, too many cas
ciently ambiguous to make coding the start and end of the war proble
question the strict categorization of an event of political violence as a
opposed to an act of terrorism, a coup, genocide, organized crime, or i
war.5 In the end, it may be difficult to study civil war without considering
4. I offer such a theoretical discussion in a book-length manuscript, currently in prog
cussion of why we should distinguish civil war from other political violence, see Sambani
5. Proceeding theoretically, rather than empirically, Tilly (2003, 14) makes a similar a
ing that civil war does not have a distinct causal logic. It is a form of "coordinated destruc
that includes various other forms of political violence that generate salient "short-run dam
petrated by highly coordinated actors (the two dimensions that describe Tilly's typology
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
816 JOURNAL OF CONFLICT RESOLUTION
in conflict shift from one form of violence to another, or it may be prof
political violence in the aggregate, rather than cut across that complex
enon with arbitrary definitions. This article is a first step in exploring t
viding an analytical review of existing coding rules that highlights th
accurately defining and measuring civil war. It also improves currentl
ing rules at the margin, so as to produce a civil war list that is more cons
core elements of most working definitions of civil war.
HOW WOULD WE KNOW A
CIVIL WAR IF WE SAW ONE?
In their seminal study Resort to Arms, Small and Singer (1982, 210) d
war as "any armed conflict that involves (a) military action internal to t
(b) the active participation of the national government, and (c) effective
both sides." The main distinction they drew between civil (internal or i
and interstate or extrastate (colonial and imperial) war was the internality
the territory of a sovereign state and the participation of the government
ant. Civil war was further distinguished from other forms of internal arm
the requirement that state violence should be sustained and reciprocated
war exceeds a certain threshold of deaths (typically more than 1,000).
This definition is deceptively straightforward. It is, I will argue, dif
impossible, to develop an operational definition of civil war without a
ad hoc coding rules to distinguish civil wars from other forms of politic
accurately code war onset and termination. First, it is often difficult t
extrastate from intrastate wars: for example, the Russian civil war in Ch
1990s might be considered as a war of decolonization, similar to Came
independence in 1954.
Second, it is unclear what degree of organization is required of the parti
guish a civil war from one-sided, state-sponsored violence. In some cases
government has ceased to exist, but we still code a civil war (e.g., Somali
In other cases, the government may be fighting a war by proxy using mil
ingly disorganized intercommunal clashes (as in Kenya's Rift Valley fr
1993), but most would not classify such a case as a civil war. Elsewhere,
zations are indistinguishable from criminal networks or ragtag militias
Third, if we focus on a numerical threshold of deaths to identify war
deal with the problem of unreliable reporting and incomplete data? Even
data, should termination be coded only on the basis of significantly redu
or should we also focus on discrete, easily coded events, such as peace
Fourth, given that violence during civil war is typically intermitten
determine when an old war stops and a new one starts? And how can we di
end of a civil war from the beginning of a period of politicide, terrorism,
of violence?
The COW project has provided a valuable public good to the profession by coding
cases of war, but it has not offered much guidance on how to answer these difficul
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 817
questions. It may also have caused some confusion. One source of concept
sion is the lack of clarity on the threshold of deaths used to distinguish civ
other violence. Small and Singer (1982, 213) used an annual death thresho
deaths in their early coding efforts, although they later seem to have aban
criterion. Many authors still operate under the assumption that the COW
an annual battle-death criterion.6 This uncertainty stems from the fact tha
ing rules have changed at least three times. The first major change was the
an annual death criterion for civil wars in Resort to Arms, whereas no such r
existed for international wars in The Wages of War 1816-1965 (Singer
1972). Then the annual death criterion for civil wars was relaxed (Singer
1994), but it was still required for extrasystemic wars, although now deat
by nonsystem members were counted (Singer and Small 1994, 5).7 In a th
changes, Sarkees (2000, 129) and Sarkees and Singer (2001) classified
extrasystemic wars as civil wars while abandoning the annual death crite
extrastate wars." It is unclear why the definition of extrasystemic war kept c
if the new rules were used to recode all armed conflicts in empires and
throughout the period covered by COW.9 More important, if data on annual
been collected to code civil wars according to the old (annual) criterion up
revision, why would COW researchers not have made use of those data, inst
ing much of that information by using a binary variable denoting the onset a
nation of civil war? To fully evaluate the research that went into compilin
lists, one needs access to COW's raw data and coding rules.'0
COW's efforts to constantly refine its data and improve its coding rules a
ble. But this process raises the following question: were the new rules co
reapplied to historical data? When the annual death criterion was abandone
COW coders determine the beginning and end of a civil war? Did they cod
year with 1,000 deaths as the onset of the war?" Or did they code the start
that the cumulative death toll surpassed the 1,000 mark? An example tha
6. See Walter (2002), Gleditsch et al. (2002, 617), and Licklider (1995, 682).
7. According to Singer and Small (see codebook, 1994), interstate system membership is
terms of "certain minimal criteria ... at least 500,000 total population and either diplomatic rec
least two major powers or membership in the League of Nations or United Nations." Extrasyst
fought between "a nation that qualifies as an interstate system member ... [and] a political ent
an interstate system member." They are subdivided into imperial and colonial wars. I refer to
and extrastate wars interchangeably, as they both refer to colonial and imperial wars.
8. As I was working on final revisions of this study, I received an e-mail message from
Sarkees (March 15, 2004), stating that the 1,000 battle-death threshold has never been aband
COW project and that poor editing of the Sarkees (2000) article gives readers this mistaken im
eral members of the COW team had read early versions of this study, and they had never correct
over, I have referred to codebooks and publications from COW that counter Sarkees's recent
9. A concern is that information on deaths of colonial subjects might have been systemat
in empires than information on deaths of insurgents in nation-states, particularly in the pos
10. The Wages of War, 1816-1965 and Resort to Arms include appendices on included and
wars, but little, if any, explanation is given for exclusions. Most cases were dropped because
meet the death threshold (Small and Singer 1982, 330).
11. This may be the implicit COW coding rule, according to a communication with Stuar
(July 31,2001). However, Gleditsch et al. (2001, 16) note that "any conflict coded by COW as
than 1,000 battle-deaths overall is recorded as a civil war for all years (even years of inactiv
before the cumulative death toll reached 1,000)"
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
818 JOURNAL OF CONFLICT RESOLUTION
strates potential errors in the application of the new rules is the Algerian
1990s. This war, which some data sets code as starting in 1992, is omi
and Small (1994), even though five other wars that started in 1992 we
revealing that their coding extended through 1992). Perhaps the war
more than 1,000 deaths in 1992 (it actually had), but the revised COW
goes up to 1997 (Sarkees and Singer 2001), includes the Algerian war
date. Because the coding rules were the same in the two COW data sets
a coding error in the 1994 version, the coded start of the war in the
gests that the war reached the 1,000 death mark only after 1992, and the
was then back-coded to the start of the violence in 1992.
The cumulative death criterion added few wars to the COW list. For the period from
1816 to 1979, 13 wars were added to the 1994 list (107 wars were included in the 1982
list). Only 5 of these 13 wars had been included in Appendix B of Resort to Arms
(Small and Singer 1982), and 3 of these 5 were excluded because of lack of system
membership of the participants or because the violence was characterized as a massacre.
Thus, shifting to the cumulative death criterion added only 2 wars to the list.'2
The cumulative death criterion introduces some problems. First, it is harder to
know when to code the start of the war. If we code it the first year the killing begins,
then we will not be able to study violence escalation because the outbreak of minor
violence will be subsumed in the period of "civil war." One research group that tries to
avoid this problem is Gleditsch et al. (2001, 2002), who code a "war" when they count
more than 1,000 deaths in a single year and do not code the beginning of the war during
the first year with more than 25 deaths. But their data set does not have a rule for coding
war termination, so we do not know precisely how to code their wars in a way that is
compatible with other data sets that use the cumulative death criterion. The problem is
created by the so-called "intermediate" armed conflict category, in which there are
more than 100 but fewer than 1,000 deaths in a given year. If we had a conflict with,
say, 600 people killed in the first year, then 3 years with virtually no deaths, and then
another year with 600 deaths, this might qualify as a civil war according to the cumulative
death criterion but would be coded as two distinct events of intermediate violence
in the Gleditsch et al. (2001, 2002) data set.'" The annual death threshold solves this
12. See, again, Sarkees's comments noted earlier. She argues that COW never abandoned the 1,000
annual death threshold. But this raises new questions. For example, the Uppsala data set and the related data
set by Gleditsch et al. (2002), which use an annual death threshold of 1,000 to code civil war, should have few
differences from COW lists if COW used an annual threshold. But, as I show later, the differences between
these data sets are large. Take the example of Cambodia: Singer and Small (1994) code two civil wars, one
from 1970 to 1975 and another from 1979 to 1991. Gleditsch et al. (2002) and Strand, Wilhelmsen, and
Gleditsch (2003) code a first war in 1967, a second war from 1970 to 1975, a third war in 1978, and a fourth in
1989 (these all have different "conflict sub-IDs" in the Strand, Wilhelmsen, and Gleditsch data set; hence,
they are considered as different "cases" or war starts).
13. The data set by Gleditsch et al. (2001, 2002) has undergone many revisions (one current and four
old versions can be found online at http://www.prio.no/cwp/armedconflict [accessed June 22, 2004]). Later
versions have addressed (although not entirely) the problem of coding war onset/termination by assigning a
"conflict ID" and "sub-ID" to each conflict and considering conflicts to be different if or when they switch
from an intrastate to an interstate war (and vice versa), when the parties and issues are different, or when
there are more than 10 years with fewer than 25 deaths per year. However, it is still not clear to me if a "same"
conflict that switches from "war" to "minor" and back to "war" should be considered a single "war" for the
purposes of comparison with war lists that use a cumulative death criterion.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 819
problem because an end to the war would be coded whenever violence dr
1,000 deaths. But this creates the opposite problem of coding too many w
what is essentially the same conflict, if levels of violence fluctuate widely
One way around these problems is to stop trying to code and analyze civ
distinct phenomenon and, instead, to code levels of violence along a cont
country and year). Wars might then be identified as periods with many vio
in addition to other characteristics of the violence, depending on the author
cal argument (e.g., the degree of organization of the parties, the presence o
violence, etc.). The analysis would then try to explain political violence fir
refining the theory, explain why violence takes different forms.
A second problem with the cumulative-death criterion is that it could simp
to code as civil wars many small conflicts that slowly accumulate deaths. In
criterion creates inherently right-censored data, making it hard to code wa
tion. By consistently applying the cumulative death criterion, Sarkees an
(2001) could code as a civil war any minor conflict or terrorist campaign th
deaths per year for more than 40 years. Fearon and Laitin (2003, 76) attemp
this with a rule that 100 deaths must occur every year on average in an on
But, in combination with the 1,000 aggregate-deaths threshold, this create
logical problem. They would not code as a civil war a conflict that caused 9
over 9 years, but they would code a conflict that caused 1,000 deaths over 10 y
A useful example to consider is Northeast India (Nagaland), where F
Laitin (2003) code an ongoing civil war since 1952. I was not able to find ev
many (say, more than 100) deaths per year occurred in armed conflict there
to 1961. According to Gleditsch et al. (2002, appendix) and Small and Sing
339), there was no war or intermediate violence during any year of the
Nagaland. This case illustrates not only the problem of how to code war te
with the cumulative threshold but also the related difficulty of how to han
chronologically overlapping insurgencies in the same country. Combining
concentrated insurgencies in India's Northeast states may be reasonable a
probably satisfy the aggregate-death threshold in the period considered by
Laitin. But a strict application of the cumulative-death rule in such cases i
atic, given that in other countries, chronologically and even geographically
ping insurgencies are often treated as separate conflicts. How to distinguis
these rebellions is not always easy. Burma, Chad, India, Ethiopia, and Zai
1960s) are all countries that pose difficulties in distinguishing among vari
lious groups and periods of violence.15 In the absence of a clear standard of ho
dle such complicated cases, a rule of thumb should be to code a "civil war" i
with many overlapping insurgencies when the violence escalates markedly
the start of low-level hostilities. In the case of India, this means that if we w
14. Moreover, that 100 is the average number of deaths per year means that a conflict wi
of deaths in the first year can be considered ongoing, even if annual deaths after the first year fal
15. For example, Gleditsch et al. (2002) and Strand, Wilhelmsen, and Gleditsch (2003)
between the Serb and Croat rebellions in Bosnia, but most others combine the two insurgencie
"Bosnian war" event. At least five separate rebellions were ongoing from 1960 to 1965 in Zair
cratic Republic of the Congo), and all data sets typically combine these events into a single c
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
820 JOURNAL OF CONFLICT RESOLUTION
bine the rebellions in the Northeast states, a civil war should be coded as sta
1980s, when violence escalated in Assam, Tripuras, and Manipur.
Despite these problems, a cumulative-death threshold becomes more de
it is combined with a criterion that military activity be sustained for the du
conflict. Both Sarkees and Singer (2001) and Fearon and Laitin (2003) appl
but it is unclear how sustained military activity is best defined. To focus
number of deaths per year reintroduces the problems of counting annual
can accurately count 100 deaths per year, we might be able to accurate
higher numbers of deaths and be better off using these numbers as our dep
able. We might instead count battles, requiring at least one battle per year
same parties. But perhaps counting "battles" would create a definitional p
severe than counting deaths, and again we would need an arbitrary "nu
battles" threshold, which would exclude cases of low-level insurgency in w
tion where battles are not easily distinguishable from terrorist activity (as i
Peru's civil war).
To reach a balance between the pros and cons of the absolute versus ann
olds, we must consider a few issues related to how we measure war mag
1. What level of violence qualifies as civil war'?
2. Should this be an absolute or relative level?
3. Should we only count battle deaths or also civilian deaths?
ARRIVING AT AN ABSOLUTE THRESHOLD
OF DEATHS TO CODE A CIVIL WAR
A characteristic of civil war that distinguishes it from other forms of vio
it causes large-scale destruction. A high threshold of deaths can set war
riots, terrorism, and some coups (although not necessarily from pogroms
But there is nothing inherently "right" about the 1,000-deaths threshold us
erature, and strict application of that threshold can cause us to drop seve
satisfy all other characteristics of civil war. A range of 500 to 1,000 dea
principle, be equally consistent with a common understanding of civil wa
that causes major destruction. Using a range rather than a single cutoff poi
better, given the highly skewed distribution of civil war deaths. In 145 c
started between 1945 and 1999, the mean number of deaths is 143,883 (wi
deviation of 374,065), and the median is 19,000. Despite the high average
have caused fewer than 1,500 deaths, and some barely reached 1,000 deat
cases-for example, the Taiwanese revolt in 1947, the Dar ul-Islam rebel
in Indonesia, or the fighting in Croatia after its independence in 1992-s
or all of the other criteria for civil wars: they were fought by well-orga
with political agendas, challenging the sovereign authority, and violence
cal. Given the poor quality of our data (recall how hard it was to count t
dead after the Twin Towers attack on September 11, 2001, in New York
how much harder it would be to measure civil war deaths in Angola in 19
1951) and the skewed distribution of deaths, we should use a more fle
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 821
rule, such as a range of deaths, instead of the 1,000-death threshold.'6 We c
sionally code a war onset at the year that we count 100 to 500 deaths and keep
as a civil war if coders count more than 1,000 deaths in total within 3 year
This incorporates the sustained fighting rule because events with low deaths in
year would be coded as wars only if armed conflict is sustained and cause
more deaths within a 3-year period. The value of this coding rule is that it
be applied to back-code events by integrating existing databases that have in
on lower level armed conflicts.
FOCUSING ON THE RELATIVE MAGNITUDE OF VIOLENCE
Another problem with the absolute threshold is that it does not reflec
magnitude as well as a relative (per capita) measure would. If we add
deaths measure to our coding rules, we would be less likely to miss ar
small nations that produced few deaths but were nonetheless dramati
for the history of those countries. A 1-year conflict that killed 500 p
number in the range suggested above to code war onset) in a country wit
inhabitants (the smallest population size allowed by most coding rules
to deaths at the level of 0.001 of the population. A conflict of equal
country of 20 million people would have caused 20,000 deaths and w
coded as a civil war in all data sets. We could use the 0.001 threshold
and wars that do not exceed the 500 to 1,000 deaths mark could still
civil wars if they meet the per capita deaths criterion.
An example that illustrates the need for greater flexibility in applyin
olds is the Dhofar Rebellion in Oman (1965-1976). This was an insurg
an ethnically organized rebel group (the Dhofar Liberation Front [D
Sultan of Oman. The rebels recruited mainly from the Qara, a small ethn
in the mountains of the Dhofar province, and the group presented itself
Maoist. Young potential recruits were sent to school to receive politica
(Price 1975, 7; Connor 1998, 156-57). There was an element of Islami
ism in the group's ideology until mass defections of Islamist soldiers
Bait Ma'ahshini massacre (Connor 1998, 159). Some scattered terroris
the populous north, but fighting was largely confined to mountainous
were supported by the People's Republic of Yemen in their campaig
mountains (Price 1975, 3). The pattern of armed conflict and the orga
rebellion are consistent with a common understanding of civil war, but t
cally excluded from civil war lists because of a low death count, even
capita terms, this conflict caused more damage than many others tha
included as civil wars.
The 1,000-deaths criterion may lead us to include more cases of civil wars in large
countries, if more populous countries can more easily produce several insurgencies
that can cause high levels of deaths. Ethnic rioting in Nigeria or state repression and
16. In my data set, I include cases that may fall just short of the 1,000 mark and identify them as potentially
ambiguous cases.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
822 JOURNAL OF CONFLICT RESOLUTION
popular dissent in China is much more likely to generate large numbe
than a coup in Fiji or Cyprus. This may be one reason that a variable co
country's population is among the most significant and robust explanat
models of civil war onset. Consider the omission from most data sets of the GrecoTurkish
war in Cyprus that started in 1963. This war was excluded by Singer and Small
(1972, 398; Small and Singer 1982, 340) due to insufficient death count. The conflict
caused around 1,000 deaths and certainly met all other criteria for civil war: there were
battles between organized military units, involving the government against rebel militias,
and parties with local recruitment and a clearly articulated political agenda. Even
if only 500 deaths had occurred in Cyprus in that period, this would have amounted to
0.001 of the population. A war with the same intensity in a country with 100 million
inhabitants would have caused 100,000 deaths-a massive tragedy that would have
been coded in all data sets.
A per capita measure would capture these cases, but we would also need to preserve
a relatively high absolute threshold not to run the opposite risk of selecting too many
small conflicts in small countries (or dropping "relatively" small wars in larger coun
tries). Creating a per capita measure is difficult and labor intensive. What we must no
do is compute per capita deaths in conflicts that have already been selected by COW or
other projects on the basis of an absolute-death threshold. Rather, coding must be
based on primary research to measure deaths on a per capita basis in all countries with
any political violence. A starting point might be the Gleditsch et al. (2001 ) list of minor
armed conflicts. So far, no death figures are available for those conflicts, and we know
only that they have caused more than 25 deaths per year and less than 1,000 for the
duration of the conflict. Getting better data for those conflicts might facilitate the con
struction of a per capita measure. An alternative (and less costly) approach is to code
per capita deaths for a shorter period (e.g., from 1980 onwards), allowing us to construct
a data sample based on the combination of absolute and per capita death measures
(post-1980) and another sample using only the absolute threshold. We could
then estimate a model on the two samples of the post-1980 period and compare the
results. We could then do a sort of out-of-sample test by estimating the same model on
each sample period to see if the model's predictions are consistent across periods using
either sample. If the results are not influenced by our coding rules, then going through
the trouble of coding a per capita measure might not be necessary.
THE EFFECTIVE RESISTANCE CRITERION AND
MEASURING BATTLE DEATHS VERSUS TOTAL DEATHS
The third crucial question is whether we should code battle deaths o
civilian deaths due to the war. The battle-deaths criterion is the legacy of
ject and usually refers to military deaths (the measure was initially us
deaths in interstate wars). Small and Singer (1982, 213) claim that they
ian as well as military deaths" in civil wars, but it is not clear if they hav
civilian deaths due to rebel attacks. Civilian deaths must be counted in death totals.'7
17. This position is gaining ground: see Sarkees and Singer (2001, 12) and Valentino, Huth, and
Balch-Lindsay (2001).
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 823
Civilians are targeted in civil war and are disproportionately affected by h
disasters created by combatants to hold civilian populations hostage and ga
of territory."8 One might also consider counting refugees and internally di
sons as a measure of the human cost of the war (cf. Doyle and Sambanis
To get an accurate measure of civilian deaths, we must also determine if
to politicide that occur in and around civil war should be counted. Small
(1982, 215) counted deaths due to "acts of massacre committed ... during
of a civil war." That seems reasonable, but a harder question is, how do w
deaths due to acts of massacre that flank periods of civil war? Should we c
having started if civilian massacres occur for a period (say, 1 year) and m
frontation ensues with deaths that exceed 1,000 in the next year? What
reverse situation, where a civil war gives way to civilian massacres? Dist
periods of civil war from periods of massacres clearly is very hard, as the c
bodia illustrates. Most data sets code civil war onsets in 1970 and 1978 in Cambodia
but no war from 1976 to 1977, which is a period that corresponds to civilian massacre
(the killing fields). Small-scale battles between the Khmer and Vietnamese-backed
troops did take place near the Thai border in that period, but it is unclear how man
people died as a result of armed conflict. Hence, most data sets code this as a period o
politicide that is distinct from the civil war. Yet, if we could have established that 10
or so people had been killed on the side of the stronger party between 1976 and 1978
we would have coded an ongoing war in that period (especially in data sets using the
cumulative-death criterion and where effective resistance is not measured as a
percentage of the total number of deaths).
In light of these difficulties, a conservative strategy is to count deaths due to massa
cres that occur right before or after the war as war-related deaths. The main difficulty
to find evidence of effective resistance. According to Small and Singer (1982, 214
15), effective resistance implies that the stronger side should suffer at least 5% of t
casualties of the weaker side. Fearon and Laitin (2003) relaxed this coding rule an
measured effective resistance by 100 state deaths. They do not, however, specify
these deaths must occur in the first year of the war or cumulatively throughout the w
Without data on the distribution of state deaths throughout the war, we are left
rudderless in trying to distinguish periods of "civil war" from periods of "politicide" o
"genocide." (Could we say with certainty that 50 government troops were not killed
the jungles of Cambodia in 1976 and another 50 in 1977?) Some find the distincti
easier to make if massacres occur before a "civil war" starts, as in the case of the 19
massacres in Nigeria, which are typically thought of as separate from the Biafran c
war, which started in 1967.19 It might also be easier to distinguish massacres from civ
war if the massacres occur after a formal end to the war is reached (i.e., a peace treaty
But if the war peters out and mostly one-sided violence starts (as in the 1984-19
18. It might even be desirable to count civilian deaths that are indirectly the result of the war as, f
example, in the case of war-related starvation deaths in Ethiopia and Somalia. However, this is a slipper
slope, as it is unclear how far one could trace the health effects of civil war (see Ghobarah, Huth, and Russet
2003 for such an effort).
19. In May-July 1966, massacres of 30,000 Ibos in the Northeast took place, including massacres dur
ing the Northern leaders' coup and murder of Ironsi and Ibo leaders. Biafra achieved de facto independen
by the end of 1966, and war started in 1967.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
824 JOURNAL OF CONFLICT RESOLUTION
massacres in Ndebele, Zimbabwe), the distinction is harder to make,
because war may restart within 1 or more years as a result of such m
Thus, the effective-resistance criterion cannot help us establish with
difference between civil war and politicide, and we would need to us
effective-resistance criterion to make a clear distinction. Just as Fearon and Laitin
(2003) require 100 deaths per year to code an ongoing civil war, we could require 10
state deaths per year to distinguish a civil war from a politicide.20 But this type of arithmetic
seems inappropriate, not least because it trivializes and imposes a petty discipline
on the complex violence that occurs during civil war. What if insurgents suffering,
say, 1,000 deaths per year can kill 60 soldiers or police officers in one year, none in
the next year, and then 5 per year for the next 8 years (for a total of 100 in 7 years)?
Would this be any more or any less of a civil war than a conflict in which 20 soldiers are
killed per year for 5 years? The problem of counting effective resistance each year
might be avoided if we require 100 deaths per year on average for the duration of the
conflict, but this would obviously require a higher threshold of the state deaths criterion
and would lead us to reclassify some civil wars as politicides or genocides.2' What
if we forget about the question of distribution of state deaths over time? Should an
absolute criterion of 100 state deaths be the single most important factor in deciding
whether to classify an insurgency as a civil war if the insurgency meets all other criteria
for a civil war? An example is the Gamaat Islamiya (Islamic Group) and Islamic
Jihad's insurgency in Egypt from 1992 to 1997. According to one source, around
1,200 people were killed over 5 or 6 years, which meets the cumulative death criterion.22
The rebels were organized and had a political ideology. Guerilla insurgency was
sustained, and attacks were directed against police, security forces, government ministers,
tourists, and the Coptic minority. Whether we code this case as terrorism or civil
war currently rests on whether the state suffered 100 or more deaths. This seems a
weak criterion on which to base the classification of any case.23
Another difficult case is Argentina's "dirty war," in which Harff (2003) codes a
politicide from 1976 to 1980, Fearon and Laitin (2003) code a civil war from 1973 to
1977, Singer and Small (1994) do not code a civil war, and Gleditsch et al. (2002) code
a war in 1975 and "possibly" in 1976 and 1977. Official statistics from the Argentine
military cite 492 deaths (including civilian officials) due to "terrorist attacks" from
1969 to 1980.24 Guerillas suffered deaths starting in 1969 until 1972 (more than 100
per year), and it is hard to know how many people were killed between 1973 and 1977,
the period sometimes coded as a "civil war." Total deaths, including guerillas and civil-
20. The 1,000 cumulative-death criterion divided by the 100 deaths-per-year rule gives a 1 0-year window
for minor conflicts to be labeled civil wars. Dividing the Fearon and Laitin (2003) 100 state-deaths
effective-resistance rule by the same 10-year window yields a 10 state-deaths-per-year rule.
21. Also, averaging state deaths might not resolve the problem, if all or most deaths occurred early in
the conflict.
22. See my online supplement for information on this and all cases discussed in the text.
23. The Kenyan "shifta" war in the 1960s is another such case that no data set includes as a civil war,
most likely because of the difficulty of finding information on the effective-resistance criterion. I found
sources listing dozens of killings of police and military but no clear evidence of more than 100, although the
number and apparent intensity of battles described in several sources suggest that there was effective
resistance.
24. See a more detailed discussion and a list of sources in the online supplement.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 825
ians, in that period seem not to have reached 1,000, particularly if we do n
deaths after the coup of 1976, which brought a new regime to power, and so a
the "war" could be coded in 1976. The new regime engaged in a massive purge
pected guerillas and their supporters; starting in 1976, the new regime killed
mated 9,000 to 30,000 people in the period from 1976 to 1983. During that
deaths incurred by the state were few, so we could not code an ongoing civil w
the 1976 coup. In the absence of clear rules on how sustained violence and e
resistance must be coded over time (i.e., how many state deaths must occur per y
is unclear if and when we can code this conflict as a civil war or rather a low-level
insurgency, combined with a coup, and followed by a politicide.
This brief discussion should reveal that most of the coding rules currently used to
measure civil war are somewhat ad hoc and that the problem is often exacerbated by
the low quality of data on deaths due to armed conflict. There is large room for measurement
error here, and such error will make it harder to establish empirically the differences
between civil war and other forms of political violence. Thus, in the presence
of these problems, one might argue that coding rules should never be applied too
strictly and that, when we are faced with an ambiguous case, we should err on the side
of caution, including such cases while making it possible to identify them at the analysis
stage. In the civil war list that I have compiled by applying a new set of coding rules
that try to address the issues raised here, I have included ambiguous cases but have
flagged them both in the data set and in a supplementary document that explains the
coding for each case. This allows analysts to make their own decisions about which
conflicts to drop or explore further to confirm that they are accurately coded.
CLASSIFYING AND ANALYZING EXTRASYSTEMIC WARS
Another issue that sometimes accounts for coding differences across
difficulty of how to code extrastate (colonial, imperial) wars. I argue
wars may reasonably be considered to be different from other civil w
from civil war lists.
The concept of a territorial state is central to the definition of civil war but creates
some problems in the classification of the so-called extrastate (or extrasystemic) wars.
The extrasystemic-wars category in COW included colonial and imperial wars. Imperial
wars were defined as wars between "system member[s] versus independent nonmember[s]
of the interstate system" (Singer and Small 1972, 382). Colonial wars were
defined as wars between a "system member versus an ethnically different,
nonindependent, nonmember of the interstate system" (Singer and Small 1972, 382).
Extrasystemic wars were classified as a distinct category due to a conceptual distinction
between wars that are "peripheral to the center of government (or the metropole)"
and therefore qualitatively different from wars that take place within the core territory
of the metropole (Sarkees 2000, 126). This was a normative definition because it interpreted
the nature of the relationship between governments and dependent or independent
regions within the state's territory. This normative element stands in contrast to
mechanistic definitions of civil war (such as the death threshold). Moreover, a considThis
content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
826 JOURNAL OF CONFLICT RESOLUTION
eration of power relationships was absent from the definition of intersta
were defined entirely based on a territorial conception of the state. O
emphasis on the state member as a territorial entity necessitated a re-thi
mate abandonment of the metropole [and periphery] distinction" (Sar
leading the COW project to redefine its coding rules and recla
extrasystemic wars as civil wars.
Identifying extrasystemic wars is more complicated as a result of th
coding rules evolved. Initially, coding rules for extrasystemic wars were
rules for intrastate war: "whereas an interstate war which qualified on
tle deaths would be included if the total fatality figure for the protagon
reached the 1000 mark, the member itself(including system member a
to sustain 1000 battle fatalities in order for the extrasystemic war t
(Singer and Small 1972, 36), and "if the war lasted longer than a year,
had to reach an annual average of 1,000" (Small and Singer 1982, 56;
2000, 129).
Eventually, that criterion was relaxed, and nonsystem member deat
into consideration. But could we assume that we would get an accurate
nized people's deaths in uprisings against the metropole given the im
contempt for human life in their colonies? If data measurement
extrasystemic wars are systematically different from data problems in o
then combining the two categories may bias analyses.25 The Mau-Ma
rebellion in Kenya, for example, was excluded from COW because it ca
deaths on the side of the United Kingdom (Singer and Small 1972, 397
added to the 1994 and 2000 updates of COW as a result of counting Ken
how many other such wars failed even to appear in the original "exclud
the first COW publications? One such example is the Rwandan revolu
Belgium), which Fearon and Laitin (2003) code as a war between 1956
this case does not appear in any COW list. A lingering concern with ex
therefore, is if the original coding rules have hampered our ability to
according to the current definition without having to engage in pain
resource-intensive historical research.
Another important question is if extrasystemic wars should be coded as taking
place in the metropole (i.e., in the territory of the system member) or in the territory of
the nonsystem member. Given our territorial conception of the state (and of civil war),
it seems clear that the war should be coded in the territory of the metropole, but this is
not established practice across the literature. Gleditsch et al. (2001 ), for example, code
extrasystemic wars under colonized states. Doyle and Sambanis (2000) and Licklider
(1995) code some extrasystemic wars as civil wars (where the country was under trusteeship,
as in the case of Namibia, or in the case of the war in Zimbabwe/Rhodesia from
1972 to 1980, given Rhodesia's independent status). Other scholars have committed
more obvious errors in coding, for example, Bangladesh's war of independence in
1971 as taking place in Bangladesh rather than Pakistan (Leitenberg 2001), when it is
25. This may be an empirical question because governments everywhere have an incentive to
underreport the number of their subjects that they kill.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 827
clear that Bangladesh did not exist as an independent state until after the c
that war. In Sarkees and Singer (2001), the intention seems to be to code
wars as having taken place in the territory of the system member.
This brings up another difficulty. On one hand, extrastate wars that are co
ing place in colonized states will have to be dropped from quantitati
because data for the "usual suspect" explanatory variables (e.g., democrac
domestic product [GDP]) are typically not available for dependent territor
other hand, if the war is coded as taking place in the territory of the syste
then explanatory variables have to be adjusted for the whole empire. Estim
age GDP per capita or the level of ethnic fractionalization for entire empir
a very difficult task and could only be done by taking shortcuts that migh
compromise the quality of the data. Consider also that averaging GDP per
entire empires will have the effect of turning highly developed countri
France or England, into middle- to low-income countries. If GDP per capit
a proxy for state strength, then averaging GDP values for empires will hav
of underestimating state capacity in the metropoles (although it may accu
ture state capacity with respect to the rest of the empire). But distance
metropole and military technology are important missing variables here
explain the metropole's reach.26
Empires were uniquely autocratic regimes, in which subjects lived under
forms of government, and one argument for setting extrastate wars apart
legal structure of empires prevented the articulation of colonized people
(voice) and left them with rebellion as their only option. Might right-han
trols for the level of democracy or autocracy capture this pressure to rebel in
Consider the example of France, which, according to Fearon and Laitin (2
several civil wars from 1945 to 1960. France scores as a "deep" democracy
relevant period in the Polity IV database (Marshall and Jaggers 2000). Ho
standard notion of democracy is that it provides checks and balances to re
putes peacefully and that it affords the right of political representation to
and groups. But this notion does not apply to the periphery of an empire
putes are resolved despotically, even when they are resolved by consensu
metropole. Fearon and Laitin attempt to correct for this by using polity
empires that are weighted by the share of the colonized subjects to the po
the metropole, thereby reducing the polity score of countries, such as Fran
land. But it is not clear to me that the combinations of regime characterist
Polity IV database uses are even applicable to cases of foreign domination o
"Watering down" France's, Britain's, or Belgium's democracy score in thi
likely to make these countries seem like so-called "anocracies" (i.e., regim
in the middle of the Polity range). But one could argue that as far as colon
were concerned, they were living in autocracies, whereas citizens in the
were living in a democracy. Thus, considering France or England as anocra
26. Moreover, the waves of decolonization wars in the 1950s and 1960s were all related t
changes that affected entire groups of countries in similar ways. This, at a minimum, calls
appropriate controls for systemic trends.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
828 JOURNAL OF CONFLICT RESOLUTION
bly wrong, and the meaning of an anocracy is here inconsistent with th
hypotheses that scholars usually test by using the anocracy measure.27
Perhaps all these problems can be addressed by adding appropriate righ
controls to the regressions by denoting, for example, if a country posse
Regional inequality, for example, may be more significant in explaining
lence in a world full of empires than in a world composed mainly of nat
But even trying to measure inequality in the British empire or in the Bel
likely to be systematically more difficult than measuring it in the averag
nation-state. If so, estimates based on the variables with systematic codin
empires are likely to be biased. Thus, abandoning the distinction between
and periphery may have been ill-advised. At its core, the earlier COW dist
lighted that some peripheries were constitutionally excluded from the p
cess, which presented insurmountable obstacles to the peaceful resolution
disputes with the metropole.
Despite these arguments, in many respects, extrastate wars are indee
civil wars: rebels are mostly locally recruited, the groups have a political
the governing authority is involved in the fighting. Thus, analysts could ad
to their civil war lists, but doing so would necessitate adding appropriate
the statistical model, as outlined above. Yet, only colonial wars should b
for inclusion.29 Imperial wars of annexation do not meet the territoriality
the nonsystem member's territory was never part of the metropole's territ
the war. Wars in East Timor, Tibet, or the Western Sahara could be cons
wars because (a) the fighting that started during a war of annexation contin
territory is annexed by an imperial power, and/or (b) the annexation i
accepted by the international community.;o
CODING CIVIL WAR
Drawing on the preceding analysis, I propose an operational definition
that resolves some of the problems we have encountered so far. Parts of m
are new, and parts are based on other, widely used coding rules. I provide
27. One hypothesis (Hegre et al. 2001) is that anocracies are prone to violent rebellion
are neither autocratic enough to preempt or crush rebellion nor democratic enough to resolve
fully. Many imperial powers were both autocratic enough (in their colonies) and democratic
metropoles).
28. A common notion of empire is that it uses the periphery to extract resources with no regard for
equality; this may also be true in some nation-states, although there is likely to be a systematic difference in
the degree of regional inequality between empires and nation-states with the same level of democracy at the
metropole.
29. Wars of decolonization may be combined with lists of secessionist wars because they are likely to
share some of the same causal logic. However, the difficulties associated with the measurement of key variables
in empires (see above) would still apply.
30. Such cases in my data set include Morocco (Western Sahara), Indonesia (East Timor), and Israel
(West Bank and Gaza). Excluded is, for example, the Malaysian Insurgency, because the level of violence
after independence was very low, although deaths during the phase of decolonization were high (see online
supplement). If there is an intervention by another state to prevent annexation of the territory, then this should
be considered an interstate conflict, even if local parties join the fighting.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 829
wars that conforms to the coding rule proposed here and later use the new c
in an empirical analysis. Extensive notes that justify the coding of each ca
able in a supplement posted online."3
An armed conflict should be classified as a civil war if
(a) The war takes place within the territory of a state that is a member of the international system32
with a population of 500,000 or greater.33
(b) The parties are politically and militarily organized, and they have publicly stated political
objectives.34
(c) The government (through its military or militias) must be a principal combatant. If there
is no functioning government, then the party representing the government internationally
and/or claiming the state domestically must be involved as a combatant.35
(d) The main insurgent organization(s) must be locally represented and must recruit locally.
Additional external involvement and recruitment need not imply that the war is not intrastate.36
Insurgent groups may operate from neighboring countries, but they must also
have some territorial control (bases) in the civil war country and/or the rebels must reside
in the civil war country.37
(e) The start year of the war is the first year that the conflict causes at least 500 to 1,000
deaths.38 If the conflict has not caused 500 deaths or more in the first year, the war is
31. Go to http://pantheon.yale.edu/-ns237/index/research.html#Data. Note that this coding rule is
more demanding than most of the others. Sometimes, the information to code each case with certainty is not
available, so I identify those cases in the notes. I usually err on the side of caution and include ambiguous
cases, identifying them in the data set so that researchers can decide if they want to include them or drop them
from the analysis.
32. This includes states that are occupying foreign territories that are claiming independence (e.g.,
West Bank and Gaza in Israel and Western Sahara in Morocco). A strict application of this coding rule could
drop those cases in which the international community (through the United Nations) rejects the state's claims
of sovereignty on the occupied territories.
33. We could include countries after their population reaches the 500,000 mark or, from the start of the
period, if population exceeds the 500,000 mark at some point in the country series. If a civil war occurs in a
country with population below the threshold, we could include it and flag it as a marginal case. Cases of civil
war close to the 500,000 mark are Cyprus in 1963 (578,000 population) and Djibouti in 1991 (450,000 population).
The per capita death measure would allow us to relax the population threshold.
34. This should apply to the majority of the parties in the conflict. This criterion distinguishes insurgent
groups and political parties from criminal gangs and riotous mobs. But the distinction between criminal
and political violence may fade in some countries (e.g., Colombia after 1993). "Terrorist" organizations
would qualify as insurgent groups according to this coding rule, if they caused violence at the required levels
for war (see other criteria). Noncombatant populations that are often victimized in civil wars are not considered
a "party" to the war if they are not organized in a militia or other such form, able to apply violence in pursuit
of their political objectives.
35. Extensive indirect support (monetary, organizational, military) by the government to militias
might also satisfy this criterion (e.g., Kenya during the Rift Valley ethnic clashes), although here it becomes
harder to distinguish civil war from communal violence. In some cases, where the state has collapsed, it may
not be possible to identify parties representing the state because all parties may be claiming the state, and
these conflicts will also be hard to distinguish from intercommunal violence (e.g., Somalia after 1991).
36. Intrastate war can be taking place at the same time as interstate war.
37. This weeds out entirely interstate conflicts with no local participation. The Bay of Pigs, for example,
would be excluded as a civil war because the rebels did not have a base in Cuba prior to the invasion.
Some cases stretch the limits of this definitional criterion-for example, Rwanda in the late 1990s, when exFAR
(Rwandan Army Forces) recruits with bases in the Democratic Republic of the Congo engaged in incursions
and border clashes against government army and civilians. If this is a civil war, then so is the conflict
between Lebanon-based Hezbollah and Israel (assuming the other criteria are met)..
38. This rule can be relaxed to a range of 100 to 1,000 because fighting might start late in the year (cf.
Senegal or Peru). Given the lack of high-quality data to accurately code civil war onset, if we do not have a
good estimate of deaths for the first year, we can code the onset at the first year of reported large-scale armed
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
830 JOURNAL OF CONFLICT RESOLUTION
coded as having started in that year only if cumulative deaths in the next 3
1,000.39
(f) Throughout its duration, the conflict must be characterized by sustained violence, at least
at the minor or intermediate level. There should be no 3-year period during which the
conflict causes fewer than 500 deaths.40
(g) Throughout the war, the weaker party must be able to mount effective resistance. Effective
resistance is measured by at least 100 deaths inflicted on the stronger party. A substantial
number of these deaths must occur in the first year of the war.4 But if the violence
becomes effectively one-sided, even if the aggregate effective-resistance threshold of
100 deaths has already been met, the civil war must be coded as having ended, and a
politicide or other form of one-sided violence must be coded as having started.42
(h) A peace treaty that produces at least 6 months of peace marks an end to the war.43
(i) A decisive military victory by the rebels that produces a new regime should mark the end
of the war.44 Because civil war is understood as an armed conflict against the government,
continuing armed conflict against a new government implies a new civil war.45 If the government
wins the war, a period of peace longer than 6 months must persist before we code
a new war (see also criterion k).
conflict, provided that violence continues or escalates in the following years. Note that in the data set, I also
code the start/end month, where possible. In some cases, my coding rules can be used to identify the start
month (e.g., in cases where the war causes 1,000 deaths in the first month of armed conflict). But in most
cases, the month only indicates the start of major armed conflict or the signing of a peace agreement, which
can give us a point of reference for the start or end of the war, respectively.
39. This rule also suggests when to code war termination if the 3-year average does not add up to 500.
In such a case, we can code the end of the war at the last year with more than 100 deaths unless one of the
other rules applies (e.g., if there is a peace treaty that is followed by more than 6 months of peace).
40. This criterion makes coding very difficult because data on deaths throughout the duration of a conflict
are hard to find. However, such a coding rule is necessary to prevent one from coding too many war starts
in the same conflict or coding an ongoing civil war for years after the violence has ended. Three years is an
arbitrary cutoff point but is consistent with other thresholds found in the literature. The data notes (see online
supplement) give several examples of cases in which the coding of war termination has been determined by
this criterion. A more lenient version would be a 5-year threshold with fewer than 500 deaths.
41. This criterion must be proportional to the war's intensity in the first years of the war. If the war's
onset is coded the first year with only 100 deaths (as often happens in low-intensity conflicts), then we would
not be able to observe effective resistance in the first year of the war if we defined effective resistance as 100
deaths suffered by the state.
42. This criterion distinguishes cases in which insurgent violence was limited to the outbreak of the
war and, for the remainder of the conflict, the government engaged in one-sided violence. A hypothetical
example is a case in which insurgents inflicted 100 deaths on the government during the first week of fighting,
and then the government defeated the insurgents and engaged in pogroms and politicide for several years
with no or few deaths on the government's side. If we cannot apply this rule consistently to all cases (due to
data limitations), then periods of politicide at the start or end of the war should be combined with war periods.
This implies that civil wars will often be observationally equivalent to coups that are followed by
politicide or other such sequences of different forms of political violence.
43. Treaties that do not stop the fighting are not considered (e.g., the Islamabad Accords of 1993 in
Afghanistan's war; the December 1997 agreement among Somali clan leaders). If several insurgent groups
are engaged in the war, the majority of groups must sign. This criterion is useful for the study of peace transitions
but may not be as important if researchers are interested in studying civil war duration, forexample.
44. Thus, in secessionist wars that are won by the rebels who establish a new state, if a war erupts
immediately in the new state, we would code a new war onset in the new state (an example is Croatia from
1992 to 1995), even if the violence is closely related to the preceding war. A continuation of the old conflict
between the old parties could now count as an interstate war, as in the case of Ethiopia and Eritrea, which
fought a war between 1998 and 2000 after Eritrea's successful secession from Ethiopia in 1993.
45. This criterion allows researchers to study the stability of military victories. Analysis of the stability
of civil war outcomes would be biased if we coded an end to civil war through military victory only when
the victory was followed by a prolonged period of peace. This would bias the results in favor of finding a positive
correlation between military outcomes and peace duration. This criterion is important for analyzing war
recurrence but not necessarily war prevalence.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 831
(j) A cease-fire, truce, or simply an end to fighting can also mark the end of a civil war
result in at least 2 years of peace.46 The period of peace must be longer than w
required in the case of a peace agreement because we do not have clear signals of th
ties' intent to negotiate an agreement in the case of a truce/cease-fire.47
(k) If new parties enter the war over new issues, a new war onset should be coded, su
the same operational criteria.48 If the same parties return to war over the same issu
generally code the continuation of the old war, unless any of the above criteria for co
a war's end apply for the period before the resurgence of fighting.
Using these coding rules, I have coded 145 civil war onsets between 1945 and
(2.08% of 6,966 nonmissing observations). Without coding new war onsets in
tries with already ongoing civil wars, the number of civil wars is 119 (1.93% o
nonmissing observations). Out of these cases, 20 are "ambiguous"-that is, they
not meet one or more of the coding rules.
DIFFERENCES IN THE CODING OF CIVIL WAR
AND THEIR SUBSTANTIVE IMPLICATIONS
The discussion thus far has established some of the sources of disagreem
civil war lists. Disagreements over the coded year of onset and termination
may matter for the inferences drawn when we analyze civil war onset, d
recurrence using different data sets. The extent of disagreement over the co
civil war onset is apparent in Table l a-d, which presents correlations b
starts during the period from 1960 to 1993-a period covered by most data
unit of analysis is the country-year. The dependent variable is civil war on
variable. All years of no war are coded equal to 0. There are two versions o
starts: in version (a), I code a 1 when a civil war starts and drop observati
ing war in that country until the war ends. Thus, if another war starts in th
try while another war is ongoing, we would not consider it. In version (
whenever a war starts, even if another war is ongoing. Country-years with
starts are coded 0, and in this way, we end up with more war starts.'5
46. Peace implies no battle-related deaths or, in a lenient version of this criterion, fewer de
lowest threshold of deaths used to code war onset, that is, fewer than 100 deaths per year.
47. These situations are different from those in which there is no violence as a result of ar
down without a cease-fire agreement, which would fall under criterion (f).
48. These incompatibilities must be significantly different, or the wars must be fough
groups in different regions of the country. For example, we would code three partially over
Ethiopia (Tigrean, Eritrean, Oromo) between the 1970s and the 1990s. New issues alone shoul
cient to code a new war because there is no "issue-based" classification in the definition of
could apply such a rule if we classified civil wars into categories-for example, secessionist w
olutions over control of the state. In addition to having new issues, most parties must also b
can code a new war onset.
49. See the online supplement for a summary of the definitions and operational criteria used by major
data projects. The period was selected to include the Collier and Hoeffler (2001) data set (1960-1999) in the
analysis, given the prominence of that study in the literature. Most of the other data sets start in 1945. I compare
results from a smaller number of data sets covering the entire post-1945 period.
50. Collier and Hoeffler (2001), Hegre et al. (2001), and Sambanis (2001) use version (a). Fearon and
Laitin (2003) use version (b).
(text continues on p. 835)
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
00
TABLE I
Correlations among Civil War Lists, 1
a. Version (a) of War Onset (3,1
Collier Fearon Doy
and and Sambani
Hoeffler Licklider Gleditsch Laitin Leite
COW 1994 COW 2000 (2001) (1995) et al. (2001) (2003) (20
warstla warst2a warst3a warst4a warst5a warst7a warst8a
warstla 1.00
warst2a 0.96 1.00
warst3a 0.82 0.83 1.00
warst4a 0.74 0.75 0.71 1.00
warst5a 0.42 0.46 0.52 0.57 1.00
warst7a 0.69 0.70 0.70 0.70 0.54 1.00
warst8a 0.60 0.66 0.56 0.66 0.46 0.59 1.00
warst9a 0.70 0.70 0.66 0.66 0.46 0.67 0.59
warstl0a 0.69 0.69 0.80 0.68 0.48 0.72 0.55
warstlla 0.76 0.76 0.75 0.79 0.53 0.77 0.61
warstnsa 0.74 0.74 0.73 0.83 0.51 0.80 0.62
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
b. Version (b) of War Onset (3,503 Obse
Collier Fearon Doy
and and Samban
Hoeffler Licklider Gleditsch Laitin Leit
COW 1994 COW 2000 (2001) (1995) et al. (2001) (2003) (20
warstlb warst2b warst3b warst4b warst5b warst7b warst8b
warstlb 1.00
warst2b 0.92 1.00
warst3b 0.83 0.83 1.00
warst4b 0.57 0.55 0.62 1.00
warst5b 0.37 0.45 0.46 0.40 1.00
warst7b 0.58 0.61 0.61 0.58 0.49 1.00
warst8b 0.51 0.55 0.49 0.50 0.37 0.56 1.00
warst9b 0.74 0.70 0.68 0.52 0.40 0.58 0.51
warstl0 0.69 0.67 0.80 0.60 0.41 0.61 0.46
warstllb 0.67 0.66 0.73 0.68 0.45 0.63 0.50
warstnsb 0.64 0.64 0.68 0.69 0.44 0.69 0.51
00
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 1 (continued)
c. Correlations among Lists of Civil War Prevalence
Collier Fearon Doy
and and Samban
Hoeffler Licklider Gleditsch Laitin Leite
COW 1994 COW 2000 (2001) (1995) et al. (2001) (2003) (20
atwarl atwar2 atwar3 atwar4 atwar5 atwar7 atwar8
atwarl 1.00
atwar2 0.94 1.00
atwar3 0.88 0.90 1.00
atwar4 0.75 0.76 0.82 1.00
atwar5 0.62 0.69 0.70 0.68 1.00
atwar7 0.67 0.70 0.73 0.75 0.65 1.00
atwar8 0.63 0.67 0.66 0.70 0.59 0.70 1.00
atwar9 0.62 0.66 0.64 0.66 0.53 0.70 0.62
atwarl0 0.70 0.73 0.77 0.74 0.64 0.76 0.66
atwarll 0.69 0.73 0.75 0.78 0.66 0.78 0.67
atwarns 0.69 0.74 0.77 0.79 0.66 0.83 0.69
d. War Onset
Data Set Number of War Starts Country
Singer and Small (1994)
Sarkees and Singer (2000)
Collier and Hoeffler (2001)
Licklider (1995) 58
Gleditsch et al. (2001)
Fearon and Laitin (2003)
Regan (1996) 116
Doyle and Sambanis (2000)
Sambanis (current study)
NOTE: Years for which the depe
civil wars in those years. Leiten
0o
44.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 835
As is evident from Table la,b, the correlation between most pairs of civi
lists is low. Among the lowest correlations (0.42 and 0.46) are those betwee
et al. (2001) and the two versions of the Correlates of War data set. These c
are even lower with version (b) of the dependent variable (0.37 and 0.
tively).5' The correlation between Gleditsch et al. (2001) and my data set is
sion (b)], and the correlation with Fearon and Laitin (2003) is 0.69 [versio
most highly correlated civil war lists are the two versions of the COW data
The correlation between war prevalence in any two lists would be high
compared wars that started within the same 2- or 3-year period (see Fearon
2003 for such a comparison). But that would be a comparison of war lists
onsets. A difference of 2 to 3 years in the coding of war onset is significant
ues of right-hand side variables, such as economic growth or political ins
would be immediately influenced by war, which would in turn influence th
results. In Table Ic, I correlate civil war prevalence (combined onset and du
the different lists. If two lists included different war starts but the war over
most of its duration, then the correlation between the two databases would
than in Table 1 a,b. Indeed, the correlations between pairs of lists are now h
between Gleditsch et al. 2001 and COW2, it rises to 0.69). Thus, there is m
agreement on the question of war onset and termination than there is over
war happened at all.
However, there is still considerable disagreement about which armed c
should be classified as civil wars. Many wars are coded in only one out of a
sets. Table 1 d lists the number of war starts and total country-years of war in
ent lists from 1960 to 1993.5" The highest number of war starts (116) is
Regan's (1996) list and the lowest two in Licklider's (1995) list (58) and Sin
Small's (1994) list (61).
Now that I have established that there is substantial variation in the codi
onset and termination in most data sets, it is appropriate to ask if these diffe
substantive implications. To answer that question in a tractable manner, I
same civil war model on all versions of the civil war onset variable and measur
tion in parameter estimates. To isolate the differences that are due to the cod
dependent variable, I use the same sources of data for the right-hand side v
analyze the same set of countries over the same period (1960 to 1993, annu
tions).54 Thus, all differences in parameter estimates should be due to differe
51. Here I consider only wars from the Gleditsch et al. (2001, 2002) list. In the regression
on, I add a second measure that included minor and intermediate armed conflicts from that l
52. If we expand the comparison to the entire period from 1945 to 1999, the correlation
smaller. My data set correlates with others as follows: 0.62 with Sarkees and Singer (2000
Gleditsch et al. (2001), 0.63 with Fearon and Laitin (2003), and 0.67 withDoyle and Samba
53. The number of observations across models differs in Table 2 because I drop observatio
ing war. Thus, any differences are directly the result of coding war onset and continuation. I
number of observations for the Collier-Hoeffler variable is slightly smaller because the lagged
is missing for the year 1960 since their data set starts in 1960, and we cannot know how th
coded the preceding year. Singer and Small's (1994) data set has fewer observations because it
events in 1993, and Leitenberg's (2001) data set ends in 1990. Other models all have the same
54. Leitenberg's (2001) data set is an exception because it ends with the end of the cold w
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
836 JOURNAL OF CONFLICT RESOLUTION
coding rules."5 I also restrict the analysis to a smaller number of lists
entire period from 1945 to 1999.
The model is very similar to those developed by Fearon and Laitin (
lier and Hoeffler (2001). Because these studies are by now well know
cuss the theory behind the selection of variables. Briefly, civil war ons
be less likely the higher the level of development, proxied by real per
(gdpll).56 Anocracies (anoc211), states at the mid-range in the Polity IV
have a higher risk of war as they are neither as effective as autocracies i
as good as democracies in peaceful conflict resolution (Fearon and Lait
et al. 2001). States with political instability (inst311) and regime transit
higher risk of war onset.57 Ethnic heterogeneity (ef) should increase
onset by pitting groups with different preferences against each other.58
ories expect a different relationship between this variable and civil
empirical results on the link between ethnicity and civil war are mixe
population (measured by the natural log of population, 1popnsll), wh
shown to be significant and positively correlated with war onset.59 I also
percentage of Muslims in the population (muslim) as a proxy of part of
civilization" hypothesis and as a measure of religious division." I cont
nomic growth, measured as annual percentage change in the level of
income (groll), because Collier and Hoeffler (2001) find this to be sig
negatively correlated with war onset. I control for countries that are
exporters (oil211).6' Such countries are thought to be at higher risk of wa
of reasons-the most commonly encountered hypothesis is that oil co
institutions or that it generates incentives for secessionist war. I contr
mountainous terrain (mtnll), following Fearon and Laitin, who view ter
the technology of insurgency (mountains provide hideouts for rebels)
for a variable measuring time at peace since the last war (pwt). All r
variables are lagged 1 year.62
55. A risk is that the degree of multicollinearity in the data may be different across civ
encing standard errors differentially. But this would be a result of differences in the coding
variable, so it does not pose a problem for the analysis.
56. I use data from Fearon and Laitin (2003) for this variable. According to civil wa
state capacity should discourage rebellion or allow the state to repress it in its early stag
ment levels should discourage rebellion by raising the economic opportunity costs of v
57. I use the "Polity 2" series from the Polity IV data project, version 2002, to const
and instability variables.
58. Ethnic heterogeneity was constructed by Fearon (2003) and is available in Fe
(2003) replication data set. This is a constant, so I do not lag it.
59. I used data from the World Bank and other sources to complete the population
supplement for more information.
60. The source for this variable is Fearon and Laitin (2003).
61. This series has several differences from the respective series in Fearon and Laitin
the same underlying sources have been used (World Bank data). I discuss all differen
supplement.
62. Fearon and Laitin (2003) do not lag all variables (e.g., oil is not lagged) and use the value for the
second period in the country series to fill in missing values at the start of the country series for those variables
that are lagged.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 837
This model captures the basic logic of prominent theories of civil war.6
include Fearon and Laitin's (2003) dummy variable for "new states" because
able perfectly predicts the outcome when I lag all independent variables.6
it is not clear to me what that variable measures that is not already captured
trols for "instability," "anocracy," and "income." "New" states are more like
civil war than "old" states, as Fearon and Laitin found, but they are also m
cally diverse than "old" states. Controlling for "new states" may artificially
significance of ethnic fractionalization (ef). I say "artificially" because th
these states are "new" may well be causally linked to their high ethnic frag
Most "new states" are former colonies with higher ethnic fractionalization
tries in which a nation-building process has reduced the effective level of e
mentation. It might therefore be the case that the "newness" of the state e
high degree of ethnic fractionalization. Alternatively, looking further upst
chain of causality, the "newness" of the state may also be a consequence of
for self-determination by ethnically distinct groups that a distant metropole
attempted or succeeded to integrate into a single nation in the predecess
Either way, if there is a causal connection between new state and ef, we can
for both in the regression, and ef is the more theoretically interesting con
We can now check if the correlations between these key variables and civ
influenced by the coding of the dependent variable. In what follows, I estim
models corresponding to 12 different coding rules and two versions of th
variable and for overall war prevalence. I review these estimates, looking f
in significance levels and very large changes in coefficient estimates (more
standard deviations), because such volatility would have important policy
tions in assessing the link between changes in the explanatory variables and
the probability of civil war onset. I present the results of these estimation
tables.66 Table 2 includes estimation results for 12 regressions using version
dependent variable, and Table 3 summarizes the range of these parameter
Table 4 includes estimation results for 12 regressions using version (b) of t
63. This is, of course, only one of several possible specifications of the model. I do not i
variables that others have found significant because I find the theoretical justification for inc
problematic or because other variables included in the model are likely to capture much of wha
variables are measuring. For example, I do not include a control for noncontiguous territory be
itself, not a good control of state strength without also controlling for military technology and a
ject military power. It may also be the case that stronger states only have the capacity to maintain
ous territories. Although this certainly does not apply to each case, an equality of means tes
countries with noncontiguous territories have significantly higher average income ($5,855) t
with only contiguous territories ($3,223). The t statistic for this test is -17.78.
64. For example, in my data set, there are only five war starts that occur in the second year
try series and for which income is nonmissing, but all five are dropped due to missing values
growth variable.
65. A logit regression of "new state" on "ethnic fractionalization" (ef) yields a statistically
and large coefficient (2.69) for ef(z value = 5.52) with 316 observations (I restricted the sampl
two observations in each country series).
66. I do not report fixed-effects results because fixed-effects estimation is very sensitive
ment error (hence we can expect the coding rule differences to influence the results). I present th
the online supplement. There are important differences across war lists, even among variables tha
in ordinary logit models.
(text continues on p. 840)
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 2
Probit Models of Civil War Onset, 19
Collier Gleditsch Gleditsch Fear
and et al et al. and Sam
Hoeffler Licklider (2001) (2001) Laitin Leiten
COW 1994 COW 2000 (2001) (1995) (Wars) (All) (2003) (200
Variable warst 1 warst2 warst3 warst4 warst5 warst6 warst7 warst8
GDP -0.109 -0.114 -0.097 -0.075 -0.039 -0.047 -0.093 -0.06
(0.032) (0.040) (0.033) (0.030) (0.020) (0.018) (0.032) (0.0
GDP growth -0.384 -0.237 0.493 0.564 -1.218 -0.560 -0.225 0
(1.015) (0.898) (0.810) (0.760) (0.798) (0.542) (0.787) (1.0
Instability 0.339 0.321 0.131 0.238 0.263 0.228 0.239 0.
(0.147) (0.134) (0.153) (0.154) (0.126) (0.104) (0.143) (0.1
Anocracy 0.212 0.233 0.199 0.234 0.217 0.287 0.251 0.2
(0.141) (0.121) (0.133) (0.147) (0.130) (0.089) (0.136) (0.1
Oil exporter 0.274 0.222 0.180 0.188 0.204 0.170 -0.060 0
(0.186) (0.157) (0.119) (0.149) (0.145) (0.150) (0.142) (0.1
Ethnic fractionalization
0.118 0.144 0.175 0.209 0.317 0.524 -0.006 0
(0.225) (0.206) (0.200) (0.217) (0.258) (0.180) (0.210) (0.2
Population
(log) 0.102 0.090 0.104 0.113 0.111 0.090 0.109 0.10
(0.032) (0.030) (0.032) (0.037) (0.033) (0.032) (0.033) (0.0
Terrain 0.005 0.005 0.006 0.003 0.004 0.002 0.004 0.0
(0.003) (0.003) (0.003) (0.002) (0.002) (0.002) (0.003) (0.0
00
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Percentage
Muslim -0.001 0.001 0.002 0.003 0.002 0.003 0.003 0.0
(0.002) (0.002) (0.001) (0.001) (0.002) (0.001) (0.001) (0.0
Peace
duration -0.003 -0.004 0.000 -0.006 -0.012 0.000 -0.002 -0.0
(0.005) (0.006) (0.006) (0.006) (0.005) (0.004) (0.005) (0.0
Constant -3.844 -3.626 -4.014 -4.147 -3.987 -3.668 -3.903 -4.0
(0.563) (0.527) (0.582) (0.718) (0.587) (0.506) (0.574) (0.5
Observations 3,770 3,861 3,815 3,765 3,926 3,530 3,611 3,3
Log
likelihood -229.21 -263.55 -254.81 -224.91 -290.74 -435.29 -250.53 -238
Wald X2(10) 67.38 75.69 72.15 73.59 102.21 82.62 89.31 4
Pseudo-R2 0.0933 0.0995 0.0876 0.0944 0.0877 0.0719 0.0934 0.0
NOTE: Coefficients (standard errors) are presented. Ongoing wars dropped afte
domestic product.
00
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
840 JOURNAL OF CONFLICT RESOLUTION
TABLE 3
Summary of Parameter Estimates from Table 2
Standard
Model 1 Variables Observations Mean Deviation Minimum Maximum
GDP coefficient 12 -0.084 0.024 -0.114 -0.039
GDP SE 12 0.029 0.006 0.018 0.040
Growth coefficient 12 -0.031 0.550 -1.218 0.816
Growth SE 12 0.776 0.152 0.542 1.026
Instability coefficient 12 0.273 0.066 0.131 0.382
Instability SE 12 0.136 0.016 0.104 0.154
Anocracy coefficient 12 0.217 0.048 0.082 0.287
Anocracy SE 12 0.127 0.014 0.089 0.147
Oil coefficient 12 0.177 0.104 -0.060 0.371
Oil SE 12 0.146 0.018 0.119 0.186
Ethnic fractionalization coefficient 12 0.255 0.139 -0.006 0.524
Ethnic fractionalization SE 12 0.211 0.019 0.180 0.258
Population log coefficient 12 0.101 0.012 0.076 0.116
Population log SE 12 0.033 0.002 0.030 0.037
Terrain coefficient 12 0.003 0.002 0.001 0.006
Terrain SE 12 0.003 0.000 0.002 0.003
Muslim coefficient 12 0.001 0.001 -0.001 0.003
Muslim SE 12 0.001 0.000 0.001 0.002
Peace duration coefficient 12 -0.004 0.003 -0.012 0.000
Peace duration SE 12 0.005 0.001 0.004 0.006
Constant coefficient 12 -3.820 0.224 -4.147 -3.370
Constant SE 12 0.582 0.053 0.506 0.718
NOTE: Mean coefficients and standard err
ent variable. Table 4 models allo
(warll) on the probability of
summarizes the range of param
One of the most robust variab
nificant, although its coefficie
from -0.039 to -0.114. In Table 4,
variation. We can see in Tables 3
2 standard deviations, depending
tionship between income and ci
Growth (groll) is never signifi
in Tables 2 and 4.6 Other nonsig
growth, and the switching s
67. In Fearon and Laitin's (2003) data s
try from 1945 to 1999. In my data set, th
variable captures some of the time dep
that I use in Table 2 models.
68. Using a 3-year growth rate did not change this. By lagging growth, we lose the first two observations
in the country series because we must lag gdpen by I year to calculate the growth rate (gdpgro), which
cannot be calculated for the first year of the series. I coded one version of the growth variable, grollm, where
(text continues on p. 843)
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 4
Probit Models of Civil War Onset, 1960-1
Collier Gleditsch Gleditsch Fearon
and et al. et al. and Samba
Hoeffler Licklider (2001) (2001) Laitin Leitenber
COW 1994 COW 2000 (2001) (1995) (Wars) (All) (2003) (2001) (1
Variable warst I b warst2b warst3b warst4b warst5b warst6
GDP -0.126 -0.134 -0.092 -0.077 -0.056 -0.054 -0.098 -0.088 -0
(0.035) (0.039) (0.029) (0.028) (0.023) (0.016) (0.029) (0.030)
GDP growth -0.295 -0.072 0.195 0.584 -1.087 -0.895 -0.351 0.431
(0.914) (0.791) (0.716) (0.738) (0.804) (0.560) (0.774) (0.896)
Instability 0.316 0.348 0.130 0.217 0.284 0.134 0.167 0.393
(0.161) (0.145) (0.145) (0.152) (0.119) (0.094) (0.131) (0.143)
Anocracy 0.314 0.289 0.226 0.215 0.204 0.247 0.239 0.170
(0.139) (0.127) (0.121) (0.143) (0.132) (0.082) (0.124) (0.123)
Oil exporter 0.358 0.291 0.194 0.222 0.240 0.302 0.111 0.373
(0.180) (0.155) (0.130) (0.156) (0.141) (0.116) (0.140) (0.153)
Ethnic fractionalization
0.095 0.157 0.295 0.260 0.391 0.591 0.084 0.07
(0.212) (0.199) (0.203) (0.222) (0.277) (0.169) (0.188) (0.228)
Population (log) 0.090 0.083 0.094 0.129 0.143 0.133 0.131 0.10
(0.031) (0.029) (0.032) (0.045) (0.035) (0.033) (0.031) (0.028)
Terrain 0.005 0.006 0.006 0.002 0.003 0.002 0.004 0.001
(0.003) (0.003) (0.003) (0.002) (0.002) (0.002) (0.002) (0.003)
Percentage -0.001 0.000 0.003 0.003 0.002 0.002 0.003 -0.00
Muslim (0.002) (0.002) (0.001) (0.001) (0.002) (0.001) (0.001) (0.001)
War at (t - 1) -0.351 -0.351 -0.447 -0.677 -0.005 -0.103 -0.274 -0.37
(0.192) (0.176) (0.200) (0.236) (0.169) (0.107) (0.124) (0.215)
40
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 4 (continued)
Collier Gleditsch Gleditsch Fearon
and et al. et al. and Samba
Hoeffler Licklider (2001) (2001) Laitin Leitenber
COW 1994 COW 2000 (2001) (1995) (Wars) (All) (2003) (2001) (1
Variable warst 1 b warst2b warst3b warst4b warst5b warst6b w
Constant -3.669 -3.575 -3.942 -4.520 -4.733 -4.353 -4.289 -3.747
(0.554) (0.516) (0.581) (0.787) (0.605) (0.530) (0.545) (0.530)
Observations 4,045 4,179 4,099 4,178 4,179 4,179 4,179 3,782
Log
likelihood -248.10 -289.52 -266.58 -237.15 -334.42 -561.33 -303.22 -270.64 Wald
X2(10) 63.13 75.12 79.28 74.27 102.19 101.28 75.07 40.53
Pseudo-R2 0.0936 0.1026 0.0863 0.0952 0.0908 0.0844 0.0838 0.0715
NOTE: Ongoing wars coded 0 if no new war starts. Coefficients (standard errors) ar
GDP = gross domestic product.
ti
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 843
TABLE 5
Summary of Parameter Estimates from Table 4
Standard
Model 2 Variables Observations Mean Deviation Minimum Maximum
GDP coefficient 12 -0.093 0.024 -0.134 -0.054
GDP SE 12 0.028 0.006 0.016 0.039
Growth coefficient 12 -0.199 0.489 -1.087 0.584
Growth SE 12 0.705 0.137 0.529 0.914
Instability coefficient 12 0.249 0.085 0.130 0.393
Instability SE 12 0.131 0.020 0.094 0.161
Anocracy coefficient 12 0.243 0.043 0.170 0.314
Anocracy SE 12 0.121 0.016 0.082 0.143
Oil coefficient 12 0.238 0.080 0.111 0.373
Oil SE 12 0.140 0.021 0.106 0.180
Ethnic fractionalization coefficient 12 0.280 0.155 0.075 0.591
Ethnic fractionalization SE 12 0.206 0.027 0.169 0.277
Population log coefficient 12 0.104 0.026 0.050 0.143
Population log SE 12 0.033 0.005 0.028 0.045
Terrain coefficient 12 0.003 0.002 0.001 0.006
Terrain SE 12 0.002 0.000 0.002 0.003
Muslim coefficient 12 0.001 0.001 -0.001 0.003
Muslim SE 12 0.001 0.000 0.001 0.002
Lagged war coefficient 12 -0.286 0.212 -0.677 0.101
Lagged war SE 12 0.164 0.041 0.107 0.236
Constant coefficient 12 -3.915 0.493 -4.733 -2.911
Constant SE 12 0.587 0.081 0.516 0.787
NOTE: Mean coefficients and standard errors are
endogeneity problem with this variab
ting the date of onset of civil war "wr
ficient sign of this variable.69 A pos
that, in the immediate postwar period,
periods of high risk of war recurren
war recurrence might also be high-g
be for growth to reduce the risk of ci
Political instability is mostly signifi
in two regressions (Collier and Hoeff
in another two cases, it is significant
lagging is delayed by one observation, so this r
Tables 2 and 4, this adds only five observation
instead of groll makes anocracy slightly more
sions in Table 2 and efa little less significant in
growth do not change.
69. Given my concern with endogeneity h
regressions. See versions of Tables 2, 3, 4, and
very similar to the ones presented here.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
844 JOURNAL OF CONFLICT RESOLUTION
less robust, being significant in only half of the regressions and nonsign
other half.
Anocracy borders statistical significance in six regressions in Table 2 an
significant only in Gleditsch et al.'s (2001) list, which combines all arm
Similarly, it is significant in only six out of the twelve regressions in Table
cient ranges from a low of 0.170 (Leitenberg 2001 list) to a high of 0.314
Small 1994 list) in Table 4, and the range widens in Table 2 from a low
high of 0.287.
Oil exports are basically nonsignificant in Table 2, although the pictu
mixed in Table 4, where it is significant in four regressions and borderlin
in another three.70 Its coefficient varies wildly from -0.060 to 0.371 (Tabl
it is less unstable in Table 5). This variable is not robust to changes in the co
onset.
Ethnic heterogeneity (ef) is positive and almost always nonsignificant with a
standard error, except for regression 6 (Gleditsch et al. 2001--all armed confli
Table 2. Efis nearly significant (at the .10 level) in Regan's (1996) model (warst
where again a number of additional wars are included because Regan uses a low
death threshold (200 deaths). In Table 4, efis more often significant (in two mod
the .05 level and in three more at the .10 level), although its coefficient has a
range, from 0.075 to 0.591 (see Table 5). That efis generally nonsignificant in Ta
and much more significant in Table 4 may have something to do with those extra
that we are able to include when we use version (b) of war onset.7'
The coefficient of population (lpopnsll) is quite stable and always positive an
highly significant, except for model 9 in Table 4, where it is not significant. This i
esting result seems to justify my earlier claim that the significance of populatio
may be an artifact of the high absolute threshold of deaths imposed to classify a
war. In model 9, Regan's lower death threshold captures more small wars in sm
countries.
In both Tables 2 and 4, mountainous terrain (mtnll) is relatively stable and is only
significant in the Correlates of War and Collier and Hoeffler (2001) data sets. Mountainous
terrain is a key variable in theories of civil war that emphasize the opportunity
structure for rebellion (e.g. Fearon and Laitin 2003; Collier and Hoeffler 2001), but it
is generally not statistically significant.
The percent of Muslims in a country (muslim) is marginally significant in two
regressions and in another two at the .10 level (see Table 2) and becomes somewhat
more significant in Table 4. There, we find it significant in the Fearon and Laitin
(2003) model (warst7b) with a positive sign, despite the findings of these authors that
religious division is nonsignificant in models of civil war onset. However, the period
analyzed seems to matter, as later (see Table 6) we will find that the significance of
70. I use the variable oil2, lagged once (oil211). If I were to use Fearon and Laitin's (2003) oilll series
(lagging it once), it would be significant in model 6 and, at the . 10 level, in models 8 and 9 and nonsignificant
in all other models, including some of those in which oil211 was significant (COW 1994 and my data).
71. Countries with more than one war at the same time have slightly higher ethnic fractionalization.
There are 26 such cases (of chronologically overlapping wars in my data set), and a means test for ethnic
fractionalization shows a statistically significant I10-point difference between them and the other cases.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 6
Probit Models of Civil War Onset, 1945-1999
Gleditsch Gleditsch Gleditsch Fearon Fearon Fearon Doy
et al. et al. et al. and and and Sambanis Sam
(2001) (2001) (2001) Laitin Laitin Laitin
(Wars) (Wars) (Wars) (2003) (2003) (2003) Expa
Variable 1 2 3 4 5 6 7
GDP -0.065 -0.079 -0.078 -0.095 -0.104 -0.099 -0.086
(0.022) (0.023) (0.023) (0.025) (0.025) (0.024) (0
GDP growth -0.962 - -1.350 -0.396 - -1.443 -0.1
(0.581) - (0.539) (0.537) (0.617) (0.3
Instability 0.305 0.279 0.278 0.222 0.176 0.186 0.22
(0.095) (0.088) (0.089) (0.101) (0.096) (0.097) (0
Anocracy 0.157 0.244 0.214 0.265 0.332 0.296 0.31
(0.108) (0.090) (0.095) (0.101) (0.093) (0.096) (0
Oil exporter 0.294 0.310 0.301 0.145 0.192 0.173 0.1
(0.127) (0.128) (0.135) (0.128) (0.119) (0.124) (0
Ethnic fractionalization 0.417 0.268 0.236 0.254 0.209 0.245
(0.217) (0.211) (0.208) (0.175) (0.166) (0.164) (0
Population (log) 0.140 0.137 0.144 0.142 0.141 0.148 0.0
(0.028) (0.025) (0.027) (0.029) (0.029) (0.030) (0
Terrain 0.003 0.002 0.002 0.003 0.003 0.003 0.002
(0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0
Percentage Muslim 0.000 - - 0.002 - (0.002)
- - (0.001) - - (0.0
War at (t - 1) 0.010 0.016 -0.007 -0.309 -0.342 -0.364 -0.30
(0.150) (0.143) (0.145) (0.097) (0.104) (0.102) (0
Constant -4.670 -4.500 -4.571 -4.581 -4.472 -4.568 -3.71
(0.487) (0.447) (0.474) (0.504) (0.489) (0.514) (0
~tl
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 6 (continued)
Gleditsch Gleditsch Gleditsch Fearon Fearon Fearon Doy
et al. et al. et al. and and and Sambanis Sam
(2001) (2001) (2001) Laitin Laitin Laitin
(Wars) (Wars) (Wars) (2003) (2003) (2003) Expan
Variable 1 2 3 4 5 6 7
Observations 5,893 6,092 6,051 5,893 6,092 6,051 5,893
Log likelihood -445.83 -491.67 -479.81 -408.14 -456.90 -442.08 -477.9
Wald X2(d/) 112.68 112.73 111.37 84.87 70.12 77.65 63.
Pseudo-R2 0.0910 0.0870 0.0943 0.0908 0.0897 0.0957 0.078
NOTE: Ongoing wars coded 0 if no new war starts. Coefficients (standard errors) are pr
GDP = gross domestic product.
o00
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 847
muslim disappears when we analyze the entire 1945 to 1999 period. But t
results suggest that newer wars (post-1960) may have some different char
from older wars (1945-1959) with respect to the role of religion.
Peace duration is significant only in a single regression (Gleditsch et a
warst5) in Table 2. We get a much more mixed picture in Table 4, where war i
vious period (warll) is significant and negative in about half of the regre
nonsignificant in the other half.72
Comparing the high- and low-death thresholds in the two regressions tha
on Gleditsch et al.'s (2001) data reveals important substantive differences
lower level violence and civil war. In model 6 (all armed conflict), we find t
heterogeneity, percentage Muslim, and oil exporter status (in Table 4) are a
nificant, whereas they become nonsignificant when we restrict the analys
wars. Instability, by contrast, is significant only for civil wars and does no
the risk of lower level violence significantly. This comparison is particularl
tive because the same research team did the coding of both lists, and any diff
parameter estimates must be attributed to the death threshold. Also, these
sions highlight the differences between the two versions of the dependen
Using version (a), instability enters significantly in both regressions 5 (civil
(all armed conflict), and oil exports are nonsignificant. Using version (b), o
become significant for regression 6, and the coefficient for instability drops
half (see Table 4).
In Table 6, I ran the same regressions, restricting the number of civil war li
so that I could analyze the entire 1945 to 1999 period and use the most recen
documented data sets. Here, again, we observe that income and popu
robustly significant, and anocracy becomes more significant (although not
sion 1). Instability is significant in most lists, though marginally nonsignif
Fearon and Laitin's (2003) list in regressions 5 and 6, where I use their cod
and do not lag the first observation in each country series. Anocracy is m
than previously. Growth is now significant and negative in three of four lists,
we do not lag the first observation, which results in artificial starting valu
country series. Adding growth to regressions 3, 6, and 10 causes us to lose s
vations, which reduces the significance of some variables (e.g., ef in regr
Although there is more agreement here than in the previous tables, we
some important differences in the results on ethnic fractionalization (now
in two lists and borderline significant in regression 1 with Gleditsch et al.'
and with respect to oil exporter status, which is positive and significant in
four lists (Gleditsch et al. 2001; Sambanis 2004). Similarly, war in the prev
is significant and negative only in two out of four lists (Fearon and Laitin
and Sambanis 2000). Other variables-percentage Muslim and mount
terrain-are consistently nonsignificant.
72. It would have been interesting to explore the effects on war onset of interactions between
able and other right-hand side variables but, because there are only a few instances of chronol
lapping civil wars in the same country, these interaction terms might have had the effect of over
model to a few observations.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
848 JOURNAL OF CONFLICT RESOLUTION
The results presented in Tables 2, 4, and 6 suggest wide-ranging agr
robustness of income and population and cast doubt on the robustness
variables used in civil war models, especially when we consider the tru
analyzed in Tables 2 and 4. The most significant differences are with
impact of oil exports, ethnic heterogeneity, and war in the previous p
there seems to be agreement that the mountainous terrain variable, wh
cant variable in Collier and Hoeffler's (2001) model, is not robust to al
sures of civil war. There is also no evidence of a robust association between civil war
and the percentage of Muslims in a country. GDP growth is generally nonsignificant
and may well be endogenous to levels of violence. If we restrict our analysis to the
three most recent data sets-Gleditsch et al. (2001), Fearon and Laitin (2003), and
Sambanis (This Study)-then we find more agreement among key variables. But
when we extend the analysis back to 1945, we must also consider if there was something
different about these earlier civil wars, particularly those that started before 1945
(a few civil wars are left-censored at 1945).
The significance of ethnic heterogeneity (ef) in some of the regressions in Table 6 is
worth exploring further, given the prominence of that variable in the civil war literature.
Both Fearon and Laitin (2003) and Collier and Hoeffler (2001 ) make strong statements
against the significance of ethnic heterogeneity as a factor leading to civil war. I
have shown that there is a very strong relationship between ethnic heterogeneity and
an aggregate indicator of armed conflict and much less so with civil war. To the extent
that violence escalates from minor to high levels, we should find ethnic
fractionalization to be significant in a dynamic model of violence escalation.
But when we look at civil wars alone, why might efnot be significant in the Fearon
and Laitin (2003) data set and analysis, given some of the results presented here? Lagging
right-hand side variables makes a difference, as we saw in Table 6. Fearon and
Laitin include in their analysis 10 wars that occur on the first year of the country series.
Using this data and dropping those wars by lagging all explanatory variables in their
replication data set brings efvery close to statistical significance (see online supplement).
Those wars are dropped in my data set (except where I note that I replace missing
observations at the start of each country series with values for the second year of
the country series, as Fearon and Laitin do). For some variables (e.g., GDP), this
approach to preserve observations makes sense and approximates what any imputation
program would do. However, other variables, such as instability, economic
growth, and anocracy, are harder to "impute" for the start of the country series, especially
for new states.73 Note also that among those dropped observations, the mean
level of ethnic fractionalization for cases of war onset (using my civil war list) is higher
than that for cases of no war. Thus, restoring those observations should not reduce the
significance of ef in my data. Indeed, I followed Fearon and Laitin's approach to preserve
those observations and recomputed the model in Table 6 (regressions 3, 6, 10).
Ethnic fractionalization is nonsignificant in the Fearon and Laitin model, but it is still
73. In Fearon and Laitin's (2003) replication data set, for example, out of a total 966 country-years of
political instability, only 4 were years of instability in a newly established state, which sounds implausible
almost by definition. In fact, the instability variable, as defined by Fearon and Laitin (a greater than 2 change
in the Polity scale) cannot be measured for a country's first year of independence.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 849
significant using my data (it is marginally above the .05 level), so differenc
rules influence the results on that variable. That said, efseems very sensitive
ing rules. Notice, for example, that by keeping 158 observations in regressio
lagging the first observation) using Gleditsch et al.'s (2001) civil war list, t
cient for efdrops from 0.42 in regression 1 to 0.24. This suggests that the r
are fragile. But the question also hinges on whether we can use this arbitr
scheme to preserve those observations.
Another, perhaps more substantive, reason for dropping those few cases
the first year of the country series may be that several of them are left-ce
example, the wars in Greece or the USSR all started before 1945, even th
data sets list them as starting on or after 1945. If violence was ongoing before
then it would have affected the right-hand side variables in 1945. Two of
Greece and South Korea (also Philippines)-are particularly importan
nonsignificance of ef in the Fearon and Laitin (2003) model because they
low efscores. In sum, there may be something different about the wars that s
ing the few years after World War II. Most were communist insurgencies w
"ethnic" dimension, and the mean value of ef for countries at war from 1
was about half that of countries at war in the 1950s, 1960s, or any other d
1999. Thus, losing those cases by lagging explanatory variables explai
comes close to significance using Fearon and Laitin's model and data.
I now turn to an analysis of civil war prevalence. Prevalence is defined as
of onset and continuation of war, so the dependent variable is coded 1 for all
war. I assume that civil war is a first-order Markov process (i.e., there is no
time dependence beyond the first period) and estimate a dynamic probit mo
ing to the previous model a set of interaction terms between a variable de
prevalence of civil war in the previous period and all right-hand side var
estimated coefficients of the interaction terms can, after a minor adjustment
preted as estimates of the relationship between the independent variables an
tinuation, conditional on the occurrence of war in the previous period. The
involves correcting the coefficient and standard errors of the interaction
example, the coefficient of gdpll with respect to war continuation is the sum
coefficients for gdpll and wlgdp: (-0.124 + 0.293). The standard error is t
root of the variance of gdpll plus the variance of wlgdp plus two ti
covariance. The estimates in Table 7 are already adjusted in this way; thus,
next to coefficient estimates for the interaction terms indicate significance w
to war continuation. Estimates of the linear terms refer to civil war onset,
The overall picture is one of substantial differences across civil war lists
prevalence brings us a step closer to analyzing war duration, and it is not
given the large differences in war onset and termination highlighted in Table
we should see differences in estimates of civil war prevalence.
GDP is again robust and negative in all regressions, although population n
always significant with respect to war onset. Most variables (except GDP) a
cant in only one or a few models, and instability is the most often significant
12 models). Growth is not significant.74
74. Using the variable grollm (instead of groll) again adds only five observations and doe
the results substantially.
(text continues on p. 853)
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 7
Dynamic Probit Models of Civil War Prevalence
Collier Gleditsch Gleditsch Fearon
and et al. et al. and Samba
Hoeffler Licklider (2001) (2001) Laitin Leitenber
COW 1994 COW 2000 (2001) (1995) (Wars) (All) (2003) (2001) (1
atwar I atwar2 atwar3 atwar4 atwar5 atwar6 atwar7 atwar8 at
GDP -0.126 -0.135 -0.098 -0.103 -0.061 -0.051 -0.100 -0.087 -0
(0.038) (0.042) (0.031) (0.034) (0.024) (0.017) (0.033) (0.031)
GDP growth -0.425 -0.281 0.418 0.460 -1.187 -0.692 -0.307 0.72
(0.995) (0.899) (0.778) (0.733) (0.854) (0.557) (0.829) (0.985)
Instability 0.361 0.348 0.122 0.259 0.289 0.258 0.259 0.403
(0.149) (0.136) (0.157) (0.154) (0.130) (0.106) (0.141) (0.155)
Anocracy 0.217 0.241 0.194 0.226 0.218 0.307 0.262 0.197
(0.142) (0.124) (0.135) (0.150) (0.135) (0.092) (0.137) (0.131)
Oil exporter 0.327 0.273 0.206 0.269 0.237 0.201 -0.051 0.421
(0.206) (0.178) (0.127) (0.172) (0.164) (0.156) (0.147) (0.187)
Ethnic fractionalization
0.033 0.062 0.130 0.077 0.257 0.483 -0.064 0.17
(0.226) (0.211) (0.196) (0.242) (0.261) (0.175) (0.206) (0.220)
Population (log) 0.053 0.043 0.058 0.038 0.054 0.057 0.077 0.02
(0.033) (0.032) (0.029) (0.031) (0.034) (0.031) (0.032) (0.031)
Terrain 0.004 0.005 0.006 0.003 0.003 0.002 0.004 0.002
(0.003) (0.003) (0.003) (0.002) (0.003) (0.002) (0.003) (0.002)
Percentage -0.001 0.001 0.003 0.003 0.002 0.002 0.003 0.000
Muslim (0.002) (0.002) (0.002) (0.001) (0.002) (0.001) (0.001) (0.001)
00
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
War(t - 1)
* Muslim -0.005 -0.003 -0.001 -0.007 -0.002 0.000 -0.003 0.001
(0.004) (0.004) (0.003) (0.004) (0.003) (0.003) (0.003) (0.003)
War(t - 1)
* GDP 0.169 0.147 0.033 0.201 -0.197 0.029 0.082 -0.003
(0.139) (0.110) (0.099) (0.111) (0.066) (0.053) (0.068) (0.058)
War(t - 1)
* Growth -0.021 -0.369 0.198 1.395 -0.451 1.036 1.200 0.419
(0.745) (0.615) (0.653) (0.838) (0.828) (0.644) (1.148) (0.592)
War(t - 1)
* Instability -0.533 -0.371 -0.266 -0.164 -0.172 -0.152 -0.256 0.158
(0.241) (0.223) (0.211) (0.217) (0.190) (0.179) (0.191) (0.173)
War(t - 1)
*Anocracy 0.077 0.144 0.339 0.115 0.336 0.019 0.112 0.405
(0.238) (0.215) (0.193) (0.194) (0.175) (0.189) (0.200) (0.154)
War(t - 1)
* Oil -1.001 -0.643 -0.508 -0.286 -0.312 -0.283 -0.088 -0.198 (0.555)
(0.514) (0.421) (0.383) (0.436) (0.334) (0.354) (0.367)
War(t - 1)
* Ethnic 1.561 1.379 0.761 1.001 1.233 0.306 1.330 0.562
(0.717) (0.587) (0.541) (0.772) (0.484) (0.443) (0.553) (0.370)
War(t - 1)
* Population 0.220 0.200 0.253 0.209 0.221 0.244 0.257 0.19
(0.043) (0.043) (0.039) (0.045) (0.040) (0.036) (0.041) (0.030)
War(t - 1)
*Terrain 0.000 0.000 0.000 0.008 0.005 0.005 0.002 0.004
(0.006) (0.005) (0.004) (0.005) (0.004) (0.004) (0.005) (0.004)
C0
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
TABLE 7 (continued)
Collier Gleditsch Gleditsch Fearon
and et al. et al. and Samba
Hoeffler Licklider (2001) (2001) Laitin Leitenber
COW 1994 COW 2000 (2001) (1995) (Wars) (All) (2003) (2001) (1
atwar I atwar2 atwar3 atwar4 atwar5 atwar6 atwar7 atwar8 atw
_cons -3.016 -2.842 -3.245 -2.898 -3.201 -3.095 -3.381 -2.683 (0.573)
(0.548) (0.486) (0.551) (0.550) (0.512) (0.549) (0.525)
Observations 4,045 4,179 4,099 4,178 4,179 4,179 4,179 3,782
Wald X2 1,050.19 1,169.87 1,270.07 952.18 988.45
Log
likelihood -337.62 -398.73 -375.13 -366.71 -431.94 -687.92 -371.13 -417.82 Pseudo-R2
0.6996 0.6834 0.7199 0.7465 0.6134 0.6510 0.7887 0.701
NOTE: Coefficients (standard errors) are presented. Bold indicates significance at .05 or higher
tion terms have been adjusted to refer to war continuation only. GDP = gross domestic p
00
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 853
The picture becomes much more unstable when we consider results with res
war continuation in the shaded portion of Table 7. Here we see that the int
between income and lagged war prevalence is often nonsignificant, but occasi
is significant with switching signs: in Gleditsch et al.'s (2001) civil war list (a
high income significantly reduces war continuation, whereas in the two mod
on Doyle and Sambanis's (2000) data (atwarlO, atwarl 1), higher income once
war is ongoing has the effect of increasing the risk of the war continuing. In
(1996) and Licklider's (1995) data, the same relationship is borderline signifi
.10).
Instability while war is ongoing significantly reduces the risk of war continuation
in two models (Singer and Small 1994; Regan 1996) and is nonsignificant in the rest.75
Similarly, anocracy is significant in one model and borderline in another two but otherwise
nonsignificant. Oil exports are generally nonsignificant. There is wide agreement
that populous countries will have longer wars and no evidence that countries with significant
Muslim populations or mountainous terrain will have longer civil wars.
Interesting results emerge again with respect to ethnic fractionalization (ef). The
interaction term with efis significant and positively correlated with war continuation
in four regressions (e.g., in regression 7, using Fearon and Laitin's 2003 civil war list)
and close to significant in one more. Thus, the results with respect to this variable are
genuinely divided: about half the models would tell us that a very diverse country
would have long wars, once a war actually started, whereas the other half would point
to no statistically significant relationship between ethnicity and war duration. There is
more support for the positive association between ethnic fractionalization and war
continuation in Table 8, where I present prevalence results from the four civil war lists
that span the entire period from 1945 to 1999, restricting the analysis to four civil war
lists. Here, we see the same sort of substantial disagreement with respect to the effects
of income on war continuation: in one model, the coefficient is significant and negative;
in another, it is significant and positive; and in the other two, it is nonsignificant.76
The more we push the data, the more the differences in coding rules will matter.
CONCLUSION
This article does three things. First, it demonstrates that there are subst
ences across civil war lists with respect to the coding of the onset and te
civil war. Exploring those differences analytically reveals some conceptu
75. I am not concerned here with the theoretical explanation of these results. How
reduce war duration is unclear. It may be that a regime transition toward significant democ
would be coded as instability) satisfies some of the rebels' demands and leads to war termin
of the instability variable may be affected by the occurrence of war, but it is not my purpose
these questions of endogeneity.
76. The differences actually become greater if we do not lag the first observations in
series (giving 6,051 observations). With respect to onset, growth is significant and negative in
regressions, instability is significant in one regression, anocracy is significant in all regres
are significant in one regression, and population is significant in one regression. With resp
tion, there is no significant change from the differences noted in Table 8 (see the result
supplement).
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
854 JOURNAL OF CONFLICT RESOLUTION
TABLE 8
Dynamic Probit Models of Civil War Prevalence, 1945-1999
Gleditsch Fearon Doyle and Sambanis
et al. (2001) and Laitin Sambanis (2000, (This
(Wars) (2003) Expanded) Study)
GDP -0.073 -0.095 -0.102 -0.085
(0.025) (0.027) (0.026) (0.025)
GDP growth -1.006 -0.204 -0.187 -0.113
(0.626) (0.473) (0.411) (0.392)
Instability 0.273 0.243 0.219 0.213
(0.108) (0.108) (0.103) (0.109)
Anocracy 0.172 0.291 0.274 0.294
(0.112) (0.114) (0.115) (0.099)
Oil exporter 0.276 0.067 0.148 0.148
(0.146) (0.151) (0.147) (0.130)
Ethnic fractionalization 0.220 0.115 0.295 0.221
(0.218) (0.192) (0.199) (0.200)
Population (log) 0.055 0.078 0.034 0.043
(0.027) (0.027) (0.026) (0.024)
Terrain 0.003 0.003 0.001 0.002
(0.002) (0.002) (0.003) (0.003)
Percentage Muslim 0.001 0.002 0.000 0.002
(0.002) (0.001) (0.001) (0.001)
War(t - 1) * GDP -0.153 0.042 0.111 0.022
(0.073) (0.048) (0.041) (0.026)
War(t - 1) ' Growth 0.108 0.288 -0.399 0.580
(0.615) (0.791) (0.495) (0.661)
War(t - 1) * Instability -0.108 -0.309 -0.169 -0.040
(0.184) (0.178) (0.153) (0.166)
War(t - 1) * Anocracy 0.319 0.054 0.159 0.021
(0.196) (0.200) (0.146) (0.177)
War(t - 1) ' Oil -0.103 -0.042 -0.228 -0.087
(0.323) (0.291) (0.302) (0.284)
War(t - 1) * Ethnic 1.004 1.052 0.700 0.639
(0.417) (0.349) (0.303) (0.353)
War(t - 1) e Population 0.211 0.265 0.197 0.226
(0.035) (0.032) (0.029) (0.032)
War(t - 1) * Terrain 0.004 0.003 0.002 0.003
(0.005) (0.003) (0.003) (0.003)
War(t - 1) * Muslim -0.001 -0.002 0.000 0.000
(0.003) (0.002) (0.003) (0.002)
_cons -3.159 -3.501 -2.689 -2.880
(0.436) (0.448) (0.429) (0.412)
Observations 5,893 5,893 5,893 5,893
Wald X2 791.89 1,736.07 1,898.86 2,022.78
Log likelihood -596.74 -548.11 -694.25 -671.34
Pseudo-R2 0.5797 0.7731 0.6963 0.7185
NOTE: Coefficients (standard errors) are presented. Bold indi
cates significance at .10. Parameters for all interaction terms h
only. GDP = gross domestic product.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 855
about the meaning of civil war and can serve as the foundation for a theoretic
sion that establishes what civil war is and how it can be distinguished from othe
of political violence. Second, it proposes a new coding rule for civil war that a
to sidestep some of the problems identified in other coding rules and offers a ne
civil wars that is based on this new coding rule. Third, it measures the subs
implications of differences in coding rules by formally comparing the emp
results that we get from a civil war model when we use 12 different rules to c
War.
The quantitative literature on civil war reveals a remarkable degree of disagr
on how to code the onset and termination of wars, and the literature is fuzzy on
distinguish among different forms of political violence. This implies the need for
rizing about civil war and then for proper measurement of the concept. Given t
ing complexities I have identified, researchers should conduct robustness tests
different civil war lists and justify their coding decisions as transparently as p
Differences in the coding of civil war have substantive implications, but perhap
large as one might have expected. The results presented here show that the es
coefficients of most variables vary widely as a result of changes in the coded o
civil war. In a few cases, the sign of the coefficient also changes, and significan
els also vary across data sets. Several variables are not robust to changes in the
of civil war onset, most strikingly in the case of lagged war prevalence,
fractionalization and oil exports, and, to a much lesser extent, anocracy and instab
Predictions of when and where civil war might occur are likely to depend cr
on which coding rule we use. More important, estimates of the substantive eff
policy interventions to reduce the risk of civil war by manipulating the level of i
political instability, or any of the other "manipulable" variables in the model w
vary widely, depending on the coding rule for war onset and termination.
At the same time, some variables are remarkably robust to coding differe
Income level and population size are very robust and significantly associated wi
war. That significance levels for these variables do not change much despite lar
ferences in coding rules is likely because these variables change slowly over ti
problem that this highlights is that we cannot rely on them to make accurate
tions of the timing of war onset. Other variables-especially mountainous ter
Muslim population, and economic growth-are consistently nonsignificant
minor exceptions).
The results from models of war prevalence suggest that predictions of civ
duration will be even less accurate than predictions of civil war onset. There
greater instability of empirical results in the prevalence model, so analyses of c
duration will be much more affected by differences in the coding rules. Robu
tests are therefore essential when analyzing war duration or termination.
Overall, civil war models seem to be good at identifying countries with lon
proclivities to civil war. But if the models are further developed to be able to pre
timing of war onset better, then we are likely to see coding differences in the dep
variable affect the parameter estimates more significantly (the changes we found
coefficient sign of economic growth, which changes substantially over time, a
indication of this). Thus, differences in coding will begin to matter if the theo
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
856 JOURNAL OF CONFLICT RESOLUTION
models "catch up" to the data by including more time-sensitive variables.
ory develops more, it seems likely that analysts of civil wars will have to go
data to develop a higher degree of consensus on the meaning of civil war
can produce accurate and credible predictions of where and when civil war
That said, I have no clear answer on whether it is better to have a single de
coding rule for civil war or if it is more beneficial to have differing definiti
hand, some standardization of our coding rules would reflect theoretical c
what civil war is, and homogeneity in measurement would rule out one pos
of difference in empirical results. On the other hand, that we have many
suggests that the concept of civil war may mean different things to diffe
and coding rules should reflect differences in subjective understandings o
cept. Analyzing the differences across coding rules can be instructive be
differences map out the space within which we might be able to fin
understanding of civil war.
A substantive result from this analysis, and to which I must return, is
fractionalization may not be as nonsignificant a correlate of civil war as ma
have argued. Its significance hinges perhaps too much on the coding rules for
but ethnic fractionalization is clearly important in explaining a broad ca
armed conflict that includes minor insurgency. This is an important result, c
how difficult it was to clearly distinguish between civil war and other forms
of violence. Ethnic fractionalization was sometimes significantly correla
war onset and war continuation, but these results were at times affected
wars that are dropped when we lag explanatory variables by 1 year. Howe
ences in coding rules seem to explain the nonsignificance of that variabl
and Laitin's (2003) results. In other runs (see supplement), religious fractio
and the size of the largest confession were significant and often dominated t
ethnic fractionalization.77 Thus, ethnoreligious identity may have been wri
quickly as a correlate of large-scale armed conflict-even "civil war"
scholars.
Several other substantive results are worth noting. Economic growth may be
endogenous to civil war, which implies the need for a different estimation strategy for
models that include that variable. The political variables of anocracy and instability,
which are central to some models of civil war, are sensitive to the coding rules but
much less so if we examine the entire post- 1945 period. The significance of the population
variable may be tied to the high threshold of violence used to identify civil war
and distinguish it from minor violence. Mountainous terrain-a key measure of the
technology of insurgency in civil war in some models-is not robust to changes in the
coding rules, and neither is the measure used to identify countries that are major
exporters of oil. Peace duration or war in the previous year is also sensitive to changes
in coding. Moreover, the significance of some of these variables is affected by whether
observations of ongoing war are dropped (e.g., anocracy is much more consistently
significant in Table 4 than in Table 2). Finally, the period analyzed influences some
77. For an analysis of the sensitivity of empirical results in the civil war literature to small changes in
the specification of the model, see Hegre and Sambanis (2004).
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
Sambanis / WHAT IS CIVIL WAR? 857
results due to the small number of war starts: for example, instability
consistently significant in Table 6 than in Table 4.
Despite these differences and difficulties, the conclusion from this study shou
be that coding wars and analyzing them quantitatively is a futile exercise. Rat
abandon these efforts, I favor redoubling them by improving the coding rule
ing them transparently to the data, and studying the implications of difference
coding rules. The legacy of the Correlates of War project is that we now have at
posal replicable data on civil war that we can analyze quantitatively to point to
oretical and empirical complexities of defining and measuring civil war. This
suggests ways of building on that legacy.
REFERENCES
Collier, Paul, and Anke Hoeffler. 2001. Greed and grievance in civil war. Policy Resear
Bank.
Connor, Ken. 1998. Ghost force: The secret history of the SAS. London: Weidenfeld and Nicolson.
Doyle, Michael, and Nicholas Sambanis. 2000. International peacebuilding: A theoretical and quantitative
analysis. American Political Science Review 94 (4): 779-801.
Fearon, James D. 2003. Ethnic and cultural diversity by country. Journal ofEconomic Growth 8:195-222.
Fearon, James D., and David D. Laitin. 2003. Ethnicity, insurgency, and civil war. American Political Science
Review 97 (1): 75-90.
Ghobarah, H., Paul Huth, and Bruce Russett. 2003. Civil wars kill and maim people, long after the fighting
stops. American Political Science Review 97 (2): 189-202.
Gleditsch, Nils Petter, HAvard Strand, Mikael Eriksson, Margareta Sollenberg, and Peter Wallensteen. 2001.
Armed conflict 1945-99: A new dataset. Unpublished paper, PRIO, Oslo, Norway.
Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg, and Havard Strand. 2002.
Armed conflict 1946-2001: A new dataset. Journal of Peace Research 39 (5): 615-37.
Harff, Barbara. 2003. No lessons learned from the Holocaust? Assessing risks of genocide and political mass
murder since 1955. American Political Science Review 97 (1): 57-73.
Hegre, Havard, 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, Havard, and Nicholas Sambanis. 2004. Sensitivity analysis of empirical results in the quantitative
erature on civil war. Unpublished manuscript, PRIO and Yale University.
Leitenberg, Milton. 2001. Deaths in wars and conflicts between 1945 and 2000. Paper prepared for the Co
ference on Data Collection in Armed Conflict, June 8-9, Uppsala, Sweden.
Licklider, Roy. 1995. The consequences of negotiated settlements in civil wars, 1945-1993. American Po
cal Science Review 89 (3): 681-90.
Marshall, Monty, and Keith Jaggers. 2000. Polity IV project [Codebook and data files]. Accessed fr
www.bsos.umd.edu/cidcm/inscr/polity/.
Mason, David, and Patrick Fett. 1996. How civil wars end: A rational choice approach. Journal of Confl
Resolution 40:546-68.
McAdam, Doug, Sidney Tarrow, and Charles Tilly. 2001. Dynamics of contention. Cambridge, UK
bridge University Press.
Price, D. L. 1975. Oman: Insurgency and development. Conflict studies. London: Institute for the Stu
Conflict.
Regan, Patrick. 1996. Conditions for successful third party interventions. Journal of Conflict Resolution 40
(1): 336-59.
Sambanis, Nicholas. 2001. Do ethnic and nonethnic civil wars have the same causes? A theoretical and
empirical inquiry (Part 1). Journal of Conflict Resolution 45 (3): 259-82.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms
858 JOURNAL OF CONFLICT RESOLUTION
. 2004. Expanding economic models of civil war using case studies. Perspectives
259-79.
Sarkees, Meredith Reid. 2000. The Correlates of War data on war: An update to 1997. Conflict Management
and Peace Science 18 (1): 123-44.
Sarkees, Meredith Reid, and J. David Singer. 2001. The Correlates of War datasets: The totality of war. Paper
prepared for the 42nd Annual Convention of the International Studies Association, February 20-24,
Chicago.
Singer, J. David, and Melvin Small. 1972. The wages of war, 1816-1965: A statistical handbook. New York:
John Wiley.
. 1994. Correlates of War project: International and civil war data, 1816-1992 [Computer file, Study
#9905]. Ann Arbor, MI: Interuniversity Consortium for Political and Social Research [distributor].
Small, Melvin, and J. David Singer. 1982. Resort to arms: International and civil war, 1816-1980. Beverly
Hills, CA: Sage.
Strand, HAvard, Lars Wilhelmsen, and Nils Petter Gleditsch, in collaboration with Peter Wallensteen,
Margareta Sollenberg, Mikael Eriksson, Halvard Bulhaug, and Jan Ketil Rod. 2003. Armed conflict
dataset codebook, version 1.2a. Accessed June 21, 2004, from http://www.prio.no/page/Project_detail//
9244/44532.html?PHPSESSID=e92e0a3b72738fae3ff8585c 12cd291b/.
Tilly, Charles. 1978. From modernization to revolution. New York: Random House.
. 2003. The politics of collective violence. Cambridge, UK: Cambridge University Press.
Valentino, Benjamin, Paul Huth, and Dylan Balch-Lindsay. 2001. Draining the sea: Mass killing, genoc
and guerilla warfare. Unpublished manuscript, Stanford University.
Walter, Barbara. 2002. Committing to peace: The successful settlement of civil wars. Princeton, NJ: Pr
ton University Press.
This content downloaded from 78.99.21.185 on Thu, 08 Nov 2018 15:30:47 UTC
All use subject to https://about.jstor.org/terms