TheStrategicLogicofEthnoterritorial …aw2951/WimmerJGSS.pdf · 2019. 8. 1. · ANDREASWIMMERANDCHRISMINER 3 definitions of the dependent variable.They are derived from either the

Journal of Global Security Studies, 0(0), 2019, 1–19doi: 10.1093/jogss/ogz009

The Strategic Logic of Ethnoterritorial

Competition: Violence against Civilians in Africa’s

Civil Wars

Andreas Wimmer1 and Chris Miner2

1Columbia University and 2University of California Los Angeles

Abstract

This article develops and tests a new theory of violence against civilians during civil wars by com-

bining geocoded data on African armed conflicts over the past two decades with a range of other

geocoded information. The theory suggests a twofold logic of ethnic targeting aimed to enlarge the

territory dominated by one’s coethnics in themost effective way. First, rebels and government fighters

kill civilians in areas populated in equal shares by their own and their adversary’s coethnics because,

in such areas, small amounts of violence suffice to tilt the local balance of power in their favor. Second,

they target places close to the border between the settlement areas of their own and their adversary’s

coethnics as this will allow expanding the contiguous area under their control. We do not find em-

pirical support for the three most prominent alternative theories, all of which assume that civilian

victimization is independent of the political conflict over which the civil war is fought. Civilians are not

more likely to be killed in areas where lootable natural resources can be found, in recently conquered

territories where fighters are supposed to eliminate enemy collaborators, or where rebel forces who

have established only weak control over their fighters operate.

Keywords: ethnic violence, Africa, civil wars, civilian victims

In the wake of Kalyvas’s (2006) seminal study of theGreek civil war, scholars started to more systematicallyinvestigate why armed organizations intentionally killunarmed civilians during the course of a civil war. Thisarticle makes a threefold contribution to this literature.First, it joins other recent work in providing evidencethat the killing of civilians is linked to their ethnicbackground, and it specifies the precise logic of ethnictargeting. Previous case studies have shown that ethnicaffiliations matter for understanding who kills whom.1

Going beyond these individual cases, this article looks atthe larger universe of African conflicts and shows that

1 For Darfur, see Olsson and Siba (2013); for Aceh,Czaika and Kis-Katos (2009); for Bosnia, Costalli andMoro (2010) and Weidmann (2011); for Guatemala,

the logic of ethnic targeting is important for understand-ing where and when civilians are killed during civil war,even when other mechanisms are taken into account andwhen we include both ethnic as well as nonethnic armedconflicts into the picture.

Second, this article theoretically disentangles and em-pirically evaluates two distinct theoretical models of whyand how ethnic targeting occurs. Fjelde and Hultman(2014) build on Valentino, Huth, and Balch-Lindsay(2004) to argue that armed groups attack their adver-sary’s coethnic civilians in order to undermine its lo-gistic and informational support base and thus gain anadvantage in winning the civil war. Armed actors kill

Sullivan (2012) and Schwartz and Straus (2018); forNorthern Ireland, White (1993).

Wimmer, Andreas, and Chris Miner. (2019) The Strategic Logic of Ethnoterritorial Competition: Violence against Civilians in Africa’s Civil Wars. Journal of GlobalSecurity Studies, doi: 10.1093/jogss/ogz009© The Author(s) (2019). Published by Oxford University Press on behalf of the International Studies Association. All rights reserved. For permissions, please e-mail:[email protected]

Dow

nloaded from https://academ

ic.oup.com/jogss/advance-article-abstract/doi/10.1093/jogss/ogz009/5487960 by C

olumbia U

niversity user on 01 August 2019

http://www.oxfordjournals.org

mailto:[email protected]

2 Ethnoterritorial Competition and Violence against Civilians

civilians, in other words, to undermine the enemy’s fight-ing capacities. They therefore should attack their oppo-nent’s coethnics where these are more numerous and thuseasier to find and target.

We outline an alternative logic of ethnic targeting,knitting together various existing arguments into an inte-grated theoretical framework. Violence is about expand-ing the territory under the political control of one’s co-ethnics, we argue, and is thus most intense where it isstrategically most useful and most effective. Rather thanweakening the enemy’s civilian supporters where theyare most numerous, violence is therefore more preciselyand strategically focused on specific areas. Border zonesbetween their own and their adversaries’ coethnics areof greater strategic interest to armed groups than eth-noterritorial enclaves because fighters are interested inexpanding the continuous territory controlled by sup-portive civilians (see, e.g., Melander 2007; Weidmann2011).Moreover, in locales where opposing groups makeup equal shares of the local population, small amounts ofviolence can tilt the local demographic and political bal-ance of power in one’s favor by intimidating the opposedgroup and encouraging them to flee, thus expanding theterritory under the political control of coethnics in themost cost effective way. This second argument is derivedfrom Balcells’s (2011) study of the civil war in Spain.2

Killing civilians in order to expand the territory con-trolled by coethnics may be part of a broader and moreradical strategy of ethnic cleansing: an attempt to createethnically homogenous regions or states by terrorizingmembers of ethnic minorities such that they would fleeor by actively deporting them (Petrovic 1994).3 Unfor-tunately, we cannot determine whether ethnic targetingis part of such a broader strategy of ethnic cleansing be-cause there is no systematic data on the war goals andstrategic visions of armed actors.

The third contribution of this article is to evaluatethree alternative theories of violence against civilians us-ing varying datasets, samples, and definitions of the de-pendent variable to mirror as closely as possible thespecific scope conditions and mechanisms foreseen byeach of the three theories. In contrast to the model ofethnic targeting outlined above, these three alternative

2 It is also impossible to explore possible symbolic andemotional forces, such as resentment, fear, and mythsdemonizing the ethnic enemy, that could motivate rank-and-file fighters to kill civilians (see Kaufman 2006;Petersen 2002).

3 Mann (2005), however, uses the termsynonymouslywithgenocides that aim at the physical destruction of agroup.

theories assume that killing civilians has little to do withethnicity or, more generally, with the macrolevel politi-cal conflict over which a civil war is fought. Greed the-ories maintain that the population in resource-rich ar-eas will be terrorized and victimized by armed groupswho seek to establish or maintain control over these re-sources (Azam and Hoeffler 2002; Eyndre 2016; Bagozzi2017; Koren and Bagozzi 2017). According to Weinstein(2006; see also Humphreys and Weinstein 2006), rebelorganizations that initially relied on funding from nat-ural resource extraction or from outside governmentsattract opportunistically motivated fighters who will bemore likely to prey upon the civilian population (exceptif the outside government is a democracy, see Salehyan,Siroki, and Wood 2014). A second theory was proposedby Kalyvas (2006): rebels and governments act upon thedenunciations of local villagers, he argues, to kill possi-ble collaborators. They are most likely to do so whenthey do not yet fully control an area and thus are al-ready able to search for and eliminate collaborators buthave not yet killed all of them. Seen from these variousperspectives, ethnicity does not define friends and foesin clear-cut terms, and “ethnic defections” are thereforethought to be frequent; armed organizations whosemem-bers share the same ethnic background sometimes fighteach other (Christia 2008) or even kill coethnic civilians(see Kalyvas 2008; Lyall 2010).

Testing these various arguments, this article im-proves over other cross-national studies that either fo-cus on only one specific theory (Wood 2010; Wood,Kathman, and Gent 2012; Fjelde and Hultman 2014;Salehyan et al. 2014) or test other propositions thatare only tangentially related to these most widely dis-cussed approaches.4 We use different, perpetrator-specific

4 A booming series of recent research explores addi-tional, more fine-grained arguments about the char-acteristics of violence-prone rebel organizations, theconditions under which they resort to indiscriminate vi-olence, and the effects of outside intervention. Moreprecisely, recent scholarship finds that democraticallyoriented rebel organizations commit fewer acts ofviolence than religiously oriented organizations in theMiddle East (Asal, Brown, andSchulzke 2015), that Com-munist organizations rape women less than other or-ganizations (Hoover Green 2016), and that more civil-ians are killed by a rebel force after a competingrebel organization has emerged (Wood and Kathman2015) or after the rebels have sustained major casual-ties at the hand of government forces (in Uganda andAfrica more generally: Wood 2014). Somewhat counter-intuitively, one-sided violence against civilians is more

Dow



olumbia U


ANDREAS WIMMER AND CHRIS MINER 3

definitions of the dependent variable. They are derivedfrom either the Armed Conflict Location and EventsDataset (ACLED; Raleigh et al. 2010) or the UppsalaConflict Data Program’s Geo-Referenced Event Dataset(UCDP-GED; Sundberg, Lindgren, and Padskocimaite2012) that cover the African continent from 1997 to2011 and from 1989 to 2010, respectively. In combina-tion with a variety of other geocoded datasets, most im-portantly the Ethnic Power Relations dataset (Geo-EPR;see Wucherpfennig et al. 2011), these data allow us toidentify which armed actor intentionally killed civiliansin which exact location, what the ethnic composition ofthat location is, and with which ethnic groups armed ac-tors themselves identify.

We specify logistic regressions with 0.25 degreehexagons (approx. 27.75 km2) as units of observationand violence as the dichotomous dependent variable.Fixed effects control for unobserved characteristics at thecountry level, such as the nature of the conflict, levels ofrule of law, and the like. We find strong and consistentsupport for the ethnoterritorial competition argument:violence against civilians is most likely in areas where thepopulation size of the coethnics of rebels and those gov-ernment elites reach parity as well as areas close to theborder between the settlement areas of these two groups.Both findings indicate that expanding the territory dom-inated by one’s coethnics is an important strategic aimwhen fighters kill civilians.5 We do not find support forthe idea—developed empirically by Fjelde and Hultman(2014)—that ethnic targeting aims at the opponent’s co-ethnics where they are most numerous and thus easiestto find and target. We briefly summarize why Fjelde andHultman, who use similar data sources and methods, ar-

frequent in conventional civil wars than in guerilla in-surgencies (Krcmaric 2018). Regarding the effects ofoutside intervention, Hultman and Kathman (2013) haveshown that United Nations peacekeeping missions re-duce civilian victimization in Africa’s civil wars whileWood and coauthors (Wood et al. 2012) find that armedintervention by outsiders decreases the killing of civil-ians by the supported civil war actor and increases vic-timization by the opponent. Finally, more civilians arekilled when many years have passed since the ratifi-cation of the ICC convention, but not where the ICC islocally present (Bussmann and Schneider 2016).

5 Since the UCDP-GED data on civilian victims are limitedto Africa, we are unable to evaluate whether these re-sults would hold for Latin America, Europe, or Asia butnote that past research on Bosnia (Weidmann 2011) andSpain (Balcells 2011) has led to similar conclusions.

rive at different results and include more detail in OnlineAppendix 3.

The results also do not support the looting, loose con-trol over fighters, and territorial control arguments andoffer some evidence that contradicts these logics: civiliansare not more likely to be victims of violence if they liveclose to lootable diamond fields, to other mining sites,or to oil wells. Weak control over rebel fighters or fund-ing from natural resource extraction decreases—ratherthan increases—armed groups’ propensity to kill civil-ians, while rebels’ external funding is not associated withcivilian victimization. Both government troops and rebelsare more likely to kill civilians close to where they lostterritory to their opponents (in line with Wood 2014),rather than where they gained it, as the Kalyvas (2006)model implies.

The article proceeds in a straightforward way. Thefollowing section discusses the various theories of vio-lence against civilians in more detail and derives empiri-cally testable hypotheses from them. Next, we introducedatasets and units of observation, define independent anddependent variables, and discuss model specifications. Asection with results follows, and a final section concludes.

Arguments and Hypotheses

The Ethnopolitical Logics of Violence

This article joins a group of authors (Valentino et al.2004; Balcells 2011; Hirose, Imai, and Lyall 2017;Schwartz and Straus 2018) who see violence against civil-ians as part and parcel of a broader conflict over politi-cal power, rather than a purely local issue of controllingrebel fighters, natural resources, or containing collabo-ration. We focus on conflicts where political goals andadversaries are defined in ethnic terms, rather than withreference to class, party, or ideology, mostly because dataon party and class affiliation or the ideological orienta-tion of the civilian population is hard to come by (for anexample of what it takes to overcome these challenges,see the study on Afghanistan by Hirose et al. 2017). Wedo not imply that there must be something in the natureof ethnicity that makes it more likely that civilians arekilled, nor that Africa is more prone to ethnic conflictsthan other regions. We also do not assume that ethnicconflicts are generally more violent than other types ofconflicts (a view disproven by Valentino et al. 2004).

We start from the assumption that ethnic conflictsare fought over the distribution of political poweramong ethnically defined alliance networks (in line withWimmer 1997; Wimmer 2013). The following figuredescribes various such configurations of ethnopoliti-cal power, modifying Tilly’s (1978) well-known polity

Dow



olumbia U



Figure 1. Ethnopolitical actors and conflicts

model. It distinguishes between ethnopolitical groupsrepresented at the highest level of government (termed“included” groups A to C) and those that are not (“ex-cluded groups” 1 to 4). This model allows to distinguishbetween two different types of conflict. First, a rebel orga-nization claims to speak in the name of excluded group1 and fights the central government, for example, overthe neglect of its home region or the lack of represen-tation in the inner circles of power. We call this type ofconflict a “rebellion” (see Wimmer, Cederman, and Min2009). A well-known example would be the various Dar-furian guerillas that fought against the Arab-dominatedSudanese government.

Second, an army faction of group A members (to givean example) attempts to overthrow the existing govern-ment, claiming that the latter has delayed the promotionof A officers in the army or disempowered the ministriescontrolled by A. We call this type of conflict “infighting”(ibid.). An empirical example are the Diola of Senegal,representatives of whom are included in a power-sharingcoalition since independence. From the mid-nineties on-ward, a separatist organization, largely dominated by Di-ola speakers, began to fight for the independence of theCasamançe region. On the basis of this basic model, wecan now outline two different theories that predict whichactors will target which civilian populations during anarmed conflict.

Ethnoterritorial Competition

According the theory proposed here, fighters seek tomax-imize the territory under control of their coethnics, andthey do so in a strategic way. Before we outline thisargument in detail, we first clarify why ethnic ties arepolitically relevant and how local struggles for politi-cal power are intertwined with the macrolevel conflictover who controls the national government. The localpopulation (group 1 in Figure 1) gains from supporting

ethnic rebels who fight in their name because this in-creases the likelihood that they will be able to win theconflict and capture the state. This in turn will producepatronage and public goods benefits for the allied eth-nic population (for evidence of ethnic favoritism in pub-lic goods provision, see Bannjerjee, Lakshimi, and So-manathan 2008, section 3.2.1; Grødeland, Miller, andKoshechkina 2000; Kramon and Posner 2016; de Lucaet al. 2015; McClendon 2016). For the same reasons, thecoethnics of ruling elites (groups A, B, and C in Figure 1)have an interest in maintaining their privileged relation-ship with the state apparatus and thus are more likely tosupport government troops than excluded populations.

These national struggles over state power are thenmirrored at the local level, where political factions andnetworks align with the ethnopolitical cleavage at themacrolevel, thus opposing representatives of group 1against those of groups A, B, and C. Consequently, lo-cal struggles over political power—over “who rules thisplace”—become intertwined with national-level conflictover “who owns the state” (Wimmer 1997).

At the local level, fighting organizations kill civiliansof the other side in order to tilt the balance of local polit-ical power and demographic weight in the favor of theirown group, intimidating the enemy population or en-couraging them to flee.They do so because they can counton the political loyalty and logistical support of their co-ethnics (cf. also Wood 2010, 602f.). The marginal utilityof violence is highest in locales where it would suffice toexpel (or kill) only a small number of civilians in orderto become a local majority. Therefore, violence should bemost likely in those areas where the population size of thecoethnics of government elites and of rebels reach parity(Hypothesis 1).

For government violence, this means maximum polar-ization between groups 1 and ABC or between A and BC,for rebel violence between groups 1 and ABC, and forviolence committed by infighters between A and BC.

Dow



olumbia U



Conversely, where the coethnics of fighters represent 90percent of the population, it makes little sense to act vi-olently against what most likely already is a quiescentand subdued minority. Where coethnics make up only10 percent of the population, the amount of violenceneeded to expel at least 40 percent of the other group isprohibitive—and it might even be undesirable if the gen-eral public perceives these locales as part of the “home-land” of their enemies.

A similar argument was first developed in Balcell’s(2010, 2011) study of the Spanish civil war fought inthe Catalan and Aragon regions. We integrate it into ourmodel of ethnoterritorial competition.6 We also go onestep beyond Balcell’s model by taking into account a sec-ond strategic interest of fighting organizations. Armedgroups are interested in establishing control over regionswhere the settlement areas of their coethnics and thoseof their opponents meet (i.e., in border areas [a morefine-grained version of this argument has been proposedfor Bosnia by Melander 2007; Weidmann 2011]). Thestrategic utility of a coethnic village in a border zone ismuch higher than that of a coethnic exclave deep in ter-ritory populated by the opponent’s coethnics. Thus, vio-lence should be more intense the closer a locale lies to aborder between the coethnics of rebels and governmentelites (Hypothesis 2; the border between 1 and ABC orA and BC for government violence, 1 and ABC for rebelviolence, and A and BC for infighters).7 These two hy-potheses are markedly different from those derived fromthe “weakening the enemy”perspective, to whichwe nowturn.

Weakening the Enemy’s Ethnic Support Base

This second theory of ethnic targeting has been proposedby Fjelde and Hultman (2014). Armed groups attacktheir enemy’s coethnics, they argue, in order to undercutits logistical, informational, and material support. Thismakes strategic sense since in ethnic conflicts, a civilian’sethnic background provides a good informational short-cut to determine their political leanings (ibid.; Olssonand Siba 2013; Sullivan 2012). In the setup described byFigure 1, government troops should therefore attack

6 Weidmann (2011) confirmed the polarization hypothe-sis with regard to Bosnia, but argued that polarizationdrives violence because of local competition for jobs,public goods, and other local resources. For a critique ofexplaining ethnic conflicts with economic competitionarguments, see Horowitz (1985,105–35); Wimmer (1997).

7 Border areas are also those where most polarized lo-cales can be found (the average polarization index thereis 0.28 compared to 0.0007 overall).

civilians of groups 1 and A (Hypothesis 3) while spar-ing civilians of groups BC. Rebels associated with group1 in turn should attack all included groups ABC (Hy-pothesis 4) and avoid groups 1 to 4 (and especially 1),while infighters from group A should attack groups BC(Hypothesis 5) and avoid all others (especially their owngroup A). Since there are no further strategic consider-ations according to this theory, rebels and governmenttroops should kill civilians of these target groups wher-ever they live. By implication, violence should be moreintense the higher the share of the target population is inan area (which is roughly what Gulden [2001] finds forGuatemala).

Fjelde and Hultman (2014), using data sources simi-lar to ours, test whether areas that contain at least somecoethnics of the adversary see more violence than ar-eas without any of the adversary’s coethnics.8 However,this does not adequately capture the supposed mecha-nism. If fighters seek to weaken their opponents’ sup-port base, they should kill more civilians the larger thelocal share of their opponents’ ethnic base is in an area.We therefore test their theory with a dummy variable forhexagons entirely populated by the adversary’s coethnics.Alternatively, we employ a continuous measurement (thepercentage of the enemy’s coethnic population), whichyields substantially similar results (for details see OnlineAppendix 3).

Other Theories: Controlling Resources, Fighters,

or Territory

We now briefly discuss the three major theories of civil-ian victimization, all of which claim that where and whycivilians are killed has nothing to do with the politicalstruggle between insurgents and governments and is thusentirely independent of civilians’ stance in that macrop-olitical conflict. In other words, violence against civiliansis supposed to follow a nonpolitical logic, determined byeither economic considerations (gaining access to naturalresources to enrich the fighters and/or fund their opera-tions), organizational logics (whether or not rebel lead-ers can keep their fighters under control), or locally spe-cific military tactics (eliminate possible collaborators inan area).

LootingThe looting argument derives from the greed schoolin the civil war literature (Le Billon 2001; Collier and

8 Using the percentage of enemy coethnics as an inde-pendent variable, Sullivan (2012) and Olsson and Siba(2013) arrive at similar conclusions for Guatemala andDarfur.

Dow



olumbia U



Hoeffler 2004), according to which governments andrebels fight to gain control over lootable resources, suchas the famed “blood diamonds” of Sierra Leone, oil re-sources in the Nigerian delta, or the rare metals foundin Eastern Congo. Azam (2002) has applied this modeof reasoning to violence against civilians (see also Hegre,Ostby, and Raleigh 2009; Querido 2009). In his game-theoretic model, violence against civilians represents aside effect of the struggle over lootable goods, which areneeded to pay fighters. Once violence against civilianshas set in, young men have incentives to join an armyor rebel force rather than to farm their land, leading toa self-reinforcing equilibrium characterized by high lev-els of violence. The hypothesis derived from this perspec-tive is straightforward: violence against civilians shouldbe more frequent in areas with abundant lootable natu-ral resources, including diamonds, oil, and minerals (Hy-pothesis 6).9

Loose Control over FightersWeinstein (2006) focuses on the initial conditions

that shape a rebel organization and thus its subsequentbehavior vis-à-vis the civilian population. Rebel orga-nizations that were funded either by natural resourcesor by external actors such as a foreign government at-tracted more opportunistic fighters interested in imme-diate economic gains rather than in achieving a politicalgoal. Such organizations will also exhibit a lower levelof control over the rank-and-file and less hierarchical in-tegration. As a consequence, these rebel forces are morelikely to prey upon the local population and kill civilians(for Sierra Leone, see Humphreys and Weinstein 2006).Three hypotheses follow from this approach: rebel or-ganizations that command less control over their fight-ers (Hypothesis 7) or rely on natural resources to fundtheir operations (Hypothesis 8) or on external financial

9 Other authors have specified the looting argument dif-ferently and maintained that civilians are killed in ar-eas with rich crop harvests (Koren and Bagozzi 2017),especially during periods of draught (Bagozzi 2017; forIndia, Eyndre 2016) or where aid organizations (Wood2015) or national governments (Khannaa and Zimmer-man 2017; for Columbia, Weintraub 2016) provide localswith lootable resources. In order not to complicate thepicture and for reasons of data availability, we don’tattempt to test these additional greed arguments butnote here, en passant, that in the statistical models dis-cussed below, violence is not less likely in places withsparse vegetation or where the soil quality is bad andthus less suitable for agriculture.

support10 (Hypothesis 9) should be more likely to com-mit violence against the civilian population.

Containing CollaborationAccording to Kalyvas’s (2006) well-known study of theGreek civil war, rebels and governments in irregular civilwars selectively kill civilians in order to eliminate collab-orators and prevent future defections. Rebel and govern-ment soldiers kill civilians when they have not yet fullyestablished control over a territory. They are less likelyto do so when they have either no control (in contestedareas) or have already fully established themselves andthus eliminated any opposition to their rule. Civilians,on the other hand, denounce each other to settle scoresfrom local-level feuds again unrelated to the macrocon-flict. This is most risky if the area is heavily contested be-tween war participants where civilians do not know whowill eventually prevail. The combination of these two be-havioral logics leads Kalyvas to expect that violence ishighest where a good supply of denunciations meets ahigh demand for identifying collaborators, that is, in ar-eas of intermediate control. Such intermediate levels ofcontrol should be found in areas where control recentlyshifted from one side to the other (Hypothesis 10) with-out, however, having shifted hands many times in therecent past. This latter situation would imply contesta-tion, which according to Kalyvas should be less violence-prone.

The set of theories discussed so far represent the mostwidely cited and empirically specific approaches to vio-lence against civilians in civil wars. Not all of them areincompatible with each other or mutually exclusive. Forexample, it is possible that a looting logic combines withthat of ethnoterritorial competition. This article does notattempt to disentangle such complex causal relationships,but seeks to evaluate whether some of the basic observ-able implications of each theory hold empirically. Futurework could certainly do more to discern possible inter-action effects or to identify the scope conditions underwhich one or the other of these theories is more likely tohold, even if a general analysis does not produce statisti-cally significant results in their support. The next sectiondiscusses the data and variables used to evaluate the var-ious hypotheses introduced above.

10 In an extension of the argument, Salehyan and coau-thors (Salehyan et al. 2014) show at the level of rebelorganizations that violence against civilians decreasesif an organization is externally supported by a demo-cratic regime and increases if the supporting state isnot democratic. For data reasons we again refrain fromtesting this more nuanced version of the argument.

Dow



olumbia U



Data and Model Specification

Dependent Variables and Units of Observation

As indicated in the opening paragraphs of this article,there are two available datasets with geocoded informa-tion on violence against civilians (ACLED and UCDPGED). For reasons discussed below, we use the UCDPdataset for all analyses except to evaluate Kalyvas’s the-ory, for which the ACLED dataset is more appropriate.The dependent variable in both datasets is defined asan event during which violence against unarmed civil-ians was committed intentionally by government sol-diers, rebel forces, or militias—thus ignoring “collateraldamage”of armed battles. Data on violent events is muchmore reliable than estimates of the number of deaths.Wetherefore define the dichotomous dependent variable asany event in which an armed group intentionally killedat least one civilian.11 For robustness purposes, we runthe main models with a count variable of the number ofcivilian deaths and report the results in Table 3.

The UCDP GED dataset is limited to Africa. It cov-ers all years from 1989 to 2010 and all countries wherearmed groups killed at least twenty-five civilians. It canbe linked to other datasets on armed conflict, conflict par-ticipants, their external support, their links to politicizedethnic groups, etc. Since all the theories discussed aboverelate to violence during civil war, we limit the universeof observations to years and countries during which anarmed conflict was active (i.e. where at least twenty-fivebattle deaths were counted in the UCDP armed conflictdataset (Gleditsch et al. 2002).

The ACLED dataset includes some conflict-pronecountries in the Middle East and Asia as well. It has beencriticized for a number of reasons, most importantly inreference to the quality of the coding (Eck 2012; the re-cent updates by ACLED seem to have addressed theseissues). Compared to UCDP, ACLED has the advantageof offering information on territorial gains and losses byrebels and government troops, which allows to evaluateKalyvas’s theory. For the rest of the analysis, however,

11 We prefer this dichotomous coding of the dependentvariable over a count version since we are not theoreti-cally interested in howmany violent episodes a hexagonexperiences, but in the more fundamental contrast be-tween peace and violence. This also represents theexplanatory focus of the theories we intend to evalu-ate. Moreover, some of the reliability problems in eventcounts based on media reports (seeWeidmann 2012)—such as how to distinguish separate episodes fromeachother—are reducedwhen using a dichotomously codedvariable.

we rely on UCDP data because, as mentioned above, it islinked with other datasets relevant for this study.

Both datasets identify perpetrators and we can there-fore distinguish between violence committed by govern-ment or rebel organizations. In order to test the two eth-nic targeting arguments, we need to further disaggregatethe dependent variable and distinguish between violencecommitted by infighters and by rebels, and we need toidentify the ethnic communities that each armed organi-zation claims to represent. This is made possible thanksto the actor identification number offered in both theUCPD-GED and the EPR datasets (Wucherpfennig et al.2012).

Both the ACLED and the UCDP datasets contain pre-cise geocoding for events. In order to make this data us-able for a cross-sectional time-series analysis, we overlaya grid of fixed territorial units in the form of hexagons of27.75 km2 size or one-quarter of a degree of latitude andlongitude. We chose hexagons as units of observation,rather than settlement units such as villages and towns,because much of Africa’s population south of the Saharalives in sparsely populated and dispersed hamlets outsideof towns and villages (Herbst 2000). More importantly,there are no data on each hamlet or village’s ethnic com-position and we thus have to rely on the coarser ethnicmaps provided by the GeoEPR dataset (see below). Therobustness section reports results from other specifica-tions of the units of observation, such as larger rectanglesor politically defined units of various size.

We disregarded the fine-grained day-by-day coding ofviolent events in both datasets and instead aggregated tothe year—the level of granularity of many other inde-pendent variables. To explore Kalyvas’s argument, how-ever, a yearly data resolution might not be fine-grainedenough. It may well be, for example, that governmenttroops gain control of a territory in December and killcivilians during January—a link that will be lost in adataset based on years. We therefore created a datasetwith month-hexagons as units of observations, using theACLED dataset for the reasons discussed above, and wethen worked with one-, two-, and three-month lags of theindependent variables.

Independent Variables

The following gives a short overview of the variousindependent variables coded for this project. We referthe reader to Online Appendix 1 for more detail oncoding procedures and data sources. Online Appendix2 provides summary statistics. To assess greed theoriesof violence against civilians, we focus on three differ-ent types of lootable resources: the number of active,

Dow



olumbia U



on-shore oil wells (from Lujala, Gleditsch, and Gilmore2005); whether or not there are diamond deposits in ahexagon (from Gilmore et al. 2005); and the numberof economically significant mining sites (e.g., for Tung-sten and other rare metals) per hexagon (data are fromthe National Minerals Information Center of the UnitedStates Geological Survey, see https://minerals.usgs.gov/minerals/pubs/country/data).Other codings of these vari-ables (e.g., the km distance to natural resource sites)are used for robustness purposes and produced similarresults.

To evaluate Weinstein’s argument, we rely onCunningham, Gleditsch, and Salehyan’s (2009) dataseton rebel organizations to code a dichotomous variable in-dicating whether or not a rebel organization received anykind of financial support from actors situated outside ofthe focal country. The second variable measures the ex-tent of control that rebel leaders exercise over the rankand file. We draw on a dataset by Rustad and MalminBinningsbo (2012) for information regarding whether arebel organization was funded by natural resources (in-cluding drugs). As with the external financing variable,there is no information on whether rebels relied on suchresources when they first organized, once the conflictwas underway, or both. We thus cannot test Weinstein’s(2006) path dependency argument. To avoid confound-ing civilians killed by other rebel organizations with dif-ferent internal organization or funding sources, we re-strict the sample in these statistical models to areas withinwhich a rebel organization operates.

Kalyvas’s theory is the hardest to evaluate empiricallybecause there is no data on actual levels of territorial con-trol by specific actors. It is perhaps not too far-fetched,however, to assume that, in recently acquired territories,an armed group has not yet established full control butalso no longer shares control with its adversary. Data onterritorial gains and losses is available in ACLED, as dis-cussed above. We create two count variables indicatinghow many times either government or rebels have gainedterritory on a hexagon, whether through battle or peace-fully. To check for robustness, we also code a distancevariable—assuming that the degree of control over terri-tory monotonically decreases with distance to the loca-tion of a territory change. The results are substantiallyvery similar.

To exclude the possibility that we are looking atcontested territories that frequently switch back andforth between rebels and government troops (where civil-ians should be safer according to Kalyvas), we excludehexagons with territorial gains in favor of both rebelsand governments in the previous twelve months. To taketwo additional scope conditions of Kalyvas’s theory into

account, we exclude conventional civil wars from consid-eration (using the war typology in Kalyvas and Balcells2010), since the theory applies to guerilla wars only. Thisreduced the number of events from 2,834 to 2,506. Wealso exclude events with more than five civilian deathsfrom the coding of the dependent variable because in-discriminate killings (such as massacres) are also not thefocus of Kalyvas’s argument (we adopt the threshold todefine indiscriminate violence from Sullivan 2012). Thismeans that we exclude an additional 983 events. Forrobustness purposes, we defined the dependent variablein a less restricted way to include contested territories,guerilla wars, and high levels of violence with more thanfive civilian victims. The results, shown in Table 3, aresimilar.

To evaluate the two ethnic targeting arguments,we rely on the Ethnic Power Relations (EPR) dataset(Wimmer, Cederman, and Min 2009), which also existsin a geocoded version for the subset of ethnic groups thathave identifiable territories of settlement (Wucherpfenniget al. 2011; we use the EPR-ETH edition of the data).The EPR dataset identifies all politically relevant ethnicgroups12 and their position in the overall power config-

12 To be more precise, an ethnic category is politically rel-evant if at least one actor (a political movement, or aparty, or an individual) with some minimal resonance inthe national political arena claims to speak for this cate-gory or if outsiders consider the category to be relevantby discriminating against its members (see for the fol-lowing, Wimmer, Cederman, and Min 2009). EPR doesnot code individuals, parties, or movements as “rep-resenting” an ethnic group if these actors cannot ac-knowledge their ethnic background in public or pub-licly pursue group interests. In line with constructivistnotions of ethnicity, the list of relevant categories canchange over time, and categories can fission or fuse.EPR is based on an encompassing definition of ethnic-ity (see, e.g., Wimmer, Cederman, and Min 2009) thatincludes groups with a distinctive religion, language,race, culture, or profession (as in caste systems).EPR lists the political status of each ethnic category foreach year by evaluating whether group members canbe found at the highest levels of executive government,such as the cabinet in parliamentary democracies, theruling circle of generals in military dictatorships, thepolitburo in communist countries, and so on. The mea-surement is thus independent from whether the regimeis democratic or not. Levels of representation are mea-sured through an ordinal scale. It ranges frommonopolypower (total control of executive government by repre-sentatives of a particular group) to dominance (some

Dow



olumbia U


https://minerals.usgs.gov/minerals/pubs/country/data


uration rendered in Figure 1. EPR also codes whether arebel organization fights in the name of a particular eth-nic group, providing a rebel identification number thatlinks EPR to other datasets (Wucherpfennig et al. 2012).

This allows identifying the local, hexagon-specificpopulation-shares needed to evaluate the two ethnic tar-geting arguments: the share of coethnics of governmentelites (ABC in Figure 1), of rebels (1), of infighters (A),of infighters’ adversaries (BC), as well as the correspond-ing polarization figures. We use Montalvo and Reynal-Querol’s (2005) well-known formula to calculate the lat-ter. Finally, Geo-EPR allows to calculate the kilometerdistance to the nearest border between the settlement ter-ritories of groups in conflict (i.e., the coethnics of rebelsor infighters on the one hand and government’s coethnicson the other hand) (for details see Online Appendix 1).Because the logic of ethnic targeting should not apply inareas where ethnicity is not a relevant factor for politicalcompetition, we restrict the sample to areas populated byat least one politically relevant ethnic group, but loosenthis restriction when running robustness models.

The probability that armed groups kill civilians de-pends on other, theoretically less interesting factors thatrelate to the sheer number and accessibility of civilians.We experimented with a large number of variables frommany different geocoded datasets and produced a se-ries of baseline models that were then tested for multi-collinearity. Online Appendix 1 describes these variablesand the data sources.

To facilitate orientation, Table 1 gives an overviewof the main arguments, hypotheses, the operational mea-surements, the datasets and their temporal resolution, aswell as the exact specifications of the dependent variable.

Model Specifications

The data are arranged as a pooled time-series cross-sectional dataset with forty-three thousand year-hexagons (UCDP), or 344,000 month-hexagons(ACLED), used to test Kalyvas’s argument. For themain analysis, we specify models as logistic regressionswith violence against civilians as the dichotomousdependent variable, standard errors clustered on thehexagon, and controls for past violence as well as forviolence in neighboring hexagons that might spill over

members of other groups hold government positions),senior partner in a power-sharing arrangement, juniorpartner, representation at the regional level (e.g., in aprovincial government), powerless, and discriminatedagainst (i.e., targeted exclusion from any level of rep-resentation).

into the focal hexagon. Country-fixed-effects modelstake into account time-invariant country characteristicssuch as levels of rule of law, the security forces’ capacityto protect citizens from violence, the type of civil war, etc.

To check for robustness of the results to different spec-ifications, we run all models without fixed effects andspecify the final model with the rare events logit estima-tor provided by King and Zeng (2001), as well as a zero-inflated negative binomial regressions with the numberof dead civilians as the dependent count variable.13 Fi-nally, we also rearranged the dataset into a cross-nestedhierarchical model (with hexagons nested into countriesand years). All these different model specifications pro-duce substantially very similar results, as briefly discussedbelow.

Results

The dependent variable changes depending on the the-ory evaluated: all violence against civilians for the lootingmodel; violence committed by rebel organizations whentesting the loose control over the rank-and-file approach;by rebels or governments to evaluate the containing col-laboration hypothesis; and by ethnic rebels, infighters,or government troops to test the two different theoriesof ethnic targeting. We therefore need to proceed in astep-wise fashion and evaluate each theory with its ownmodels.

In a final step, we present two models that integrateall significant variables from either the UCDP or theACLED datasets and define any violence against civiliansas the dependent variable. This will allow testing whetherthe results of the previous, actor-specific models hold upwhen regressing on all violence committed by any kind ofactor. It will also allow us to determine whether the eth-nic targeting models can be confirmed when taking intoaccount areas where ethnicity is not relevant and whencontrolling for mechanisms identified by other theoreti-cal approaches.

Looting

The first model in Table 2 shows no support for the ideathat violence against civilians is driven by competitionfor lootable natural resources (in line with Bellows andMiguel [2006] on Sierra Leone and Hegre et al. [2009]on Liberia). Civilians are not more likely to become vic-tims of violence when they live in hexagons with manyoil-producing wells, that contain a diamond deposit, or

13 Test statistics indicated that a zero-inflated model is tobe preferred over others.

Dow



olumbia U



Table 1. Arguments, hypotheses, and measurements

Looting Loose controlContainingcollaboration

Weakening theenemy

Ethnoterritorialcompetition

Main argument Civilians who arein the way ofaccessing andexploitinglootable resourceswill be killed

Rebels relying onexternal aid ornatural resourcefunding will haveless control overrank and file andwill kill morecivilians

Civilians are killedwhere rebels orgovernmentsoldiers haven’tyet established fullcontrol

Civilians arekilled if they shareethnic ties withthe attacker’sopponent

Civilians arekilled in localitiesclose to settlementborders betweengroups in conflictand where thesereach parity

Main motivation forkilling civilians

Control overresources

Looting by rankand file

Eliminatingcollaborators

Weaken civiliansupport of enemy

Expanding areaunder control ofcoethnics

Do politicalaffiliations of victimsmatter?

No No No Yes Yes

Does ethnicbackground ofvictims matter?

No No No Yes Yes

Where the theorypredicts civilianvictims

1) Areas withmanyoil-producingwells

1) Area withrebels that receiveoutside financialsupport

1) Areas in whichgovernmentacquired territoryare targeted bygovernmenttroops

1) Areaspopulated by thecoethnics of rebelsare targeted bygovernmenttroops

1) Areas close tothe settlementborder betweengroups in conflict

2) Areas withdiamond deposits

2) Areas withrebels that arefunded by naturalresources

2) Areas in whichrebels acquiredterritory aretargeted by rebels

2) Areaspopulated bycoethnics ofgoverning elitesare targeted byrebels

2) Areas with highlevels ofpolarizationbetween coethnicsof governing elitesand of rebels

3) Areas withmanyeconomicallysignificant miningsites

3) Areas withrebels thatmaintain weakcontrol over rankand file

Dataset tocode-dependentvariable

UCDP UCDP ACLED UCDP UCDP

Temporal resolution Yearly Yearly Monthly Yearly YearlySample restrictions Only countries

and years with anongoing civil war

Only countriesand years with anongoing civil war

Only countriesand years with anongoing civil warfought as aguerilla (notconventional) war



Only geographicareas where rebelsare active

Only geographicareas with at leastone politicallyrelevant ethnicgroup

Only geographicareas with at leastone politicallyrelevant ethnicgroup

Dependent variable All violenceagainst civilians

Violencecommitted byrebelorganizations

Violencecommitted byrebelorganizations orgovernments(separate models)

Violencecommitted byethnic rebels,infighters, orgovernmenttroops (separatemodels)

Violencecommitted byethnic rebels,infighters, orgovernmenttroops (separatemodels)

Dow



olumbia U



Table

2.Logisticregressionsonevents

ofviolenceagainstcivilians

12

34

56

78

9

Dependentvariable

Allviolence

Rebelviolence

Low

-levelgov.

violence

(ACLED)

Low

-levelrebel

violence

(ACLED)

Gov.violence

Rebelviolence

Infighter

violence

Allviolence

Allviolence

Independentvariables

Presen

ceof

oil-prod

ucingwells

(dum

my)

−0.175

8

(0.480

)Presen

ceof

diam

ondde

posit

(dum

my)

0.05

37

(0.297

)No.

ofecon

omically

sign

ificant

miningsites

0.03

48

(0.080

)Streng

thof

controlo

ver

rank

-and

-filerebe

lfigh

ters

0.03

43

(0.073

)Outside

finan

cial

supp

ortforrebe

ls(dum

my)

0.05

11

(0.140

)Reb

elsfund

edby

naturalresou

rces

(dum

my)

−1.330

5***

−1.174

6***

(0.159

)(0.171

)No.

oftimes

gov.ga

ined

territory

0.27

060.34

96**

(0.217

)(0.156

)No.

oftimes

rebe

lsga

ined

territory

0.46

68**

*0.13

08(0.125

)(0.193

)Distanc

eto

anyterritorych

ange

−0.553

6***

(0.057

)Entirehe

xpo

pulation

inrebe

llion

−0.252

8(0.656

)

Dow



olumbia U



Table

2.Continued

12

34

56

78

9

Dependentvariable

Allviolence

Rebelviolence

Low

-levelgov.

violence

(ACLED)

Low

-levelrebel

violence

(ACLED)

Gov.violence

Rebelviolence

Infighter

violence

Allviolence

Allviolence

Entirehe

xpo

pulation

includ

edin

governmen

t−0

.870

3***

0.11

680.08

90

(0.318

)(0.123

)(0.125

)Entirehe

xpo

pinclud

edan

dno

tin

rebe

llion

0.54

80**

(0.263

)Po

larization

btw.inc

lude

dan

drebe

lliou

spo

pulation

0.55

66**

*1.54

86**

*3.06

95**

*0.44

94**

0.50

50**

(0.191

)(0.298

)(1.016

)(0.224

)(0.216

)Km

distan

ceto

nearest

border

betw

eengrou

psin

confl

ict(log

ged)

−0.090

3**

−0.101

1*−0

.660

6***

−0.105

3***

−0.123

0***

(0.037

)(0.055

)(0.118

)(0.032

)(0.029

)

Cou

ntry

fixed

effects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Con

trol

variab

les

AA

BB

AA

AA

AObserva

tion

s40

,739

12,685

249,16

225

9,60

029

,397

19,305

11,676

24,805

32,675

Notes

onnu

mbe

rsof

observations

Drops

coun

triesan

dyearswitho

utviolen

ce

Con

flict

areas

only;d

rops

coun

triesan

dyearswitho

utan

yrebe

lviolen

ce

Drops

coun

tries

and

mon

ths

with

anon

going

conv

ention

alwar,h

exes

witho

utan

yviolen

ce,

and

hexe

sthat

chan

ged

hand

sbe

tweenrebe

lsan

dgo

vernmen

tpa

styear

Drops

hexe

switho

utat

least1

politicallyreleva

ntethn

icgrou

p;drop

sco

untriesan

dyearswitho

utan

yviolen

ce

Onlyhe

xes

withbo

thACLED

and

UCDPda

ta;

drop

sco

untriesan

dyearswitho

utan

yviolen

ce

Drops

coun

triesan

dyearswitho

utan

yviolen

ce

Notes:(1)

Rob

uststan

dard

errors

inpa

rentheses;statisticalsignifican

celevels:***

p<

0.01

,**p

<0.05

,*p

<0.10

.(2)

Ainclud

esthefollo

wingcontrolv

ariables:v

iolencelast

year,v

iolencein

neighb

oringhexa

gons,close

to

border

(dum

my),d

istanceto

city,d

istanceto

capital,road

leng

th,u

rban

popu

lation

,rou

ghness

ofterrain.

(3)Binclud

esthefollo

wingcontrolv

ariables:d

istanceto

battle

witho

utch

ange

interritory,violence

last

mon

th,v

iolence

inneighb

oringhexa

gons,u

rban

popu

lation

size,close

tobo

rder

(dum

my),traveltim

eto

city,lengthof

road

s,sparsely

vegetatedsurface,roug

hnessof

terrain.

Dow



olumbia U



where economically significant minerals are mined. Thesame holds true if we use alternative specifications ofthese variables: distance to production sites, number ofsuch sites within a radius of 50 or 100 km, or a codingof conflict minerals (results not shown).14

Loose Control Over Fighters

Models 2 and 3 reduce the universe of observations tothose geographic areas where rebel organizations havebeen active in order not to misattribute violence com-mitted by a specific rebel group to another one with dif-ferent organizational characteristics and sources of fund-ing (using the UCDP GED Conflict Polygon dataset, seeCroicu and Sundberg 2012). According to Model 2, thestrength of hierarchical control over fighters is not asso-ciated with the propensity of a rebel organization to killcivilians.15 Financial support by outsiders is also not cor-related with the chances of civilian victimization—whileadditional results (not shown) demonstrate that provid-ing rebels with troop support from the outside, a sanctu-ary, or weapons increases their propensity to kill civilians.In other words, rebels’ military capacity to kill is moreimportant than the source of their revenues, as main-tained by the loose control theory.With regard to naturalresource funding, we find the opposite of what the loosecontrol argument predicts: rebel organizations that profitfrom the sale of natural resources are less likely to killcivilians than other organizations. Perhaps this is becausesuch rebel organizations are simply not interested in es-tablishing control over local populations because there isless need to rely on their economic support.

Containing Collaboration

In order to evaluate Kalyvas’s theory of violence againstcivilians, we now shift to the ACLED hexagon-monthdataset, for the reasons explained above. The number ofobservations is larger because the units of observationsare now hexagon-months, rather than years. Model 3 in-dicates that government troops do not kill more civiliansas they gain territory in a hexagon during the last month.However, they do cause more civilian deaths where theylost territory to rebels, possibly killing civilians whileretreating or in areas adjacent to positions they lost

14 We also do not find any association with opium orcannabis (data from Buhaug and Lujala 2005).

15 We obtain the same results if we include the relativemilitary strength of rebels (in contrast to Wood 2010) ora variable codingwhether or not they effectively controlterritory—two other variables that capture the militarycapacity of rebel organizations to kill civilians.

(perhaps out of revenge, frustration, or rage). Model4 suggests that rebels also kill where they lose terrain,rather than where they gain it. Note that these resultsremain similar if we use a two- or three-month lag ora hexagon-year version of the dataset without any lags(results not shown).

As discussed above, the codings of independent anddependent variables used in Models 3 and 4 take thescope conditions of Kalyvas’ theory into account as wellas possible. Specifically, they exclude contested territoriesthat shift “sides” more than once during the last twelvemonths (and where civilians should be safer accordingto the theory); events with more than five civilian deaths(which are unlikely to represent selective targeting); andthose occurring in the midst of conventional wars. Mod-els 1 and 2 in Table 3 show the results when we loosenthese restrictions and take all violence in all kinds of civilwars into account. The results remain broadly similar,except that rebel forces now also target civilians wherethey recently acquired territory (in line with Kalyvas’stheory; but they still kill civilians where they lost terri-tory as well).

What emerges from these results is that civilians aremore likely to become victims the closer they live to ac-tual fighting, and the more accessible they therefore areto fighters. All signs of the coefficients of the territorychange variables are positive if they are significant. Evenmore importantly, the coefficient for the number of bat-tleswithout territory changes is also associated with civil-ian deaths in statistically significant ways (not shown). Inthe final models, we therefore need to control for the dis-tance to any kind of territory change to account for thesimple effects of proximity to troops and rebels that fighteach other.

Ethnic Targeting

Models 5 to 7 in Table 2 evaluate the two competingethnic targeting arguments. Each model refers to a dif-ferent perpetrator-victim dyad, and, accordingly, we usethree different dependent variables: government violenceagainst civilians (Model 5); violence committed by rebels(Model 6); and violence by “infighters” (Model 7).We donot find consistent support for the idea that armed groupsattack their adversaries’ coethnics where these are mostnumerous, in contrast to Fjelde and Hultman (2014).Government troops do not target hexagons entirely pop-ulated by the coethnics of rebels or infighters more thanthey do other hexagons (Model 5). Rebel fighters evensignificantly avoid, rather than target, hexagons exclu-sively populated by the coethnics of government elites(Model 6). We arrive at the same conclusions if we use

Dow



olumbia U



Table

3.Differentmodelspecifications

Diff.specificationsofModels3and4in

Table2

Diff.specificationsofModel9inTable2

12

34

Dependentvariables

Gov.violence

(ACLED)

Rebelviolence

(ACLED)

Rareevents

logit

0-inflatednegative

binomialregressionDV:

numberofeventsduring

which

civilians

were

killed

Independentvariables

>0

0No.

oftimes

governmen

tga

ined

territory

0.17

050.25

71**

*(0.151

)(0.098

)No.

oftimes

rebe

lsga

ined

territory

0.33

44**

*0.21

69**

*(0.074

)(0.070

)No.

ofoilp

rodu

ctionsites

−0.854

2*−1

.511

8***

0.30

12(0.505

)(0.539

)(0.564

)Natural

resource

fund

ingof

rebe

ls−0

.773

8***

0.24

951.64

89**

*(0.170

)(0.342

)(0.270

)Po

larization

btw.inc

lude

dan

drebe

lliou

spo

pulation

0.71

04**

*2.24

12**

*0.08

94(0.149

)(0.385

)(0.348

)Distanc

eto

border

betw

eenethn

icgrou

ps(inkm

,log

ged)

−0.103

8***

−0.189

1***

0.07

26**

(0.022

)(0.054

)(0.035

)Cou

ntry

fixed

effects

Yes

Yes

No

No

Con

trol

variab

les

AA

BB

Observa

tion

s31

4,83

230

7,47

633

,106

33,106

33,106

Notes

onno

.ofob

servations

Drops

coun

triesan

dyearswitho

utan

yviolen

ce

Notes:(1)

Rob

uststan

dard

errors

inpa

rentheses;statisticalsignifican

celevels:*

**p

<0.01

,**p

<0.05

,*p

<0.10

.(2)

Ainclud

esthefollo

wingcontrolv

ariables:v

iolencelast

mon

th,v

iolencein

neighb

oringhexa

gons,close

tobo

rder

(dum

my),d

istanceto

city,lengthof

road

s,urba

npo

pulation

,rou

ghness

ofterrain,

distan

ceto

battleswitho

utterritorychan

ge.(3)

Binclud

esthefollo

wingcontrolvariab

les:violence

last

year,v

iolencein

neighb

oring

hexa

gons,close

tobo

rder

(dum

my),d

istanceto

city,d

istanceto

capital,road

leng

th,u

rban

popu

lation

,rou

ghness

ofterrain.

Dow



olumbia U



a share of the enemy’s coethnics as an independent vari-able (results not shown). Both results together allow usto conclude that neither rebels nor governments attacktheir enemy’s coethnics where they are more numerousand thus easiest to target. Only in models of violencecommitted by infighters does the strategic logic of at-tacking one’s enemy’s support base appear to be at work(Model 7).

In Online Appendix 3, we offer a detailed explana-tion for why our findings diverge from those of Fjeldeand Hultman (2014), who use similar data sources. Inthe case of government violence, Fjelde and Hultman’somission of polarization and distance from their mod-els explain the different results. Regarding rebel violence,the findings become consistent with the ones presentedhere if we drop some of their (highly collinear) controlvariables, or if we use the substantially more meaning-ful share of the included population as an independentvariable, rather than a dummy for the presence of anymembers of the included group.

The ethnoterritorial competition argument, on theother hand, predicts that levels of polarization betweengroups in rebellion and those in power should determinethe likelihood of violence against civilians. Indeed, thesign of the polarization measure is significant in the ex-pected direction and at the highest level in all models. Theeffect is also substantially important: increasing polariza-tion by one standard deviation increases the probabilityof violence by about 0.45. We also find evidence for thesecond mechanism driving ethnoterritorial competition:competing fighting organizations attempt to expand thecontinuous territory under the control of coethnics by fo-cusing violence on the borders between ethnic settlementareas. The variable that measures the kilometer distanceto the nearest boundary between ethnic group(s) repre-sented in government and those in rebellion is statisti-cally significant in all models. A one standard-deviationincrease in this distance decreases the probability of vi-olence by approximately 0.2—a magnitude in the orderof some of the more important control variables. Thus,in line with the ethnoterritorial competition model, civil-ians are killed in areas that allow fighting organizationsto expand the territory under control of coethnics wherethis is strategically most useful and at the same time mosteffective.

Spuriousness and Endogeneity

It could be, however, that the relationship between polar-ization and violence is spurious because we do not haveany information on the ethnic background of victims andthus cannot exclude that an armed group attacks its own

coethnics, rather than those of its adversary. Additionalanalysis (not shown here) demonstrates that “neutral”populations (groups 2–4 in Figure 1 for rebellion andgroups 1–4 for infighting) are significantly less likely to betargeted by government troops and rebels (these resultsdo not hold up for infighters, however). We also find thatneither rebels nor infighters specifically target their owncoethnic civilians. It is therefore less likely that the asso-ciation between polarization and violence is spurious.

Additional support for the idea of ethnic targetingcomes from country-level studies for which informa-tion on the ethnic background of villages, which are of-ten ethnically homogenous, is available. They show thatarmed organizations rarely attack their coethnic popula-tions (for Olsson and Siba 2013; for Guatemala, Sullivan2012). One of the rare datasets with information onthe ethnic background of individual victims comes fromNorthern Ireland. A full 74 percent of all individualskilled by Protestant paramilitaries, for example, wereCatholic civilians. By contrast, only 16 percent of the vic-tims were Protestant civilians (e.g., family members of ri-val paramilitaries or “traitors”).

It is also possible that polarization as well as the dis-tance to settlement border simply pick up the differencebetween any mixed territory and monoethnic hexagons,rather than specifically a polarized ethnodemographicconfiguration and the continuous distance to where ter-ritories overlap. To evaluate this possibility, we codedmore empirically intuitive variables indicating whether0–10 percent of the population belongs to rebelliousgroups,11–20 percent, 21–30 percent, etc. The resultsshow that most of the government violence occurs whenits coethnic population in a hexagon reaches between 50and 60 percent of the total population. For rebel vio-lence, a share of the coethnic population, between 60 and70 percent of the total population yields the most vio-lence. It thus seems that both government and rebels aremore likely to attack civilians where their coethnic popu-lation is already in a slight demographic majority—quitein line with the general thrust of the ethnoterritorial com-petition argument. Note, however, that these results lookdifferent without country fixed effects, and we thereforeshould perhaps not rely too much on them.

Finally, we briefly address concerns about reverse cau-sation. Could it be that violent cleansings produce apolarized local settlement pattern by driving out mem-bers of the opposed ethnic group until they approx-imate half of the population—a relationship that ismasked by the yearly resolution of the data? The tem-poral order of how independent and dependent vari-ables were measured makes this rather unlikely: much ofthe African geocoding of EPR relies on maps drawn by

Dow



olumbia U



Soviet ethnographers in the 1960s, long before the re-cent wave of civil wars and long before the data series onviolence against civilians started in 1989. In some casesof massive population shifts, Geo-EPR offers differentmaps for different periods. We made sure that our eth-nic settlement maps refer to periods before the conflictdata series sets in. In other words, temporal ordering mit-igates against some of the most obvious reverse causationproblems.

Final Models of All Violence Committed by Any

Actor in Any Conflict

The preceding analysis refers to violence committedby either government, or rebels, or infighters. Corre-spondingly, the universe of cases varies quite dramat-ically. How can we bring the three different ways ofbehaving toward civilians into one model that uses theentire range of observations? And how can we also bringback into the picture the other variables that proved tobe significantly associated with violence against civiliansin previous models, in order to see whether the logic ofethnoterritorial competition actually operates even whentaking other mechanisms into account? And finally, doesethnic targeting show up as a significant predictor of civil-ian victimization even if we include areas where ethnicityis not politically relevant, thus extending the sample to allcountries and areas under civil war?

Models 8 and 9 in Table 2 offer such a more en-compassing view. The dependent variable is all violenceagainst civilians. In Model 8, we include a variable thatmeasures the distance to a territory change from theACLED dataset because we have seen above that anykind of territory change (in whoever’s favor) is associ-ated with an increased likelihood of violence. Unfortu-nately, combining ACLED with UCDP data would re-duce the number of observations since ACLED containsdata from 1997 onward only.We therefore run a separatemodel with the ACLED-based territory change variable(Model 8) and another one (Model 9) with an additionalUCDP-based independent variable that proved to be sig-nificant in previous models: the natural resource fundingof rebels.

In bothModels 8 and 9,we evaluate the logic of ethnictargeting with a term that measures the degree of polar-ization between any group in rebellion (whether rebelsor infighters) and the included population. This variableallows us to capture the targeting logic of governments,rebels, and infighters alike. We now code 0 on the polar-ization variable for all those hexagons without any po-litically relevant ethnic groups—while these areas weredropped in the ethnic targeting models described above

(Models 5 to 7). The same goes for the distance to a bor-der between groups in conflict variable, which for thepurposes of Models 8 and 9 is coded as the distance be-tween included groups and any group in rebellion.

The results of both models are encouraging. Ethnopo-litical polarization and distance to ethnic settlement bor-ders are significant factors to understand violence againstcivilians, even if we include in our analysis countries andperiods where ethnic politics is not a major driving fac-tor,when we control for rebels’ capacity to kill, the effectsof natural resource funding, and the distance to territorychanges between rebels and government troops.

Robustness Checks

For robustness purposes, we run Model 9 as a rare eventlogistic regression as well as a zero-inflated negative bi-nomial with a count of the number of killed civilians asthe dependent variable. The results (see Models 3, 4, and5 in Table 3) are substantially very similar. We conduct asecond robustness check by using 50 km squares as unitsof observation and recoding all variables accordingly(on the modifiable areal unit problem, see Schutte andWeidman 2011). Results (not shown here) remain sub-stantially identical—except that in Models 6 and 8 thepolarization variable does not achieve standard levelsof significance since coarser units of observation loosemuch of the empirical information contained in the eth-nic settlementmaps.A third robustness tests reconstitutesthe entire dataset with third-level administrative units(or counties) as units of observation in order to checkwhether using more politically meaningful units changesany of the main results. We find that this is not the case(results again not shown).

Finally, we evaluate whether the results presented inTable 2 rest on a small number of observations only. Formodels that are restricted to hexagons in countries andyears in which ethnicity is politically relevant (Models 5to 7 in Table 2), we find that five hexagons from Mali,Uganda, Rwanda, and the DRC had either high residualor high leverage values. Removing them from the datasetdid not change anything, nor did running the analysisdeleting individual countries from the analysis.

Conclusions

A booming strand of research on civil wars seeks to un-derstand when civilians will be killed by combatants.Geocoded datasets allow linking the occurrence of suchevents to other features of the same localities and the ac-tors operating in them. This article shows that, contraryto popular perceptions and much of the policy discourse,

Dow



olumbia U



civilians in areas rich in diamonds, oil, or conflict min-erals are not more likely to become victims of violence.Furthermore, rebels funded by natural resources are lesslikely to kill civilians, while rebels with higher orga-nizational capacity seem to use it to kill civilians aswell—rather than to constrain the looting of their rank-and-file fighters. Finally, containing and preventing col-laboration in areas not yet fully controlled by armedgroups, as during the Greek civil war and in Vietnam(Kalyvas and Kocher 2009), does not seem to motivatearmed groups in Africa—at least as far as we can tellgiven the crude measurement of territorial control thatwe encountered in existing data.

We do find, however, consistent support for a morepolitical view of violence against civilians. Specifically,violence is part of a strategic struggle over local andnational-level political power—rather than over who cancontrol natural resources, collaborators, or fighters. Inconflicts where political cleavages align with ethnic divi-sions, both local populations and national-level political-military organizations fight over which group is entitledand empowered to dominate the political arena. In suchenvironments, governments and rebels have incentives totarget the strategically most useful areas where the settle-ment areas of their affiliated populations meet and wherethey can therefore expand the territory continuously pop-ulated by their own coethnics. Armed groups do so espe-cially when these populations are roughly of equal sizeand the local balance of power can thus be tilted in one’sfavor with the least effort. This twofold dynamic of eth-noterritorial competition is an important part of the puz-zle to understand violence against civilians.

There are ample opportunities for future research togo beyond what has been achieved in this article. Mostimportantly, causal inference is limited because we donot have information on the ethnic background of theindividuals killed. As such, we do not know whether ornot government and rebels are indeed targeting those ex-act populations foreseen by our theoretical model. Whilethere are strong indications that they are in fact doing so,as discussed in the spuriousness and endogeneity section,producing a dataset with a coverage similar to ACLEDor UCDP-GED that contains more information on the in-dividuals killed and those who perpetrated the violencewould represent an important step forward. Obviously,assembling such a dataset amounts to a herculean taskgiven the paucity of information in the news reports onwhich event datasets rely.

Finally, our theory of the political dynamics of eth-noterritorial competition was limited to ethnic con-flicts. However, as Laia Balcell’s work on the Spanishcivil war has shown, very similar mechanisms are present

where political cleavages and rebel-civilian relationshipsare based on shared political ideology (see also onAfghanistan,Hirose et al. 2017). It would represent a sig-nificant, but very demanding improvement if the EthnicPower Relations dataset, including its geocoded version,could be expanded to include other politically relevant,relatively stable political cleavages. The analysis couldthen go beyond deciphering the logic of ethnic targetingand reach for a more general understanding of the polit-ical dynamic underlying violence against civilians.

Acknowledgements

The authors would like to thank audiences at the work-shop “Bridging Micro and Macro Approaches to CivilWar and Political Violence” organized in Barcelona byLaia Balcells, at the Graduate Institute in Geneva, and atthe department of sociology at Princeton University.

Supplementary Information

Supplementary information is available at the Journal ofGlobal Security Studies data archive.

References

Asal, Victor, Mitchell Brown, and Marcus Schulzke. 2015. “‘KillThem All–Old and Young, Girls and Women, and Little Chil-dren’: An Examination of the Organizational Choice of Tar-geting Civilians.” Political Science Research and Methods 3(3): 589–607.

Azam, Jean-Paul. 2002. “Looting and conflict between ethnore-gional groups. Lessons for state formation in Africa.” Journalof Conflict Resolution 46 (1): 131–53.

Azam, Jean-Paul, and Anke Hoeffler. 2002. “Violence againstCivilians in Civil Wars: Looting Or Terror?” Journal of PeaceResearch 39 (4): 461–85.

Bagozzi, Benjamin E. 2017. “Droughts, Land Appropriation, andRebel Violence in the Developing World.” Journal of Politics9 (3): 1057–72.

Balcells, Laia. 2010. “Rivalry and Revenge: Violence againstCivilians in Conventional Civil Wars.” International StudiesQuarterly 54 (2): 291–313.

———. 2011. “Continuation of Politics by Two Means: Directand Indirect Violence in Civil War.” Journal of Conflict Reso-lution 55 (3): 397–422.

Bannjerjee, Abhijit, Iyer Lakshimi, and Rohini Somanathan.2008. “Public Action for Public Goods.” InHandbook of De-velopment Economics, Vol. 4, edited by T. Paul Schultz andJohn Strauss, 3117–54. Amsterdam: Elsevier.

Bellows, John, and EdwardMiguel. 2006. “War and Institutions:New Evidence from Sierra Leone.” American Economic Re-view 96 (2): 394–99.

Dow



olumbia U



Buhaug, Halvard, and Päivi Lujala. 2005. “Accounting for Scale:Measuring Geography in Quantitative Studies of Civil War.”Political Geography 24 (4): 399–418.

Bussmann, Margit, and Gerald Schneider. 2016. “A Porous Hu-man Shield: The Law of War, the Red Cross, and the Killingof Civilians.” Review of International Organizations 11 (3):337–59.

Christia, Fotini. 2008. “Following the Money: Muslim versusMuslim in Bosnia’s Civil War.” Comparative Politics 40 (4):461–80.

Collier, Paul, and Anke Hoeffler. 2004. “Greed and Grievance inCivil War.”Oxford Economic Papers 56 (4): 563–95.

Costalli, Stefano, and Francesco Moro. 2010. “The Dynamicsof Violence in the Bosnian War: A Local-Level QuantitativeAnalysis.”Occasional Papers, Università degli Studi di Siena,Diapartimento di Scienze Storiche, Guiridiche, Politiche eSociali 24. https://www.circap.org/uploads/1/8/1/6/18163511/occasional_24_costalli_moro2.pdf.

Croicu,Mihai, and Ralph Sundberg. 2012.UCDPGEDConflictPolygons Dataset Version 1.1-2011. Uppsala, Sweden: UCDP,Department of Peace and Conflict Research.

Cunningham, David E., Kristian S. Gleditsch, and Idean Sale-hyan. 2009. “It Takes Two: A Dyadic Analysis of Civil WarDuration and Outcome.” Journal of Conflict Resolution 54(4): 570–97.

Czaika, Mathias, and Krisztina Kis-Katos. 2009. “Civil Conflictand Displacement: Village-Level Determinants of Forced Mi-gration in Aceh.” Journal of Peace Research 46 (3): 399–418.

de Luca, Giacomo, Roland Hodler, Paul A. Raschky, andMichele Valsecchi. 2015. “Ethnic Favoritism: An Axiom ofPolitics?” Center for Economic Studies & Ifo Institute Work-ing Paper No. 5209.

Eck, Kristine. 2012. “In DataWe Trust? A Comparison of UCDPGED and ACLEDConflict Events Datasets.”Cooperation andConflict 47 (1): 124–41.

Eyndre, Oliver Vanden. 2016. “Targets of Violence: Evidencefrom India’s Naxalite Conflict.”Economic Journal 128 (609):887–916.

Fjelde, Hanne, and Lisa Hultman. 2014. “Weakening the En-emy: A Disaggregated Study of Violence Against Civilians inAfrica.” Journal of Conflict Resolution 58 (7): 1230–57.

Gilmore, Elisabeth, Nils Petter Gleditsch, Päivi Lujala, and JanKetil Rød. 2005. “Conflict Diamonds: A New Dataset.”Con-flict Management and Peace Science 22 (3): 257–92.

Gleditsch, Nils Petter, Peter Wallensteen, Mikael Erikson,Margareta Sollenberg, and Håvard Strand. 2002. “ArmedConflict 1946–2001: A New Dataset.” Journal of PeaceResearch 39 (5): 615–37.

Grødeland, Ase B., William L. Miller, and Tatyana Y.Koshechkina. 2000. “The Ethnic Dimension to BureaucraticEncounters in Postcommunist Europe: Perceptions and Expe-rience.”Nations and Nationalism 6 (1): 43–66.

Gulden, Timothy R. 2001. “Spatial and Temporal Patterns inCivil Violence: Guatemala, 1977–1986.” Politics and the LifeSciences 21 (1): 26–36.

Hegre, Håvard, Gudrun Ostby, and Clionadh Raleigh. 2009.“Poverty and Civil War Events: A Disaggregated Study ofLiberia.” Journal of Conflict Resolution 53 (4): 598–623.

Herbst, Jeffrey. 2000. States and Power in Africa: Compara-tive Lessons in Authority and Control. Princeton: PrincetonUniversity Press.

Hirose, Kentaro, Kosuke Imai, and Jason Lyall. 2017. “CanCivilian Attitudes Predict Insurgent Violence? Ideology andInsurgent Tactical Choice in Civil War.” Journal of Peace Re-search 54 (1): 47–63.

Hoover Green, Amelia. 2016. “The Commander’s Dilemma:Creating and Controlling Armed Group Violence.” Journal ofPeace Research 53 (5): 619–32.

Horowitz, Donald L. 1985. Ethnic Groups in Conflict. Berkeley:University of California Press.

Hultman, Lisa, and Jacob D. Kathman. 2013. “United NationsPeacekeeping and Civilian Protection in Civil War.”AmericanJournal of Political Science 57 (4): 875–91.

Humphreys, Macartan, and Jeremy M. Weinstein. 2006. “Han-dling and Manhandling Civilians in Civil War.”American Po-litical Science Review 100 (3): 429–47.

Kalyvas, Stathis N. 2006. The Logic of Violence in Civil War.Cambridge: Cambridge University Press.

———. 2008. “Ethnic Defection in Civil War.”Comparative Po-litical Studies 41 (8): 1043–68.

Kalyvas, Stathis N., and Laia Balcells. 2010. “International Sys-tem and Technologies of Rebellion: How the End of the ColdWar Shaped Internal Conflict.”American Political Science Re-view 104 (3): 415–29.

Kalyvas, Stathis N., andMatthew AdamKocher. 2009. “The Dy-namics of Violence in Vietnam: An Analysis of the HamletEvaluation System (Hes).” Journal of Peace Research 46 (3):335–55.

Kaufman, Stuart J. 2006. “Symbolic Politics or Rational Choice?Testing Theories of Extreme Ethnic Violence.” InternationalSecurity 30 (4): 45–86.

Khannaa, Gaurav, and Laura Zimmerman. 2017. “Guns andButter? Fighting Violence with the Promise of Development.”Journal of Development Economics 124: 120–41.

King, Gary, and Langche Zeng. 2001. “Logistic Regression inRare Events Data.” Political Analysis 9 (2): 137–63.

Koren, Ore, and Benjamin E. Bagozzi. 2017. “Living Off theLand: The Connection Between Cropland, Food Security, andViolence Against Civilians.” Journal of Peace Research 54 (3):351–64.

Kramon, Eric, and Daniel Posner. 2016. “Ethnic favoritism ineducation in Kenya.”Quarterly Journal of Political Science 11(1): 1–58.

Krcmaric, Daniel. 2018. “Varieties of Civil War and MassKilling: Reasessing the Relationship Between Guerrilla War-fare and Civilian Victimization.” Journal of Peace Research55 (1): 18–31.

Le Billon, Philipp. 2001. “The Political Ecology of War: NaturalResources and Armed Conflicts.” Political Geography 20 (5):561–84.

Lujala, Päivi, Nils Petter Gleditsch, and Elisabeth Gilmore. 2005.“A Diamond Curse? Civil War and Lootable Resource.” Jour-nal of Conflict Resolution 49 (4): 538–62.

Lyall, Jason. 2010. “Are Coethnics More Effective Counterin-surgents? Evidence from the Second Chechen War.”AmericanPolitical Science Review 104 (1): 1–20.

Dow



olumbia U


https://www.circap.org/uploads/1/8/1/6/18163511/occasional_24_costalli_moro2.pdf


Mann,Michael. 2005. The Dark Side of Democracy: ExplainingEthnic Cleansing. Cambridge: Cambridge University Press.

McClendon, Gwyneth C. H. 2016. “Race and Responsiveness:An Experiment with South African Politicians.” Journal of Ex-perimental Political Science 3 (1): 60–74.

Melander, Erik. 2007.Ethnic Cleansing in Bosnia-Herzegowina,1992–1995. Paper given at the Conference on Disaggregatingthe Study of Civil War and Transnational Violence, Essex.

Montalvo, José G., andMarta Reynal-Querol. 2005. “Ethnic Po-larization, Potential Conflict, and Civil Wars.”American Eco-nomic Review 95 (3): 796–816.

Olsson, Ola, and Eyerusalem Siba. 2013. “Ethnic Cleansing orResource Struggle in Darfur? An Empirical Analysis.” Journalof Development Economics 103: 299–312.

Petersen, Roger D. 2002. Understanding Ethnic Violence: Fear,Hatred, and Resentment in Twentieth-Century Eastern Eu-rope. Cambridge: Cambridge University Press.

Petrovic, Drazen. 1994. “Ethnic Cleansing–An Attempt atMethodology.” European Journal of International Law 5(3):342–59.

Querido, Chyanda M. 2009. “State-Sponsored Mass Killings inAfrican Wars–Greed Or Grievance?” International Advancesin Economic Research 15 (3): 351–61.

Raleigh, Clionadh, Andrew Linke, Håvard Hegre, and JoackimKarlsen. 2010. “Introducing ACLED: An Armed Conflict Lo-cation and Event Dataset.” Journal of Peace Research 47 (5):1–10.

Rustad, Siri Aas, and Helga Malmin Binningsbo. 2012. “A PriceWorth Fighting For? Natural Resources and Conflict Recur-rence.” Journal of Peace Research 49 (4): 531–46.

Salehyan, Idean, David Siroky, and Reed M. Wood. 2014. “Ex-ternal Rebel Sponsorship and Civilian Abuse: A Principal-Agent Analysis of Wartime Atrocities.” International Orga-nization 68 (3): 633–61.

Schutte, Sebastian, and Nils B.Weidmann. 2011. “Diffusion Pat-terns of Violence in Civil Wars.” Political Geography 30 (3):143–52.

Schwartz,Rachel, and Scott Straus. 2018. “What Drives Violenceagainst Civilians in Civil War? Evidence from Guatemala’sCivil War.” Journal of Peace Research 55 (2): 222–35.

Sullivan, Christopher M. 2012. “Blood in the Village: A Local-Level Investigation of State Massacres.”Conflict Managementand Peace Science 29 (4): 373–96.

Sundberg, Ralph, Mathilda Lindgren, and Ausra Padskocimaite.2012. UCDP GED Version 1.0-2011. Uppsala, Sweden: De-partment of Peace and Conflict Research, Uppsala University.

Tilly, Charles. 1978. FromMobilization to Revolution. Reading,MA: Addison-Wesley.

Valentino, Benjamin, Paul Huth, andDylan Balch-Lindsay. 2004.“Draining the Sea: Mass Killing and Guerrilla Warfare.”International Organization 58 (2): 375–407.

Weidmann, Nils B. 2011. “Violence ‘From Above’ or ‘From Be-low’? The Role of Ethnicity in Bosnia’s Civil War.” Journal ofPolitics 73 (4): 1178–90.

———. 2012. Biases in Conflict Event Data. Paper given at theEthnic and Conflict, Uppsala, April 2012.

Weinstein, Jeremy. 2006. Inside Rebellion: The Politics of Insur-gent Violence. New York: Cambridge University Press.

Weintraub, Michael. 2016. “Do All Good Things Go Together?Development Assistance and Insurgent Violence in Civil War.”Journal of Politics 78 (4): 989–1002.

White, Robert W. 1993. “On Measuring Political Violence:Northern Ireland, 1969–1980.” American Sociological Re-view 58 (4): 575–85.

Wimmer, Andreas. 1997. “Who Owns the State? UnderstandingEthnic Conflict in Post-Colonial Societies.”Nations and Na-tionalism 3 (4): 631–65.

———. 2013.Waves of War. Nationalism, State-Formation, andEthnic Exclusion in the Modern World. Cambridge: Cam-bridge University Press.

Wimmer, Andreas, Lars-Erik Cederman, and Brian Min. 2009.“Ethnic Politics and Armed Conflict: A Configurational Anal-ysis of a New Global Dataset.”American Sociological Review74 (2): 316–37.

Wood, Reed M. 2010. “Rebel Capability and Strategic Violenceagainst Civilians.” Journal of Peace Research 2010 (47): 5.

———. 2014. “From Loss to Looting? Battlefield Costs andRebel Incentives for Violence.” International Organization 68(4): 979–99.

———. 2015. “Doing Harm by Doing Good? The Negative Ex-ternalities of Humanitarian Aid Provision During Civil Con-flict.” Journal of Politics 77 (3): 736–48.

Wood, Reed M., and Jacob D. Kathman. 2015. “Competing forthe Crown: Inter-Rebel Competition and Civilian Targeting inCivil War.” Political Research Quarterly 68 (1): 167–79.

Wood, Reed M., Jacob D. Kathman, and Stephen E. Gent. 2012.“Armed Intervention and Civilian Victimization in IntrastateConflicts.” Journal of Peace Research 49 (5): 647–60.

Wucherpfennig, Julian, Nils Metternich, Lars-Erik Cederman,and Kristian Skrede Gleditsch. 2012. “Ethnicity, the State, andthe Duration of Civil War.”World Politics 64 (1): 79–115.

Wucherpfennig, Julian, Nils B. Weidmann, Luc Girardin, Lars-Erik Cederman, and Andreas Wimmer. 2011. “Politically Rel-evant Ethnic Groups Across Space and Time: Introducing theGeoepr Dataset.”Conflict Management and Peace Science 20(10): 1–15.

Dow



olumbia U


Documents

TheStrategicLogicofEthnoterritorial …aw2951/WimmerJGSS.pdf · 2019. 8. 1. · ANDREASWIMMERANDCHRISMINER 3 definitions of the dependent variable.They are derived from either the