The Geography of Civil War*

03buhaug (ds) 20/6/02 8:36 am Page 417

HALVARD BUHAUG Department of Sociology and Political Science, Norwegian University of Science & Technology (NTNU) & International Peace Research Institute, Oslo (PRIO)

SCOTT GATES Department of Political Science, Michigan State University & International Peace Research Institute, Oslo (PRIO)

Geographical factors play a critical role in determining how a civil war is fought and who will prevail. Drawing on the PRIO/Uppsala Armed Conflict dataset covering the period 1946–2000, the authors have determined the location of all battle-zones for all civil wars in this time period, thereby identifying the geographic extent and the center point of each conflict. Using ordinary least squares (OLS) and three-stage least squares (3SLS) estimation techniques, factors are analyzed that determine the scope of the conflict (area of the conflict zone) and the location of the conflict relative to the capital. It is found that in addition to geographical factors such as the total land area of the country, scope is strongly shaped by such factors as the adjacencies of a border of a neighboring country, the incidence of natural resources in the conflict zone, and the duration of the conflict. The distance of the conflict zone from the capital is influenced by the scope of the conflict, the size of the country, whether or not the objective of the rebels is to secede, and whether or not the rebel group has a religious or ethnic identity. Also, evidence is found of an endogenous relationship between scope and location.

When I took a decision, or adopted an alterna- Introduction tive, it was after studying every relevant . . . factor. Geography, tribal structure, religion, Lawrence of Arabia’s observation is as true social customs, language, appetites, standards – today as it was in his time. In recent years, all were at my ﬁnger-ends. our theoretical and empirical understanding T. E. Lawrence, Letter to B. H. Liddell Hart, of the factors identiﬁed by Lawrence as June 1933 (Collins, 1998: 3) related to the onset and duration of civil war

* An earlier version of this article was presented at the Lein and Jan Ketil Rød have made essential contributions. annual North American meeting of the Peace Science Furthermore, we are grateful to the Department of Peace Society (International), Atlanta, GA, 26–28 October and Conflict Research, Uppsala University, for collabor- 2001. We thank the Research Council of Norway, the ation on the conflict data, Anke Hoeffler for providing Norwegian University of Science and Technology data on mountainous terrain, and Nils Petter Gleditsch, (NTNU), and the World Bank for funding various parts Mansoob Murshed, and five anonymous referees for of the data collection. The work on locating the conflicts insightful comments on an earlier draft. The data on was conducted as part of a cross-disciplinary project at which the article is based will be posted at http: NTNU on geographic diffusion of conflict, where Haakon //www.prio.no/cwp/datasets.asp.

417 03buhaug (ds) 20/6/02 8:36 am Page 418

418 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

has progressed tremendously.1 Yet, despite with the terrain: how mountains rise, how important insights gained from this research, the valleys open out and plains spread out, we have very little systematic knowledge as well as with the characteristics of rivers about the actual fighting of civil wars. Ironi- and swamps’. Keegan (1993) in his History cally, one reason for the general lack of of Warfare features the role of geographic understanding in this regard is that there is variables and distinguishes between ‘perma- little or no actual fighting or war in these nently operating’ and ‘contingent’ factors. models of war onset or duration. There are Permanent factors include terrain and no battles, no deaths, no weapons, no guer- climate. These factors have long been the rilla tactics, and no counter-insurgency focus of military tacticians and military activities. Territory and resources are never historians.2 And presumably because such lost or gained. There are no victories and stories are more interesting, the majority of there are no defeats. Yet, motivations regard- such studies have featured the catastrophic ing peace and war are clearly linked to the blunders caused by commanders ignoring prospects of winning or losing a civil war. In geography, including the likes of Major addition to securing wealth through the General George McClellan at Antietem capture of resources, civil wars are often (1862), General Baron Levin Bennigsen at fought over a political objective – control Friedland (1807), General Ludwig Benedek over the apparatus of the state or the creation at Könnigratz (1866), and Lieutenant of a new sovereign state. Clearly, different General Mark Clark at Rapido in the Battle objectives will alter the way a civil war is for Italy (1944). In contrast, gifted generals, fought. A war over control of the state will of course, consistently have taken such fundamentally differ from a war of secession. factors into account and, where possible, Military historians and strategists have long have used them to their advantage. understood how geographical factors play a Keegan’s second concept, contingent critical role in influencing how a civil war is geographic factors, relates to the constraints fought and who will prevail. Taking military on logistics and intelligence. To ignore these history as a departure point, this paper exam- factors is to ignore Clausewitz’s admonition, ines how strategic objectives and geographi- ‘the end for which a soldier is recruited, cal factors affect the location, relative to the clothed, armed and trained, the whole capital, and scope (measured conflict area) of object of his sleeping, eating, drinking, and armed civil conflict. marching is simply that he should fight at the right place and the right time’ (Murray, 1999: Geographic Factors and Armed Civil 210). Geography is not just important on Conflict the battlefield, but at the operational level as well. With ruinous results, such Physical Geographical Factors renowned military minds as Napoleon and The earliest military strategists understood the Oberkommando des Heeres (OKH) of the role of geography and conflict. In his Nazi Germany neglected these factors in Discoursi, Machiavelli (1517/1988: 52–53) their attempts to conquer Russia and the wrote that a soldier must become ‘familiar Soviet Union respectively. It seems that social scientists too, in their analyses of civil 1 The other articles included in the special issue serve as war, also have tended to ignore the role of testimony to the tremendous strides we have made in understanding the onset and nature of civil war in recent 2 In addition to Keegan (1993), see Collins (1998) and years. See Sambanis (2002) and Gates (2001) for reviews Murray (1999) for discussions regarding the role of geog- of the literature. raphy in warfare. 03buhaug (ds) 20/6/02 8:36 am Page 419

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 419

geographic variables with regard to the only to a limited extent been included in quan- fighting of civil war. titative conflict studies, and always as exogen- Influenced by classical theorists, most ous factors affecting outbreak or duration. applications of geography to war have been Fearon & Laitin (1999) and Collier & Hoeffler geo-political analyses applied to military strat- (2001) both include crude measures of moun- egy or global security issues. The basic premise tainous terrain and forest cover as predictors of though almost seems too obvious – that the onset of conflict. And in an effort to determine location and size of a country affect the design relevant South American dyads, Lemke (1995) and nature of military strategy. Mackinder acknowledges the relevance of terrain by con- (1904) speaks of the pivot area, or heartland, structing a time–distance measure of interstate while Ratzel (1896) draws on Darwin’s sur- transportation cost. vival of the fittest when arguing for a ‘law’ of territorial growth.3 For the most part, geopo- Natural Resources litical strategic analysis has been global in Recent work on the economics of civil war perspective, applied to explain the inter- clearly demonstrates that to understand civil national politics of Germany, Great Britain, war today we need to understand the role and the United States. Yet, even for civil war, played by natural resources in financing the such factors as the size of a country, its purchase of arms. Theoretical studies such as location, and the nature of its borders are Addison, Le Billon & Murshed (2000) push extremely important, influencing temporal our theoretical understanding of the role of and spatial domain as well as the potentiality resources with regard to conflict (particularly for diffusion. as a source of loot to finance a war), while Empirical studies addressing the geo- empirical studies by Auty (1998), Collier & graphic aspect of conflicts typically consist of Hoeffler (2001), de Soysa (2000), Ross statistical, nation/dyadic-level analyses (2001) and others provide evidence of the regarding interstate war, and rest on different wide array of problems associated with measures of proximity of states (Bremer, resource abundance. Resource extraction is 1992; Diehl, 1991; Gleditsch, 1995; for the most part spatially fixed. Businesses Vasquez, 1995). Following Boulding (1962), engaged in such activities cannot choose Sprout & Sprout (1965), and Richardson where the natural resources are located, and, (1960), geography is treated primarily as a unlike enterprises in other types of economic concept of contiguity and distance, affecting activity, they cannot relocate. As a business, interstate interaction. These studies have you must decide not to invest or to disen- demonstrated that inter-capital distance and gage. To sustain access to the resources and number of borders are essential predictors of protect their investments, natural resource conflict proneness. However, such measures extraction businesses generally rely on paying are less interesting when studying internal ‘whoever is in power’ (Le Billon, 2001: 569). conflicts. This makes natural resources extremely When it comes to exploring determinants of amenable to taxing and to looting. the location of conflicts, little or no systematic Natural resources differ a great deal with effort has been made. Attributes of the conflict- regard to their concentration.4 Point ridden countries (topography, climate, popu- 4 The distinction between point and diffuse resources in lation) as well as characteristics of the conflicts Addison, Le Billon & Murshed (2000), Addison & (severity, duration, goal of rebel group) have Murshed (2001), and Auty (2001) lies not in their geographic dispersion, but the fact that point-sourced resource 3 For more modern literature of this nature, see Pepper & rents are concentrated and capturable. We, like Le Billon Jenkins (1985) and Kliot & Waterman (1991). (2001), feature the geographic characteristics. 03buhaug (ds) 20/6/02 8:36 am Page 420

420 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

resources tend to characterize oil-drilling to conflict risk if related at all (Collier & operations and pit mining. Illegal drugs Hoeffler, 2001; Fearon & Laitin, 1999). The (cocaine, hash, heroin, etc.), timber other dimension is polarization or domi- resources, and alluvial diamond mining are nance. As defined by Collier & Hoeffler, more widely dispersed and are therefore dominance occurs if the largest ethnic group more difficult for a government to control. constitutes 45–90% of the population. Alluvial diamonds in particular, regarded as There is a broad consensus that this variable the ultimate loot, have served to finance civil is positively related to conflict (Collier & wars in Sierra Leone, Angola, Liberia, and Hoeffler, 2001; Elbadawi & Sambanis, 2002; the Democratic Republic of Congo. Timber Ellingsen, 2000; Hegre et al., 2001; Reynal- and drugs, too, have financed many rebel Querol, 2002). Also, Collier & Hoeffler groups. Timber has played a big role in argue that ethnic and religious diversity financing the conflicts in Myanmar, Cam- within a region reduces the opportunity for bodia, Liberia, and the Philippines, while rapid rebel recruitment.5 They find that drugs have financed conflicts in Afghanistan, minorities that have a rural base are far more the Caucasus, Colombia, Kurdistan, and likely to see large-scale ethnic violence than Tajikistan (Le Billon, 2001: 573). urban and widely dispersed minorities. Geographical location and the concen- Accordingly, separatist wars typically emerge tration of the resources are critical with where the ethnic groups are located in clearly regard to the opportunities of belligerents to defined regions of the states. seize or retain control of the resource rev- As with geographic factors like topogra- enues (Addison & Murshed, 2001; Auty, phy and resources, no study of identity and 1998; Le Billon, 2001). One crucial aspect of civil war has been concerned with the location is the proximity to the decision- physical attributes of the conflict. Whether making center. Natural resources located and how identity-related conflicts differ from closer to a country’s capital should be easier non-identity conflicts with respect to for the state to control (Le Billon, 2001). location and scope has never been explored. Poor data on location and significance have Herein lies a huge challenge. so far prevented scholars from including natural resources in cross-national conflict Hypotheses studies. Location and scope geographically define the The Human Geography of Identity characteristics of a war. Indeed, with regard A substantial number of today’s civil wars are to the capabilities, limitations, and vulnera- related to identity; i.e. they are fought bilities of armed forces, ‘it seems safe to between different ethnic or religious groups. predict that the pertinence of spatial relation- As a consequence, there is a widely held belief ships will remain undiminished indefinitely’ that ethnic (and religious) diversity causes (Collins, 1998: 11). As noted by Starr (2001) conflicts. Reality may not be so simple. in his presidential address at the Peace Ethnic composition may be operationalized Science Society (International) meetings, along two dimensions. The first dimension is absolute and relative space are relevant to the fragmentation: the more groups, or the study of conflict. higher the probability that two individuals The objective of this article is to examine drawn at random are from different groups, factors that determine location and scope of the higher the level of fragmentation. There 5 See Gates (2002) for a related discussion regarding geo- is a consensus that this is negatively related graphical distances, ethnicity, and rebel recruitment. 03buhaug (ds) 20/6/02 8:36 am Page 421

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 421

civil wars. This rather unconventional choice Figure 1. Deﬁning Location and Scope

of dependent variables contrasts most conﬂict Conflict zone / area controlled studies’ focus on outbreak, incidence, dura- by rebels tion, or ending of war. As a consequence, our Conflict center point ﬁndings may not be immediately comparable Location : with other studies on civil war. Having said conflictÐcapital that, most of our research questions do distance involve elements of generality too – if

physical characteristics of nations, such as Scope: size, topography, and natural resources, are measured conflict zone Capital found to influence the spatial domain of civil wars, they probably affect the temporal dimension as well. Moreover, future work relationships can be expressed in terms of may involve exploring to what extent our geo- two hypotheses: graphic data manage to predict outbreak, duration, and recurrence of civil wars. H1: Rebel groups that aim to seize power A civil war is inherently about armed con- from the state will tend to fight their flict between the state and an organized rebel wars closer to the capital city than group. Hence, a key reference point is the secessionist groups, ceteris paribus. capital of a country; after all, the capital is H2: Rebel groups with an ethnic/religious where the state is based. We can therefore identity will tend to fight their wars presume that state power is centered at the further away from the capital city than capital. The other main reference point is the non-identity groups, ceteris paribus. geographic center of the zone of conflict. From these two reference points we define We argue that, by the very nature of seces- the location of the conflict, measured in sionism, rebel groups aiming to create a new terms of the distance between the capital city state will focus their fighting against the state and the conflict center point (hence we treat in the territory that they are trying to liber- location as a relative concept). The other key ate. Such territory is presumed to be not concept, scope, is defined as the geographic proximate to the capital, the seat of state domain of the conflict zone, measured as the power. A similar reasoning applies to ethnic circular area centered around the conflict groups engaged in armed rebellion against center and covering all significant battle the state. The presumption is that an ethnic zones (see Figure 1). We discuss the particu- group engaged in armed rebellion against the lars of these operationalizations in the next state will not be the group that is concen- section. trated in the capital. Rather, the identity of aggrieved groups tends to be made with Location of Conflict respect to the dominant ethnic groups con- The distance between the capital of a centrated in the capital. country and the zone of conflict in a civil We control for two key variables, the size war is hypothesized to be related to a of the country and the scope of the conflict. number of factors. Two primary explanatory At the margins, the size of the country will factors are identified, the strategic objective limit the conflict–capital distance. Small of the rebellion (or the nature of the incom- countries in armed civil conflict will by their patibility) and whether the rebellion is based very nature find the location of the conflicts around an ethnic or religious identity. These to be nearer the capital than the location of 03buhaug (ds) 20/6/02 8:36 am Page 422

422 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

conflicts in countries with large areas. The terrain will be larger than the scope of scope is also hypothesized to be associated those that are not, ceteris paribus. with the location from the capital. After all, the zone of conflict is a measurable area, Duration is hypothesized to be related to which is related to the total size of the war, given that time increases the possibilities country and the distance between the capital for a rebel army to increase its zone of and the zone of conflict. This relationship, activity. Also short-lived conflicts, such as a too, at the margins exhibits a certain deter- coup d’état, are concentrated in the capital ministic quality. Though given the irregular city. Insurgency movements that endure year shapes of countries and the influence of other after year tend to encompass a broad terri- geographical features, the extent of mathe- tory. matical determinism is minimal.6 International borders are hypothesized to be related to the size of a conflict zone Scope of Conflict because of the value of such borders to a rebel The scope of conflict is also hypothesized to army. Rebels will push to gain access to an be associated with a number of factors. international border because neighboring Several explanatory variables are featured, the countries often provide a safe refuge away duration of the conflict, whether the conflict from governmental troops, but also because zone abuts an international border, whether weapons and natural resources are traded and a resource is present, as well as the extent of transported across these borders. Control of mountainous and forested terrain. These international borders thus ensures that the relationships can be expressed in the form of rebel army will fight another day. the following hypotheses: Natural resources, whether point or dispersed resources, provide revenue for a rebel H3: The scope of an armed conflict is posi- army. A rebel army has an interest in expand- tively associated with the duration of a ing its zone of control to capture these conflict, ceteris paribus. resources and thereby derive financial gain H4: The scope of conflicts that abut an from them, regardless of whether secession- international border will be larger ism or state power is the ultimate political than the scope of those that do not, goal. ceteris paribus. Rough terrain is ideal for guerrilla warfare and difficult for a government army to H5: The scope of conflicts with natural control. Mountain areas, giving advantage to resources present will be larger than rebel troops, allow the rebels to expand the the scope of those that are without, scope of conflict, whereas forests provide ceteris paribus. cover, particularly against detection or aerial H6: The scope of conflicts in mountainous attack. This aids in the freedom of movement terrain will be larger than the scope of and shipment of arms, thereby associated those that are not, ceteris paribus. with a wider zone of conflict. We control for two variables to account H7: The scope of conflicts in forested for the scope of conflict, the total size of the country and the location of the conflict 6 If all countries were circles and if physical geographical factors were uniform, then there might be some utility in relative to the capital. These relationships deriving such a mathematically deterministic relationship. follow the same pattern as in the location But given the wide distribution of country shapes and geographical features, this relationship can be modeled proba- model, except that with scope we are bilistically. working in two dimensions rather than in 03buhaug (ds) 20/6/02 8:36 am Page 423

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 423

one. At the limit, the relationships are math- concerned with studying geographic attrib- ematically deterministic, but given the wide utes of conflict zones, data on disaggregated variety of shapes of countries, the relation- conflicts are not necessary. Consequently, the ship remains probabilistic and can be esti- ‘subconflicts’ were merged if they consisted mated statistically. of identical actors, incompatibility, and geographic location, and less than three years passed from one conflict unit to the next. In Description of Data total, our dataset includes 265 civil conflicts The unit of analysis in our study is armed for the 1946–2000 period. These conflicts civil conflicts, as defined by the are displayed in Figure 2. PRIO/Uppsala Armed Conflict dataset (Gleditsch et al., 2001). However, in this Dependent Variables dataset, quite a few intrastate conflicts are Our major contribution to this dataset is the subdivided into several separate conflict inclusion of variables representing location units, either reflecting varying severity from and scope of the conflicts. Quite a few con- one year to the next (from ‘minor’ to ‘inter- flicts are limited to one specific place, either mediate’ to ‘war’, or vice versa) or because a city or an administrative region. These con- violence has temporarily decreased below the flicts were assigned conflict center points threshold of conflict (i.e. less than 25 battle- equaling the geographic coordinates for the deaths per year).7 As our analysis is mostly specific city/region. As for larger conflicts (in

Figure 2. Civil Conﬂicts, 1946–2000

Adapted from Gleditsch et al. (2001), using software developed by Jan Ketil Rød. For full details on speciﬁc armed conﬂicts, see the Appendix to that paper.

7 For example, the conﬂict in Cabinda between the coded as three separate conﬂicts, one in 1992, one in 1994, Angolan government and FLEC (1992–97) is originally and one in 1996–97. 03buhaug (ds) 20/6/02 8:36 am Page 424

424 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

geographical terms), we first identified the significant battle zones were covered. Con- major battle zones – i.e. places where the flicts that took place within a single city were fighting resulted in loss of lives – and areas assigned a 50-km radius.10 All estimations controlled by rebel groups.8 The conflict reported below were run with logged and center was then defined as the mid-point of unlogged variables. The models estimated these locations. Determining the center of with logged variables produced much each conflict is not trivial, since different stronger results. sources frequently diverge on the exact whereabouts (and severity) of the battles. Explanatory and Instrumental Variables Moreover, as a number of conflicts are To control for identity-based conflicts, we located along dispersed border zones, some constructed a dummy identity variable, given conflict centers actually refer to areas quite the value 1 if the rebels originate from unaffected by the fighting. Future work on different ethnic and/or religious groups than the conflict data will be dedicated to reduc- the government. The main sources for this ing this problem, most likely by utilizing variable were the ‘wartype’ variable of GIS. Sambanis (2000)11 and various volumes of Our first dependent variable – the opera- Keesing’s Record of World Events. According to tionalization of location – is measured as the our data, 59% of the conflicts are related to distance from the conflict center to the identity. capital city. The values on this variable were Our second explanatory variable is the calculated using a geodetic distance calcula- PRIO/Uppsala dataset’s dichotomous incom- tor, estimating the conflict–capital distances patibility variable, indicating whether terri- with accuracy far superior to our require- tory (secession) or governance (control over ments.9 The conflict–capital distances vary the state) is the incompatibility between the between 0 km (the capital is the center of government and the rebels. Some 40% of our conflict) and 3,360 km (from West Papua to conflicts concern territory. This variable is Jakarta). The location variable was log- closely related to the identity indicator (r = .6), transformed prior to use. as conflicts over territory almost by definition Additionally we constructed two proxies are related to identity. However, 35% of the of scope, the absolute area of the conflict conflicts over state power were also fought zone (log-transformed) and the conflict area between different ethnic/religious groups. as a proportion of total land area. For sim- In this dataset, all conflicts have been plicity, we defined the conflict zones as being ascribed a start- and end-year. As we expect circular and centered around the conflict conflicts in remote areas – typically provid- center point; the radius of the conflict zone ing rebel hideouts in forests, mountains, or equals the distance from the center to the behind international boundaries – to endure most distant battle zone, rounded upwards to longer than conflicts of a more urban nature, the nearest 50-km interval to ensure that all we constructed a variable on conflict duration (end-year minus start-year). 8 Data on the location of battle zones were drawn from the PRIO/Uppsala Armed Conflict Project’s archive at the 10 A negative consequence of the assumed circular shape of Department of Peace and Conflict Research, Uppsala Uni- the conflict zones is that the measured scope inevitably versity, and various volumes of Keesing’s Record of World covers some areas not affected by the conflict, thus overes- Events. timating the total area of the civil war. Future work on 9 This program will be made available on the website of the improving the geographic data will reduce this problem. Department of Geomatics at NTNU: http: //www.geomatikk. 11 Sambanis’s (2000) ‘wartype’ variable is again made up ntnu.no/prosjektarbeid/geo-hoved98/Applet/b-program. from several sources, most notably Licklider (1995) and the htm. State Failure Project (Esty et al., 1998). 03buhaug (ds) 20/6/02 8:36 am Page 425

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 425

Table I. Descriptive Statistics

Observations Mean Std. dev. Minimum Maximum

Location 258 5.23 1.82 1.6 8.1 Absolute scope 258 4.60 1.65 2.3 9.2 Relative scope 246 44.41 38.03 0 100 Land area 251 6.04 1.64 0.7 9.7 Identity 262 0.59 0.49 0 1 Incompatibility 265 1.59 0.49 1 2 Duration 265 5.45 7.93 1 53 Border 258 0.51 0.50 0 1 Resource 262 0.21 0.41 0 1 Mountain 256 23.19 21.78 0 81 Forest 247 30.23 21.56 0 96

In order to control for rebels hiding between 0 and 94%.12 Although these beyond national borders or conflicts that for measures describe the type of terrain for each other reasons frequently involved neighbor country, they do not indicate the extent of territory, we constructed a dummy indicator mountains and forests specifically for each on whether or not the conflict zone abuts a conflict zone. An important improvement border with another country. Roughly one- will be to construct these variables from half (51%) of the conflicts in our sample gridded data on topography through the aid extend to (or across) the national border of of GIS tools, thus facilitating a comparison the conflict-ridden country. between the terrain in the conflict zone and Previous conflict studies seem to confirm the terrain in the rest of the country. that primary commodities serve as a major Table I summarizes the descriptive statis- source of rebel finance (Collier & Hoeffler, tics for each of these variables. Table II is the 2001). Thus we include a dichotomous correlation matrix for these variables. resource variable, simply indicating whether or not the conflict zone contains essential Method of Analysis and Results natural resources such as fossil minerals, metals, or diamonds. This variable was OLS Single Equation Models constructed by comparing the scope of the In order to test our three hypotheses regard- internal conflicts with maps on resource ing conflict location, we first specify two distribution from Kunnskapsforlagets Store models to fit the OLS regression of the con- Verdensatlas (1997), Oxford Economic Atlas of flict–capital distance. As Table III demon- the World (1972), as well as descriptive data strates, both models yield very similar results, from the CIA (2001). and their explanatory powers are very high. Data on country area (log-transformed Both conflict area (model 1) and conflict area prior to use) and forests were drawn from as a proportion of land area (model 2) are World Bank (2000). The forest variable gives important determinants of the conflict the proportion of land area covered by forest, location. Accordingly, the more distant the varying between 0 and 96%. We also include 12 See Gerrard (2000) for a theoretical discussion of defi- a measure of mountainous terrain, identical nition and operationalization of the mountainous terrain to Collier & Hoeffler (2001); the values vary variable. 03buhaug (ds) 20/6/02 8:36 am Page 426

426 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

Table II. Correlation Matrix Location scope Absolute scope Relative Land area Identity Incompatibility Duration Border Resource Mountain Absolute scope .60 Relative scope .008 .48 Land area .54 .44 –.46 Identity .54 .27 –.06 .28 Incompatibility –.55 –.09 .31 –.31 –.60 Duration .15 .25 .20 .04 .19 –.18 Border .35 .33 .17 .12 .32 –.30 .27 Resource .11 .27 .10 .14 .05 .16 –.06 –.03 Mountain –.006 –.01 .02 –.04 .08 –.10 .07 .06 –.32 Forest .09 .10 –.002 .10 –.13 .06 .02 –.12 .12 –.12

conflict centers, the larger the battle-zones – second part of the OLS regression section, in both relative and absolute terms. Not sur- addressing the hypotheses regarding the prisingly, the conflict location is also posi- scope of conflict. As we have two endogen- tively associated with the size of the country. ous variables of scope – measuring absolute As noted above, all models were estimated and relative conflict area respectively – with logged and unlogged distance and area models 3 and 4 generate estimations of variables. Not only did the logged variables absolute scope of conflict, while models 5 produce higher levels of statistical signifi- and 6 consist of the relative scope of conflict. cance, they also performed better with regard At first glance, we see that most of the to the diagnostic analyses (omitted variable exogenous variables generate very robust tests, Cooks’ distance, lvr2plots and avplots) results, being significant in all four equa- that accompanied the OLS estimations (but tions. Moreover, the R2 are quite high are not reported here). (though not in the league of models 1 and 2). According to Hypothesis 1, secessionist Hypothesis 3 states that the duration of conflicts should generally be located further armed conflict is positively associated with away from the capital city than conflicts over the scope of conflict. This is supported in all state power. The findings confirm our pre- four models; the estimate is statistically diction – the incompatibility variable is significant although the coefficient is quite statistically very significant in both models. small. We are thus led to the rather intuitive Supporting our second hypothesis, we see conclusion that longer-lasting conflicts that identity is positive and significant generally involve a larger geographic area. regardless of model, although the effect is less In line with Hypothesis 4 (another seem- impressive than that of the incompatibility ingly obvious relationship), internal variable. This will be more thoroughly dis- conflicts that abut – or cross – international cussed in the final section. boundaries are also associated with higher- Table IV reveals the results from the than-average conflict zones. The impact is 03buhaug (ds) 20/6/02 8:36 am Page 427

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 427

Table III. OLS Estimation of Location, not overwhelming, but always significant at 1946–2000 the 0.10 level. According to Hypothesis 5, Location Model 1 Model 2 the presence of natural resources increases the conflict area. Table IV demonstrates Absolute scope 0.49 that this clearly is the case; the resource (.062)** dummy is robust throughout. The only Relative scope 0.02 variables that do not fare well in the OLS (.003)** estimation are the physical geography indi- Land area 0.21 0.61 (.065)** (.058)** cators: mountain and forest. Neither of Identity 0.54 0.61 these had any impact on the scope of con- (.199)** (.216)** flict, regardless of choice of endogenous Incompatibility –1.27 –1.41 variable.13 (.190)** (.208)** Constant 3.45 2.65 (.476)** (.490)** 13 We also found no evidence that population was related to the geography of civil war. Population density, total N 243 243 population, and dispersion of the population (similar to R2 .620 .575 Collier & Hoefﬂer, 2001) – tested but not reported here – all proved to be insigniﬁcant with respect to both scope and * p ≤ .10, ** p ≤ .05, and standard errors are in parentheses. location.

Table IV. OLS Estimation of the Absolute (3–4) and Relative (5–6) Scope, 1946–2000

Model 3 Model 4 Model 5 Model 6 Absolute Absolute Relative Relative scope scope scope scope

Location 0.39 0.38 5.64 5.62 (.048)** (.050)** (1.235)** (1.299)** Land area 0.18 0.16 –14.88 –15.14 (.050)** (.050)** (1.226)** (1.283)** Duration 0.03 0.03 0.77 0.73 (.010)** (.010)** (.302)** (.306)** Border 0.39 0.43 9.49 9.34 (.192)** (.204)** (4.545)** (4.868)* Resource 0.83 0.92 17.51 17.47 (.213)** (.233)** (5.533)** (5.989)** Mountain 0.004 0.09 (.004) (.098) Forest 0.003 0.06 (.004) (.094) Constant 0.96 0.86 91.62 90.43 (.271)** (.295)** (6.897)** (8.070)**

N 246 230 246 230 R2 .466 .474 .374 .375

* p ≤ .10, ** p ≤ .05, and standard errors are in parentheses. 03buhaug (ds) 20/6/02 8:36 am Page 428

428 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

Results of the Three-Stage Least Squares squares (2SLS) in that it uses the covariance Estimations matrix of disturbances, which improves the In this section, we check for robustness of efficiency of estimation leading to smaller our results with respect to the endogeneity of standard errors. However, this improvement area and distance, while examining the depends on the consistency of the covariance relationship between our explanatory vari- matrix estimates, since with 3SLS the mis- ables and the two endogenous variables, specification of one equation affects the esti- scope and location. The distance between the mates in all other equations. In sensitivity capital city and the center of the conflict zone analysis, we have therefore tested the system is highly correlated with the area of the con- of equations with 2SLS instead and found no flict zone (r = 0.602). Accordingly, we assess substantial changes. the estimated effect of the explanatory vari- Table V reports results for our three-stage ables controlling for this relationship least squares simultaneous equation estima- between the two endogenous variables. tions. From the regressions and their associ- In order to test this, we have developed a ated R2 values, one can see that our simultaneous equation model, in which the instruments work well in explaining cross- distance between the capital and the center sectional differences in distance (location) of the conflict zone, the size of the area of the and area (scope). For the most part, our conflict zone (measured both as the natural results do not vary too much from the single log of the area and the percentage of the total equation OLS results. A simultaneous 3SLS land area), and the total logged area of the version of each pair of the single equation country are endogenized and are explained as OLS models was analyzed. We tested four a function of exogenous variables, so-called models (models 7–10). Models 7 and 8 pro- instrumental variables (which are the same as duced the best results in terms of overall the independent variables reported in our measures of fit. Models estimated in terms of OLS results section). All estimations were the relative area of the conflict zone (per- undertaken with three-stage least squares centage of total land area in the combat zone) (3SLS) in STATA for all variables analyzed in produced markedly worse results than the our single equation OLS estimations. Three- models in which the actual area of the con- stage least squares involve three steps. First, flict zone was measured (the natural log of predicted or instrumented values of the the square kilometers). In terms of the two endogenous variables (scope and distance) equations being estimated, the estimation of are generated, using all exogenous variables the distance produced much stronger results in the system. Second, a cross-equation than the estimation of the size of the conflict covariance matrix is estimated. Third, the zone. The specific z-scores associated with simultaneous equation with the two en- each coefficient across each model are quite dogenized variables is estimated with robust. No single estimation challenges our generalized least squares using the instru- evaluation of each hypothesis. mented variables and other exogenous vari- With regard to Hypothesis 1 (H1), we find ables, as well as the estimated covariance strong support. Rebel groups whose objective matrix.14 The estimation technique 3SLS has is to secede from the country do tend to fight the important advantage over two-stage least further away from the capital than groups intending to take control of the state. The 14 This third stage can also be used to estimate an equation purported objectives of different rebel groups for a specific dependent variable using the instrumented variables, the other exogenous variables and the estimated are related to the Uppsala coding of the covariance matrix. nature of the incompatibility. We find that 03buhaug (ds) 20/6/02 8:36 am Page 429

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 429

Table V. Three-Stage Least Squares Estimation of Absolute (7–8) and Relative (9–10) Scope, and Location, 1946–2000

Model 7 Model 8 Model 9 Model 10 Absolute Absolute Relative Relative scope scope scope scope

Scope Location –.13 –.070 –7.55 –9.61 (.135) (.125) (3.190)** (3.492)** Land area .45 .41 –7.84 –6.82 (.091)** (.090)** (2.302)** (2.360)** Duration .040 .037 .95 1.05 (.012)** (.012)** (.295)** (.302)** Border .86 .90 22.18 22.87 (.231)** (.236)** (6.02)** (5.955)** Resource .94 1.02 19.73 19.50 (.227)** (.239)** (6.08)** (5.816)** Mountain .005 .11 (.0046) (.113) Forest .005 .11 (.0044) (.112) Constant 1.71 1.31 104.22 114.29 (.406)** (.416)** (10.625)** (10.549)**

Location Absolute scope .49 .50 (.119)** (.116)** Relative scope .023 .022 (.0060)** (.0058)** Land area .21 .21 .65 .64 (.070)** (.070)** (.078)** (.075)** Identity .50 .60 .56 .44 (.206)** (.216)** (.230)** (.216)** Incompatibility –1.31 –1.35 –1.63 –1.59 (.199)** (.206)** (.251)** (.238)** Constant 3.53 3.47 2.50 2.69 (.484)** (.510)** (.629)** (.602)**

Scope R2 .261 .314 .123 .034 Location R2 .620 .647 .598 .571

* p ≤ .10, ** p ≤ .05, and standard errors are in parentheses.

across models, incompatibility is significantly battles further away from the capital than associated with the distance between the non-identity groups. Across models, this capital and the center of the conflict zone. relationship is always statistically significant These results reflect what we found with the at p < .05, though the effect is not half as single equation OLS estimations. strong as for the incompatibility variable – Hypothesis 2 (H2) is also confirmed. the single strongest predictor of location. Identity-based groups tend to fight their Hypothesis 3 is strongly supported. 03buhaug (ds) 20/6/02 8:36 am Page 430

430 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

Duration and the size of the conflict zone are with respect to the results of the single positively and statistically significantly equation OLS estimations and the 3SLS associated across models and measurements estimations. of the size of the conflict zone. We can con- clude that longer-lasting conflicts in general Conclusion encompass larger areas. With regard to Hypothesis 4, conflict To our knowledge, this is the first systematic zones that abut an international border will inquiry into the scope and location of civil tend to be larger than conflict zones that do conflict. Drawing on the PRIO/Uppsala not. This finding is extremely robust. Armed Conflict dataset for the 1946–2000 Hypothesis 5 is also strongly supported. The period (Gleditsch et al., 2001), we have presence of natural resources in the conflict determined the location of all battle-zones zone is positively and strongly statistically for all conflicts in this time period, thereby significantly related to the size of the conflict identifying the geographic extent and the zone across all the models. Hypotheses 6 and center point of each conflict. With this data, 7 are not supported in any of the models. we are able to analyze the factors that deter- As for the endogenous effects, we find mine the scope and location of civil conflict. that when the scope of the conflict is Using ordinary least squares (OLS) and measured as the natural log of the area, dis- three-stage least squares (3SLS) estimation tance is statistically significantly related to techniques, we have analyzed the factors that absolute scope, but absolute scope is not determine the scope (measured in terms of related to distance. When the models are logged square kilometers, and as the propor- estimated measuring the conflict area as a tion of a country that is covered by the scope) proportion of the total area of the country and the location of the conflict relative to the (relative scope), the relationship is com- capital (measured in terms of logged kilome- pletely endogenized with both variables ters). Given the interdependence between associated with one another in the simul- location and scope, we modeled these taneous estimation. The total land area of relationships as simultaneous equations, the country is also treated endogenously which were estimated with 3SLS. and is entered in both equations. It is Geographical factors are indeed import- robustly statistically significant. As ant. From our analysis, we found that the expected, the relationship between the scope of conflict is associated with such geo- scope of the conflict area and the size of the graphical factors as the total land area of the total area of the country changes with the country, whether or not the conflict zone is way in which the battle-zone is measured. If adjacent to the border of a neighboring measured in absolute terms, the relationship country, and whether there are natural is positive. The relationship follows from resources in the conflict zone. Interestingly, the expectation that larger countries simply two geographical factors considered to be have more room within which to fight. The critical to combat, mountains and forest relationship is negative when the size of the cover, were found not to be statistically associ- battle-zone is measured relatively. As ated with the scope of conflict. This is most expected, smaller countries tend to have a likely due to poor data. Our data on moun- higher percentage of land involved in contains and forests pertain to the country as a flict. Despite finding support for these whole. We do not have precise enough infor- endogenized relationships, our conclusions mation to inform us as to what extent the regarding the hypotheses do not change zone of conflict is forested or mountainous. 03buhaug (ds) 20/6/02 8:36 am Page 431

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 431

When military strategists and historians same shape, these relationships would be discuss the central role of terrain, they are mathematically deterministic. What is more thinking in terms of specific battles and the interesting is that the nature of the rebellion specific nature of local geographical features. and the nature of the rebel group affect where Our overly general data does not allow us to the conflict will be located. The pattern is assess such a specific role. quite strong. The nature of the incompati- A non-geographical factor, the duration of bility (basically the reason for the war) plays the conflict, is also found to be associated the biggest substantive role in determining with the area of the zone of conflict. Given the location of conflict, far exceeding the the lower R2 values associated with the OLS effect of other variables. This result indicates results and the relevant portion of the 3SLS that researchers examining other aspects of estimations, it is clear that our findings with civil war, such as onset and duration, should respect to the scope of conflict are not as well be looking at this variable. The strategic developed as our understanding of the objective of the rebel group is important. location of conflict. We consider the primary We have studied a number of factors that problem to be related to imprecise measure- play a role in determining the scope and ment. Our method of calculating the area of location of conflict. We have found com- the zone of conflict imposes a circular pelling evidence indicating that many of the measure, where the actual shape is more factors at T. E. Lawrence’s fingertips do likely to follow the contours of international indeed affect the nature of conflict. The vari- boundaries, seashores, rivers, etc. Discon- ables in our analyses relate to most of tinuous shapes or fragmented zones of con- Lawrence’s, including geography (location flict, so common in guerrilla warfare, are and scope of conflict, and total land area), merely approximated by our measurement tribal structure, religion, language (rebel technique. By imposing a circular zone of group identity), and appetites (rebel objec- conflict, we in many ways impose a favorable tive or nature of the incompatibility). The operational structure to the battle area factors that affected the nature of the Arab (Collins, 1998: 18). In the future, we hope Revolt in 1916–18 seem to apply to civil to improve our measurement of the area of wars in general. conflict. GIS technology may be useful for improving this measurement. We are more confident of our results References regarding the location of conflict. We have Addison, Tony; Philippe Le Billon & S. Mansoob more confidence in how this variable is Murshed, 2000. ‘Conflict in Africa: The Cost measured, and our estimations of models of Peaceful Behaviour’, paper presented to a with conflict–capital distance as the depen- conference on the Economics of Political dent variable produced much stronger find- Violence, Center for International Studies, ings. Our OLS and 3SLS results show that Woodrow Wilson School of Public and Inter- the location is influenced by the size of the national Affairs, Princeton University, and the Development Research Group of the World scope, the size of the country, the nature of Bank, 18–19 March (http: //www.wws.prince- the rebellion (whether or not the objective of ton.edu/~cis/murshed.pdf). the rebels is to secede), and whether or not Addison, Tony & S. Mansoob Murshed, 2001. the rebel group has a religious or ethnic ‘From Conflict to Reconstruction: Reviving identity. The relationships between location the Social Contract’. Mimeo, UNU/WIDER, and scope and total area of the country are 11 July (http://www.wider.unu.edu/publi- not very surprising. If all countries were the cations/dps/dp2001-48.pdf). 03buhaug (ds) 20/6/02 8:36 am Page 432

432 journal of P EACE R ESEARCH volume 39 / number 4 / july 2002

Auty, Richard M., 1998. ‘Resource Abundance Fearon, James D. & Davis D. Laitin, 1999. ‘Weak and Economic Development: Improving the States, Rough Terrain, and Large-Scale Ethnic Performance of Resource-Rich Countries’, Violence Since 1945’, paper presented at the Research for Action 44, UNU/WIDER, 95th Annual Meeting of the American Helsinki (http://www.wider.unu.edu/publi- Political Science Association, Atlanta, GA, cations/rfa44.pdf). 2–5 September. Auty, Richard M., 2001. Resource Abundance and Gates, Scott, 2001. ‘Empirically Assessing the Economic Development. Oxford: Oxford Uni- Causes of Civil War’. Mimeo, International versity Press. Peace Research Institute, Oslo, 9 December. Boulding, Kenneth, 1962. Conflict and Defense. Gates, Scott, 2002. ‘Recruitment and Allegiance: New York: Harper & Row. The Microfoundations of Rebellion’, Journal Bremer, Stuart A., 1992. ‘Dangerous Dyads: of Conflict Resolution 46(1): 111–130. Conditions Affecting the Likelihood of Inter- Gerrard, John, 2000. ‘What is a Mountain?’ state War, 1816–1965’, Journal of Conflict Res- Mimeo, Development Economics Research olution 36(2): 309–341. Group, World Bank. CIA, 2001. The World Factbook 2001 Gleditsch, Nils Petter, 1995. ‘Geography, (http://www.cia.gov/cia/publications/fact- Democracy, and Peace’, International Inter- book/index.html). actions 20(4): 297–323. Collier, Paul & Anke Hoeffler, 2001. ‘Greed and Gleditsch, Nils Petter; Peter Wallensteen, Mikael Grievance in Civil War’, manuscript, World Eriksson, Margareta Sollenberg & Håvard Bank (http://www.worldbank.org/research/ Strand, 2001. ‘Armed Conflict 1946–2000: A conflict/papers/greedandgrievance.htm). New Dataset’, paper presented at the Euro- Collins, John M., 1998. Military Geography for conference on conflict data, Uppsala, 8–9 June Professionals and the Public. Washington, DC: (http://www.pcr.uu.se/pdf/nilspetterpapper. Brassey’s. pdf). de Soysa, Indra, 2000. ‘The Resource Curse: Are Hegre, Håvard; Tanja Ellingsen, Scott Gates & Wars Driven by Rapacity or Paucity?’, in Mats Nils Petter Gleditsch, 2001. ‘Toward a Demo- Berdal & David Malone, Greed and Grievance: cratic Civil Peace? Democracy, Political Economic Agendas in Civil Wars. Boulder, CO: Change, and Civil War, 1816–1992’, Ameri- Lynne Rienner (113–135). can Political Science Review 95(1): 33–48. Diehl, Paul F., 1991. ‘Geography and War: A Keegan, John, 1993. A History of Warfare. New Review and Assessment of the Empirical Litera- York: Knopf. ture’, International Interactions 17(1): 11–27. Keesing’s Record of World Events 1960–1994, Elbadawi, Ibrahim & Nicholas Sambanis, 2002. 1996. London: Cartermill (http://www. ‘How Much War Will We See? Estimating the keesings.com). Prevalence of Civil War in 161 Countries, Kliot, Nurit & Stanley Waterman, eds, 1991. The 1960–1999’, Journal of Conflict Resolution, Political Geography of Conflict and Peace. forthcoming. London: Belhaven. Ellingsen, Tanja, 2000. ‘Colorful Community or Kunnskapsforlagets Store Verdensatlas, 1997. Oslo: Ethnic Witches’ Brew? Multiethnicity and Kunnskapsforlaget. Domestic Conflict During and After the Cold Le Billon, Philippe, 2001. ‘The Political Ecology War’, Journal of Conflict Resolution 44(2): of War: Natural Resources and Armed Con- 228–249. flicts’, Political Geography 20(5): 561–584. Esty, Daniel C. et al., 1998. ‘The State Failure Lemke, Douglas, 1995. ‘The Tyranny of Dis- Project: Early Warning Research for US tance: Redefining Relevant Dyads’, Inter- Foreign Policy Planning’, in John L. Davies & national Interactions 21(1): 23–38. Ted Robert Gurr, eds, Preventive Measures: Licklider, Roy, 1995. ‘The Consequences of Building Risk Assessment and Crisis Early Negotiated Settlements in Civil Wars, Warning Systems. Boulder, CO: Rowman & 1945–1993’, American Political Science Review Littlefield (27–38). 89(3): 681–690. 03buhaug (ds) 20/6/02 8:36 am Page 433

Halvard Buhaug & Scott Gates G EOGRAPHY OF C IVIL WAR 433

Machiavelli, Niccolò, 1517/1988. The Prince, Sprout, Harold & Margaret Sprout, 1965. The edited by Quentin Skinner & Russell Price. Ecological Perspective on Human Affairs. Cambridge: Cambridge University Press. Princeton, NJ: Princeton University Press. Mackinder, Halford J., 1904. ‘The Geographical Starr, Harvey, 2001. ‘The Power of Place and the Pivot of History’, Geographic Journal 23 Future of Spatial Analysis in the Study of Con- (April): 421–437. flict’, Presidential Address at the 35th North Murray, Williamson, 1999. ‘Some Thoughts on American meeting of the Peace Science Society War and Geography’, Journal of Strategic (International), Atlanta, GA, 26–28 October. Studies 22(2–3): 201–217. Vasquez, John A., 1995. ‘Why Do Neighbors Oxford Economic Atlas of the World, 1972. Fight: Proximity, Interaction, or Territoriality’, London: Oxford University Press. Journal of Peace Research 32(3): 277–293. Pepper, David & Alan Jenkins, eds, 1985. The World Bank, 2000. World Development Indicators Geography of Peace and War. Oxford: Black- 2000. Washington, DC: World Bank. well. Ratzel, Friedrich, 1896. ‘The Territorial Growth of States’, Scottish Geographical Magazine 12: HALVARD BUHAUG, b. 1972, Cand. Polit. 351–361. (NTNU, 2001); PhD candidate in Political Reynal-Querol, Marta, 2002. ‘Ethnicity, Political Science, NTNU (2002– ). Current main Systems and Civil Wars’, Journal of Conflict interest: geography and armed conflict. Resolution 46(1): 29–54. Richardson, Lewis, 1960. Statistics of Deadly Quarrels. Pittsburgh, PA: Boxwood. SCOTT GATES, b. 1957, PhD in Political Ross, Michael Lewin, 2001. ‘Does Oil Hinder Science (University of Michigan, 1989); Democracy?’, World Politics 53(3): 325–361. Associate Professor, Michigan State Uni- Sambanis, Nicholas, 2000. ‘Partition as a Solution versity; Research Professor and Programme to Ethnic War’, World Politics 52(4): 437–483. Leader, PRIO. Recent articles on civil war Sambanis, Nicholas, 2002. ‘A Review of Recent have appeared in Journal of Conflict Resolution Advances and Future Directions in the Quan- and American Political Science Review. titative Literature on Civil War’, Defence and Peace Economics, forthcoming.