Roving and Stationary Bandits in African Armed Conflicts. Kyle C. Beardsley Kristian Skrede Gleditsch Nigel Lo

Transcription

1 Roving and Stationary Bandits in African Armed Conflicts Kyle C. Beardsley Kristian Skrede Gleditsch Nigel Lo Working draft of March 11, 2013

2 Introduction Recent research on civil war has highlighted the specific locations or zones where conflict take place. Civil wars often take place in confined and atypical parts of a country, and researchers have emphasized the need to look at the local characteristics in conflict zones to evaluate claims about possible motives and opportunities for conflict (Buhaug and Rød 2006; Buhaug, Cederman and Rød 2008; Cederman, Buhaug and Rød 2009; Cederman, Weidmann and Gleditsch 2011; Cunningham and Weidmann 2010; Hegre, Østby and Raleigh 2009; Raleigh and Hegre 2009; Weidmann 2009, 2011; Weidmann and Ward 2010; Zhukov 2012). Although this line of research has been helpful and productive, most of the existing research has treated the conflict zone as a fixed attribute. In reality, however, conflict zones can change considerable over the course of a conflict. There are a number of reasons why actors may deliberately induce changes in the conflict zones or conflict characteristics that will make conflicts more or less likely to involve changes in geographic scope. While a number of studies have considered the potential for intrastate conflict to spread between countries (Beardsley 2011; Braithwaite 2010; Buhaug and Gleditsch 2008; Salehyan 2009; Salehyan and Gleditsch 2006), and recent work has examined the spread of conflict between more geographically disaggregated locations (Schutte and Weidmann 2011; Zhukov 2012), existing work has yet to focus on the movements of whole conflict zones, and why some rebel-government dyads are more or less likely to consistently fight in the same locations. Rebels are mobile, and may seek to evade attacks by the government or enact strikes against government forces in different locations. Theories of guerrilla warfare place a 1

3 great deal of emphasis on how groups should find the optimal locations for armed attacks. Ernesto Che Guevara s 1961 theory of foco warfare, for example, argues in favor of fighting in the periphery or parts of the countryside where the repressive apparatus of the state is relatively weaker. Although such guerrilla warfare by itself cannot overthrow the government, sustained fighting activity can allow converting people to the rebel s cause and grow the movement to a point where it eventually take on the government. Conversely, governments counterinsurgency strategies typically advocate attacks against rebel strongholds to force rebels to flee, to force them to fight under less favorable circumstances, and to try to turn civilians against the rebels. The Cuban revolution provides a clear example of a moving conflict zone. After the rebels led by Castro arrived by boat on Cuba on 2 December 1956, the group was initially almost squashed during intense fighting with the army on the plains before making their way into the Sierra Maestra mountains. The rebels later established themselves in the mountains, and successfully fought off government attacks, helped both by the increasingly unpopularity of the Batista regime and a US arms embargo against Cuba. In December 1958, following an unsuccessful government counteroffensive, the rebels started moving North and West. After the capture of the city of Santa Clara, Batista fled the country, and the rebels marched on the capital without facing significant resistance. The conflicts in the Great Lakes Region of Africa demonstrate that such examples of large movements in conflict zones are common and not limited to individual incidents such as the Cuban Revolution. Rebels have operated from neighboring countries and sought to move the theater of fighting to places where the government was unable to retaliate. As the central government in Zaire collapsed, Tutsi rebels in the East of the country, sup- 2

4 ported by Rwanda and Uganda, joined with local forces and marched on the capital more than 1500 kms to the West. The above discussion makes it clear how many aspects of conflict constellations that can make rebels or conflicts more or less likely to move over time, including the relative strength of actors, the resource endowments of rebels, and their relationship with local population are likely. In this manuscript we develop new theory to distinguish between roving and stationary rebels in civil wars. We develop new measures, based on the shifts in the centroids of conflict polygons derived from the new event data provided by the Georeferenced Event Dataset (GED) developed by the Uppsala Conflict Data Program (UCDP). Our results show that rebels are more likely to be mobile when they are quite weak relative to the government, as they are easily displaced. We also find that groups are more mobile when a third-party provides direct military support as a conflict participant, which is consistent with the notion that outside support decreases the interdependence that rebel groups might form with local communities. Moreover, we observe that the mobility of rebel groups matters to the nature of the conflict waged, as mobile rebel groups that are relatively strong are more likely to be more brutal toward civilian populations. Such groups are likely to be mobile because they lack the incentives to form something like a social contract with local communities; instead, they are prone to prey on local populations. 3

5 Theory Why Rove? We consider two types of explanation for why some rebel groups are more mobile than others. Like many other theories in international politics and conflict research (Most and Starr 1989), we organize our approach around opportunity and willingness as necessary conditions for rebel movement. In terms of opportunity, we focus on the constraints created by the strength of the government against which the rebels fight. In terms of willingness, we consider the heterogeneity in rebel group types and posit that opportunistic rebels will tend to share a lot of the same characteristics as Mancur Olson s 1993; 2000 roving bandits. Starting with opportunity, there are two sides to how government power is likely to shape the mobility of rebel groups. On the one hand, a rather weak government will find it more difficult to fight and contain rebel activities, especially in the periphery of a country (Buhaug and Rød 2006; Buhaug, Cederman and Rød 2008; Cederman, Buhaug and Rød 2009; Cunningham, Gleditsch and Salehyan 2009). Rebels in these situations will tend to have more opportunity to roam than groups that must contend with a government capable of countering rebel advances. In the extreme, collapsing governments present an opportunity for rebel groups to advance on and take the capital. On the other hand, a strong government can decrease the opportunity for rebel groups to remain stationary. That is, stronger governments can enforce their laws and keep a rebel group from establishing its own hierarchical relationship in local communities (Kalyvas 2006; Skaperdas and Syropoulos 1995). With the threat and use of force, such govern- 4

6 ments can force rebel groups to be on the run, lest they face major losses and capture. Just as Cunningham, Gleditsch and Salehyan (2009) found that relatively strong rebel groups tend to be associated with shorter wars because they can impose enough costs on the government to settle or be vulnerable to defeat, such strong rebel groups are better able to stand their ground in the areas they control. This line of argument competes with the one above, as both make different assumptions about what mobility looks like. If mobility can be quiet and go often undetected even by strong governments, then we are more likely to see strong governments have a type of displacement effect. If mobility entails massive shifts in people and resources that strong governments are likely to prevent, then we are more likely to see rebel groups dig in when there is a strong government. Turning to willingness, we begin with Mancur Olson s notion of roving and stationary bandits and then apply it to Jeremy Weinstein s 2007 opportunistic rebellions. As a thought experiment, Olson considers two different types of bandits in anarchy. The roving bandits rely on brute force to move from place to place, extract rents, and then move onto the next place after they have drained the available resources. The stationary bandits, in contrast, set up hierarchical relationships and extract resources less by force and more by some form of a social contract, where the bandits provide some public goods such as security and in return receive rents. The stationary bandit does not have an interest in destroying or in extracting all the available resources, for fear that there will be no more resources in the future. There is thus an incentive for the stationary bandit to invest in the conquered area rather than completely exploit it. With this imagery in mind, we argue that opportunistic rebellions are more likely to 5

7 involve a type of roving banditry. Those groups that are able to finance their activities from lootable resources or from external support will be less dependent on local support and more likely to attract the type of participants that are prone to loot and abuse the civilian population. Because they have alternative income streams, they do not have to worry as much about the consequences of leaning on local communities until they effectively collapse. Opportunistic rebellions should thus be more likely to move from community to community and less likely to establish strong roots in a given area. In contrast, activist rebellions are more likely to involve substantial investment in local support. They are more likely to resemble the types of actions expected by a stationary bandit. This discussion on willingness thus provides additional insight into both when we should expect groups to be mobile and how the mobile groups are likely to behave. Groups that have ample access to resources outside of local revenues are more likely to be mobile and also more likely to be rather brutal to the local populations that they encounter. These groups have less incentive to invest in the long-term viability of the local civilian population. Groups that are more dependent on local communities, however, will be less likely to roam and less likely to treat local civilians with brutality. We turn now to our first two hypotheses, which are partially competing (or suggest a curvilinear relationship) and deal with the relative strength of the rebels and government. On the one hand, we might expect that rebel groups which are much weaker than the government will need to be more mobile, since they lack the resources to hold territory or to otherwise be able to inflict sufficient costs on the government to deter it from enforcing its control over the area. In essence, the stronger the government becomes relative to the rebels, the more likely it is to displace conflict, as the rebels only means of survival is to 6

8 stay mobile. That is, flight will be much more attractive when choosing to fight is likely to be disastrous. Another way to see this argument is to return back to Olson s notion of roving bandits in anarchy. When the government is strong, existing hierarchical relationships are well entrenched, and the rebels will not find it feasible to become stationary bandits and compete for their own hierarchical relationships. Roving may be their only option. When the government is weak, rebels will have more of an option to set up their own local structures of order in the absence of such order provided by the government. more. H1a (Displacement): When rebels are relatively weak, the conflict space will move On the other hand, a weak government government relative the rebels is less capable of keeping the rebels from expanding their control of territory and/or moving on the capital. For example, one early indication that the Gaddafi government in Libya was losing control in 2011 was the relative easy with which the rebels primarily based in the east of the country were able to expand the theater of conflict toward Tripoli in the west. more. H1b (State Collapse): When rebels are relatively strong, the conflict space will move The second set of hypotheses speaks more directly to Olson s roving and stationary bandits, and deal with issues related to the social contract. Rebels that depend more on local support will stay in place, while those that have other sources of resources will more likely rove. There could be two causal pathways at work here, which need not be competing. First, scarce resources could leave rebel groups with no other option but to depend on 7

9 the local population for provisions and cover, which means that the group becomes tied to specific locations of support. Second, groups that arise out of local grievances and become a source of local protection from state infringement or the providers of other public goods neglected by the state are not likely to need other resources to stay viable, given abundant local support. If Beardsley and McQuinn (2009) are correct, however, and rebel groups will tend to adapt toward investment in the most lucrative resources similar to the Revolutionary Armed Forces of Colombia (FARC) this second mechanism might not have as much explanatory power as a simple explanation about groups being forced by scarcity to invest locally. That is, given a greater availability of lucrative non-local support with a high potential return on investment, groups that started as locally-based groups are likely to become less tied to local communities, where their return on investment is likely to be lower given the need to provide public goods in return for the support. We form two hypotheses from this argument, with expectations on two different types of non-local-community resources. One type of resource is external support from abroad. Many rebel groups receive substantial funding from diasporas and from third-party patron states or other allied rebels. The more that a group depends on such transfers, the less need they have to invest in local communities in return for such local support. H2 (External Support): When rebels get substantial outside support, the conflict space will move more. Another type of non-local resource is access to lootable minerals such as diamond deposits. Groups that can invest in controlling lootable resources also will have less need to invest in their ties to local communities. 8

10 H3 (Lootable Resources): When rebels have access to lootable resources, the conflict space will move more. Why Roving Matters In addition to explaining why some conflict zones move more than others, we posit a testable hypothesis regarding a crucial behavioral distinction between mobile and stationary groups. We expect roving rebels to be more likely to engage in violence against civilians because the rebels are less dependent on their voluntary support. The groups mobility affords them the ability to essentially pillage local resources without having to worry about how that damages their long-run ability to form an equilibrium symbiotic relationship. As long as there are other communities to exploit, mobile rebel groups will not damage their own viability by currently devastating a particular community. Note that this logic parallels the work by Weinstein (2007) on opportunistic rebellions, which highlights how groups that depend on lucrative sources of resources will be less likely to depend on local support, and more prone to indiscriminate violence. Groups that tend to be more stationary, in contrast, do depend on observing constraints on how much they can abuse the local population. Not only would overstepping their bounds lead to local opposition that would make their activities more costly, but also demand too much of the available resources at a time could lead to a collapse of the local economy and thus the ability to continue reaping benefits from the community. There is an important distinction between groups that are roving because of displacement and groups that are roving because they are opportunistic rebellions. One way to approach this is to consider how the relative strength of the rebels conditions the expected 9

11 relationship between mobility and violence against civilians. We should expect rebels that are relatively strong and yet roving to be the most likely to inflict violence against civilians. These are the groups that have more freedom in their choice to remain stationary or to rove they are not compelled to rove for survival borne out of relative weakness. By implication of the theory above, we can surmise that this set of rebel groups is likely to consist of a high proportion of opportunistic rebellions that do not depend on civilian support. If they actually depended on local support, then they would simply stay put, as long, again assuming that they are strong enough to resist government action. In contrast, the mobile groups that are rather weak relative to the government are likely to have a higher proportion of groups that are not as prone to severely exploit local resources for two reasons. First, many of these groups are likely to be those with origins in more community-based grievances that recruited more ideological types than opportunist types. Were it not for government strength, these groups would stay in place, and as they become forced to move, they still remain dependent on more sympathetic voluntary support than other sources of funding. Second, even weak rebel groups that have attracted more opportunist rebels and that depend on other sources of funding will not likely be strong enough to pillage. Their limitations in being able to sufficiently harm the government are likely to also limit them in being able to move from one community to the next and loot. H4 (Violence): As the conflict-space movement increases, rebels are more violent against civilians when the rebels are not relatively weak. 10

12 1 Illustrative Cases We provide three selected examples to illustrate our propositions and the data sources, highlighting variation in actor characteristics and roaming. We take al-qaida in the Islamic Maghreb (AQIM) in Algeria as an example of a small, relatively weak rebel group that is highly mobile. We consider the Rwandan Patriotic Front (RPF) as an example of a strong and more stationary rebel group. Finally, the Lord s Resistance Army (LRA) in Uganda provides an example of a relatively strong group that is mobile and has a history of brutality against non-combatants. Algeria - al-qaida in the Islamic Maghreb The AQIM emerged from the civil war in Algeria in the 1990s, as a splinter group from the Armed Islamic Group (GIA), initially calling itself The Salafist Group for Preaching and Combat. The AQIM first appears as an active non-state armed group in the Uppsala Conflict Data Program s Armed Conflict Data in The group is generally believed to be small (the estimates of the number of troops range from 2000 to as few as 30, with 500 as a best estimate) and is clearly much weaker than the government. The main mode of attack of the AQIM has been car bombs or quick ambushes rather than any form of sustained guerrilla warfare. Consistent with our notion that weak rebels are more likely to roam, the location of attacks is widely scattered, and can change considerable from one year to another. Figure 1 displays the actual geographical spread of the individual conflict events (shown as dots) and the conflict polygons in the UCDP Georeferenced Event Data (GED) for Algeria vs. AQIM for the years 2000 and 2009 (we will explain these data in greater 11

13 detail later). As noted in the UCDP GED Conflict Polygons Dataset Codebook (Version 1.1, p. 24), since incidents in the conflict took place in two distinct theaters or regions, this is one of a small number of conflicts in the UCDP GED data that has been split up into two distinct polygons. We consider the two polygons separately in our analyses. AQIM AQIM km 1000 km Figure 1: Algeria - al-qaida in the Islamic Maghreb The first polygon (4A) covers the northern, more densely populated are of the country. Although this displays some year-on-year variation, since the attacks are frequent and dispersed all across the Northern coastline and populated parts of Algeria, the conflict polygon moves little from one year to another. By contrast, the second polygon (4B) covers a much smaller number of attacks in the Sahara desert in the interior of the country. These attacks are scattered over a very large era, and since the attacks can be few and very far apart, the conflict zone can move considerably from one year to another. 12

14 Rwanda - Rwandan Patriotic Front Whereas small and weak rebel groups such as the AQIM are likely to engage in terrorist attacks or surprise attacks over a wide area, we have argued that stronger rebel groups should be less likely to roam. The Rwandan Patriotic Front (RPF) provides one example of an exceptionally strong rebel group. Although numerically smaller than the Hutu dominated Rwandan army, the RPF were much better equipped and well-trained, with the support of the Ugandan army. Following the invasion from their bases in Uganda in 1990, the RPF made quick headway and their initial military offensive only failed due to military interventions from France and Zaire on the side of the government. Following the initial setback, the RPF engaged in guerrilla warfare in smaller parts of the north of Rwanda, until it renewed its offensive and marched on the capital in the wake of the Rwandan genocide in RPF RPF km 100 km Figure 2: Rwanda - Rwandan Patriotic Front Figure 2 displays the individual conflict events and the conflict polygons in the UCDP GED data for Rwanda vs. the RPF for the years 1990 to Although the conflict 13

15 polygon shifts so that there is little overlap between the two, the small geographical extent or scope of the conflict implies that there is much less movement here than in the case of AQIM. One might argue that the small geographical size of Rwanda would make roaming less likely in the first place. However, we note that the attacks are concentrated around the RPF core area in the Virunga mountains, where the government was unable to effectively target the rebels. Moreover, the the conflict certainly could have spread across national boundaries, as many other conflicts in the Great Lakes region have, and the RPF clearly had a transnational presence at the outset in this case. Uganda - Lord s Resistance Army The Lord s Resistance Army (LRA) in Uganda provides an example of a relatively strong group with significant roaming and a history of targeting civilians. is first coded as active in 1989 in the UCDP data. The group has its original roots among ethnic Achoilis, who dominated the military prior to the coup that brought Museweni to power, but the LRA has later acquired an increasingly obscure religious ideology and aims. The LRA first engaged in guerrilla warfare in the North of the country and targeted civilians for supplies. However, following government efforts to erode LRA support and the organization of pro-government armed groups among the local population, the LRA turned to increasingly more brutal attacks against civilians. As retaliation for Ugandan support to the Sudan People s Liberation Army (SPLA), the Sudanese government allowed the LRA to established bases in Sudan, and the conflict has also spilled over into the Democratic Republic of the Congo. As can be seen in Figure 3 in 1989 the conflict polygon is confined to Acholiland, but 2009 the battle events are large outside Uganda proper. 14

16 LRA LRA km 500 km Figure 3: Uganda - Lord s Resistance Army Research Design To test these hypotheses, we use as the unit of analysis the dyad-polygon-year from the UCDP GED data (Croicu and Sundberg 2012).1 The GED data comprise geo-referenced event data of battles within armed conflicts, where an armed conflict has at least 25 battlerelated fatalities per year. From these individual events, the GED data has also estimated geo-referenced polygons, based on calculations of the convex hull of the individual battles, to provide a definition of the battle areas in each year of active armed conflict. A few armed conflicts have multiple theaters of activity and thus yield multiple polygons. Moreover, the UCDP exclude certain isolated outliers that would lead to unreasonably 1 We chose to use the GED data instead of the Armed Conflict Location and Event Data (ACLED) because the conflict polygons have already been drawn and for ease of integration with many of the other variables that pertain to UCDP armed conflicts. We chose to use the polygon-year as the unit of analysis instead of a grid-square-year because we either do not have sufficiently disaggregated information for some of our explanatory variables (e.g., relative strength), and other explanatory variables are at the level of the rebel group as a whole (e.g., outside support). 15

17 definer or overly large conflict polygons. 2 We also only focus on the state-rebel violence in civil wars covered by the non-state actor data (Cunningham, Gleditsch and Salehyan 2009), since the scope of our theory applies to cases of rebellion against the state but not necessarily to all cases of violence or non-state conflict or fighting between rebel groups or cases of one-sided violence by the state against civilians. As a result, we choose each polygon of these government-rebel dyads in each year to serve as our unit of analysis. This gave us 371 initial cases. Our approach is to see how a number of covariates help explain movements in those polygons from one year to the next. We estimate random effects OLS models, with the dyad-polygon as the unit. We choose to run random effects models because of the potential for conflict-area change to be a function of idiosyncratic characteristics that vary across the units. Without accounting for such unit heterogeneity, our models would be susceptible to strong amounts of autocorrelation. Moreover, random effects is preferable to fixed effects here, since much of the interesting variation to explain is across the units and not within the units. That is, we are more interested in why some conflict areas are more prone to shift than others and not as much the question of when particular conflict areas are likely to experience the most movement. 2 More precisely, events are considered outliers for defining polygons if they alone would account for more than 20% of the area and decrease the density of the data points in the polygon by more than 20%, and the events account for less than 5% of the total number of events and deaths in the conflict (Croicu and Sundberg 2012, 9). 16

18 Dependent Variables The dependent variable to test the first three hypotheses is the distance change in the polygon centroid from the previous period. This captures the extent to which the rebel groups become active fighters in new areas. In this way, our measure is more closely tied to the actual relocation of conflict, as opposed to the escalation of conflict (Schutte and Weidmann 2011), which might be captured by the changing area of the conflict zone. When unresolved conflicts drop below the 25-fatality threshold and become dormant for a little while, the distance change calculated for the first observation of the conflict when it again rises above the threshold is calculated by taking the centroid distance change across the dormant period and dividing by the difference in years between the two observations. All cases of armed conflict that started and ended in the same calendar year were dropped from the analysis, since these conflicts did not have an opportunity to move at a scale that we can register. This resulted in the loss of 114 observations. The dependent variable to test H4 is the one-sided casualty estimates of rebels against civilians. To calculate these numbers, we took the UCDP GED data on one-sided violence between the rebels and the civilians and added up all the one-sided casualties attributed to each rebel group in the same location country in the same year. We then merged that with the other information on the government-rebel dyads. Explanatory Variables The first set of hypotheses (H1a & H1b) focus on the relative strength of the rebels to the government. There are a number of ways to measure that, and we choose three indicators here. The first is a measure coded in the non-state actor data itself as a categorical measure 17

19 of relative rebel strength, based on a number of factors including troop sizes and fighting capacity. While most intrastate conflicts involve rebels that are weaker than the government, we are interested in cases in which they are extremely disadvantaged and have to rely on flight for survival. As a result, we code a dummy variable for whether this variable indicates that the rebels are much weaker than the government. This variable, as well as all others that can change from year to year in a conflict, is lagged one year because the dependent variable is movement from the previous year to the current year and there is a potential for simultaneity bias without the lags. A second measure that we use is the nearest distance between the conflict-polygon borders and a state boundary. High values on this variable imply that the conflict area is not in the periphery, involve fewer opportunities for rebels to benefit from operating across borders and from neighboring states, and that the central state is likely to be relatively strong. Existing research has shown that governments have a much more difficult time projecting their power into areas near the borders (Buhaug and Rød 2006; Buhaug, Cederman and Rød 2008; Cederman, Buhaug and Rød 2009; Salehyan 2009). Third, we use GDP per capita as a measure of the resources available to the government. GDP per capita has been used in existing studies as a proxy for state capacity (Fearon and Laitin 2003). As proxies go, however, GDP per capita can capture a number of different phenomena, and it is often difficult to pinpoint what it is measuring. At a minimum, we thus make sure to control for the local productivity within each conflict area. Nordhaus (2006) has collected data on the domestic production of 1deg-latitudeby-1deg-longitude geographic cells in five-year increments, in units of 1995 purchasingpower adjusted US dollars. We use those data and average across the cells within our 18

20 conflict areas. By controlling for local production, the national GDP per capita measure becomes more of an indicator of wealth that the central state can capture and convert into military strength as needed. We also go one step further and interact the local productivity with the national productivity. Since we want a measure of relative power, the local productivity can inform how much wealth the rebels in the area are able to capture. So, we would expect there to be large disparity in government-rebel power when the national GDP per capita is high but local productivity in the conflict areas is low. This can be assessed by simply interacting the two variables. The second and third hypotheses pertain to sources of funding that can make rebel groups less dependent on local support. For H2, we use two measures of external support from the UCDP External Support Data (Högbladh, Pettersson and Themnér 2011). The first is whether there was direct military participation as a warring party by external actors for the rebel side in each government-rebel dyad. The second is whether there was monetary support by external actors for the rebel side. For H3, we focus on diamond deposits as a lootable resource. Using the CSCW data on the geographic location of diamond deposits, we calculate whether there are diamond deposits in each conflict polygon (GIlmore, Gleditsch, Lujala and Rød 2005). For H4, the dependent variable for the first models becomes the key explanatory variable. We also interact this variable with each of the measures of relative rebel strength, so that we can see if one-sided violence is most likely when rebel groups are quite mobile yet not very weak. 19

21 Control Variables There are a number of variables that are other potential causes of our dependent variables, and that are also plausibly linked to some of the key explanatory variables. For our base models, we take a parsimonious approach and only control for those factors over which we have strong theoretical priors. in other tests for robustness, we include the other variables described below. Each model includes the area of the polygons that define the conflict areas. The magnitude of the movement in the polygon centroid is affected by the overall size of the conflict, where the onset of hostilities in new locations will tend to move the centroids of smaller conflicts further than the centroids of larger conflicts. Moreover, the overall scope of the conflict is likely to relate to rebel capabilities and the possibility for diamond mines to be present in ways that are not part of the hypothesized mechanisms. Related, we also control for the number of GED events used to comprise each polygon, since these will affect the shape of the polygons formed. Polygon shape will have some bearing on the extent to which the centroid is likely to move in the next period, as shapes with fewer vertices are likely to be more prone to large shifts as battle locations come and go from year to year. We also control for the duration of the conflict in years in model robustness tests because longer conflicts may have become entrenched and have less space into which to move. We include an indicator of whether a peacekeeping force which might impede the spread of conflict or help displace it (Beardsley 2011) is present, based on the International Military Interventions (IMI) data (Pickering and Kisangani 2009). 3 We control 3 See Beardsley (2011) for a description of how the IMI data were used to select out the peacekeeping events. 20

22 for the percent of the conflict-area polygons that encompass urban space, from the Global Rural-Urban Mapping Project (GRUMP) data (GRUMP 2011), because the spread of conflict is likely to play out differently in urban and rural environments. We also include the number of distinct ethnic-group homelands encompassed by the polygons, where the ethnic mapping comes from the GeoEPR data (Wucherpfennig, Weidman, Girardin, Cederman and Wimmer 2011), because the impact of ethnic fractionalization on conflict dynamics has been much debated. Since conflict in the mountains faces geographical constraints in the potential to spread, we also control for the mean elevation of the conflict polygons. We additionally expect secessionist rebels to have more limited geographic objectives than other types of groups, so we control for whether the government-rebel dyads are rooted in a secessionist struggle, from the non-state actor data. For H4, we also control for the severity of the hostilities, since an association between mobile conflict areas and one sided violence could also be explained by high levels of hostilities in general and have nothing to do with the types of groups that become mobile. We thus control for the estimated number of battled deaths related to government-rebel fighting from the UCDP GED data, as well as the number of social conflict events that occurred in each conflict polygon from the Social Conflict in Africa Dataset (SCAD) (Hendrix and Salehyan 2011). This latter measure is included to control for lower-level potential for mass upheaval. We also include both of these measures as control variables in the robustness checks for the models testing the first three hypotheses. 21

23 Results Table 1 shows the results of two models of the determinants of movement in the centroids of the conflict polygons from the previous year. Model 1 includes minimal control variables, while Model 2 includes many potential confounding variables. Since the results are similar across the models, we use Model 1 to inform the substantive effects. These substantive effects can be seen in Figures 4 through 7. The results support H1a. When rebels are much weaker than their government opponents, they are more likely to be mobile (see Figure 4). Moreover, rebels that operate far from state boundaries, away from the periphery and international havens, are more likely to be mobile (see Figure 5). In addition, assessing GDP per capita and its interaction with local economic production indicates that rebels are most mobile when the overall economic health of the state is strong but the local resources available are scant. That is, from Figure 6, we see that GDP per capita only has a positive and statistically significant effect when the local economic production is below 5,000 USD per capita. Each of these relationships is consistent with the notion that strong governments will tend to displace weak rebellions which must often flee to remain viable and do not have much of an option to form deep roots in particular communities. We do not see much support for H1b, which expected conflict zones to rapidly shift when the state is weak. In results not shown, we additionally tested for curvilinear effects, where perhaps conflict movement is highest when either the state is really strong or really weak, but we did not find any evidence to support this. It is likely that our yearly data are too aggregated to capture many of the cases in which the state rapidly collapsed and rebel movements or fronts took control of the capital. For one, many such collapses do 22

24 Table 1: Random Effects Models of Centroid Distance Change IV 1 2 Rebels much weaker 64.81* (38.51) 66.55* (38.69) Distance to state border * ( ) * ( ) GDP per capita, lag ** (0.0186) ** (0.0186) Local economic production (GCP), 5 yrs (0.0148) (0.0153) GDPpc x GCP -8.51e-06** (3.50e-06) -8.75e-06** (3.62e-06) External direct miltary participation with rebels, lag 204.6* (107.1) 175.5* (104.4) External monetary support for rebels, lag (29.69) (30.38) Diamond deposits (50.31) (52.96) Conflict area, lag -9.63e-08 (9.12e-08) -7.62e-08 (9.46e-08) Number of GED events, lag (0.320) * (0.327) Conflict duration (3.113) Peacekeeping, lag (27.98) Pct. of area that is urban (80.09) Average elevation * (0.0273) Number of ethnic group homelands (4.676) Conflict fatalities, lag ( ) Secessionist conflict * (42.61) Constant (35.21) (47.37) Observations Standard errors in parentheses * p <.05, ** p <.01 in a one-tailed test 23

25 Expected Centroid Distance Change (km) Rebel Group Much Weaker Note: 90% confidence intervals Figure 4: Expected Centroid Distance Change, by Rebel Strength not cross over arbitrary year boundaries. For another, many such collapses involve state abdication that does not necessitate an invasion of the capital. Turning to H2, there is modest evidence that outside support matters we see that actual participation as a secondary warring party by an outside actor is associated with more movement in the conflict space (see Figure 7). This is consistent with the notion that rebels which receive outside assistance have less need to rely on consistent local support from specific communities. The indicators of monetary support, however, does 24

26 Expected Centroid Distance Change (km) Distance to State Border (m) Note: 90% confidence intervals Figure 5: Expected Centroid Distance Change, by Distance to State Border not have a positive association with conflict-zone movement, and this is arguably a better measure than the measure of outside direct military support as a warring-party participant. The effect of actual third-party participation in hostilities could have more to do with high levels of escalation than anything about the social contract and local interdependence. 4 We do not observe much support for H3. The presence of lootable resources as 4 See Schutte and Weidmann (2011) for an approach that distinguishes conflict relocation from conflict escalation. 25

27 measured by whether there are diamond deposits in the conflict areas do not appear to have much of an effect on the movements of the conflict zones. One limitation of our test is that we do not formally assess whether lootable resources are present in areas outside the conflict zone that rebels could move to. It is also possible that lootable resources such as diamond mines might have countervailing effects, as they could encourage armed actors to stay in place and protect the mines even if there is little dependence on local support. We find support for H4 in the analysis of one-sided violence against civilians. Table 2 presents three models of rebel violence against civilians with each of the measures of relative strength interacted with the centroid distance change. The expectation is that rebels will be more violent against civilians when they are more mobile and not particularly weak. In Model 3, we observe that the relationship between the centroid distance change and rebel one-sided violence is strong and positive only when the rebels are not much weaker than the government (see Figure 8). When they are weaker, the relationship goes away, presumably because they are likely moving because they are displaced and not because they are opportunistic rebels. Similarly, from Model 4, movement in conflict zones is associated with one-sided violence only when the state is resource poor and thus less likely to maintain a daunting security sector (see Figure 9). Mobile rebellions in such situations are likely to be opportunistic types that roam not because they have to, but because they do not have incentives to become interdependent with local communities. In Model 5, we do not see the same pattern with regard to the interaction with border distance. This is worthy of further investigation in its own right, and may indicate that the mobile groups in the peripheries are more heterogenous in terms of whether they are 26

28 opportunistic or activist. Even activist groups might be quite mobile in borderlands, especially if there are porous borders that allow them to seek haven and the potential to spread their cause elsewhere. These one-sided violence findings intersect well with the importance of displacement found in the tests for H1a. Strong governments tend to displace rebel groups, but these groups are not necessarily like other groups that are mobile because of their opportunistic tendencies and are averse to local interdependence. Mobile groups may or may not be dangerous to civilians; those that are mobile more by choice than by need present the greatest problem. Conclusions we have examined a series of propositions on factors that would make conflicts more likely to move geographically. Many studies that have looked at the specific locations where conflict occurs and the relationship of conflict events to the local context, or the shape and size of conflict zones. However, ours is the first to consider how characteristics of rebel groups and their relationship to the state influences changes in the conflict zone over time. Rebels are mobile, and can chose over a menu of options in terms of specific attack targets and locations. We find that weak rebels tend to move more, consistent with the notion of displacements. Third party support is also associated with more mobile rebels, consistent with a resource-autonomy tradeoff where rebels with outside support become less dependent or connected to communities. Mobile and strong rebel groups are more likely to target civilians as they lack the incentives to form a social contract with local communities and enlist active support, and instead prey on their resources. 27

29 Table 2: Random Effects Models of One-Sided Fatalities Attributed to Rebels IV Centroid distance change 0.769** (0.187) 0.832** (0.196) (0.133) GDP per capita (0.0304) Centroid change x GDPpc ** (3.83e-05) Rebels much weaker (77.66) Centroid change x Rebels much weaker ** (0.203) Distance to state border 4.64e-05 ( ) Centroid change x Distance to border 4.05e-07 (4.15e-07) Local economic production (GCP), 5 yrs (0.0115) Conflict fatalities (0.0146) (0.0153) (0.0157) Conflict area 4.04e-07** (1.07e-07) 4.37e-07** (1.11e-07) 5.11e-07** (1.15e-07) Number of GED events (0.411) (0.425) * (0.442) Number of SCAD events (3.256) (3.394) (3.564) Constant (56.13) (65.64) (55.82) Observations Standard errors in parentheses * p <.05, ** p <.01 in a one-tailed test 28

30 References Beardsley, Kyle Peacekeeping and the Contagion of Armed Conflict. Journal of Politics 73(4): Beardsley, Kyle and Brian McQuinn Rebel Groups as Predatory Organizations: The Political Effects of the 2004 Tsunami in Indonesia and Sri Lanka. Journal of Conflict Resolution 53(4): Braithwaite, Alex Resisting Infection: How State Capacity Conditions Conflict Contagion. Journal of Peace Research 47(3): Buhaug, Halvard and Jan Ketil Rød Local Determinants of African Civil Wars, Political Geography 25(3): Buhaug, Halvard and Kristian Skrede Gleditsch Contagion or Confusion? Why Conflicts Cluster in Space. International Studies Quarterly 52(2): Buhaug, Halvard, Lars-Erik Cederman and Jan Ketil Rød Disaggregating Ethno- Nationalist Civil Wars: A Dyadic Test of Exclusion Theory. International Organization 62(3): Cederman, Lars-Erik, Halvard Buhaug and Jan Ketil Rød Ethno-Nationalist Dyads and Civil War: A GIS-Based Analysis. Journal of Conflict Resolution 53(4): Cederman, Lars-Erik, Nils B. Weidmann and Kristian Shrede Gleditsch Horizontal Inequalities and Ethno-Nationalist Civil War: A Global Comparison. American Political Science Review 105(3): Croicu, Mihai Catalin and Ralph Sundberg UCDP GED Conflict Polygons Dataset Codebook version Department of Peace and Conflict Research, Uppsala 29

31 University. Cunningham, David, Kristian Skrede Gleditsch and Idean Salehyan It Takes Two: A Dyadic Analysis of Civil War Duration and Outcome. Journal of Conflict Resolution 53(4): Cunningham, Kathleen Gallagher and Nils B. Weidmann Shared Space: Ethnic Groups, State Accommodation, and Localized Conflict. International Studies Quarterly 54(4): Fearon, James D. and David Laitin Ethnicity, Insurgency, and Civil War. American Political Science Review 97(1): GIlmore, Elisabeth, Nils Petter Gleditsch, Päivi Lujala and Jan Ketil Rød Conflict Diamonds: A New Dataset. Conflict Management and Peace Science 22(3): GRUMP Global Rural-Urban Mapping Project, Version 1 (GRUMPv1): Urban Extents Grid. Palisades, NY: NASA Socioeconomic Data and Applications Center (SEDAC), Center for International Earth Science Information Network (CIESIN), Columbia University; International Food Policy Research Institute (IFPRI), the World Bank, and the Centro Internacional de Agricultura Tropical (CIAT). Accessed 4 July Guevara, Ernesto (Che) Guerrilla Warfare. New York: Monthly Review Press. Hegre, Håvard, Gudrun Østby and Clionadh Raleigh Poverty and Civil War Events: A Disaggregated Study of Liberia. Journal of Conflict Resolution 53(4): Hendrix, Cullen S. and Idean Salehyan Social Conflict in Africa Database (SCAD). 30

32 Accessed on 23 August Högbladh, Stina, Therése Pettersson and Lotta Themnér External Support in Armed Conflict Presenting new data. Presented at the 52nd Annual International Studies Association Convention, Montreal, Canada. Kalyvas, Stathis The Logic of Violence in Civil War. New York: Cambridge University Press. Most, Benjamin A. and Harvey Starr Inquiry, Logic, and International Politics. Columbia, SC: University of South Carolina Press. Nordhaus, William D Geography and Macroeconomics: New Data and New Findings. Proceedings of the National Academy of Sciences 103(10): Olson, Mancur Democracy, Dictatorship, and Development. American Political Science Review 87(3): Olson, Mancur Power and Prosperity. New York: Basic Books. Pickering, Jeffrey and Emizet F. Kisangani The International Military Intervention Dataset: An Updated Resource for Conflict Scholars. Journal of Peace Research 46(4): Raleigh, Clionadh and Håvard Hegre Population size, concentration, and civil war. A geographically disaggregated analysis. Political Geography 28(4): Salehyan, Idean Rebels without Borders. Ithaca, NY: Cornell University Press. Salehyan, Idean and Kristian Skrede Gleditsch Refugees and the Spread of Civil War. International Organization 60(2): Schutte, Sebastian and Nils B. Weidmann Diffusion Patterns of Violence in Civil Wars. Political Geography 30(3):

33 Skaperdas, Stergios and Constantinos Syropoulos Gangs as Primitive States. In The Economics of Organised Crime, ed. G. Fiorentini and S. Peltzman. Cambridge: Cambridge University Press. Weidmann, Nils B Geography as Motivation and Opportunity. Journal of Conflict Resolution 53(4): Weidmann, Nils B Violence from above or from below? The Role of Ethnicity in Bosnias Civil War. Journal of Politics 73(4): Weidmann, Nils B. and Michael D. Ward Predicting Conflict in Space and Time. Journal of Conflict Resolution 54(6): Weinstein, Jeremy Inside Rebellion. New York: Cambridge University Press. Wucherpfennig, Julian, Nils B. Weidman, Luc Girardin, Lars-Erik Cederman and Andreas Wimmer Politically Relevant Ethnic Groups across Space and Time: Introducing the GeoEPR Dataset. Conflict Management and Peace Science 28(5): Zhukov, Yuri M Roads and the Diffusion of Insurgent Violence: The Logistics of Conflict in Russia s North Caucasus. Political Geography 31(3):