Incentives for Restraint

A situation of repeated interaction can facilitate collective action if the future discounted returns to cooperation sufficiently outweigh the gains that can be achieved from defection in a given period. For civilians living in subsistence economies with low growth rates, local differences in productivity can determine whether a cooperative arrangement that maintains peasant production along with a regular transfer of surplus is sufficiently attractive to an armed group that it will refrain from abusive behavior. The condition for restraint is that the returns to civilian labor are such that there exists a surplus of income, over and above civilian subsistence requirements, that makes it worthwhile to refrain from total extraction, in anticipation of future gains from extraction. Such conditions, we expect, are more easily met in wealthy areas. This logic gives rise to our first hypothesis:

The logic is somewhat altered when there are nonpecuniary costs (or benefits) to abusive behavior. In environments where coercive behavior is costly, cooperative relations may be maintained even if the discounted future returns to restraint do not outweigh the immediate material returns that can be achieved through abuse. This intuition is consistent with a rich literature in political science that emphasizes how dense community ties ease the resolution of collective action problems and allow for the sanctioning of misbehavior. Classical studies of revolution identified the autonomy of peasant communities and the strength of horizontal networks as necessary conditions for mobilization (Moore 1966; Skocpol 1979). A more recent variant identifies three key aspects of community as essential for helping groups to overcome collective action problems (Taylor 1988). Individuals embedded in dense social networks experience repeated interaction with one another, they enjoy greater facility in monitoring the behavior of others (Fafchamps and Minten 2002; Ghosh and Ray 1996), and they can employ a rich array of positive and negative sanctions to police contributions. Taken together, community ties create a situation in which conditional cooperation is possible and generalized reciprocity emerges as a norm. Individual members of fighting units thus may face higher costs if they employ abusive tactics in their own communities. These considerations give rise to our second hypothesis:

The role of community ties in achieving collective action has also been emphasized in the study of mobilization and cooperation within ethnic groups (Fearon and Laitin 1996). Ethnic identities often link individuals to a host of informal institutions and networks that may facilitate trust by promoting the flow of information about reputations, enabling sanctioning, and generating expectations that cooperative overtures will be reciprocated (Ostrom and Ahn 2002). This generates our third hypothesis:

Again, we note that these two hypotheses are not uncontested. Azam (2006), for example, argues that warlords have particular incentives to wreak havoc in their own communities. By making it costly for civilians to produce, it becomes more attractive for them to join the warlord's armed group as combatants. The abuse of one's own community depresses the wages that have to be paid to fighters. Other logics link violence to even more micro-level processes: violence by armed groups may be oriented largely toward settling private scores (Kalyvas 2003), a dynamic that may be more likely to occur within communities or within ethnic groups than across them.

Contestation

Whether the benefits to restraint are sufficient to translate into less abusive behavior depends on a set of between-group strategic considerations that arise whenever authority over territory is fragmented into zones of insurgent control, government control, and areas where control is contested (Tilly 1978). Although there are many different ways of interpreting the notion of control, we consider only two.

Consider first a situation in which in each period groups of different sizes encounter citizens in the same zone probabilistically. Here control can be taken as a measure of the likelihood with which a civilian encounters a given group. In such cases, the greater the control a group has over a given territory, the more confident its leaders can be that they (and not others) will benefit in future periods from restraint they exercise today. Groups that control a particular territory can expect to benefit from discipline, even in situations where smaller groups elect to engage in abusive actions in the same area. Greater levels of control then are associated with lower levels of abuse. This argument, we note, is consistent with a number of theories in the literature. Kalyvas (2006) suggests that by protecting individuals from other warring parties, by giving rise to a socialization process, by rendering threats of punishment credible, and by facilitating monitoring and the collection of information, uncontested sovereignty reduces the need to resort to coercion and abuse. Hultman (2005) offers a different argument that also produces predictions consistent with ours: where and when they are weak, abusive tactics, she suggests, are used by rebel groups to send signals of their resolve, thereby improving their position at the negotiating table.

Consider next a situation in which civilians may encounter multiple groups in a given period, but possibly with different frequencies for different groups. Here the relative frequency of contact with different groups can be taken as a measure of their control. In such cases, and unlike in the previous case, a cooperative arrangement depends on the ability of groups to engage in implicit collusion. For collusion to be compatible with the incentives of fighters from different groups, however, there must be a sufficiently large margin between the quantity of extractable resources and the subsistence requirements of civilians such that each group, after taking some share for themselves, can still leave enough on the table that future groups have an incentive to refrain from abusive behavior. Sustaining cooperative relationships with civilians under such conditions becomes more difficult as the number of groups increases. This feature of competition and collusion in coercive environment is somewhat surprising if compared with standard economic logics of competition. Whereas in market contexts, more competition among potential consumers for access to a producer's output leads to greater returns for producers, the logic is reversed in coercive environments. When armed groups can apply violence in turn, competition among them limits rather than enhances the scope for cooperative agreements with civilians. As developed here, the logic implies a form of tacit coordination across rival groups, a phenomenon reported often in accounts of “new wars” (Kaldor 1999). These arguments suggest a clear relationship between the degree of control a given group exercises over territory and its propensity to engage in abusive behavior and give rise to our fourth hypothesis:

But as before, this hypothesis is contested. Other arguments that emerge from models that allow greater agency to civilians than does ours suggest that levels of violence may be particularly low when control is contested. Gates (2002), for example, argues that rebel groups do not attempt to recruit soldiers from contested zones; because such recruits can more easily defect to the other side, they become too expensive to maintain. In a study of drug gang behavior, Levitt and Venkatesh (2000) find that gangs treat civilian populations better in periods of high contestation (by pricing drugs below cost) in a bid to bind consumers to their side. The logic, supported by theory in industrial organization (Klemperer 1995), depends on the freedom of civilians to choose the side which they want to benefit. Finally, for the special case of selective violence, Kalyvas (2006) examines a game in which civilians can use denunciation strategies to provide armed groups with the information they need to punish defectors. In a logic similar to that underpinning mutually assured destruction during the Cold War, he argues that lower levels of selective violence are observed in the most hotly contested zones because civilians, fearing retribution, do not provide information to fighting groups that can be used as a basis for selective violence.

Internal Structures of the Factions

Within-group collective action failures may also contribute to abusiveness. If the social rewards to restraint (for members of a fighting group) outweigh the private gains from abuse, then levels of abusive behavior committed by a unit will depend on the ability of the group's members to resolve this collective action problem. In such cases, a unit's aggregate returns depend on the group's ability to ensure that individual members do not overemploy violence for their individual gain.

The logic of the argument is similar to that supporting our claim about the advantages of monopoly power. We argued earlier that tacit collusion across groups can result in cooperative arrangements, but only if the number of groups is not too large. By the same logic, internally divided groups in which individual fighters extract benefits for themselves can collectively achieve cooperative relationships, but only if their numbers are sufficiently small relative to the level of wealth of noncombatants. In contrast, if a group is cohesive, in the sense that it operates like a unitary actor, maximizing the sum of its members' utilities, then nonabusive behavior can be optimal even in situations in which tacit collusion cannot be achieved by noncohesive groups.

In practice, a diversity of group characteristics and formal structures may affect the ability of a group to coordinate and police the actions of its membership. Examples include the existence of common goals, preexisting social networks, ethnic ties, and formal codes of discipline (Gates 2002; Sambanis 2001). Miller (1992) emphasizes the importance of common goals—whose achievement depends on the success of collective action rather than on the pursuit of private material gains—in explaining variation in the success of private organizations. Case studies of effective military organizations where ideology is paramount often identify similar group characteristics as essential for motivating and sustaining participation (e.g., Elliot 2002). However, across civil wars, there can be considerable variation in the extent to which fighters are motivated by collective rather than private goals. This logic motivates our fifth hypothesis:

Other arguments focus less on goals and more on the role of community ties (Taylor 1988). Insurgent groups will have more effective systems to police defection if they are built on the foundation of preexisting social networks or powerful ethnic ties that facilitate monitoring and punishment (Weinstein forthcoming; Wood 2004). Analogous to our second and third hypotheses, we then have:

Finally, theorists of collective action have identified formal, internal mechanisms that promote compliance and control as essential for ensuring group solidarity including clear rules, procedures, and avenues for punishing indiscipline (Hechter 1987). Such considerations motivate our final hypothesis:

The logic linking abusiveness to within-group cohesion depends on two critical assumptions. One relates to the returns to violence: our model assumes that abusive behavior produces private rewards, but at a social cost to the group. The literature on genocide and mass killing offers a compelling alternative view. Valentino (2004), for example, argues that mass killing is an intentional tactic employed by political and military leaders in pursuit of their strategic objectives. In such cases, the benefits to abuse may accrue to the group rather than to the individual. Indeed, if it is also the case that the cost of committing abuse is applied to the individual rather than to the group, the logic of our argument is reversed: mechanisms of discipline and group solidarity would then enable, rather than restrain, the abuse of noncombatant populations. From this perspective, an organization that functions effectively in pursuit of its overall objectives can thus serve to increase abusiveness if its leaders are motivated by a desire to kill or destroy particular groups of civilians. The second assumption relates to the nature and goals of a fighting group. Our logic assumes that if a group is cohesive, then its goals will reflect the interests of its individual members. This assumption—which can be derived, for example, from a rational actor model in which players have full information and quasilinear utility—may be difficult to maintain in a context where leaders have considerable control but are ignorant or deaf to the interests of the rank and file. In such cases, the fact that restraint may collectively benefit the members of a cohesive group may not be sufficient to induce restraint.

Violence in Sierra Leone

The conflict in Sierra Leone was known around the world for the vicious treatment of its civilian population by combatant factions (for a recent history of the war in Sierra Leone, see Keen 2005). Human Rights Watch issued a series of reports describing these atrocities: sexual violence against women, the forcible recruitment of child soldiers, and campaigns of killing, amputation, and pillaging perpetrated by the different groups (Human Rights Watch 1998, 2001, 2003). Although estimates range widely, Human Rights Watch asserts that the war left over 50,000 civilians dead and more than 50% of the population displaced from their homes. The horrors of the war have been widely reported in international media, and shock at the extent of the massacres, amputation, and pillaging led to the organization of an international war crimes tribunal based in Freetown.

But despite the uniform image of abuse described in journalistic accounts, there was significant variation in how civilians were treated in different parts of the country. Rates of death and displacement varied dramatically, with displacement rates close to zero in some chiefdoms and rising to 80% in others. Even before the massive displacements of the late 1990s, there was significant variation in how civilians experienced the conflict, both across factions and within the warring groups in different geographic zones. The most evident differences in violence are across the five factions. The Revolutionary United Front or RUF (the main insurgent group), the Armed Forces Revolutionary Council (AFRC), a military junta, and the smaller West Side Boys (WSB) group, have been associated with the highest levels of abuse. The Sierra Leone Army (SLA) is often believed to be responsible for lower levels of abuse, although in many areas their behavior was considered indistinguishable from that of the rebels they were fighting against. The Civil Defense Forces (CDF), an offensive paramilitary force significantly larger in number than the RUF, was also responsible for some violations throughout the country, but is commonly associated with significantly less abusive behavior.

But it is unsatisfactory to account for variation in the abuse of civilians simply by saying that the CDF acted one way while the RUF acted another. At best, this is description not explanation. As an account of the war, it is particularly unsatisfactory because the demographics of the membership of the RUF and CDF in particular, including age, ethnicity, regional origin, languages spoken, and livelihood, are not all that different (Humphreys and Weinstein 2004). In our analysis, we look inside the factions—to the level of the fighting units—to uncover the determinants of abusive behavior.

The Data

Analyses of violence in civil war typically draw on accounts provided by a small number of key informants. Often, this approach generates rich insights; in many cases, however, the representativeness of the information gathered can be easily called into question. We sought to collect more systematic information about the structures, tactics, and strategies of the warring factions using a nationally representative sample of ex-combatants. The survey was conducted between June and August 2003, slightly more than a year after the war came to an end. The study targeted a sample of 1,000 ex-combatants; a total of 1,043 surveys of ex-combatants were completed. The main method for gathering information was through the administration of a closed-ended questionnaire by an enumerator in the respondent's local language. Interviews were conducted at training program sites and in community centers around the country.³

To ensure as unbiased a sample as possible, the survey employed a number of levels of randomization. First, teams enumerated surveys in geographic locations and chiefdoms that were randomly selected, such that each combatant had an equal chance of being selected. Estimates of the population of ex-combatants presently residing in the chiefdoms were made based on data from the National Commission on Demobilization, Disarmament, and Reintegration (NCDDR) and the National Statistics Office. The estimates of the population distribution were used to draw 63 clusters of 17 subjects throughout the country, with each cluster drawn with an equal probability. These clusters fell within 45 chiefdoms or urban localities, and these 45 localities formed the basic enumeration unit.⁴

Within each enumeration unit, sites were also randomly selected, with both urban and rural areas represented. For each enumeration unit, specific numerical targets were set for the major factions (including gender and age distributions), based on the randomization and the estimated national distribution of faction members. Within each enumeration unit, enumerators worked through both official (UN and government) contacts and local community leaders to develop lists of ex-combatants. Teams identified pools of candidates from more than one source: some from the town or village Chief, some from the village youth coordinator, some from various DDR and NCDDR skills training centers, and so on. In every case, the teams aimed to identify two to three times the targeted number of potential respondents and then to randomly select respondents using a variety of methods. In most instances, Chiefs and DDR staff asked a number of ex-combatants to meet at a public location and teams selected candidates randomly from that pool (by choosing every third person or selecting numbers from a hat). Although this method worked well in most parts of the country, in some areas less than twice the target population was identified. This challenge tended to arise in remote rural areas, enumeration units with small ex-combatant populations, and regions where communities remain highly polarized.

Capturing Variation in the Course of the Conflict

The war in Sierra Leone was long and complex. It lasted for over a decade and involved five primary factions, numerous subfactions, and various external actors. Over the course of the conflict, the government changed hands four times and two peace accords were negotiated and failed. Individual experiences of the fighting were also complex. Some ex-combatants were involved in the conflict for short periods of time, whereas others entered early in the conflict and stayed to the end. Some changed subfaction or primary faction during the conflict, and almost all moved locations.

The survey asked detailed questions about characteristics of armed groups that varied among time periods, across factions, and in different locations. Asking questions about these aspects without making explicit reference to time periods would yield a set of “average” answers that would mask the temporal and geographic variation in the conflict. To ensure that the survey collected accurate information from specific time periods with particular factions, we developed randomization protocols within the survey.

Respondents were asked to map their involvement in the conflict by giving their location and faction membership for seven designated time periods, marked by major events in the history of the conflict. For each respondent, the survey recorded the number of periods in which the respondent was active and the enumerator selected one of these periods of activity using a randomization protocol. Enumerators were trained to remind the respondent throughout the survey that they were to answer questions about the specific time period selected by the randomization protocol.

The Unit of Analysis

A critical issue in evaluating various theories of violence is determining the appropriate unit of analysis. Data on the specific actions of individual combatants are sensitive and thus difficult to gather; as a result, the survey does not record an individual's acts of violence. Rather, it collects information about how the fighters in a respondent's unit behaved on a day-to-day basis. To assess the behavior of units within the fighting factions, we use detailed information gathered about the location of fighters at different points in time, as well as their factional and subfactional affiliations, to construct what we call “quasi-units.”⁵

Two types of quasi-units were constructed. In type FCH quasi-units, we group subjects from the same faction (F) and the same chiefdom (CH) at a given point in time. Type SFD quasi-units are based on more finely grained factional information but more general geographic information—grouping together individuals from the same subfaction (SF) operating in the same district (D). Typically, both methods result in groupings of between 5 and 15 respondents in a given quasi-unit at a given point in time. The existence of two distinct (although not independent) definitions of units allows us to check for the dependence of our results on any particular characterization of a unit.

With quasi-units as our unit of analysis, we generated variables to describe the characteristics of each quasi-unit based either on demographic information of the full set of members of the quasi-unit, or on statements made by individuals describing their own unit's behavior for the time period, faction, and region used to define membership in the quasi-unit.⁶

Measuring the Extent of Civilian Abuse

The dependent variable is an index of everyday policies and practices employed by groups that reflect the extent to which these groups engaged in abusive or cooperative relations with civilian populations. Recall that in the context of our model, the decision to abuse is analogous to a choice between engaging with civilians in a way that does not adversely affect their productive capacity, or, predating on them to the extent that they flee or their ability to produce is otherwise destroyed. In bringing theory to data, we interpret the behaviors that correspond to these choices broadly, identifying features that capture restraint in relations with civilian populations both by the absence of permissive norms—such as the arbitrary use of sexual as well as nonsexual violence—and by the presence of policies that may be considered minimally supportive of communities. This broad interpretation, although consistent with our model, treats different forms of abuse as if they follow similar logics. We recognize that this approach may be contested, and therefore disaggregate our dependent variable in a subsequent section.

The index is created using factor analysis of respondents' answers to eight related questions in which they are asked to describe patterns of interaction with non-combatants (for question wording, see Appendix). The weights derived from the factor analysis were then used to create a single measure, the extent of civilian abuse, which ranges from 0 to 1. The measures used to construct the index include three distinct types of questions. First, we include questions about the ways in which food was collected, including whether food was taken forcibly from civilians, whether it could be collected peacefully on demand, and whether a system was in place whereby food would be delivered regularly in fixed amounts. Second, we add responses to questions that assess the likelihood that an individual in a fighting unit would be punished for stealing, amputating, or raping a civilian.⁷

Besides our theoretical rationale for thinking about abusiveness as a broader set of soldier–civilian interactions, we elected not to ask direct questions about violence because a war crimes tribunal and truth and reconciliation commission were just beginning their work. Although the tribunal promised to hold accountable only those who bore the greatest responsibility for atrocities, many former soldiers worried about the prospect of punitive action for their past behavior. Asking about the actions of specific individuals would have put our survey teams in a difficult position: respondents would have been reluctant to participate, incentives to lie would have been higher, and communities might have confused our work with investigative activities of the tribunal. So unlike event data, our measure is not intended to capture the magnitude or frequency of abuse—these depend on both the numbers of fighters and the number of civilians present in any location as well as the efficiency of organizations—but rather on the manner in which individual units interact with civilians, conditional on such interactions taking place.

The dependent variable displays substantial variation, both over time and across space (summary statistics for this and for all other variables used in the analysis are provided in Table 4). Average levels of abuse exhibit a gradual decline over the course of the conflict, with a slight rise following the AFRC/RUF attack on Freetown in 1998. However, there are sharp differences in temporal patterns across factions. The CDF employed less abusive strategies than the RUF overall, but its abusiveness remained fairly consistent over time even as its relative size increased. The RUF, on the other hand, after an initial spike at the start of the war and subsequent fall, exhibited a gradual increase in abusiveness as the war progressed after 1992.

Figure 1 displays the geographical variation of the index for each of the two major factions in the conflict (the RUF and the CDF). The figure on the left, which displays patterns of abusive behavior by the RUF, shows relatively low levels of abuse in the southwest of the country—where the conflict began—and in the diamond areas around Kono, and very high levels of abuse in the regions north of Freetown, from where the attacks on the capital were launched by a joint AFRC/RUF force. By contrast, the figure on the right, representing the behavior of the CDF, shows lower levels of abuse overall, with levels at their peak in the capital, Freetown, and in highly contested regions south of the capital.

Spatial Variation in Civilian Abuse

Incentives for Restraint

We first examine the hypothesis that poorer regions suffer higher levels of abuse. To test this hypothesis, we employ district-level data on poverty levels in Sierra Leone that were collected by the Sierra Leone Central Statistics Office. Extracted from the 1989 Household Survey, completed before the war, our measure of Poverty represents the share of the population in each district living below a nationally defined poverty line. In the bivariate analysis, we find the expected positive correlation between poverty and abuse, but for neither type of quasi-unit is the relationship statistically significant. No discernible relationship exists either once fixed effects are introduced or once we control for other variables. We examine some possible reasons for these nonfindings in the concluding section.

The second hypothesis suggests that a group's abusiveness is negatively related to the strength of the preexisting ties that link combatants to community members. These ties, we argued, could facilitate voluntary exchange and make abusive behavior more costly. The survey instrument recorded the respondents' chiefdoms of origin as well as areas of operation; we constructed a variable Home, for each quasi-unit, indicating the proportion of combatants who originated from the region in which the group was operating. The left hand panel of Figure 2 reveals a statistically significant negative correlation between the Home effect and levels of abuse. The size of the circles in the graph indicates the number of members of each quasi-unit and serves as a weight that reflects our confidence in the estimates of the values calculated for each variable at the level of the quasi-unit.

Levels of Abuse When Units Fight in Their Own Neighborhood

Although the bivariate relationship is a strong one, the relationship between Home and levels of abuse is driven in large part by systematic differences between the two main factions, the RUF and CDF. As the right-hand panel of Figure 2 illustrates, CDF fighters were more likely (on average) to operate from a home base and also were less likely (on average) to commit acts of abuse against civilian populations. If the overall relationship is driven by this major division between the CDF and the RUF then, the strength of the bivariate relationship notwithstanding, being at home may be irrelevant for understanding patterns of abuse. Other features of the CDF, such as differences in command structure, ideology, or leadership, could be responsible for the observed differences in combatant behavior.

Indeed, as the second panel of Figure 2 demonstrates, the negative relationship is substantively weaker across quasi-units within a given faction and is largely driven by between faction variation.⁹

The third hypothesis focuses on shared ethnic identities as a constraint on abusive behavior, or alternatively, as a mechanism for generating collaboration without coercion. In the survey instrument, respondents were asked to report the principal ethnic group of their faction (if any) and the principal ethnicity of the civilians living near their base. To test this hypothesis, we created a dummy variable, Co-ethnicity, taking a value of 1 if these two groups are the same and a 0 otherwise. The bivariate analysis reveals a strong relationship between co-ethnicity and levels of abuse. Units that share ethnic ties with local civilians exhibit abuse levels approximately one standard deviation lower than those that do not. However, this characteristic of solider–civilian interaction is much more common in CDF units than in the RUF. As a result, when we try to discern an effect of co-ethnicity independent of faction-fixed effects, we find no evidence for this hypothesis. In the multivariate analysis, we again fail to reject the null of no effect when we control for the other explanatory variables. Thus support for the influence of local community ties, whether regional or ethnic, is weak. Although both relationships have face validity, once we account for other characteristics of factions, local community ties offer no additional explanatory power.

Contestation

Hypothesis 4 suggests that abusiveness is negatively related to the extent of a unit's control. Where one group is dominant, lower levels of abuse should be apparent. One challenge in testing this relationship involves defining “dominance,” because control is likely to depend on many unobservable features of the quasi-units as well as the prize they are fighting for. We adopt a simple and transparent approach; we measure the relative number of troops present in a given locality. By tracking the movements of a representative sample of fighters, we have good estimates of troop levels in the chiefdoms throughout the war. We develop a measure for each quasi-unit of the extent to which their group is dominant in a given area. The measure of Dominance records the estimated size of the quasi-unit relative to the estimated total number of troops in the zone.

For the more geographically precise quasi-unit—FCH—the bivariate relationship between dominance and the measure of abuse is strong and in the expected direction. The simple bivariate relationship is not statistically significant, however, for the less geographically precise measure. Strikingly, there is some evidence of the relationship in the fixed-effects bivariate analysis: it is observed in the less precise measure and enters just short of significance at the 10% level for the more precise measure, although in both cases the estimated magnitude of the coefficient is greatly reduced. But, as with the measures of community ties, we find that the results disappear in the multivariate context. The coefficient on Dominance is insignificant in each of the eight models presented in Table 2. It turns out that many of the cases of dominance and low levels of abuse occur in CDF factions, and their behavior can be accounted for by other characteristics shared by CDF units.

Internal Structures of the Factions

A final perspective focuses on the mechanisms factions can employ to punish defection within the group. It suggests that abuse is more likely when groups lack the tools they need to prevent individual combatants from committing abuses that are costly to the group. Our fifth hypothesis relates to recruitment practices: warring factions that recruit combatants with the promise of private benefits are more likely to exhibit high levels of civilian abuse. To test the hypothesis, we collected data on the incentives used to induce people to take part in a faction. Although many combatants, particularly within the RUF, were abducted and kept in through the use of force, private material enticements were also used. Using reports of promises of money and diamonds made to individual combatants, we constructed a measure of the propensity of each quasi-unit to be staffed with combatants that were recruited via Material Incentives.

The bivariate relationship between material incentives and our measure of abuse is large and significant. A one-standard deviation shift in the measure of material incentives is associated with an increase in abusiveness of approximately one full standard deviation. This effect is, in part, the result of differences across factions: the CDF were less likely to make offers of material gains to potential recruits. But the results with fixed effects demonstrate that this relationship also obtains within factions, a result driven particularly by variation across units within the RUF. In the multivariate analysis, the coefficient on material incentives is statistically significant across six of our eight models; significance is lost only in the models that include our measure of internal ethnic fragmentation.

A second aspect of the internal structure of groups is the Density of Social Ties within the unit. Hypothesis six suggests that groups with strong internal networks will be less likely to commit abuses against civilians because those social ties can be employed to prevent defection. To create a measure of social ties within the factions, we created an average, weighted by factor loadings, of three measures of social connectivity in the unit: the share of individuals in each quasi-unit that had friends in the unit at the time of entering, had family in the unit, or knew fellow community members who were participating. Factions that recruited by using force or by offering selective incentives are less likely to integrate or make use of preexisting network structures. As a result, they are also less likely to benefit from these networks as they attempt to resolve various collective action problems. We find a strong, negative relationship between the density of social ties within quasi-units and the level of abusiveness. However, this is explained mainly by the fact that recruits to the CDF were more likely to join factions in which they already had friends, family, or community members participating. In the multivariate analysis, we find no support for this measure of social ties.

Ethnic ties within the unit, represent another measure of the cohesiveness of a fighting group. The seventh hypothesis posits that ethnically heterogeneous units should exhibit lower levels of civilian abuse. To measure the heterogeneity of units, we created a standard measure of Ethnolinguistic Fragmentation. For every quasi-unit that had at least three members, we estimated the relative size of each ethnic group within the quasi-unit and then took the sum of squares of these shares. This provides a concentration index. The difference between this index and 1 is the index of fragmentation. The index takes a value of 0 if all members are from the same ethnic group, a value close to 1 if all members are from different ethnic groups, and values between 0 and 1 for less extreme structures. The distribution of the fragmentation index for each of the two major factions is striking. Although the CDF as a whole is no less heterogeneous than the RUF, its quasi-units were significantly more homogeneous. CDF fighting units appear to have formed along ethnic and regional lines; RUF units reflect the heterogeneity of the overall faction. The bivariate relationship between our measure of fragmentation and the abuse index is strong and positive, both without and with faction fixed effects. This relationship is also robust to the inclusion of additional explanatory variables in the multivariate analysis. There is also an effect on the variance of outcomes—less diverse units exhibit a predictably low level of abuse, with very few exceptions among the highly homogenous groupings. If we disaggregate, we find that, although much of the effect of ethnic fragmentation is accounted for by differences between the ways in which the CDF and the RUF organized their units, there is a strong, statistically significant effect within the RUF grouping alone. A one-standard deviation shift in RUF fragmentation levels increases the abuse rate by about 0.1 points, and this accounts for more than 20% of the variation between RUF units.

The final hypothesis proposes a positive relationship between a group's structures to maintain internal discipline and the degree to which it abuses civilians. To create a measure of Internal Discipline, we weighted a series of responses to questions about reported norms of behavior within the factions. We used information about the extent to which individuals would face disciplinary action, if, without the consent of a commander, they were drunk at the base, killed someone from their own group during combat, were drunk while in combat, killed someone from their own group at the base, stole from someone in their own unit, amputated someone in their own unit, and raped someone in their own unit. None of these measures of disciplinary action within units involves any information that relates directly to interaction with civilians. So, in theory, the measure is analytically (although not necessarily empirically) independent of our measure of civilian abuse.¹⁰

The bivariate relationship between the index of internal discipline and our measure of abuse is strong and exists across both major factions. It indicates that the higher the level of indiscipline permitted inside the faction, the greater the abuses that were allowed outside of the faction. The relationship survives the inclusion of faction-fixed effects and emerges as a robust, significant predictor of abusiveness in all models presented in the multivariate analysis. These results suggest that internal factional attributes including the characteristics of a group's membership, how they were recruited, and how they relate to one another are key factors that help to explain variation in levels of abuse in the Sierra Leone conflict.

Robustness Tests

In this section, we explore the robustness of our findings by examining the extent to which our results depend on how the dependent variable is operationalized. In particular, we address three distinct concerns relating to the construction of our measure of abuse. The results of our investigation are presented in Table 3. Column I in Table 3 reproduces columns 2 and 4 from Table 2 as a baseline.

The first concern relates to the fact that the measure of abuse combines assessments of the presence of abusive behavior with assessments of the absence of constructive relations with noncombatants. In column II, we replicate the baseline with a version of the dependent variable that excludes all measures of positive, nonabusive behavior.¹¹

A third concern is that one key independent variable—the measure of a unit's internal discipline—may be too closely related in the minds of our respondents to a component of the dependent variable, the use of disciplinary measures in response to abusive behavior towards civilians. In column V, we exclude all disciplinary measures from the index of abuse. Consistent with the results in columns II, III, and IV, our core findings related to in-group discipline are essentially unchanged, suggesting that the relationship between cohesion and abuse cannot simply be attributed to a conflation of group practices on the part of respondents. The relationship between material incentives and this measure of abuse is, however, somewhat weaker and fails to reach significance in one of the two models. One additional difference is noteworthy in the final model: the coefficient on dominance enters significantly in the equation employing FCH quasi-units; however, the effect is substantively small.