THE TIES THAT BIND COMMANDERS

(1)

THE TIES THAT BIND COMMANDERS

A STUDY OF SEXUAL VIOLENCE AND RESTRAINT

BY REBEL GROUPS IN AFRICA

HERMAN WIESELGREN

Master's Thesis in Peace and Conflict Studies

Spring 2020

Department of Peace and Conflict Research, Uppsala University

Supervisor: Lisa Hultman

(2)

ii

ABSTRACT

An emerging literature on rebel groups’ restraint in the use of sexual violence has begun to explore the role of the commander. This study seeks to build on this literature by investigating under what conditions rebel commanders will enforce restraint. I argue that perpetrating sexual violence against civilian communities to which a rebel group has ethnic ties is highly costly for the rebel commander. Thus, commanders will strive to enforce restraint in the use of sexual violence against co-ethnic civilians. However, their ability to enforce restraint, I argue, hinges on the authority and influence they have over their combatants. Hence, I hypothesize that rebel groups with strong commanders are more likely to enforce restraint toward co-ethnic civilians. This argument is examined through unique spatial data on the geographic patterns of rebel sexual violence combined with spatial data on rebel groups’ co-ethnic civilian populations in Africa 1989-2009. Regression analysis finds that rebel groups with strong commanders and ethnic ties are less likely to perpetrate sexual violence against co-ethnics, yet also that groups with weak commanders are more likely to victimize co-ethnics.

(3)

iii

ACKNOWLEDGEMENTS

(4)

iv

Over the past two decades, the field on sexual violence in armed conflict has made significant strides. Scholarly research has identified a wide range of possible explanatory factors for why combatants in rebel groups perpetrate sexual violence. Some literature has taken to examining individual and group motives for rebel combatants to perpetrate sexual violence, i.e. a bottom-up perspective of combatant motives. Other literature has focused on motives for rebel grobottom-ups to deliberately engage in sexual violence, i.e. a top-down perspective of the commander motives. Concordantly, scholarly research has up until recently principally theorized about factors which have a direct effect on sexual violence prevalence. However, an emerging literature on commander restraint has added an important dimension by investigating the intersection between the commander and combatant level.

In an influential piece on the typology of rape and sexual violence, Elizabeth Wood (2018) argues that combatant motivations are conditioned by the stance of commanders. Depending on the stance and the ability of the commander to enforce that stance, combatants may conform to the stance of their commander. If commanders have a preference against sexual violence and are able to enforce it, combatants will conform to their stance and prevalence will be low. Thus, combatant-level motives for sexual violence can be superseded by commander-level motives against sexual violence, given that the commander is able to enforce restraint. This gives cause for examining what motives there are for commanders to refrain from and restrain sexual violence, and under what conditions commanders are able to enforce restraint. A few studies have investigated why and how commanders of armed organizations are able to enforce prohibition (Hoover Green 2016; Muvumba Sellström 2019). Yet, there is little literature on what factors motivate commanders to enforce restraint, i.e. why they would refrain from ordering sexual violence and prohibit their combatants from perpetrating it. Whitaker, Walsh and Conrad (2019) and Elizabeth Wood (2009) theorize that dependency on civilians may motivate rebel groups refrain from perpetrating sexual violence. These studies lay at the foundation of this study as they prompt questions of what role the commander has in enforcing restraint among his combatants, what factors may condition rebels’ restraint, and how patterns of sexual violence may be shaped by the identity of potential victims. Thus, this study is guided by the research question:

(6)

2

To address the research question, I draw on literature from the field of ethnicity and one-sided violence. I argue that ethnicity and the social bonds between civilian members of the same ethnic group impacts the outcome of cost-benefit analyses of rebel commanders. I argue that ordering, authorizing, or tolerating the perpetration of sexual violence against co-ethnics would be disruptive of the social embeddedness of the rebel group. Disruption of the social embeddedness in turn reduces the civilians’ willingness to voluntarily support the rebel cause, a highly costly outcome for the commander. Therefore, the commander should wish to refrain from ordering sexual violence and prohibit any individually motivated sexual violence by rebel combatants. Under the condition that the commander has a strong command over his combatants, he will enforce restraint among his combatants in the use of sexual violence. Thus, I hypothesize that rebel groups with strong commanders are more likely to enforce restraint in the use of sexual violence against co-ethnic civilians. Through connecting literature on civilian killings and literature on sexual violence, I contribute to our understanding of the impact of ethnic ties between rebels and civilians and how it can affect rebel behaviour.

I test the proposition through a series of binomial logistic regression models using spatially disaggregated data in Africa 1989-2009. Data on patterns of sexual violence is recoded from the Sexual Violence in Armed Conflict (SVAC) dataset and used to geographically pinpoint the locations of incidents of sexual violence. The findings of the study suggest that ethnicity plays a role in shaping rebel behaviour. Rebel groups with strong central command are found to be less likely to perpetrate sexual violence against co-ethnic civilians than against others. However, rebels with weak command are found to be more likely to perpetrate sexual violence against civilians with which they have ties. These findings prompt further examination into the interplay of ethnicity and civilian dependency in shaping rebel behaviour, as well as inspires the collection of geographic data on patterns of sexual violence.

(7)

3

2. PREVIOUS RESEARCH

2.1. Drivers of sexual violence

In the previous section, it was stated that previous research has identified key factors for the perpetration of sexual violence on both combatant- and commander-level. It was also stated that there have been advances into understanding the conditions under which commanders may strive and be able to enforce restraint among their combatants. The main arguments of these studies will be reviewed in detail in this section, first identifying key drivers of violence at the combatant-level and next drivers and conditions for restraint at the commander-level.

Explanations at the combatant-level tend to view sexual violence as a consequence of either a presence of gendered structures in armed conflict or some underlying preferences among combatants. Studies of the former viewpoint treat sexual violence as the result of masculine norms and ideals. Some argue that combatants’ failure to fulfil hyper-masculine ideals in an armed conflict setting makes combatants more prone to attempt to reassert their masculinity through perpetrating sexual violence (Eriksson Baaz and Stern 2009; 2013; Meger 2015). Others argue that gender inequality at a broader level within a society ties into combatants’ behaviour. In their view, systematic gendered oppression and discrimination prior to conflict normalizes and motivates extreme sexual violence against women in armed conflict. Gender inequality also reduces the reporting frequency of sexual violence, which acts as an enabling condition for perpetrators of sexual violence (Davies and True 2015).

Studies focused on the underlying preferences of combatants assume that some combatants’ preferences for sexual violence are shaped prior to participation in an armed organization. In his seminal work on rebel recruitment, Weinstein (2007) argues that resource-endowed rebel groups attract combatants who are more abusive to civilians. As resource-endowed groups will rely on selective incentives for participation, they are argued to attract higher numbers of opportunistic and rent-seeking recruits. These combatants are argued to be more prone to perpetrate violence against civilians in general, but authors such as Cohen (2013) have applied the theory to sexual violence specifically.

(8)

4

(9)

5

also factors at the commander-level which may reduce the incentives of allowing sexual violence by combatants. Thus, potential conditions and motives for restraint are reviewed next.

2.2. Conditions for restraint

Having identified the main factors which are argued drive sexual violence at the combatant- and commander-level, we turn to the existing scholarly research related to restraint. The recent piece by E. Wood (2018) provides a fundamental framework on how combatants and commanders collectively shape patterns of sexual violence. It should be noted that E. Wood’s study focuses on armed organizations in general and only on perpetration of rape, whereas this study is aimed at investigating only rebel groups as well as the broader category of sexual violence. However, the cost-benefit thinking of all armed organization commanders should be fairly similar to the thinking of rebel commanders. Furthermore, since sexual violence prevalence is largely driven by prevalence of rape (Cohen and Nordås 2014)1 the framework she proposes should work in the broader category of sexual violence as well.

E. Wood argues that combatant motivations are conditioned by the stance of commanders. Commanders are argued to assume either a stance of promotion, toleration or prohibition in relation to sexual violence. Where a commander promotes or prohibits sexual violence, individual or unit motives and preferences are argued to be largely irrelevant, as combatants will conform to the commander’s stance. In these cases, sexual violence prevalence will vary with the commander’s stance; prevalence being high with promotion and low with prohibition. In contrast, where a commander assumes a stance of toleration, a neutral stance, sexual violence prevalence will vary with combatant-level motives. Since the commander assumes a stance of no-action toward sexual violence, individual and unit motives will be the driving and unchecked force behind prevalence. However, E. Wood also recognizes that commanders may tolerate sexual violence which they would ideally wish to curtail because of an inability to prohibit it. That is, prohibition may be too costly in terms of organizational resources or the commander might not have the required control over his combatants.

Hence, E. Wood suggests that motives for or against sexual violence by commanders should supersede combatant motives, conditional on the commander’s ability to restrain. As mentioned in the introduction, some studies have examined this. Hoover Green (2016) discusses why some commanders are able to restrain their combatants and argues that sexual

1_{Approximately 90 percent (671/744) of prevalence observations in the SVAC 1.0 dataset include rape as a}

(10)

6

violence often occurs due to non-alignment of combatants’ and commanders’ preferences for various forms of violence. Formal political education is argued to align combatants’ preferences to those of their commanders, thereby lowering the frequency of unordered sexual violence. Johansson and Sarwari (2019) propose that a relative loss of power diminish the resources available to maintaining good relations with civilians and restraining combatants from perpetrating unordered sexual violence. Without mechanisms for restraint in effect, opportunistic elements of rebel groups are argued to be free to perpetrate sexual violence at their own will. Finally, Muvumba Sellström (2019) argues that rebel commanders may rely on stigmatization of sexual violence to prohibit their combatants from perpetrating sexual violence, collecting evidence from rebel groups in Burundi.

These studies all find that commander authority and influence over rebel combatants plays a significant part in his ability to prohibit sexual violence. Formal political education can be seen as a way of increasing commander influence over his combatants, by aligning the combatants’ goals to his. Reduced organizational resources for mechanisms of restraint limits the authority that the commander has in the eyes of his combatants. Stigmatization as a strategy of prohibiting sexual violence strengthens these theories, as implementing stigma is only feasible when the commander has influence over the combatants.

2.3. Motives for restraint

(11)

7

While the article by Whitaker, Walsh and Conrad (2019) is the first study fully devoted to rebels’ incentives for restraint in the use of sexual violence, some potential determinants of rebel restraint have been mentioned previously. For example, it has been proposed that commanders may wish to restrain sexual violence as it could undermine commander’s control over their combatants, alienate civilians they depend on or cause a clash with the commander’s ideological beliefs (E. Wood 2018). Dependency on civilians as a possible determinant resonates with the findings by Whitaker, Walsh and Conrad. A similar argument can also be found in an early piece by E. Wood (2009), in which she briefly mentions that armed groups who are dependent on civilian and expect to be for an extended time, are less likely to tolerate sexual violence. Dependence on civilians for intelligence, she argues, is especially important, as it is difficult to accrue through coercion long-term.

2.4. Research gap

Previous research recognizes the potential impact that commanders’ preferences for restraint can have on the perpetration of sexual violence by their combatants. Primarily it identifies dependency on civilians as a potential determinant of commander preference for restraint. Furthermore, the ability to restrain is central in shaping the perpetration of sexual violence. Commanders with strong authority and influence over their subordinate combatants are able to take a stand against sexual violence, should they wish to. The idea that commanders assume a stance on sexual violence based on their preferences and organizational ability to do so represents a new keystone in our understanding of sexual violence. Yet critical questions remain concerning how the commander and combatants interact and how patterns of sexual violence are affected by victim identity. Furthermore, commander ability to restrain as a potentially conditional factor to factors which shape preference has yet to be explored in previous research. This research gap motivates posing the question:

Under what conditions will rebel commanders enforce restraint in the use of sexual violence against civilians?

(12)

8

3. THEORY

3.1. Key theoretical concepts

Before presenting the theoretical argument of this study, theoretical definitions for the key concepts which will be used frequently should be given.

An armed conflict refers to, unless specifically stated, an internal contestation over territory or government between a state actor and a non-state actor which result in at least 25 battle-related deaths in a calendar year2. A rebel group is thus defined as the non-state actor involved in an internal armed conflict. Within rebel groups, a distinction is made between combatants and commanders. A combatant is a rank-and-file member of a rebel group; the individual who does the fighting and at most has operational freedom. Individuals in charge of squad or platoon-sized unit are thus also regarded combatants, as they should not make strategic decisions. A commander is an individual in the rebel group who commands a sizeable unit of combatants and has a certain degree of strategic freedom. There are layers of commanders within a rebel group, ranging from mid-level commanders in charge of company-sized units, to high-level commanders in charge of a division or branch.

Following the definition by E. Wood (2018, 515), sexual violence refers to “rape; sexual slavery and forced marriage; forced prostitution, pregnancy, and sterilization; sexual mutilation; and sexual torture”. Enforcing restraint in the use of sexual violence refers to a commander’s action of a) refraining from ordering or authorizing his combatants to perpetrate sexual violence, and b) prohibiting his combatants from perpetrating unordered sexual violence. The terms civilians and civilian population/community are used interchangeably to refer to unarmed populations which are not active participants in the armed conflict.

3.2. Ethnic ties and social embeddedness

(13)

9

wars” have been theorized to be particularly deadly for civilians, due to campaigns of ethnic cleansing aimed at weakening rival ethnic groups (Hägerdal 2019). The same has been argued in relation to sexual violence, ethnic cleavages motivating use of rape as a strategic weapon-of-war (Jaleel 2013). However, for the purposes of this study, previous findings which identify determinants of absence from violence against civilians and co-ethnicity are more relevant. Fjelde and Hultman (2014) find that rebel one-sided violence is more prevalent in areas inhabited by the adversary’s co-ethnics than in areas of one’s own co-ethnics. The scope of their argument encompasses all armed groups, but for the purposes of this study I focus on the implications for rebel groups. Firstly, they argue that ethnic ties between an armed group and a civilian population may act to facilitate civilian support to the armed group. Shared ethnicity between a rebel group and civilians can reduce mobilization costs and increase internal cohesion within the rebel group. Moreover, the pre-existing social networks between co-ethnic rebel groups and civilians can increase trust and a sense of solidarity, thus facilitating collaboration in the armed struggle. Hence, civilian support for armed groups should frequently occur along ethnic lines. Secondly, because of the strategic advantages of strong civilian support, they argue that rebel groups thus have incentives to target the adversary’s civilian support base in order to deplete their support. They argue that abuse of the adversary’s co-ethnics may deter civilians’ collaboration with the adversary and reduce the adversary’s legitimacy among their civilian constituents.

(14)

co-10

ethnic civilians from Fjelde and Hultman (2014). In the following sub-section, I suggest that rebels should also face strong incentives to enforce restraint in the use of sexual violence.

3.3. The costs of sexual violence

(15)

11

and humiliates the community of the victim. Hence, the pain and suffering caused by sexual violence should cause a similar disruption of rebels’ social embeddedness as killings.

Therefore, I argue that rebel commanders face high costs of sexual violence against co-ethnic civilians. When faced with these costs, commanders will have strong incentives for enforcing restraint. The first step of the commander would be to refrain from ordering or authorizing sexual violence against co-ethnics. The feasibility of refraining from ordering or authorizing sexual violence should not hinge on the strength of the commander, since the perpetration of sexual violence is driven by commander-level motives. Of course, the commander will not always choose to refrain, since the strategic benefits of sexual violence may sometimes outweigh the costs of losing civilian support3. But considering the high value of civilian support, the strategic benefits of committing to sexual violence against co-ethnics would have to be considerable. Furthermore, the implementation of restraint would not need apply to all sexual violence, only violence directed at co-ethnics. For example, if rape is authorized as a substitute for pay or sexual slavery orchestrated at the group level as a source of revenue, the commander could enforce restraint with regard to co-ethnics while ordering or authorizing sexual violence against other civilians. Thus, I derive the following hypothesis:

H1: Rebel groups are more likely to enforce restraint in the use of sexual violence against civilian communities to which they have ethnic ties.

The second step of the commander after refraining, would be to prohibit his combatants from perpetrating sexual violence. The feasibility of prohibition, however, should very much hinge on the strength of the commander. Perpetration of sexual violence in that case is driven by combatant-level motives, meaning that the preferences of combatants would have to be subdued or aligned with the commander’s preferences. As argued by E. Wood (2009), when combatant preferences are distinct from those of the armed group, the disciplinary capacity of the armed group will decide which prevails. Hence, I argue that for the cases where combatant-level motives are present, only commanders with authority and a strong influence over his combatants will effectively prohibit sexual violence against co-ethnic civilians. Thus, I derive the following hypothesis:

H2: Rebel groups with a strong commander are more likely to enforce restraint in the use of sexual violence against civilian communities to which they have ethnic ties.

3_{One such case could be sudden shifts in power balance, as theorized by Johansson and Sarwari (2019). As the}

(16)

12

Importantly, strong and weak commanders alike may still order, authorize and tolerate sexual violence against other civilians. The theoretical argument only assumes ethnic ties to exert an effect on sexual violence against the rebels’ co-ethnic civilians. No claim is made regarding the general behaviour or tendencies of rebel groups which have ethnic ties. Furthermore, ethnic ties to civilians are argued to increase the likelihood that commanders enforce restraint, rather than the extent to which they enforce restraint. I assume that the impact of sexual violence on disruption of social embeddedness is non-linear, i.e. that the effects of initial transgressions are higher than the transgressions that follow. Due to the shame and humiliation of sexual violence, co-ethnic civilians should be more likely to reject the rebels after the first few transgressions than after the hundredth. Thus, if the costs of sexual violence are the highest for initial occurrences of sexual violence, then there should be little incentives for half-hearted restraint. If rebels find the costs of sexual violence bearable compared to the strategic or politico-organizational benefits that are gained, then they should commit fully to policies of sexual violence. Concordantly, the hypotheses state that enforcement of restraint is more likely, rather than it is more frequent or more intense.

(17)

13

4. RESEARCH DESIGN

4.1. Scope and aim of analysis

The theoretical proposition put forth will be examined through a quantitative analysis. The argument suggests that rebel commanders’ perceived costs of sexual violence play a central role in determining the probability and targets of sexual violence in conflict. As such, the analysis aimed at examining this proposition must be performed at a more fine-grained level of analysis than country-level. Indeed, contemporary conflict research has a strong need for disaggregated data (Gleditsch, Metternich, and Ruggeri 2014; Raleigh et al. 2010) and currently there is a strong trend to fill that need (Fjelde, Hultman, and Nilsson 2019; Hegre et al. 2019). The use of disaggregated data in this study thus fits the needs and trend of current research, providing an empirical contribution. Ideally, individual-level data would be used in order to draw a connection between civilians’ ethnic identity and victimhood of sexual violence. However, the limited availability of global data on ethnicity of victims of sexual violence makes this impossible. Even at an aggregate conflict or country level, global data on ethnicity and sexual violence is only focused on selective violence (Cohen and Nordås 2014). Thus, I need to identify a common denominator between data on ethnicity and sexual violence which would allow me to measure both. While ethnic dimensions of data on sexual violence are not available, geographic dimensions frequently are. Additionally, data on the geographic patterns of ethnic groups is also available. This allows for measuring sexual violence and ethnicity through geographic location, thus identifying the common denominator required. Hence, this study relies on spatially disaggregated data on ethnic groups and sexual violence prevalence to measure the theorized relationship.

(18)

14

This section on research design will begin by presenting the structure of the data used for the quantitative analysis of this study. Next, it will specify the data and operationalization of the dependent and independent variables as well as outline the control variables used in the analysis. Finally, the method of estimation used in the empirical analysis will be presented along with motivations of methodological choices.

4.2. Structure of data

The scope of this study is limited to rebel groups in Africa during the time-period 1989 to 2009. The geographic and temporal parameters are determined by the availability of data on sexual violence at the location-level. The choice to subset the sample to one continent was made in order to make the workload of coding sexual violence manageable. Africa was selected due to the high variation that would generate on the dependent variable. 54 percent of sexual violence prevalence observations in the SVAC data occur in Africa, 68 percent for only sexual violence by rebel groups (Cohen and Nordås 2014). The time-period 1989-2009 is decided by the temporal scope of the data on sexual violence. Limiting the geographic scope of the study to Africa reduces the generalizability of the findings. Since the sample only includes Africa, it cannot be assumed that the findings hold outside of Africa.

(19)

15

they suggest that there may be a preference for reporting battle-related and civilian fatalities at the first-order level (Högbladh 2019).

As was mentioned in the previous sub-section, variation across actor, space and time is necessary to examine the relationship between ethnic ties to civilians and restraint in the use of sexual violence against them. The GADM data provides a structure for attaining variation across space for both ethnic ties and rebel sexual violence. Next, the dataset is expanded by rebel groups to account for variation of ethnic ties and sexual violence across actors. The final data contains all rebel groups in Africa during the time-period 1989-2009, amounting to a total of 116 rebel groups in 27 countries. The unit of analysis is rebel group – administrative unit – year and the total number of observations is 17,547 group-admin-years.

(20)

16

4.3. Data and operationalization of sexual violence

(21)

17

co-ethnic civilians. Through measuring sexual violence given certain circumstances, the empirical analysis will attempt to identify patterns which can provide evidence for or against the hypotheses. Thereby, the dependent variable of this analysis is a prevalence measure of rebel sexual violence.

In order to operationalize rebel sexual violence, two elements are needed: the prevalence and location of rebel sexual violence. To capture the prevalence of sexual violence, i.e. who perpetrated sexual violence and when, the three ordinal prevalence measures of the SVAC dataset are combined to a single binary variable. If either the US State Department, Amnesty International or Human Rights Watch have reported any level of sexual violence for a rebel group in a year, this binary variable is assigned a value of one. The use of a binary prevalence measure is motivated through three reasons. First, there is a substantial amount of missing data on the location of sexual violence and the analysis cannot afford discarding prevalence observations on the basis of non-triangulation. Second, the location variable from the SVAC dataset aggregates all the known locations for sexual violence perpetrated that year. As such, it does not contain data on the level of sexual violence in any specific location. Third, the theoretical argument posits that the ethnic ties affect the probability of sexual violence, so recording the intensity of sexual violence would be theoretically superfluous.

Next, to capture the location of sexual violence, the location text variable from the SVAC dataset is used. For the observations which have a value of one on the binary prevalence measure of sexual violence previously constructed, the location text variable is examined. It includes all known locations of sexual violence by a rebel group in the given year. These locations are derived from the country reports and may read, for example, “in detention”, “in the fields” or “in X district, X province”. For years in which a rebel group was reported to perpetrate sexual violence in multiple instances, the instances are delimited with a semicolon or a colon in the location text variable. Some group-year observations are completely blank and others contain information which does not yield any knowledge on the geographic location of sexual violence, such as that it occurred in detention. However, for most cases of sexual violence the location text variable contains enough information to pinpoint a first-order admin unit in which some sexual violence occurred per year.

(22)

18

locations where sexual violence had occurred in a year were separated by a semi-colon in SVAC. However, occasionally commas were used to separate locations, which added some confusion since a single location could contain commas as well. In these instances, the original country reports were consulted to identify and distinguish between the individual locations. Secondly, when the name of the individual location was known, different resources were used to unveil the geographic location. As many of the locations of sexual violence were obscure villages or the like, no single source could be used to identify in which admin unit they existed. Most frequently Google Maps was used and if not then local weather forecasting websites were used as well as maps issued along country reports. Identifying and consulting the country reports from which SVAC coded the variable was of much assistance. Thirdly, once the geographic locations were ascertained, these were cross-referenced with GADM maps to identify the first-order admin unit/units in which sexual violence was recorded.

The geographically identifiable admin units of each observation where specified in a new text variable. For group-year observations where some or all instances of sexual violence could not be geographically identified, a notation was made so as to know that these observations have missing data. Next, this data on rebel groups’ sexual violence was merged with GADM data. Thereby, all year observations in the SVAC dataset where disaggregated to group-admin-year observations. Since each rebel group is only reported to operate in one country, the admin units are the admin units in the active country for each rebel group. All admin units and years in which rebel sexual violence has been reported where then identified. Thus, the dependent variable is Rebel sexual violence, which is assigned a value of 1 if a rebel group has perpetrated sexual violence in a specific admin unit and year, and 0 if not.

4.4. Data and operationalization of ethnic ties and strong commanders

(23)

19

In order to operationalize ethnic ties, two elements are needed: the identity of civilian communities to which a rebel group has ethnic ties and the location of those communities. To identify the ethnic groups to which the rebel group has ties, the ACD2EPR dataset is used. It contains information on whether rebel groups claim to pursue political goals on the behalf of an ethnic group, whether they recruit fighters from it and whether they receive support from it. Rebel groups which claim to specifically represent an ethnic group should generally perceive themselves as having ties to this ethnic group. Rebel groups which have made claims to fight on behalf of an ethnic group, evidence of which can be direct or indirect (e.g. name of the group), are therefore coded as having ethnic ties to this group. Additionally, support from civilians to co-ethnic rebels is evidence of reciprocity in the perceived ethnic ties. Rebel groups which receive support from at least 50 percent of the members of an ethnic group are therefore also coded as having ethnic ties to this group. The actors from the ACD2EPR data overlap to a very high degree, with only 4 out of 116 rebel groups in the SVAC data not being present in the EPR data. Of these, only one group was reported to perpetrate sexual violence and only in one year. Therefore, the missing data should not significantly impact the findings.

To identify the location of the ethnic groups to which the rebels have ethnic ties, spatial data on the settlement distribution of ethnic groups is gathered from the Geo-EPR dataset. The spatial overlap between an ethnic group and an administrative area can be calculated to determine whether an ethnic group is present in the area. However, as this study uses first-order administrative units, which are often fairly vast, a threshold for ethnic presence is introduced. Only administrative units which have an overlap of above 5 percent with an ethnic group are coded as being inhabited by that ethnic group. Ethnic ties are measured through overlap between ethnic groups present in the admin unit and ethnic groups which share ties with the rebels. Therefore, the independent variable Ethnic ties is coded as 1 if at least one of the ethnic groups present in the admin unit has ties to the rebel group and 0 if not.

(24)

20

command control in the NSA data, strengthcent. The original variable has a three-point ordinal scale, low-medium-high. However, this study is only interested in identifying which rebel commanders have the ability to restrain their combatants if they wish. Commanders which have either medium or high control should have the prerequisites to restrain their combatants and should therefore be coded as strong rather than weak. Therefore, the NSA variable is dichotomized to the binary variable Command strength which is assigned a value of 1 if the rebel group has a medium or high command strength and 0 if has low command strength. The actors in the NSA dataset overlap fairly well with the actors in the SVAC dataset, with only 13 out of 116 rebel groups missing. However, the NSA data only includes active conflict years. This means that the five post-conflict years included in the SVAC data receive missing values. To overcome this, the last reported value for the rebel group was imputed to the subsequent observations. This introduces some uncertainty, since it may not be safe to assume that rebel groups retain the same level of command strength during the post-conflict years. However, this is expected to have an insignificant effect on the performance of the analysis, since the original NSA data is largely time-invariant.

4.5. Control variables

The control variables included in this analysis were selected based on their possible confounding effect on the relationship between ethnic ties and rebel restraint in the use of sexual violence. From a theoretical perspective, rebel ethnic ties to civilians should be largely resistant to external factors. Considering that neither individual rebels nor individual civilians can alter their ethnicity, variation on their ethnic ties could only be caused by changes to the composition of the groups. If a rebel group begins recruiting from a different ethnic group for example, the ethnic identity of the rebel group may change over time as a result. This would by extension cause a shift in their ties to the original ethnic group. However, it can be deemed an extremely rare event that a rebel group should undergo these types of identity changes while retaining identifying characteristics such as their name. Substantial changes to the name or stated purpose of the group would result in the rebel group being coded as a new actor in most datasets, including SVAC and EPR. Considering that rebel groups who change their core ethnic identity would most likely also change other identifying markers, it is safe to assume that they would emerge as new actors in these cases.

(25)

21

claims as well as civilian support variables from the ACD2EPR dataset. The ethnic ties variable is coded from the presence of an ethnic group in an admin unit and year, to which the rebel group has ties of ethnic claims or support. Therefore, factors which may induce rebels to make claims or receive support from an ethnic group as well as factors which affect the location of ethnic groups may act as confounders. These factors can be traced to the characteristics of the admin unit, the rebel group and the conflict.

Firstly, the characteristics of the admin unit may exert an effect, specifically the ethnic composition of the admin unit. The more politically relevant ethnic groups that are present in an admin unit, the higher the statistical likelihood that one group in the admin unit has ties to a rebel group. Ethnic fractionalization and diversity in an admin unit may also be conducive to increased contention by the rebel group. The higher the number of ethnic groups, the higher the likelihood that that one of them has ties to the government. As is found in the article by Fjelde and Hultman (2014), rebel groups are more likely to engage in one-sided violence in the government’s ethnic constituencies. If the same holds for sexual violence, or if sexual violence occurs as a by-product of other violent behaviour, admin units which house ethnic groups who support the government should have a higher risk of sexual violence than others. Based on this, the dichotomous variable Ethnic diversity is created. It assumes a value of one if there is more than one politically relevant group in the admin unit. In addition, the Shared ties government variable is created. Unfortunately, the ACD2EPR dataset does not contain information on ethnic groups’ individual relationship to the government. Therefore, there exists no good measure of an ethnic groups ties to the government. To attempt to capture some of the intended variation the shared ties government variable is coded based on the support and recruitment variables from the EPR dataset. The variable is assigned a value of one if any of the present ethnic groups supports both the government and rebels or is used as a pool for recruits by both government and rebels. This should give some indication of which ethnic groups the government has ties to. However, since this measure captures ethnic groups from which rebels also gather support or acquire recruits, it should be expected that it will covary to a large degree with the independent variable ethnic ties.

(26)

22

perpetrate violence in the beginning of a conflict repeat the behaviour throughout the span of it. Rebel groups who have victimized civilians in the past should be more likely to perpetrate sexual violence, as well as less likely to receive support from or to represent an ethnic group. Based on this, the variable Rebel One-Sided Violence (OSV) is created. It is the natural log of civilian fatalities caused by a rebel group in an admin unit and year. It is assigned a value based on the sum of civilian fatalities for all UCDP events of OSV by the rebel group reported in the specific admin unit and year4. Furthermore, the rebels may not have the prerequisites for ethnic ties with a civilian community. Rebels without a distinctive ethnic composition would not be able to have ethnic ties to any group, as they cannot invoke shared ties with any ethnic group. Therefore, the binary variable Any ties is created. It records whether or not the specific rebel group has ties to any ethnic group.

Thirdly, the characteristics of the conflict may exert an effect, specifically the location and intensity of fighting. Conflict intensity is expected to exacerbate the risk of sexual violence (e.g. Johansson and Sarwari 2019). Further, conflict intensity is also expected to increase ethnic mobilization. Ethnic groups only become politically relevant and thus included in the EPR data, through having an organization represent their interests. Thus, the intensity and location of conflict may serve as a cause for rebel groups to claim to represent an ethnic group, or for the ethnic group to support the rebels. In her article on ethnic mobilization, Eck (2009) finds that the reverse holds; ethnic mobilization increases conflict intensity. While her design controls for reverse-causality she raises questions regarding the theoretical possibility of an endogenous relationship and to what extent rebel groups are path dependent in mobilizing along ethnic lines. Based on this, two variables on intensity are created. Conflict intensity is the natural log of the number of battle-related and civilian deaths in an admin unit and year. It is assigned a value based on the sum of UCDP best fatality counts for all violent events reported in the specific admin unit and year. Rebel intensity is the natural log of the number of battle-related and civilian deaths with connection to a rebel group in an admin unit and year. It is assigned a value based on the sum of UCDP best fatality counts for all violent events reported in the specific admin unit and year, in which the rebel group was an active party.

4_{To identify the admin unit in which violent events took place, I employed the same procedure as Cil et al.}

(27)

23

4.6. Method of analysis

The method used for the quantitative analysis of this study is binomial logistic regression. Since the dependent variable rebel sexual violence is dichotomous, along with the independent and conditional variable, this regression method is preferable. The software R version 3.6.2 was used to estimate the regression models and perform subsequent computations (R Core Team 2019). Several R packages were used for the data manipulation and analysis, of which three were instrumental. The packages dplyr, version 0.8.5 (Wickham et al. 2019), and sf, version 0.9.3 (Pebesma 2018), were necessary for managing and merging the spatial data. The package stargazer, version 5.2.2 (Hlavac 2018), was necessary for exporting tables.

In a first step the relationship between the independent and dependent variable will be tested using the full sample, i.e. all admin units in African countries with internal armed conflict 1989-2009. In this step the independent and conditional variables are to be tested both independently and in conjunction with the control variables outlined previously. This first step of the analysis will seek to determine the overall effect of ethnic ties, with and without strong central command, on rebel sexual violence prevalence.

In a second step, the relationship will be tested using sub-samples based on the risk sample variables, specified in the following sub-section. In this step only key control variables will be included. The other variables from the full model of step 1 are removed in step 2 to avoid multicollinearity and overfitting of the regression models, which would reduce the confidence that can be placed in the results. Specifically, the variable Rebel intensity is not included in the second step since it is expected to capture much of the same variation as Conflict intensity and Rebel OSV. Additionally, the variable Any ties was also cut as it is expected to capture the variation of the independent variable Ethnic ties. Removing these two variables reduces the risk of multicollinearity and improves the confidence one can place in the accuracy of the standard errors, which indicate how representative the sample is of the population. The variable Shared ties government is removed as it is highly inflated and insignificant throughout all model specifications, thus only increasing the risk of overfitting if included.

(28)

24

(29)

25

5. ANALYSIS

5.1. Descriptive statistics

First presenting some descriptive statistics, table 1 holds a few key insights into the distribution of the data. As can be seen in table 1, the distribution of the dependent variable Rebel sexual violence is highly skewed. A mean value of 0.009 for a binary variable indicates that less than one percent of observations in the full sample has a value of 1. Comparably, Ethnic ties is reported as having a mean of 0.219, with an observation taking either a value of 1 or 0, meaning that approximately 22% of observations have a value of 1. Command strength has the lowest sample size of 13,130 compared to most other variables’ 17,574. This indicates that there is a substantial amount of missing data. Seeing as there is a slightly skewed distribution on the variable ethnic ties and a lot of missing data on the command strength variable, the distribution of values across these variables should be examined more closely in a contingency table.

Table 1: Summary statistics

Variable name N Mean St. Dev. Min Pctl (25) Pctl (75) Max Rebel sexual violence 17,547 0.009 0.094 0 0 0 1 Ethnic ties 17,547 0.219 0.413 0 0 0 1 Command strength 13,130 0.799 0.401 0 1 1 1 Conflict intensity 17,547 1.189 1.978 0 0 2.20 12.70 Rebel intensity 17,547 0.362 1.171 0 0 0 10 Rebel OSV 17,547 0.095 0.583 0 0 0 10 Shared ties government 17,547 0.123 0.328 0 0 0 1 Any ties 17,547 0.600 0.490 0 0 1 1 Ethnic diversity 16,677 0.704 0.457 0 0 1 1 Conflict Full (risk sample 1) 17,547 0.842 0.365 0 1 1 1 Conflict Year (risk sample 2) 17,547 0.332 0.471 0 0 1 1 Area of Operations Full (risk sample 3) 17,547 0.345 0.475 0 0 1 1 Area of Operations Year (risk sample 4) 17,547 0.109 0.311 0 0 0 1 SV Full (risk sample 5) 17,547 0.397 0.489 0 0 1 1 SV Year (risk sample 6) 17,547 0.178 0.383 0 0 0 1

(30)

26

command. Since there are so few observations where rebels have weak command and ethnic ties, there is a concern that the variation might be driven by just a few actors. However, looking at the data, these 551 observations are spread across eleven rebel groups and five countries. This can be compared to thirty-five rebel groups and sixteen countries for the 2,642 observations of rebels with ethnic ties and strong command. This gives us confidence, as there is high variation on actors and countries within each category, proportional to their size.

Table 2: Ethnic ties and Command strength contingency table

N=13,130 Ethnic ties

Command strength 0 1 Sum

0 2,090 (15.9%) 551 (4.2%) 2,641 (20.1%) 1 7,847 (59.8%) 2,642 (20.1%) 10,489 (79.9%) Sum 9,937 (75.7%) 3,193 (24.3%) 13,130 (100%)

Table 3 presents how many observations of sexual violence there are for rebel groups with ethnic ties and strong command compared to other rebel groups. It can be seen here that for the 2,642 observations of strong groups with ties, only 48 also report sexual violence. This demonstrates the severe skewedness of the distribution of sexual violence observations mentioned previously. Looking at the data, these 48 observations are, however, spread across ten actors, alleviating some concerns. That said, it should be noted that the findings of any analyses on this data will be driven by a very small amount of observations, impacting the confidence we can place in them negatively.

Table 3: Ethnic ties * Command strength and Sexual violence contingency table

N=13,130 Ethnic ties * Command strength

Sexual violence 0 1 Sum

(31)

27

5.2. First step of analysis

Next, I present the results of the first step of the analysis. The theoretical proposition of this study holds that ethnic ties between rebels and civilians should increase the costs of ordering, authorizing and tolerating sexual violence for rebel commanders. The first hypothesis suggests that when rebels have ethnic ties to a civilian community, they are more likely to enforce restraint. The second hypothesis suggests the same but adds that this effect should be conditional on strength of the rebel commander.

The first model is a simple bivariate model which estimates the impact of ethnic ties on rebel sexual violence. As reported in table 4, the main independent variable Ethnic ties is positive and significant at the 95% confidence level. This suggests that rebels are more likely to perpetrate sexual violence in regions which host co-ethnics than in regions which do not, counter to the first hypothesis. In the second model, the variable Command strength is added to the analysis along with Ethnic ties * Command strength, the interaction term between the two independent variables. In model 2, ethnic ties remains positive and command strength is also positive. The interaction term of ethnic ties and command strength, however, is negative and all three are statistically significant. This suggests that rebels with a strong central command are less likely to perpetrate sexual violence in regions where there are co-ethnics than in other regions. Statistical significance of the coefficients suggest that the null hypothesis can be rejected, to a certain confidence level. These confidence levels are reported at the bottom of each table. Being able to reject the null hypothesis means that the results found in this analysis are likely to be representative of the actual relationship in the full population. The negative direction and statistical significance of the interaction term in model 2 thus provides support for the second hypothesis.

(32)

28

Table 4: The effect of ethnic ties on rebel sexual violence, using full sample

Model 1 Rebel SV Model 2 Rebel SV Model 3 Rebel SV Model 4 Rebel SV Ethnic ties 0.979*** _1.761*** _1.386** _1.908** (0.162) (0.411) (0.582) (0.774) Command strength 0.736** _1.221** _1.717** (0.337) (0.520) (0.723) Ethnic ties * Command strength -1.150** _-1.532** _-2.058***

(0.451) (0.606) (0.795)

Any ties 0.301 0.486*

(0.236) (0.279)

Ethnic diversity 0.166 0.006

(0.236) (0.239)

Shared ties government -15.625 -15.418

(398.500) (413.201) Conflict intensity (ln) 0.043 (0.056) Rebel intensity (ln) 0.314*** (0.070) Rebel OSV (ln) 0.276*** (0.066) Conflict intensity (ln) t-1 -0.004 (0.063) Rebel-conflict intensity (ln) t-1 0.291*** (0.080) Rebel OSV (ln) t-1 0.299*** (0.074) Constant -5.009*** _-5.338*** _-6.352*** _-6.753*** (0.105) (0.317) (0.583) (0.778) Observations 17,547 13,130 12,391 10,913 Log Likelihood -884.653 -809.481 -640.048 -575.583 Akaike Inf. Crit. 1,773.307 1,626.962 1,300.097 1,171.166

(33)

29

To test the robustness of these findings, I also fit model 4 with country and time fixed effects. Country fixed effects should account for any variation caused by country-specific factors while time fixed effects should capture any variation in reporting of sexual violence over time. The results are reported in table 5. Looking at model 4b, time fixed effects do not affect the significance or direction of ethnic ties or the interaction variable. There is a slight change to the coefficient estimates for ethnic ties and the interaction term, but these are too small to hold any substantive meaning. Looking at model 4a, however, country fixed effects decrease the significance of both ethnic ties and the interaction term. The estimates here are only significant at the 90 percent confidence level, compared to the 95 and 99 percent level they were in model 4. They do, however, maintain their direction and approximate magnitude. Command strength also loses its significance when accounting for country fixed effects.

Table 5: The effect of ethnic ties on rebel sexual violence with fixed effects, using full sample

Model 4a Rebel SV Model 4b Rebel SV Ethnic ties 1.558* _1.736** (0.891) (0.793) Command strength 0.954 2.321*** (0.819) (0.730)

Ethnic ties * Command strength -1.517* _-2.124***

(0.894) (0.813)

Any ties -0.406 0.617**

(0.549) (0.297)

Ethnic diversity 0.295 -0.142

(0.261) (0.255)

Shared ties government -22.987 -16.666 (683.902) (607.871) Conflict intensity (ln) t-1 0.099 0.006 (0.068) (0.067) Rebel intensity (ln) t-1 0.362*** 0.371*** (0.089) (0.092) Rebel OSV (ln) t-1 0.035 0.323*** (0.086) (0.086)

Country fixed effects Yes -

Time fixed effects - Yes

Observations 10,913 10,913

Log Likelihood -479.787 -446.246

Akaike Inf. Crit. 1,021.574 950.492

(34)

30

(35)

31

Table 6: The effect of ethnic ties on rebel sexual violence, using sub-samples w/o missing data

Model 4c Rebel SV (excl. group-years) Model 4d Rebel SV (excl. countries) Ethnic ties 1.160 2.267*** (0.809) (0.869) Command strength 1.316* _0.659 (0.725) (0.758) Ethnic ties * Command strength -1.300 -1.212 (0.829) (0.849)

Any ties 0.871*** _-1.301**

(0.314) (0.564)

Ethnic diversity -0.218 16.690

(0.253) (885.684)

Shared ties government -15.570 -17.039

(404.256) (979.503) Conflict intensity (ln) t-1 -0.087 0.161** (0.076) (0.078) Rebel intensity (ln) t-1 0.391*** 0.322*** (0.094) (0.091) Rebel OSV (ln) t-1 0.356*** 0.149* (0.086) (0.085) Constant -6.389*** _-22.461 (0.781) (885.685) Observations 9,528 7,571 Log Likelihood -479.563 -239.951

Akaike Inf. Crit. 979.127 499.901

Note: *_p<0.1;**_p<0.05;***_p<0.01

5.3. Second step of analysis

(36)

32

In model 8, the coefficient estimates and standard errors for ethnic ties, command strength as well as the interaction term are all inflated. This could be due to the lower sample size decreasing the amount of variation that can be observed. When looking at the data, however, it suggests that the problem lies elsewhere. Within the sample of 1,436 observations used in model 8 there are 82 positive observations of sexual violence, of which 31 are positive for both ethnic ties and command strength. This amount should be sufficient considering the small number of variables included in the model. Instead, the cause of the inflated coefficients and standard errors could be increased multicollinearity. Small sample size can exacerbate existing collinearity to make coefficient estimates and standard errors less accurate (Lavery et al. 2019). Interestingly, the estimates for rebel OSV are significant and positive in all models5, while conflict intensity is positive and significant in five out of six models and ethnic diversity in four out of six models. This suggests that the characteristics of rebel groups as well as the characteristics of the conflict may affect the rebels’ use of sexual violence. Descriptively, these findings suggest that admin units in which there are more conflict fatalities and admin units in which more than one ethnic group resides are more likely to experience sexual violence. The findings also suggest that rebel groups who perpetrate OSV in a region are more likely to also perpetrate sexual violence.

Having presented the results of the analysis, I now turn to how they relate to the theoretical proposition posed in this study. Also, additional interesting findings are identified and discussed, alternative explanations for observing these results are presented and the limitations of the research design and analysis are discussed.

(37)

33

Table 7: The effect of ethnic ties on rebel sexual violence, using risk samples

Model 5: Rebel SV Model 6: Rebel SV Model 7: Rebel SV Model 8: Rebel SV Model 9: Rebel SV Model 10: Rebel SV Risk sample: Conflict Full Risk sample: Conflict Year Risk sample: AoO Full Risk sample: AoO Year Risk sample: SV Full Risk sample: SV Year Ethnic ties 2.640*** _1.964** _2.197*** _16.324 _2.192*** _2.240*** (0.775) (0.780) (0.780) (596.653) (0.771) (0.780) Command strength 1.996*** _1.249* _1.601** _15.627 _2.005*** _2.033*** (0.723) (0.745) (0.740) (596.653) (0.724) (0.727) Ethnic ties * Command strength -2.321*** _-1.497* _-1.772** _-15.971 _-2.413*** _-2.644*** (0.801) (0.829) (0.818) (596.653) (0.796) (0.804) Ethnic diversity 0.196 1.310*** _0.675* _1.563*** _0.097 _0.080 (0.265) (0.482) (0.359) (0.536) (0.241) (0.244) Conflict intensity (ln) t-1 0.127*** 0.126* 0.149*** 0.118 0.141*** 0.153*** (0.048) (0.065) (0.057) (0.079) (0.046) (0.048) Rebel OSV (ln) t-1 0.497*** 0.497*** 0.433*** 0.369*** 0.353*** 0.265*** (0.061) (0.066) (0.066) (0.078) (0.064) (0.065) Constant -7.216*** _-7.390*** _-7.117*** _-21.031 _-5.979*** _-5.229*** (0.732) (0.821) (0.755) (596.653) (0.746) (0.752) Observations 9,425 3,832 4,669 1,436 4,512 2,279 Log Likelihood -490.792 -272.269 -330.793 -196.223 -509.267 -435.538 Akaike Inf. Crit. 995.584 558.539 675.586 406.446 1,032.535 885.077

(38)

34

6. DISCUSSION

6.1. Interpretation of findings

Throughout the results of the analysis, the ethnic ties variable is consistently reported to exert a positive and significant effect on sexual violence. Regardless of what other variables are included, it remains significant through all model specifications, aside from the robustness test in model 4c (table 6) and the narrowest risk sample in model 8 (table 7). This finding is in direct contradiction of the expected relationship specified in the first hypothesis. This implies that rebel groups, in general, do not enforce any particular restraint with regard to co-ethnics. Instead they suggest that rebels are more likely to perpetrate sexual violence in admin units where they have ethnic ties, indicating that rebels are more prone to sexual violence against co-ethnics. As such, evidence in support of the first hypothesis is not found.

As reported in table 4, the interaction term of ethnic ties and command strength is reported to exert a negative and significant effect on sexual violence. These findings are also robust to accounting for country and time fixed effects (see table 5). These findings provide evidence of ethnic ties having a decreasing effect on the probability that rebels will perpetrate sexual violence. This is an important finding for the theorized relationship, as it is indicative of the hypothesized relationship between ethnic ties and restraint. The results from the second step of the analysis further provide support for H2. The interaction term of ethnic ties and command strength on sexual violence was consistently negative and significant throughout all but one of the risk samples (table 7, models 5-7 and 9-10). Considering particularly the results of model 9, they suggest that rebels with strong command who have or will perpetrate sexual violence do so less in regions in which they have ethnic ties. The results from model 10 suggest that rebels with strong command who perpetrate sexual violence in a given year also do so less in regions where they have ethnic ties. As such, these findings indicate that all else equal, rebels are less likely to perpetrate sexual violence against co-ethnic civilians. This points to strong support of the second hypothesis, that rebel groups with strong commanders are more likely to enforce restraint in the use of sexual violence against co-ethnic civilians than others.

(39)

35

in the use of sexual violence too costly, would act in a manner which produces precisely the findings reported. As sexual violence is in some way beneficial or restraining it infeasible, these groups perpetrate sexual violence. However, as they do, they take into consideration the potential costs of alienating co-ethnic civilians and disrupting their social embeddedness. Therefore, they perpetrate sexual violence against other civilian communities. Enforcing restraint with one hand and sanctioning violence with the other, commanders may reap the supposed benefits of sexual violence while limiting the costs incurred.

It is very curious and unexpected that ethnic ties are found to exert a positive effect on the probability of sexual violence. This would suggest that there are motives either for the commander or combatants to perpetrate sexual violence against co-ethnic compared to other civilians. However, considering that the interaction effect is negative, ethnic ties only increases the probability of sexual violence for rebel groups with weak central command. If the theorized effect of command strength holds, this has two implications for the theoretical argument. Firstly, that strong commanders are central in shaping the sexual violence perpetration of rebel groups, as was suggested in the above paragraph. Secondly, it suggests that when commanders are not able to restrain their combatants, their individual or unit preferences drive the perpetration of sexual violence. Since strong commanders reverse the impact of ethnic ties, the findings indicate that combatant incentives are the driving factor in sexual violence perpetration. This suggests that when commanders can enforce restraint on the basis of ethnic ties, they will; when they are not able to enforce restraint, combatants will seek to perpetrate sexual violence against co-ethnics rather than other civilians.

(40)

36

6.2. Substantive meaning of findings

To further examine the substantive meaning of the findings from the second step of the analysis, the predicted probabilities of sexual violence for different values on ethnic ties, command strength and the interaction term for models 7, 9 and 10 are calculated. The motivation for only calculating predicted probabilities for these models lies in that ethnic groups are expected to be clustered. When rebel groups make claims on behalf of some co-ethnic constituents, they are likely to do some or most of the fighting close to these constituents. Therefore, having an rebel group fight for an ethnic group in the admin unit is predictive of the admin unit being inside the conflict zone. By being inside the conflict zone, the admin runs a higher risk of experiencing rebel sexual violence. Hence, the predicted probabilities for sexual violence when there are ethnic ties are automatically and falsely higher when using a broad sample. Only the narrower risk samples will be fit for calculating predictive probabilities, since they exclude a greater amount of the admin units which are entirely outside the conflict zone.

The predicted probabilities for models 7, 9 and 10 are presented in figure 2. The results show the expected probability of rebel sexual violence in an admin unit, the values on the vertical axis, given different values for ethnic ties and command strength. To calculate the predicted probabilities, all control variables were held at their mean, including the dichotomous variable ethnic diversity6_{. In figure 2, command strength is held at its mean for the darkest shaded}

columns, Mean-value command without ties. For the other two columns it is held at zero in the medium-shaded columns, Weak command with ties, and at one in the brightest shaded columns, Strong command with ties. For the two latter columns, the ethnic the ethnic ties variable is set to one. This gives us the predicted probabilities of sexual violence for the average rebel group without ethnic ties compared to rebel groups with weak versus strong command and ethnic ties.

6_{Setting ethnic diversity to its mode value instead had little impact on the results, on average increasing the}

(41)

37

Consider the values of the columns on the left-hand side for model 7, which uses the Area of Operations (AoO) Full risk sample. These show that the probability of sexual violence in an admin unit with ethnic ties is higher than admin units without ties, for weak and strong command alike. However, among admin units with ethnic ties, the risk of sexual violence is slightly higher when rebel command is weak compared to when it is strong. The absolute probability for this model and sample is not particularly relevant, since the sample encompasses a very large amount of admin units in which sexual violence is not overly likely to occur. Looking at the columns in the middle and on the right-hand side, the results are very different. For model 9, which uses the SV Full risk sample, admin units in which rebels with weak command have ethnic ties have a much higher probability of sexual violence than those with ties to strong command rebels or without ties. The probability of sexual violence in this sample is 1.9 percent for admin units which do not have ethnic ties, 3.1 percent for admin units with ties to weak command rebels and 2.1 percent for admin units with ties to strong command rebels. This suggests that the risk of sexual violence increases substantially when weak command rebels have ties to the admin unit and only slightly when strong command rebels have ties. For model 10, which uses the SV Year risk sample, the probability of sexual violence

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

Model 7: AoO Full Model 9: SV Full Model 10: SV Year

Figure 2: Predicted probabilities of rebel sexual violence

(42)

38

is 4.3 percent for admin units which do not have ethnic ties, 6.9 percent for admin units with ties to weak command rebels and 3.9 percent for admin units with ties to strong command rebels. This suggests that the risk of sexual violence increases substantially when rebels with weak command have ties to the admin unit but decreases when rebels with strong command have ties to the admin unit.

These findings are very interesting and provide mixed evidence for the second hypothesis. As the sample becomes narrower, the effect of rebels with ethnic ties and strong command on the probability of sexual violence turns from positive to negative. As was previously discussed, the narrower samples have a lower risk of ethnic ties capturing unintended variation. This suggests that the predicted probabilities for model 10 may be the most relevant in substantive terms. In this model, the risk of sexual violence increases with 2.6 percentage points, a sixty percent relative increase, when an admin unit has ties to rebels with weak command compared to when it has no ties. Although, the risk of sexual violence decreases with 0.4 percentage points, a relative decrease of about ten percent, when an admin unit has ties to strong command rebels compared to when it has no ties. This provides evidence against the first hypothesis and in support of the second hypothesis. However, predicted probabilities should also be computed for rebels with weak command and rebels with strong command separately. The results are reported in figure 3 for rebels with command strength set to zero and in figure 4 with command strength set to one. All control variables are set to their mean value.

(43)

39

In figure 4, the results are somewhat mixed. In model 7, the risk of sexual violence in an admin unit increases by 0.6 percentage points, a relative increase of fifty percent. In models 9 and 10, the risk of sexual violence decreases when there are ethnic ties. In model 9, the probability decreases with 0.5 percentage points, a twenty percent relative decrease, and in model 10 the probability decreases with 1.9 percentage points, a more than thirty percent relative decrease. Taking into account that the probabilities for model 7 are likely less accurate, these findings suggest that rebel groups with strong command are less likely to perpetrate sexual violence in areas where they have ethnic ties compared to areas where they do not. This points to support for the second hypothesis.

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

Model 7: AoO Full Model 9: SV Full Model 10: SV Year

Figure 3: Predicted probabilities of sexual violence for rebels with weak command

THE TIES THAT BIND COMMANDERS