Department of Peace and Conflict Research Spring 2017

Uppsala University

Department of Peace and Conflict Research Spring 2017

_____________________________________________________

DANCING AFTER THE REBELS ’ ^PIPE ?

_____________________________________________________

Explaining the Relationship between Rebel Violence Against Civilians and the Onset of Peace Negotiations

BENEDIKT WIMBAUER

Master Thesis in Peace and Conflict Studies Supervisor: Kristine Eck

22 May 2017

Abstract i

A

This study seeks to contribute to a lively debate on the relationship between one-sided violence and the outcome of armed conflict by examining whether the degree of one-sided violence, committed by rebel groups, affects how long it takes until peace negotiations are initiated. It is argued that rebels attack civilians to increase the government’s costs of war and enforce concessions in negotiations. In light of these attacks, governments quickly convene negotiations to prevent more violence in the future. Furthermore, democratic governments are expected to start negotiations more quickly than autocracies, as one-sided violence affects democratic leaders’

chances of reelection. These arguments are tested in a quantitative analysis of 104 African conflict dyads from 1989 to 2009, using a Cox Proportional Hazards Model. The empirical analysis provides no support for the claim that higher levels of rebel one-sided violence lead to a quicker onset of negotiations. There is weak evidence that negotiations begin more quickly in democracies compared to autocracies, but no violence at all would result in the quickest onset.

The analysis also indicates that negotiations have a pacifying effect on violence in that they seem to reduce the number of civilians killed in rebel attacks, even though conflict continues.

Key words: Violence Against Civilians, Peace Negotiation Onset, Costs of War, Duration

Analysis, Cox Proportional Hazards Model

Acknowledgements ii

A

I wish to express my sincerest gratitude to my supervisor Kristine Eck, who supported me

throughout the process of writing this thesis by providing invaluable input and guidance, and

who always encouraged me that I was on the right track. I also want to thank all of my fellow

students in the Master’s programme, who made me have a wonderful time not only during these

four months of writing the thesis, but throughout the almost two years we were studying

together and debating over conflicts, political crises and other things that one might often forget

about when living in a peaceful and progressive country such as Sweden. A special thanks also

goes to Nicole Carter for commenting on an earlier version of this study. Last but not least I

want to thank my parents who supported me during all the years of my studies, and who

encouraged me in studying what I’m good at and passionate about.

Table of Contents iii

T

C

Abstract ... i

Acknowledgements ... ii

Index of Tables and Figures ... iv

1. Introduction ... 1

2. Previous Research on Violence Against Civilians ... 4

3. Theoretical Framework ... 7

3.1. Conceptualizing Violence Against Civilians ... 7

3.2. Explaining Rebel Violence Against Civilians ... 7

3.3. Explaining Why Governments Seek Negotiations ... 10

3.4. Conditioning the Argument: Regime Type ... 13

4. Research Design ... 15

4.1. Statistical Method ... 15

4.2. Data ... 18

5. Empirical Analysis ... 23

5.1. Findings ... 23

5.2. Extended Analysis ... 26

6. Critical Discussion ... 34

7. Conclusions ... 42

References ... 44

Appendix ... 51

Index of Tables and Figures iv

I

T

F

Table 1 List of Variables ... 22

Table 2 Descriptive Statistics ... 23

Table 3 Cox Proportional Hazards Model Estimates ... 24

Table 4 Separate Models for Democracies and Autocracies ... 25

Table 5 Statistics for Each Stratum ... 26

Table 6 Model Specifications I (Influence Statistics) ... 28

Table 7 Conflicts and Negotiation Onset Over Time ... 30

Table 8 Frequency Table Regime Type ... 51

Table 9 Model Specifications II (Dependent Variable) ... 51

Table 10 Model Specifications III (Independent Variable) ... 52

Table 11 Model Specifications IV (Alternative Testing of Hypothesis II) ... 52

Table 12 Model Specifications V (Model Choice) ... 53

Table 13 Test of Proportional Hazards Assumption ... 55

Table 14 Multicollinearity Test ... 55

Figure 1 Causal Mechanism ... 12

Figure 2 Cox Hazard Functions ... 29

Figure 3 Influence Statistics ... 54

Introduction 1

1. I

When rebels and governments decide to wage war, the main victims from violence are often civilians. There is evidence that armed conflict negatively impacts a country’s economy, hampers efforts to reduce poverty and hunger, and deteriorates the public health sector (Gates et al. 2012;

Plümper and Neumayer 2006). Civilians not only suffer from these indirect effects of war, but are often direct victims from violence. In the twenty years after the end of the Cold War, African governments and rebel groups killed approximately 650.000 civilians, averaging more than 30.000 casualties every year

. International law may classify such attacks as war crimes and crimes against humanity, thereby illustrating that these attacks are considered unacceptable in at least most parts of the world. Yet, they still occur frequently.

In the past two decades, peace and conflict literature has focused on explaining the motivations behind one-sided violence committed by parties to conflict, mostly arguing that one- sided violence should be seen as a strategy rather than as a side effect of conflict. Based on this assumption and the question of whether targeting civilians helps governments or rebels achieve their strategic goals, recent studies have started to examine the relationship between violence against civilians and the outcome of conflict. However, there is a lack of consensus in the literature on the implications of these studies. For example, Thomas (2014) and Wood and Kathman (2014) have found evidence that rebels are more likely to negotiate with and get concessions from governments if they resort to violence against civilians, while Fortna (2015) has concluded that rebels who attack civilians are less likely to achieve either military goals or get concessions in bargains.

Despite the increased focus in previous research on the link between one-sided violence and war outcomes, three issues have not been sufficiently taken into account. First, scholars have not paid attention to the factor of time, as violence against civilians can accelerate or delay a certain war outcome. Second, studies on the link between one-sided violence and peace negotiations–which is the most common way how contemporary conflicts end–have only looked at the outcome or occurrence, but not at the onset of negotiations. Third, the extent of violence:

previous studies have either examined the effect of terrorism, or only measured whether there was one-sided violence, while not taking into account the extent of violence. These three gaps in the existing literature are the point of departure for this present study, which investigates whether

1 This number excludes civilian casualties caused by direct military confrontations between parties to conflict. The vast majority of people died in the Rwandan genocide and subsequent tensions in neighboring countries.

Introduction 2

the extent of rebel one-sided violence in armed conflict affects how long it takes until peace negotiations are initiated.

This research question is based on the assumption made in previous studies that rebel groups involved in armed conflict with governments attack civilians in order to increase the government’s costs of war and subsequently enforce concessions. Incumbents fear to lose popular support and territorial control as a consequence of these attacks, since they know that similar violence may again happen in the future. Prompted by the urgency to respond to the crimes, governments are likely to concede to the convening of peace negotiations with the enemy, and they expectedly decide to do so as fast as possible to ensure that the enemy is not going to attack even more civilians. Whether rebel one-sided violence significantly increases the costs of war on the government side depends on the extent of violence: the more civilians the rebels kill, the higher should be the pressure on the government to engage in negotiations; if it does not, the population is likely to lean up against their leaders. A second argument further conditions the first one: in democratic states, the population can remove unpopular leaders, who fail to protect them from rebel attacks, from power in upcoming elections. Since electoral victory becomes less likely with any additional rebel attack against civilians, democratic governments should be particularly inclined to initiate negotiations as soon as possible. The effect is expected to be weaker in autocracies.

The two arguments are tested empirically in a quantitative analysis of 104 African conflict dyads, each consisting of a government and a rebel group, between 1989 and 2009, with UCDP data for rebel one-sided violence and information on negotiation onsets based on a previous study (Thomas 2014). I use a Cox Proportional Hazards Model, a special type of duration model

, to examine the relationship between the extent of rebel one-sided violence and the time until peace negotiations begin. The results in this study show no support for the argument that higher levels of rebel violence against civilians lead to a quicker onset. The coefficient even points in the opposite direction, but fails to achieve statistical significance. With regard to the conditional hypothesis, there is at least weak evidence that peace negotiations begin sooner in democracies compared to autocracies. However, no rebel one-sided violence at all would result in the quickest convening of peace talks. The findings in this study confirm one strain in previous

2 The focus on rebel one-sided violence is one of the scope conditions of this study. The implications for government violence against civilians are dealt with in the sixth chapter.

3 Duration models investigate the history of an event, rather than the simple linear relationship of variables. For instance, in this study the dependent variable is time until onset of peace negotiations, and the duration model measures the likelihood of negotiation onset separately for every point in time, rather than whether A leads to B in general terms. Duration models and the Cox Model in particular will be introduced more in detail in the fourth chapter.

Introduction 3

research that holds that rebels do not benefit from attacks that they carried out against the civilian population.

This paper proceeds as follows. The second chapter will provide an overview of previous

research conducted on the link between violence against civilians by parties to armed conflict and

war outcomes, such as negotiated settlement, introduce the research gaps and present the

research question derived from these gaps. In the third chapter I will define the concept of one-

sided violence and lay out my theoretical argument explaining why increasing levels of rebel one-

sided violence are expected to lead to a quicker onset of peace negotiations. In the fourth

chapter, I test this claim empirically and conclude by a critical discussion of the findings in the

sixth chapter. The final chapter presents a conclusion and offers points of departure for further

research.

Previous Research on Violence Against Civilians 4

2. P

R

V

A

C

Throughout history, violence against civilians “was always a companion of human warfare” (Schütte 2015, 105). Nevertheless, it has long remained a minor research field within peace and conflict studies or social sciences more broadly. Prompted by conflicts such as the genocide in Rwanda and ethnic cleansing during the post-Yugoslavian wars, a growing number of researchers have started to focus on violence against civilians in the past two decades, mainly arguing that violence against civilians should be seen as a strategic choice in conflict. Scholars have also increasingly investigated the effect of such violence on the outcome of conflict to understand whether it is an effective strategy for parties to conflict or not. None of these studies have, however, taken into account the factor of time to account for the temporal effect of violence against civilians on war outcome, more specifically on the convening of peace negotiations.

Motivations behind violence against civilians

There are two main approaches generally considered to explain why conflict actors target civilians. The first approach views violence against civilians as a strategic tool to achieve war aims, while the second approach holds that such violence does not follow a strategic logic. With regard to the second approach, Humphreys and Weinstein (2006) find that groups that are unable to monitor the behavior of their members are more violent towards the civilian population.

Under such circumstances, attacks against civilians are possibly due to motives such as personal revenge (Kalyvas 2006, 58–59), a greedy attitude related to the looting of natural resources (Weinstein 2007, 7), or they are a manifestation of sadist tendencies of individual combatants (Mueller 2004, 9). Yet, these motives do not explain variance among groups and the large-scale use of violence against civilians in some cases (Valentino 2014, 98). Therefore, most scholars agree that attacks against civilians in armed conflict are primarily strategic in nature

. Several authors argue that parties to conflict use violence against the unarmed to improve their own bargaining position or push the opponent out of the war (Downes 2006; Hultman 2007, 2009;

Thomas 2014; Wood and Kathman 2014). Hultman (2007), for instance, concludes that rebels at the brink of defeat resort to one-sided violence to pressure the government into concessions, since violence imposes costs on the enemy. According to Fjelde and Hultman (2014), rebels tend to attack civilians in areas inhabited by the opponent’s constituency to weaken the enemy’s

4 These theories usually derive from the assumption that parties to conflict are rational actors. According to Jackson and Morelli (2011, 36), an action is rational when it maximizes the expected payoff out of the options available to an actor and relative to the actor’s believes about the potential consequences of the action. Applied to the issue of violence against civilians, this means that actors choose to attack the unarmed because they believe it brings them benefits as compared to not carrying out attacks. See more about rationalist explanations for war in Fearon (1995).

Previous Research on Violence Against Civilians 5

support base. Similarly, Balcells (2010) explains one-sided violence with conflict actors’

competition over local power in contested areas.

One-sided violence and outcome of conflict

Based on the assumption that violence against civilians in armed conflict is mostly strategic, scholars have raised the question of whether such violence is effective. A couple of studies have looked at its effect on the outcome of conflict. The findings from these investigations are contradictory. For example, Fortna (2015) finds that groups that systematically attack civilians with small-scale bombs are less likely to achieve their war aims, both in terms of military victory and concessions granted on the bargaining table. This is in line with further studies that conclude that terror tactics

are ineffective if rebel groups want to achieve their war goals (Abrahms 2006;

Ashworth et al. 2008). Alternatively, several scholars have argued that violence against civilians is an effective tool for conflict actors. According to Downes (2007), indiscriminate violence by government forces can help the incumbents defeat insurgencies under certain, rather restrictive, conditions. Another argument is that rebel one-sided violence instills fear among the population and causes reactions from governments that are beneficial for the rebels’ cause rather than detrimental (Kydd and Walter 2006). Wood and Kathman (2014) find a curvilinear relationship between rebel one-sided violence and the likelihood of a negotiated settlement, suggesting that the chances for a negotiated settlement increase with rising levels of violence up until a certain point, at which the chances for settlement decrease. It has also been concluded that suicide attacks help insurgents achieve strategic goals in more than 50 percent of the cases (Pape 2003).

Similarly, Thomas (2014) examines how civilian victimization is related to the participation of rebel groups in peace negotiations. She finds that the more terrorist attacks domestic rebel groups commit, the more likely they are to participate in bargains and to get concessions.

Identifying the Research Gap

The studies described above measure the effect of one-sided violence on the occurrence of a certain war outcome. What these studies leave unanswered, however, is the temporal dimension of the relationship. What if violence against civilians increases the likelihood of a certain war

5 Terrorism is a contested concept and often emotionally loaded. Schmid (2004) showed that in 1988, there were already more than 100 definitions of terrorism. Because of this variety of definitions, studies on terrorism are often difficult to compare. While some authors include non-civilian targets, most notably security forces, in their analyses (Abrahms 2006; Pape 2003), others define terrorism simply as violence against civilians aimed to influence a wider audience (Fortna 2015; McCormick 2003). Yet, only roughly one fourth of political science articles dealing with terrorism include civilians in their definitions (Weinberg, Pedahzur, and Hirsch-Hoefler 2004). This shows that terrorism and violence against civilians are generally understood as different concepts. I will nevertheless use some terrorism literature in this study because there are aspects in which both forms of violence might exert a similar effect (for instance the increased costs of war due to the violence).

Previous Research on Violence Against Civilians 6

outcome, but at the same time delays its entry into force? It should therefore be of interest to know something about the temporal effect of rebel one-sided violence to get a more differentiated picture of the consequences of such violence. This is the main research gap that this study seeks to fill. Moreover, previous literature has linked one-sided violence to war outcomes such as negotiated settlement and military victory, but none of them have investigated the effect of one-sided violence on the onset of negotiations. I argue that mechanisms explaining why parties to conflict sign an agreement (or not) do not necessarily explain why the negotiations start in the first place. Examining the start of negotiations may provide new insights regarding the relationship between one-sided violence and peace negotiations. A third aspect in which the present study differs from previous research is the measure of one-sided violence. Most of the mentioned studies investigate the effect of terrorism instead of violence against civilians, or they use a binary indicator answering only whether violence has occurred or not. Thomas (2014, 811) claims that the terrorist assaults recorded in her dataset mostly affected civilians, but it is uncertain whether these attacks were primarily aimed at harming civilians. It could be deemed possible that rebels sought to target security forces and government institutions, while accepting some harm inflicted on civilians in the form of ‘collateral damage.’ Moreover, the number of attacks does not tell us much about the number of victims; some rebel groups may conduct only a few attacks, but very well planned and therefore highly deadly, while others commit many attacks without causing significant damage. As will be explained more in detail in the third chapter, large-scale attacks with a high number of casualties are usually subjected to intense media coverage, and therefore put pressure on governments to react (for example by starting peace negotiations), while small-scale attacks do not attract the same attention. For this reason, the extent of violence should matter.

This study seeks to contribute to the scholarly debate by filling these gaps in previous

literature. The main purpose underlying this investigation is to shed light on the importance of

the factor time in studies on one-sided violence. The scope of this study is limited to rebel

violence against civilians. While I acknowledge that state violence is equally relevant, as

motivations for violence against civilians are usually different for governments and for rebels,

especially in current civil wars, an inclusion of state violence in this study may result in a less

powerful argument. Deriving from the research gaps identified in previous debate, the present

study aims to answer the following research question: Does the extent of rebel one-sided

violence in armed conflict affect the period of time until peace negotiations are initiated?

Theoretical Framework 7

3. T

F

This paragraph is dedicated to the development of a theoretical argument explaining the relationship between rebel one-sided violence in armed conflict and the timing of peace negotiations. I argue that rebels resort to attacks against the civilian population to raise the costs of war for the government. By doing so, rebels hope to coerce the opponent into negotiations and benefit from concessions in the bargains. Since governments fear further costs associated with additional rebel violence against civilians in the near future, they feel an urgent need to quickly convene negotiations in order to prevent more violence from occurring. This response is expectedly faster in democracies than autocracies, since democratic leaders need to guarantee the safety of the population in order to maintain popularity for reelection. In the following sections I define the concept of one-sided violence and then lay out the theoretical argument more in detail.

3.1. C

V

A

C

According to Wood (2016, 3), one-sided violence is “a wartime strategy in which a party to the conflict (or its agents) intentionally attempts to target and/or kill civilians and other noncombatants.” Compared to other conceptualizations of violence against civilians, this is a rather narrow one. The Uppsala Conflict Data Program (UCDP), for instance, does not make assumptions about the intentional (or not) character of violence (Eck and Hultman 2007). Yet, the UCDP definition excludes civilian casualties caused by direct battles between conflict actors, so that it only considers attacks directed against civilians. It is often impossible to distinguish between intentional and accidental violence, especially in attacks involving the employment of bombs or missiles. Therefore, violence against civilians, as referred to in the present paper, shall be understood as the use of armed force by parties to conflict against the civilian population, which occurs outside of direct battles between the conflict parties. Whether one-sided violence is intentional or not will not be of importance. The terms ‘violence against civilians’ and ‘one-sided violence’ will be used synonymously.

3.2. E

R

V

A

C

As mentioned before, many scholars agree that acts of one-sided violence perpetrated in armed

conflict are mostly of strategic nature. This means that rebel groups believe that the resort to

violence against civilians has strategic value and will prove effective to attain their goals. To

understand why rebels believe in the effectiveness of one-sided violence, it is important to know

what goals they pursue through conflict. Civil conflict is usually characterized by large capability

asymmetries between a more powerful government and weaker rebels (Wood and Kathman

2014, 689). Although there are certainly exceptions to this rule, in most cases governments have

Explaining Rebel Violence Against Civilians 8

more resources at their disposal to finance stronger and better equipped military forces than the rebels. Under such circumstances, in which the rebels are significantly weaker than the government, rebels are unlikely to attain a military victory. The second best outcome is then a negotiated settlement, through which the insurgents seek concessions. Following this logic, the intention of rebel groups should mainly be to increase the pressure on the government to negotiate or make concessions (Stanton 2009, 31–33). However, governments are generally averse to the idea of granting concessions to rebels, which they often consider illegitimate.

Therefore, rebels first need to convince the governments of the necessity to convene negotiations. In other words, rebels must achieve some process-oriented goals first before they can pursue their primary objective (Wood and Kathman 2014, 689). If the primary objective is to get concessions in bargains (which may for instance consist of impunity for former crimes, a transformation of the economic or the political system, and political participation) then getting the government to the bargaining table would be a process-oriented goal.

The key for rebel groups to convince governments of the necessity to negotiate lies in a manipulation of the costs of war balance. According to Hultman (2007, 209), “[t]he government’s assessment of the costs for fighting is of great importance throughout the conflict and determines whether it is willing to continue fighting or if it is better off by making concessions and settling the conflict with the rebels.” One way in which insurgents can increase the government’s costs of war is by posing a serious challenge to its military on the battlefield. Battles between parties to conflict reveal the relative power of both sides (Reiter 2003, 33); if rebels manage to impose high costs on the enemy by ambushes and attacks, the government may be convinced that the costs for defeating the opponent may be too high (Hultman 2007, 209). Under such circumstances, governments should be more inclined to accept negotiations compared to scenarios in which the rebels cannot demonstrate that they are a serious military challenge.

However, insurgents do not always have the capacity to significantly hurt government

forces in combat. I argue that rebels can increase the costs of war for a government by resorting

to violence against the civilian population. The killing of civilians does not require strong military

power; in fact, even small and weak groups can cause high numbers of casualties among the

population. Attacks against civilians are less risky than attacks against military targets since

civilians are typically unarmed and poorly prepared to defend themselves against armed

aggressors. By carrying out such attacks, rebels also demonstrate that they are willing to commit

the most ruthless atrocities to hurt the opponent. While one-sided violence inflicts immediate

costs on the population and its government, the notion of expected costs of a government

reaction is arguably even more important (Wood and Kathman 2014, 691). In order to make sure

Explaining Rebel Violence Against Civilians 9

that civilians are safe from attacks in the future, governments would have to end the conflict with the rebels in some form (Fortna 2015, 526).

Many authors have referred to the ‘costs of war’ argument without a thorough discussion of what these costs are and how they arise. In order to facilitate a better understanding of the concept, I present three kinds of costs: political costs, military costs, and economic costs. The argument that one-sided violence is politically costly takes root in the general acceptance that governments are responsible for the security and protection of their civilian population (Hultman 2007, 209). Citizens may punish a government that failed to guarantee security by removing it from power in the next elections (Hultman 2012, 167). Moreover, violence against civilians may give rise to hardliners within the government, resulting in power struggles and eventual removal of more moderate incumbents. Rebel one-sided violence also arguably increases the military costs of war for governments. By spreading fear and terror, rebels impede governments from maintaining territorial control (Hultman 2007, 209–210). Civilians, intimidated by attacks, will align their behavior with rebel orders since they have learned that disobedience will be punished (Kydd and Walter 2006, 50). This intimidation can help rebels gain control over large areas. They may also use territorial control to intermix with the civilian population, making it difficult for government forces to distinguish between rebels and noncombatants (Engeland 2011, 132). The government’s territorial control is further eroded by forced displacement, which is mainly caused by one-sided violence committed by conflict actors (Combaz 2016, 2). Situations of massive human mobility are often characterized by chaos and the inability of governments to control people flows. It is costly for governments to regain military control over these areas. The ‘costs of war’ argument can be expanded further beyond the political and military sphere. One aspect so far neglected by scholars is how one-sided violence relates to (expected and immediate) economic costs of war. It may be necessary to raise the military budget, for instance, to achieve a breakthrough on the battlefield; or if a government decides to strive for a peaceful settlement, concessions such as adjustments to the political economy of a country may be necessary to force the opponent to lay down their arms. Thus, attacks against civilians may as well raise the economic costs of war on the government side.

To summarize, I have proposed that rebels use one-sided violence to increase the

government’s political, military, and/or economic costs of war and to eventually get concessions

in subsequent bargains.

Explaining Why Governments Seek Negotiations 10

3.3. E

W

G

S

N

Despite the increased costs due to rebel one-sided violence, governments do not have to favor negotiations; they could instead rely on another strategy to defeat the rebels or prevent one-sided violence. It is therefore important to shed light on the types of mechanisms governments have at their disposal, and to explain why rebel violence against civilians makes the regime choose negotiations and not other strategies. When confronted with non-state groups using one-sided violence, states have various options to react: 1) act with restraint; 2) use military force indiscriminately against rebels; 3) use military force selectively against rebels; 4) propose or agree on negotiations with the opponent. Each of these options have specific advantages and disadvantages and will now be discussed separately.

First, a government can act with restraint. This can be a way of demonstrating to the opponent that violence against civilians does not guarantee a government response. Government restraint also means that the rebels can continue to operate with impunity; they may simply increase their attacks to a level where the costs for a government have exceeded the benefits of inaction (Siqueira and Sandler 2006, 879). Impunity could then lead to more violence against civilians until the regime switches to another strategy. Moreover, a government that fails to react sufficiently to violence against their civilian population loses popular support, as it seems the government is unable or unwilling to provide protection (Rosendorff and Sandler 2004, 659).

Second, a government can respond with indiscriminate use of military force to rebel one-sided violence to defeat the opponent. Theoretically, the chances of a military victory for the government might increase the larger the counteroffensive is (Lake 2002, 22). At first, the civilian population may support such a reaction, especially when it had experienced high levels of violence by the rebels. However, while a harsh reaction can help defeat the enemy, it has serious downsides. Since large military operations often include the violation of human rights, restrictions of civil freedoms, and unintended collateral damage including civilian harm, they risk creating additional grievances among the civilian population (Rosendorff and Sandler 2004, 658).

Indiscriminate state violence may thus drive formerly moderate civilians into the hands of rebels, as it makes the government itself appear extremist (Lake 2002, 20; Thomas 2014, 816).

Furthermore, the international community might increase pressure on governments whose reaction to rebel one-sided violence is considered disproportionate. Besides sparking diplomatic pressure, massive retaliation campaigns by the authorities also increase the likelihood of external intervention (Stanton 2009, 59). Indiscriminate use of force is also economically costly.

Governments may have to cut social programs and other services to finance this strategy, risking

the impoverishment of great segments of their population (Siquiera and Sandler 2006, 880). It has

Explaining Why Governments Seek Negotiations 11

been found that indiscriminate government violence only succeeds in defeating insurgencies under certain conditions, for instance when rebels draw their support from areas with small populations (Downes 2007, 440). Third, governments facing one-sided violence within their territories can increasingly use selective violence against the rebels and their supporters. By targeting individuals who provide the rebels with information or basic goods, a government indirectly restricts the opponent’s fighting capacity in two ways: it discourages civilians from joining the rebels, and deprives the enemy of recruits, supplies, and intelligence (Downes 2007, 421). But selective violence is time consuming (Ibid., 441; Jackson 2007, 74). It requires fine- grained information about the enemy gathered by intelligence agencies, which often have to be reinforced before selective violence can be effectively employed against an opponent (Kitson 1971; Jackson 2007). A government that is entangled in conflict with rebels that target civilians is pressured to prevent further attacks as soon as possible. Failure to do so may result in the loss of popular support, territorial control, and the replacement of government officials by hardliners.

Furthermore, selective violence is a very expensive measure in economic terms (Kaplan 2012).

None of these three strategies solves one dilemma governments are trapped in after rebel attacks against civilians occur: the urgency to prevent further attacks of the same kind in the near future. Herein, I argue, lies the main comparative advantage of the fourth strategy governments can turn to: peace negotiations. Following Höglund and Druckman (2013, 416), peace negotiations shall be defined as “a process of dialogue and bargaining between adversaries aimed at reaching a joint decision to bring an end to or solve a violent conflict.” Negotiations are different from ceasefire agreements, which regulate the conflict actors’ abdication of violence. Ceasefires are often a basis or prerequisite for more substantial peace negotiations (Chounet-Cambas 2011, 19). Rebel violence against civilians puts pressure on the government because it imposes costs on the regime. Due to the possibility to implement a ceasefire, negotiations are the only strategy by which governments can, if implemented effectively, prevent further rebel attacks against the civilian population in the short term. The urgent nature of a government response to rebel one- sided violence not only explains why governments favor negotiations over other strategies, but also why they act quickly: politicians know that rebels may again target civilians, by which they would impose further costs on the government. Since rebels could commit further attacks at any day, incumbents consider it important to prevent more violence as soon as possible. For this reason, governments favor negotiations, and I expect them to initiate those quickly in order to implement a ceasefire and prevent more civilian casualties.

Negotiations are not without costs for governments either. They may reward rebels for

violent behavior (Thomas 2014, 806), signal the government’s weakness (Pruitt 2006, 381),

Explaining Why Governments Seek Negotiations 12

encourage political hardliners to overturn the regime (Ibid.), and even perpetuate violence against civilians committed by other groups that seek concessions from the government (Toros 2008, 408; Thomas 2014, 806). However, I argue that the advantages from negotiations outweigh their disadvantages in the context of rebel violence against civilians, mainly because peace talks are probably the most effective way to quickly reduce or prevent violence, as compared to alternative strategies.

My argument entails yet another important nuance: the level of one-sided violence. I posit that it matters less whether there is violence against civilians or not, but rather the extent of violence. Parties to armed conflict usually accept some civilian losses during conflict; a few civilian deaths do not raise the costs of war significantly for any side (Stanton 2009, 39). Cases with very few casualties may not even be reported by the media, or at least not to a degree that they would raise public or international pressure on the government considerably. Such pressure is, in turn, likely to be exerted when rebels employ massive one-sided violence. Moreover, governments may consider rebels targeting only a low number of civilians as too weak to actually threaten their power or to raise their costs of war. Instead, previous literature has argued that civilian suffering and the threat of future damage must be severe to impose sufficient costs on the opponent (Ibid.). In order for one-sided violence to signal rebel resolve and strength to the government, low levels of violence are thus insufficient. Rebels demonstrate their power to hurt a government only if they employ relatively high levels of one-sided violence. The causal mechanism for the whole argument is illustrated in Figure 1.

Fig. 1: Causal Mechanism

Recapitulating my argument, rebels attack civilians strategically to enforce negotiations. Facing increased costs of war and therefore an urgent need to prevent further attacks, governments are expected to concede and convene negotiations as soon as possible. The more violence rebels use,

Violence against civilians by rebel

groups

Time until onset of peace negotiations Increase of

government’s costs of war

Urgency for government

reaction

Increase IV

Decrease DV

IV DV

Conditioning the Argument: Regime Type 13

the higher the costs of war on the government side and the more likely are governments to agree to negotiations. This leads to my first hypothesis:

Hypothesis I: Higher levels rebel violence against civilians in armed conflict lead to a sooner onset of peace negotiations.

3.4. C

A

: R

T

Having explored the strategic logic of why rebel groups attack civilians and why governments agree to negotiate in such a context, one important question still remains: why do some governments favor a negotiated solution when confronted with rebels that target civilians, while others pursue another strategy? I argue that a government’s regime type largely conditions the choice of a strategy to respond to rebel one-sided violence.

Democratic leaders depend on public support for reelection (Hultman 2012, 167). When a government is not able to protect its civilian population, this support is likely to decrease.

Citizens will then punish their incumbents for the lack of protection by removing them from power in upcoming elections (Ibid.). Democratic regimes are vulnerable to domestic public pressure and rebels increase the political costs of war for democratic leaders by attacking those the government relies upon to stay in power (Stanton 2009, 41). Since free, fair, and transparent elections are seldom held in autocratic states, autocratic incumbents do not have to fear the population’s grievances at the ballot box. Thus, rebel violence against civilians is expected to be more politically costly for governments in democratic countries, in comparison with autocratic ones. Autocratic regimes are also, to some extent, constrained in their decision on how to react to rebel violence against civilians. However, they have two main mechanisms at their disposal to enforce the loyalty of those that may threaten their power, thereby reducing the political costs resulting from such attacks. First, autocratic governments can use repression in the form of military force against anti-regime protesters, imprison or silence critics, and censor the media to prevent information on rebel violence against civilians from spreading. Second, autocratic governments can buy off powerful individuals through patronage networks to make sure that such individuals will not challenge their power (Escribà-Folch and Wright 2015, 165). I argue that none of these options presented could be effective mechanisms in full democracies, and are only of limited value in unstable hybrid regimes.

The main characteristics of democracies also explain why democratic governments prefer

negotiations to alternative responses to rebel one-sided violence. Acting with constraint would

most likely lead to the government’s removal from power through elections, especially if

competing candidates propose better security concepts. Since the media in democratic states is

Conditioning the Argument: Regime Type 14

less restricted than the media in autocratic states, a government’s indiscriminate use of violence against rebels, collateral damage, and war crimes might be better covered in a democracy. Going further, democratic leaders can be held accountable for such actions, not only through elections, but also through legal conviction. The use of more selective violence should also be rather unpopular for democratic governments, since such measures take time, probably longer than a democratically elected leader remains in office. Conversely, autocratic regimes seldom have these time constraints. What remains for democracies is the path of negotiations. By initiating negotiations with the rebels, democratic leaders have both the opportunity to reduce or stop violence in the short term and to portray themselves as heroes who brought peace to the country, which would increase their popularity and better secure their reelection.

Thus, rebel violence against civilians is expected to impose higher costs of war on democratic governments than on autocratic ones. The urgency of responding to the violence is therefore higher in democracies, which makes their governments concede to the initiation of peace talks more quickly than their autocratic counterparts. This leads to the following, conditional hypothesis:

Hypothesis II: Higher levels of rebel violence against civilians in armed conflict lead to a

sooner onset of peace negotiations in countries with a democratic regime

compared to states governed by an autocratic regime.

Research Design 15

4. R

D

The hypotheses elaborated in the previous chapter will be tested using a quantitative method.

While qualitative research is often exploratory, used to examine phenomena previous literature has overlooked and to test or build a new theory, quantitative methods are often confirmatory, i.e. employed to test hypotheses derived from theory (Gerring 2007, 39–40; Teddlie and Tashakkori 2009, 363). The main research gap this study seeks to fill is the relevance of time:

previous studies have not investigated whether one-sided violence accelerates or delays negotiations. The focus is therefore mainly on the relationship between the two variables and less on the causal mechanism linking these, although the theoretical background is important to understand the link. The decision for a quantitative study was also driven by the availability of accurate and precise data on one-sided violence, which reduces many of the problems quantitative research is associated with, especially measurement errors. Below I present the concrete method employed, before describing the data used in the empirical analysis.

4.1. S

M

I use a Cox Proportional Hazards Model to test the hypotheses elaborated in Chapter 2. The Cox Model is a special type of duration models

. Initially, duration models were used in biomedical research. Later, they have increasingly attracted the attention from scholars in other fields such as social sciences (Danacica and Babucea 2010, 339; Allison 2009, 369). Duration models are different from conventional linear regression models in a couple of aspects. First, they analyze the event history. Like duration models, linear regression models focus on the relationship between two variables of interest. What conventional models do not take into account, however, are differences of the relationship over time

(Allison 2009, 370). Duration models investigate the event history and changes over time because their dependent variable is ‘time until event’, not

‘event’ as in most other models. Second, duration models can deal with censored observations.

Censoring refers to situations in which a subject’s full event history is unobserved, which usually occurs in two ways: left-truncation, which disregards event history prior to the first observation under study, and right-censoring, which leaves event history following the last observation in the study unobserved (Box-Steffensmeier and Jones 2004, 16). Most regression models do not

6 Besides “duration analysis”, the literature also often refers to the approach as “survival analysis”, “event history analysis”, “failure time analysis”, “hazard analysis”, and “transition analysis”.

7 Assume that a study is conducted on the effect of financial aid on the relapse of former inmates into criminal activities. 25 percent of the participants in the study relapse. It is reasonable to assume that those subjects who relapse quickly after their release have a higher propensity towards crime than those that relapse later. A linear regression treats the subjects equally, while duration models take into account the event history and therefore detect differences (Allison 2009, 369–370).

Statistical Method 16

distinguish between cases where right-censoring happens, and cases that simply never experience a ‘failure’ (Eck 2009, 375)

. Failure in the context of duration models means the occurrence of the event under study (applied to this study the onset of peace negotiations). The time until the failure, or event, is referred to as the time the subject is at risk (of experiencing the failure), or survival time. Censoring problems, both right-censoring and left-truncation, are automatically introduced by the scope conditions set by the researcher: by limiting the time period under study, for instance, observations both before and after the period under investigation are ‘censored’.

Third, duration models can account for situations in which a subject experiences an event more than once. Peace negotiations often break down and start up again. If the study was limited to the occurrence of a first event only, a considerable amount of information would get lost because all observations after the first failure would be excluded from the analysis. Not accounting for this could, depending on the purpose and scope of a study, lead to biased results.

Cox Models are the most commonly used duration models (Singh and Mukhopadhyay 2011, 147; Allison 2009, 375). This is due to the semiparametric nature of Cox Models: they do not make prior assumptions about the baseline hazard but instead leave the particular form of the duration dependency unspecified. This makes it a particularly flexible tool compared to parametric duration models, for which the researcher needs to specify how the hazard rate will change over time prior to the analysis

(Griffin 1995, 316; Box-Steffensmeier and Jones 2004, 21–22). If the real baseline hazard is different from the assumptions made by the researcher, parametric models may lead to incorrect results (Bergström and Edin 1992, 22). A further advantage of the Cox Model is its handling of tied data (Box-Steffensmeier and Jones 2004, 53).

In large datasets it is not uncommon to have simultaneously occurring failure times. Cox allows for the addition of different treatments to handle such tied data. The most accurate of these treatments is the Efron (1977) method, which will therefore be used in this study (Box- Steffensmeier and Jones 2004, 59).

8 To understand the concept of censorship, additional information may be useful. The dependent variable in duration models is “time until event”. The complete time is not always observable, however; subjects may be at war already before the time period under investigation (left-truncation). Similarly, the event may happen after the end of the study period (right-censorship). Also, there may never be an event. Linear models “think” that censored observations and observations that never experience an event are the same; duration models distinguish between these two scenarios.

9 What parametric model is chosen depends on the expected baseline hazard: a Weibull model assumes a monotonic baseline hazard, while log-logistic or exponential models assume a non-monotonic distribution. It has been proposed that parametric models be preferred over the Cox Model if the researcher is primarily interested in the shape of the hazard rate over time, e.g. to see whether the effect of the IV is much stronger or weaker at a later point in the study period than at the beginning, and not in the overall relationship. Since the focus of this study lies on the general relationship between IV and DV and not the particular shape of the correlation curve, the Cox Model is the more appropriate choice (Griffin 1995, 316).

Statistical Method 17

One of the features of duration models is the possibility of investigating multiple-failure data.

Cox models facilitate the analysis of both ordered and unordered multiple-failure events. In an unordered events setup, the time order of the events does not matter. Whether a failure occurs first or after a previous failure is not important. This is not the case in the present study:

negotiations at time k

+1 would potentially not take place without previous negotiations that had started at time k

. For this purpose, I use an extended version of the Cox model that allows for an analysis of ordered events, the so-called conditional risk-set model proposed by Prentice, Williams and Peterson (1981). It is conditional in the sense that it assumes that a subject is not at risk of experiencing a later event (k

+1) until the previous event (k

) has occurred (Box- Steffensmeier and Jones 2004, 161; Cleves 1999, 35). In such a model setup, the observations are stratified according to the number of failures. I thus created a variable that assumes values of 1 for observations before a first failure, 2 before a second failure, and so on. In addition to the stratification of events, I will make use of robust standard errors to solve potential problems with time dependency of the variables included in the analysis; such time dependency could arise because some variables are time-varying, and therefore observations could be conditionally dependent. Robust standard errors are more accurate in this case (Eck 2009, 378; Lin and Wei 1989, 1074–75).

The Cox model provides hazard ratios, which indicate the risk a subject incurs of experiencing a failure at any moment, given that the failure has not occurred yet (Box- Steffensmeier and Jones 2004, 15). They are similar to odds ratios in linear regression in the sense that the values refer to probabilities. A hazard higher than 1 means that the risk of failure increases, or similarly, that the time until failure decreases. If the hazard ratio is lower than 1, the time until failure gets longer. A hazard ratio of 1 indicates that the covariate has no effect on the likelihood of failure (Ibid., 63). Applied to this study, the hazard ratio reflects the likelihood that negotiations are initiated between rebels and a government at any point in the observed period of time, given that negotiations have not happened yet or that the subject has not left the risk pool because war has not ended after previous negotiations

. Due to the fact that hazards do not indicate how long it takes until subjects “fail”, the term ‘duration model’ can be slightly misleading. From the hazards themselves, it can only be inferred if an event is likely to happen sooner or later with varying levels of the independent variable, but not by how much.

10 The hazard ratio is interpreted as a one-step increase in the independent variable. For example, if the hazard is 1.5, then a one-step increase in the independent variable leads to a 50 percent higher chance that the event occurs (that there is a ‘failure’).

Data 18

To test the second hypothesis (examining the conditional effect of regime type), I will use a split- set version of my dataset. The Cox Model is then run for democracies and autocracies separately to compare the effects

.

4.2. D

The empirical analysis is based on disaggregated time-series data on African intrastate armed conflicts between 1989 and 2009. Both the spatial and temporal frame were chosen on the basis of data availability

. The unit of analysis is dyad-month. A dyad consists of a government and a non-state actor that are entangled in armed conflict against each other. Every dyad-month is an observation, while the whole set of observations for one dyad add to one case (or subject). Peace negotiation onsets are frequent and sometimes occur more than once per year and case, which is why dyad-year data would be too inaccurate and facilitate problems with endogeneity, i.e. the difficulty to determine whether negotiation onset is a consequence of rebel one-sided violence, or vice versa.

Dependent Variable

This study is the first attempt to investigate the onset of peace negotiations with a quantitative approach. The gap in previous literature also implies a lack of data on negotiation onset.

Therefore, I had to create the dependent variable used in this study myself, based on data provided by Thomas (2014). Thomas’ dataset includes information on whether governments and rebels were engaging in negotiations or not for African conflicts between 1989 and 2009, measured for each month in the study period, and represents the most comprehensive dataset for peace negotiations so far with a total of 403 months of negotiations. Since I rely on this data, my study uses the same geographical and temporal limits; due to time and resource constraints, information gathering for another continent was beyond the possibilities and the scope of this study.

The original data encompasses formal negotiations, which require the presence of the main conflict dyad, but exclude secret back-channel negotiations and cases of third-party mediation in which no actual talks between the conflict actors have taken place (Thomas 2014). To obtain the dependent variable as used in my study, I first created a variable indicating the onset of

11 I considered adding an interaction term to investigate the conditional relationship of regime type instead of splitting up the data, but decided against it. The interpretation of the results is more straightforward in the split- model, and it is more parsimonious.

12 While data for the independent variable (UCDP) was available on conflicts worldwide between 1989 and 2015, the data used to create the dependent variable (Thomas 2014) was limited to African conflicts between 1989 and 2009. Therefore, Thomas’ dataset also sets the spatial and temporal frame in my study.

Data 19

negotiations. The negotiation variable in Thomas’ dataset reflects the holding of negotiations for every month and not the onset. Conflict actors sometimes negotiate for several months or even years; thus, not every month that is coded as experiencing negotiations actually refers to the onset of talks. I decided to classify as a new onset only observations coded as experiencing negotiations that occurred after at least two months without negotiations. The reason for this is that while it is likely that negotiations are postponed for a month, a two-month break would be quite a long time for a simple break; the more time has passed, the more likely it is that negotiations have been disrupted and are reconvened, indicating a new onset of negotiations. Some kind of a threshold must be applied here, and I consider the two-month threshold as a plausible middle ground

. Based on the information from when negotiations started, three new variables were coded as required for the Cox Model set-up described earlier, reflecting the complete time a subject is in the risk set, the time from the previous event

, and the number of failures (stratification). While all this information is integrated in the Cox Model, the actual dependent variable is time until peace negotiation onset.

Independent Variable

One-sided violence committed by rebel groups is the independent variable used in this study. I extract the variable from the UCDP Georeferenced Event Dataset (GED), which contains temporally and spatially disaggregated data for different event types for the period from 1989 to 2015, including for violence against civilians (Sundberg and Melander 2013). UCDP defines one-sided violence as

“the use of armed force by the government of a state or by a formally organized group against civilians which results in at least 25 deaths” (Pettersson 2012, 2). The data excludes casualties resulting from battles between the conflict actors as well as cases with less than 25 deaths per calendar year. To obtain monthly data for every rebel group in my dataset, I aggregated single observations for every month. Only African rebel groups involved in armed conflict with a government were considered, in line with the scope conditions of this study referred to earlier. A few adjustments were made to the UCDP data. When two or more rebel groups jointly conducted attacks against civilians, the number of casualties was split up among the perpetrating groups. The same strategy was chosen when attacks were not attributed to a specific actor, but to a collective entity comprising several rebel groups

. The time period was aligned with the dependent variable, and

13 Robustness checks will be carried out, using both a one-month and a three-month threshold for a new onset.

14 I followed the steps for the conditional risk set model measuring the time from entry (Cleves 1999, 37-38).

15 In Burundi, over a couple of years attacks against civilians perpetrated by different Hutu rebel groups were attributed to “Hutu rebels” instead of a concrete group. Following the Rwandan genocide, several Hutu militias were active in neighboring Burundi and DRC, which complicated the attribution of events to a specific actor (Pettersson 2014).

Data 20

therefore restricted to 1989 to 2009. The location of the attacks was further disregarded. In some cases, rebels attacked civilians in a country other than the one where they were involved in conflict with a government. For most facets of the ‘cost of war’ argument it does not matter whether the civilians killed by rebels were citizens of the State having a conflict with the insurgents or of another one. The role of one-sided violence as a signaling device for increased expected costs of war on the government side, for instance, should not depend on the nationality of the victims. Overall, the dataset contains 104 rebel groups, of which 34 resorted to violence against civilians within the 21-year time frame under investigation.

Control Variables

Seven control variables were added to the models. The inclusion of control variables followed thorough theoretical considerations to avoid creating what Achen (2005, 336) called “garbage-can lists of variables”. Only variables that could possibly act as confounders, i.e. that are expected to influence both the dependent and independent variable, were included.

First, some governments show little concern for the physical safety and security of their

population, or parts of it, and carry out attacks against civilians. Under these circumstances, rebel

one-sided violence is not expected to significantly increase the government’s costs of war, since

the authorities will not be concerned about the power of rebel groups to hurt civilians when they

carry out such attacks themselves. Also, governments using violence against their own civilians

are probably less willing to compromise, and therefore peace negotiations should be less likely. I

include a dummy variable for government one-sided violence, indicating whether governments use one-

sided violence or not. The data comes from UCDP GED. Second, rebel groups suffering losses

in combat have been found to more likely attack civilians (Hultman 2012). Governments facing

rebels that are on the brink of defeat might strive for a military victory rather than a negotiated

settlement because chances for a government victory increase the more rebel soldiers the

government forces kill. I therefore include a measure for rebel battle-related deaths, also taken from

the UCDP GED. Third, competition among several rebel groups over territorial control or

resources could lead to more violence against civilians, since groups may punish civilians

collectively if they collaborate with the enemy. Similarly, governments hesitate to negotiate with a

rebel group when also involved in fighting with other groups because these other groups might

then demand similar concessions as the first group. To account for this effect, I count the number

of rebel groups involved in conflict with a government for every country and month. This

information is extracted from the UCDP GED. Fourth, the stronger an insurgency, the more

important is it for the group to establish good relations with the civilian population, whom it

Data 21

intends to govern in the future. One-sided violence to signal resolve is counterproductive for strong rebel groups because it makes these groups lose credibility. Moreover, performance on the battlefield signals the group’s strength to the government, and the stronger the rebels, the more incentives a government has to rapidly initiate negotiations before government forces are defeated by the insurgents. To control for relative rebel strength, I use a measure from the Non-State Actor dataset (Cunningham, Gleditsch, and Salehyan 2009). The variable is categorical and indicates whether a rebel group is (i) much weaker, (ii) weaker, (iii) at parity, (iv) stronger or (v) much stronger than the government. Fifth, rebel groups exerting control over territory should be more restrained in their attacks against civilians. Instead of attacking civilians, it would be more beneficial for the rebels to act as a de-facto government to obtain some form of income through taxes and to draw support from the population under their control. Governments are hesitant to negotiate with rebels controlling territory since they fear that the opponent might claim territorial independence, and therefore threaten the State’s territorial integrity. A measure of territorial control from the previously mentioned Non-State Actor dataset is used to account for this effect. Sixth, third parties mediate between conflict actors to stop violence, for instance when these parties fear that the conflict spills over into other countries or across a region. When mediation efforts take place, rebels should be expected to reduce attacks against civilians to gain the sympathies of the mediators, which then may propose more or better concessions for the rebels. Mediation aims to make parties negotiate and should therefore increase the likelihood of negotiations. I include a measure of third-party mediation from the UCDP Categorical Variables dataset (Cederman, Min, and Wimmer 2010). Finally, rebel groups that incur the majority of their income from natural resources often attract opportunistic recruits who are more likely to target civilians than more ideological fighters (Weinstein 2007). Governments may hesitate to negotiate with these groups because they fear losing access to the resources, which they aim to regain military, through concessions with rebels. To account for lootable resources, I use replication data from Wood and Kathman (2014), who aggregated original data from Lujala, Gleditsch and Gilmore (2005) as well as Lujala, Rod and Thieme (2007). The variable includes gemstones, such as diamonds, as well as drugs.

For the split-model, which examines autocracies and democracies separately, I include a regime type measure from the Polity IV project (Marshall, Gurr and Jaggers 2016). This is, strictly speaking, not a control variable, but an auxiliary variable to split up the dataset to test the

16 The UCDP Categorical Variables dataset ran only until 2008, and Wood and Kathman’s compiled data for lootable resources only until 2003, which led to many missing observations in my dataset. After comparing the values for previous years, I noticed that there was hardly any change over time, which made it possible to add the missing values for each subject, based on the values from previous years. For a few subjects I conducted own research, for example when information on territorial control was missing completely.