Comparing conflict intenstiy measures using time-series analysis and random forest

(1)

Comparing Conflict Intensity

Measures Using Time-Series

Analysis and Random Forest

A.M. Rentier

Univ

ersi

ty

of

Amster

dam

(2)

Comparing Conflict Intensity

Measures Using Time-Series

Analysis and Random Forest

by

A.M. Rentier

Master of Science

Political Science, International Relations

Graduate School of Social Sciences

University of Amsterdam

Supervisor:

Dr. M. Medeiros

Second assessor:

Dr. U.E. Daxecker

Student number:

11851937

Date final version:

21-06-2019

(3)

Abstract

While civil conflict has attracted much attention over the past decades, the share of research that has been dedicated to conflict intensity is relatively small. Knowing the devastating effects conflict can have on states and their civilian societies, the need to study why some conflicts escalate into full-blown civil wars while others do not seems clear. Two conflict intensity measures from the literature are reviewed, analyzed and compared: protests and rebellions (MAR) and battle deaths (UCDP). A thorough evaluation and comparison of the data shows differences in the spread of conflicts across geographical regions and through time. In the analysis Random Forest, a machine learning method relatively new to Political Science, will be introduced to investigate the determinants of conflict intensity in addition to OLS regression. Results from both methods show that population size and the number of groups are positively correlated to conflict intensity. GDP per capita is negatively related to conflict intensity. While regime type seems to be insignificant in the OLS regression, the Random Forest shows that it is actually an important determinant for conflict intensity, but possibly not in a linear way. Random Forest has proven to be a valuable addition to the analysis and is recommended for further use in the Political Science field. In conclusion, both differences and similarities between the two different conflict intensity measures have been identified using OLS regression and Random Forest, showing the importance of justification when choosing between these two measures.

(4)

Preface

This thesis is the last requirement to fulfill the graduation requirements for the degree Master of Sci-ence in International Relations at the University of Amsterdam.

First of all, I would very much like to thank Dr. Mike Medeiros, my supervisor from the University of Amsterdam, for his supervision. When I met Dr. Medeiros, I was looking for an appropriate thesis project that would allow me to incorporate my knowledge from my previous education in my thesis for Political Science. Having talked already to a number of other supervisors, Dr. Medeiros was the first to respond directly with a yes to my ideas and to even start giving me ideas. I am very grateful for the opportunity he has given me to pursue this methodological thesis in which I was able to incorporate a method that has not been used much in Political Science.

Secondly, I would like to thank my family and partner for their support during the project. I would like to thank my parents, Alex en Karen, for always supporting me throughout my university journey. Lastly, I would like to thank my second reader, Dr. Ursula Daxecker, for her time to read and review my work.

I wish you much pleasure reading my Master Thesis.

A.M. Rentier Amsterdam, June 2019

(5)

3 Research Design 16 3.1 Data. . . 16 3.1.1 Conflict Intensity. . . 16 3.1.2 Regime Type. . . 19 3.1.3 Poverty. . . 20 3.1.4 Population Size. . . 20 3.1.5 Ethnic Fractionalization. . . 20 3.1.6 Region . . . 20 3.2 Descriptive Statistics. . . 21 3.3 Method of Analysis . . . 23 3.3.1 Linear Regression . . . 23 3.3.2 Random Forest . . . 23 4 Results 26 4.1 Comparing Measures. . . 26 4.2 Linear Regression. . . 29 4.2.1 Robustness Checks . . . 30

4.2.2 Discussion of OLS Results . . . 30 iv

(6)

Contents v 4.3 Random Forest. . . 32 4.3.1 Discussion of Random Forest Results . . . 34

5 Conclusion 36

A Appendix 39

A.1 World Map Conflict Intensity . . . 39 A.2 Definition of Armed Conflict. . . 40 A.3 Countries. . . 41

(7)

List of Figures

3.1 Example of Data for Colombia . . . 22

3.2 Example of Data for United Kingdom . . . 22

3.3 Decision Tree predicting the number of battle deaths . . . 24

4.1 Total worldwide conflict intensity 1989-2003 . . . 26

4.2 Conflict Intensity per Region . . . 27

4.3 Accumulated conflict intensity per country (UCDP) 1989-2003 . . . 28

4.4 Accumulated conflict intensity per country (MAR) 1989-2003 . . . 28

4.5 OOB error rate for Random Forest on Battle Deaths data . . . 32

4.6 OOB error rate for Random Forest on MAR data . . . 33

4.7 Two Variance Importance Plots of Random Forests on Battle Deaths Data. . . 33

4.8 Variance Importance Plot of a Random Forest on MAR data . . . 34

4.9 Distribution of the minimal depth for a Random Forest on Battle Deaths data . . . 35

4.10 Distribution of the minimal depth for a Random Forest on MAR data . . . 35

A.1 Accumulated conflict intensity averaged per country (MAR) 1989-2003 . . . 39

(8)

List of Tables

3.1 Coding Conflict Intensity Measure Minorities at Risk. . . 17

3.2 Descriptive Statistics UCDP data 1989-2003 . . . 21

3.3 Descriptive Statistics MAR data 1989-2003. . . 21

4.1 Top five countries with highest accumulated conflict intensity (1989-2003) . . . 27

4.2 Results OLS . . . 29

4.3 Results OLS Robustness Checks . . . 31

A.1 Overview Number of Years with Conflict per Country (A-D) . . . 41

A.2 Overview Number of Years with Conflict per Country (E-P) . . . 42

A.3 Overview Number of Years with Conflict per Country (N-U). . . 43

A.4 Overview Number of Years with Conflict per Country (V-Z) . . . 44

(9)

1

Introduction

1.1. General Introduction

At first glance, the list of ongoing civil wars shows that a shocking number of countries are still torn by fighting. Syria, Afghanistan, Darfur: some of these wars are known throughout the world, and have had devastating effects on millions of people who are forced to seek refuge in neighbouring countries or even on other continents. The state of South-Sudan, independent from Sudan since 2011, has known civil war for almost its entire existence. The Rohingya have fled Myanmar after ethnic violence against the minority escalated again in 2017, leading to more than half a million people crossing the border with Bangladesh. Hundreds of villages were destroyed and thousands of Rohingya were killed or died during their flight across the border (BBC2018). Furthermore, some civil wars have been left almost totally unrecorded, such as the civil war in Yemen, where three quarters of the population is in need of humanitarian aid (UNHCR 2019). While civil conflict nowadays seems to be concentrated in African and Asian countries, all regions in the world have experienced their share of conflict; some are even teetering on the verge at present, such as Venezuela. A new generation of European citizens is growing up without any remembrance of the civil wars that savaged the former Yugoslav states, while their coevals in Afghanistan have yet to experience worry-free peace in their country.

Civil conflict has been a subject of great interest to the community of Political Scientists all over the world for the past few decades. Since 1946, civil war has been the most ubiquitous type of war, which has led to an enormous increase of literature on civil war. Sambanis(2002:218) tries to summarize and structure the quantitative literature on civil conflict and divides the literature in three parts, related to the main phases of war: civil war onset, duration, and peacebuilding after the war. Even though

Sambanis(2002:219) does not claim his literature review to be exhaustive, it is clear that the literature on conflict intensity is underrepresented. Civil conflict comes in different sizes, from small rebellions with only a handful of victims to wars that kill millions of people and that will have long lasting effects for generations, such as the Syrian civil war or the Rwandan genocide. It is important to find out what causes certain conflicts to be far deadlier than others, but the attention that has been paid to explaining and analyzing conflict intensity is limited.

The most devastating conflicts in the world did not happen in one day, but slowly evolved over time. Some conflicts have lasted several years, cumulatively resulting in an extremely high number of battle deaths, while others, like the Rwandan genocide, took place within a few months. The severity of con-flict is an important aspect of civil concon-flict and has not yet received the attention it merits. Knowledge of the factors that determine whether a conflict stays local or evolves into a full-scale war is needed, and can help both state and non-state actors to actively engage in de-escalating conflict.

(10)

2 1.Introduction

1.2. Why do we need to look at conflict intensity?

The impacts of civil conflicts are numerous and long lasting. Countries that suffered from a civil war tend to have lower incomes and more people living in absolute poverty (Collier et al. 2003:2). Large numbers of refugees fleeing their war-torn native countries also have an impact on neighbouring coun-tries (idem:2). While battle related deaths are directly incurred by a conflict, mortality rates post-war can still be higher than the baseline (idem:24). Black et al. (2003:2227) show that of the ten countries with the highest under-5-year child mortality rates, seven have actually suffered from civil conflict in the recent past.Ghobarah et al. (2003:200) conclude that civil war causes long-term damage to public health-care systems with effects observable even long after the war ended. The risk of death or dis-ability caused by infectious diseases such as malaria or tuberculosis is significantly higher in post-war countries (ibid).

Not all effects of civil conflict are immediately observable, and some consequences of conflict will re-main hard to measure. Psychological traumas caused by conflict are common and permanent. High rates of depression, PTSD and suicides can all be related to the previous occurrence of war. Ghobarah et al. (2003:199) find an (indirect) effect of civil conflict on the suicide rate for women between 15-44, which might reflect the trauma of rape. Rozée and Boemel(1990:36) find a high rate of older Cambo-dian refugee women to suffer from depression and somatization years after the horrific civil war that took place in Cambodia in the 1970s.

The intensity of civil conflict may be a determining factor for the severity of these effects. Knowing about the many consequences civil conflict causes, it stands to reason that the intensity of conflict matters as well. Happiness and life-satisfaction are affected by civil conflict and related to the number of battle deaths per 1,000 people (Welsch 2008:330). More deadly wars are likely to recur, since the animosity is incited by a higher death toll and reconciliation becomes harder (Fortna2004:287). What makes some wars more severe than others? Is it a result of ethnic cleavages, political institutions or is it related to the presence of natural resources? Finding out what factors determine the intensity of conflict is important, as is explaining duration or onset of civil conflict. However, before these questions are answered, conflict intensity has to be defined.

1.3. What is conflict intensity?

One of the difficulties in explaining the severity of a conflict is how to define conflict intensity. Deter-mining what conflict intensity comprises of, and how to measure it, is an important step for research on conflict intensity. Conflict can have long lasting effects on multiple levels of society and some of these effects are hard to measure, let alone finding an appropriate means in which to explain these effects. Defining the intensity of a conflict itself is equally as difficult. Destroyed landscapes, demolished towns and villages, spread of diseases, battle deaths, military expenditure, the number of people that have fought: arguably all of these things could be looked at when determining the intensity of a conflict. However, the goal of this thesis is not to find the best measure possible for conflict intensity, but rather to review the measures at hand and evaluate their applicability and explanatory power. Even with an imperfect measure, useful results can still be produced when new insights are gained and relations are determined.

(11)

1.3.What is conflict intensity? 3

One common practice is to use the number of battle deaths as a measure for the intensity of a con-flict (see for example Lacina2006:277,Gurr1968:1107). While this may seem like an objective and straightforward measure of conflict intensity, some important questions need to be addressed. First, how are battle related deaths defined? Second, how can we measure the number of battle deaths? And lastly, is the number of battle deaths an appropriate measure for conflict intensity? For example a number of 1,000 battle deaths in a country with millions of inhabitants has a different impact than 1,000 battle deaths in a country with 500,000 citizens (in that case, 0.2 % of the population has died).

In order to use battle related deaths, this concept must be specified first. The Correlates of War (COW) project uses”the number of battle-connected fatalities among military personnel”(Sarkees2000:128). While this number can give us insight into the military capacity of warring parties and their success, this definition does not include non-combatant deaths. The commonly used data-set that was introduced byLacina and Gleditsch(2005:148) uses”all people, soldiers and civilians, killed in combat”. This mea-sure accounts for all deaths in a civil conflict, whether they were legitimate or not, and shows the”scale and scope” of the military confrontations. Including non-combatants deaths is important to correctly estimate the number of battle deaths, since the distinction between combatants and non-combatants is not always entirely clear (for example when private military firms are hired by a state, and thus the people fighting for the state are not defined as combatants) (ibid). However, this measure is incomplete as well, if one wishes to describe the total number of people who died during a civil conflict. Effects of conflict, such as the collapse of the public health system, the increase of violence and criminal activities and destruction of infrastructure and economy can increment the total death-toll. To account for these deaths as well, Lacina and Gleditsch (2005:149) distinguish between war deaths and battle deaths. War deaths comprise all people that died during or after the war, in a battle or as a result of changed social circumstances. However, this still requires some estimation, as it is hard to determine whether certain events would not have happened without war (for example famines, epidemics or the scale of impact of natural disasters) and how long the effects of war last.

The data-set provided byLacina and Gleditsch(2005) is used frequently to analyse conflict intensity as for example Balcells and Kalyvas(2014),Eck(2009),Heger and Salehyan(2007),Lacina(2006) and

Lujala(2009) did, and is consequently viewed as an appropriate measure for conflict intensity. Since it is used widely, correctness of data is very important. While the number of battle deaths seems like an objective measure to look at conflict intensity, the method for data collection is rarely reviewed by authors using the Lacina and Gleditsch (2005) data-set. The codebook for the UCDP battle-related deaths data-set that is closely related to the Lacina and Gleditsch(2005) data-set describes the data collection method (Allansson and Croicu 2017:13). Three sets of sources are used: global newswire reporting, global monitoring and translation of local news performed by the BBC and secondary sources such as local media, NGO and IGO reports, field reports and books. Subsequently, a”two-pass”system is used, where the first pass looks at global newswire reporting and BBC Monitoring data which are both sourced from the Dow Jones Factiva aggregator using a search-string that looks for words as kill or deaths (Croicu and Sundberg2017:12). The second pass adds local news-sources which is decided by project leaders and UCDP coders (idem). It becomes clear that even an ’objective’ measure on battle deaths, uses ’subjective’ data from media and is compiled by human coders.

(12)

4 1.Introduction

Sambanis(2004:821) addresses the question of relative versus absolute magnitude of a conflict. The impact of a certain number of deaths is dependent on the countries’ population and the dispersion of the victims over geographical areas and ethnic groups. For example a conflict that caused 1,000 deaths in a state with one million inhabitants (for example Djibouti or Cyprus) would mean 0.1 % of the population has been killed. A similar part of the population for a country as the United States (roughly 330 million inhabitants) would be 330,000 deaths. Counting absolute numbers of battle deaths, as is common in the PRIO and UCDP data-sets, can cause a distorted image of the intensity or impact of a conflict. This also relates to the minimum number of battle-deaths a conflict needs to reach before it is included in the data (for UCDP and PRIO this is 25 battle deaths). The Holy See currently has a population of 800 people, and a number of 25 battle deaths would account for 3.1% of the popula-tion. Sambanis(2004:822) proposes to set a per capita measure of battle deaths but this has its own drawbacks.

Another measure for conflict intensity is based upon the Minorities at Risk (Minorities at Risk Project

2009) data that is a project based at the University of Maryland’s Center for International Development and Conflict Management (CIDCM) since 1988. While the MAR data-sets are well known and often used in political science research, using the data to create a conflict intensity measure is a relatively new idea introduced byRegan and Norton(2005:327). Combining different data on the level of protest and rebellion creates a new conflict intensity variable (for a detailed explanation, see section3.1). Saideman et al. (2002) analyse protest and rebellion intensity separately. The variables used to express protest and rebellion intensity are coded by students and reviewed by senior staff (Minorities at Risk Project

2009). The codebook provides little information on the sources that are used by the coders to determine the value of a variable. Even though the MAR project assures users that all coders are well-trained, that information is gathered from open-sources and that multiple sources are consulted for each code, it is still questionable how reliable the coding is. The coding procedures have been refined throughout the different versions of the MAR-data, especially after problems with internal consistency were detected (Minorities at Risk Project2009). Evaluating inter-coder reliability and checking indicators for internal consistency has yet to be done.

1.4. Thesis Outline

The two conflict intensity measures that are introduced above have both been used by various authors. Keeping in mind the different procedures used to collect data, the choice of conflict intensity measure can impact the results of an empirical analysis. It is therefore important to find out if both measures lead to comparable results when using them for research on conflict intensity. This thesis hopes to provide new insights in two conflict intensity measures as well as show the importance of justifying the choice to use either one of them instead of the other. In addition, this thesis hopes to contribute in a methodological way by introducing and using a method uncommon to the Political Science field. In this thesis, two different measures for conflict intensity are analyzed using regression and ensemble methods, to find out what influences the intensity of conflict. Furthermore, a comparison will be drawn between results of the analysis of the two different measures to find out if the factors that influence conflict intensity are comparable. This study will be structured into six chapters. In the next chapter,

(13)

1.4.Thesis Outline 5 academic literature on civil conflict and conflict intensity will be examined. Factors that are found to influence conflict intensity in the literature review will be used to formulate hypotheses. The third chapter introduces the research design, where all data-sets and coding decisions are discussed and some descriptive statistics are presented. The data used for the independent variables is reviewed thoroughly and it is explained why some commonly used data is flawed. The two methods used in the empirical analysis, OLS regression and Random Forest, are examined before moving on to the results. Where OLS regression is a well-known method in the field of Political Science, Random Forest is not yet. This thesis illustrates the applicability of Random Forest and shows the possibilities that this method can have for the Social Sciences.

The results from the different analyses are discussed and the robustness is reviewed through robustness checks. As expected, population size is positively related to conflict intensity for both measures. The same result is confirmed for the number of groups. In general, the OLS results show more significant variables for the protest and rebellion conflict measure. Random Forest results show that for both variables population, GDP, regime type and number of groups are more important than the region a conflict is taking place. However, the most important regions are very different for the two measures. The thesis ends with a conclusion chapter to summarize the contribution of this thesis and possibilities for future work.

(14)

2

Theoretical Framework

2.1. Literature on Civil Conflict

The literature on civil conflict has expanded over the last decades and will be hard to summarize. However, a short overview of important topics in the academic study of civil conflict will be provided. First, the very basics of civil conflict, were the definition of civil conflict itself will be treated. Then the different causes of civil conflict, such as ethnicity or inequality, will be introduced in short and the effect that they have on the duration of civil conflict. Lastly, the aftermath and resolving of conflict will be discussed. This short literature will serve as a starting point for the theoretical framework to study conflict intensity.

2.1.1. The Basics of Civil Conflict

Before discussing the specifics of civil conflict, a definition of civil conflict has to be provided. A defi-nition of civil conflict is a very important condition for any researcher wanting to do quantitative work in the field of civil conflict, since it determines the data that will be selected for a certain analysis. This can impact conclusions resulting from diverging results following from the empirical analysis, es-pecially when a lot of dissimilarities exist between different data-sets based on competing definitions.

Sambanis(2004) describes and compares different formulations of civil war and the results following from the analyses before introducing a 12-item list of properties that any armed conflict should meet to be classified as a civil war. While this thesis focuses on all civil conflict and not just civil war, the 45 pages long search for an operational definition of civil war shows the difficulties and complexities of formulating an unequivocal concept. Protests and rebellions exist in various sizes and shapes and are all forms of civil conflict. According to the Minorities at Risk users manual, protracted civil war is the highest form of rebellion that can be coded for any group in the database (Davenport2003:90). Other forms of civil conflict are for example terrorist campaigns and insurgencies. Important characteristics of civil conflict concern the actors and the location of the conflict. Kalyvas and Balcells (2010:418) describe four different ”technologies of rebellion”, based on the military capacity of the rebels and the state. If both sides possess advanced military technologies, the authors name this a conventional civil war; if both sides lack military technologies a symmetrical nonconventional conflict (SNC) takes places. When the state is better equipped than the rebels, an irregular war is observed. The last possi-bility, when the state lags capacity compared to the rebels, is that a successful military coup takes place.

Sambanis (2002:218) gives multiple characteristics of a civil war that are generally used in the liter-ature: a civil war causes at least 1,000 deaths1_{; it takes places within the territory of a recognized}

state; the state is involved as a principal combatant; the sovereignty of the state is challenged; rebels

1_{Note that this is an absolute threshold, which brings us back to the problem described in the introduction of absolute versus}

relative deaths.Sambanis(2004:829) recognizes this problem and introduces an additional criterion for a civil war that a state needs to have 500,000 inhabitants or more.

(15)

2.1.Literature on Civil Conflict 7 able to set up organized armed opposition are involved. Three comparable criteria are introduced by

Fearon and Laitin(2003:76), who define civil war as a conflict that involves combat between state and nonstate actors, with at least 1,000 deaths in total, more than 100 per year and a number of 100 deaths on both sides. The Correlates of War project defines civil war as an armed conflict, with at least 1,000 battle deaths per year during the war, active participation of the national government and an effective resistance of both the state and insurgents (Sarkees 2010:5). Some of these characteristics are also applicable to civil conflict in general, such as the state being an important actor and the conflict taking place within the geographical area of an established state. However, the threshold of 1,000 battle deaths is clearly not applicable to all civil conflicts and any research considering all different forms of civil conflict should include both large-scale and small-scale events. TheUppsala Conflict Data Program

takes into account conflicts with more than 25 deaths, thus excluding very small scale activities. Even when a definition of civil conflict could be agreed upon, more challenges lie ahead. Special cases will always encourage political scientists to further refine the definitions of civil war, for example what happens to an ongoing conflict when new parties enter a conflict or when states are created as a result of civil conflict. The onset and termination of conflict are hard to pinpoint to a certain time, because they usually do not start at a single moment in time but slowly evolve over time. Sambanis(2004:830) codes the start year of a civil war as the first year 500 or more deaths are caused or if the next three years cause 1,000 deaths together. The MAR project takes into account the number of armed fighters and the frequency of attacks in order to discern the different forms of guerrilla activity. The end of a civil war of rebellion is perhaps even harder to define. Ideally, civil conflict would be resolved with a peace treaty that effectively stops the fighting and marks the beginning of peace but unfortunately this is not always the case. Conflicts can end with truce, cease-fire, treaties or because the fighting slowly demises, but can also erupt again after peace agreements are signed. Therefore, the termination year of a conflict can only be determined a few years after the fighting has ceased when a certain period of peace has been passed.

2.1.2. Causes and Duration of Civil Conflict

What causes civil conflict to originate? Much research has been executed in order to find an answer to this question, and it can be concluded that there is no simple answer to this question. While this thesis concerns conflict intensity rather than conflict causes, both studies are highly interconnected. A short summary of findings in the literature on civil conflict causes will be provided in order to serve as a starting point for the quest to find the determinants of conflict intensity. Since some important factors determining onset or duration of conflict are also related to conflict intensity, they will not be treated extensively in this section but in section2.2.1. When civil conflict has erupted, multiple actors will focus on resolving the conflict and make the duration of the war as short as possible. Where some wars are ended within a few months, such as the Dominican Civil War in 1965, other civil wars endure more than a decade or even longer (for example the Salvadoran Civil War, which lasted more than 12 years). Three different conceptualizations of civil conflict that provide rival explanations of the duration of the conflict will be discussed (Collier et al. 2004:254).

The greed versus grievance debate centers around two influential and competing arguments of the causes of civil conflict. The greed argument focuses on the economic gains that drive rebels to engage

(16)

8 2.Theoretical Framework

in conflict with the purpose of enriching themselves. This argument has been put forth byCollier and Hoeffler(2004) in their influential paper ”Greed and grievance in civil war”. Causes of conflict are, in their opinion, for example dependency on primary commodity exports, because of the opportunities these exports create for extortion, low per capita income, a slow or negative growth rate of the econ-omy and a high percentage of uneducated or unemployed men. These arguments assume that rebels are rational actors who rebel based on a cost-benefit analysis of violence which includes the probability of a victory (Sambanis 2002:221). Fearon and Laitin (2003:76) agree with Collier and Hoefflerthat financing is an important determinant for rebels to keep fighting, but they disagree that this is not because of greed but because economic variables proxy for state capacity.

Proponents of the grievance argument, on the other hand, argue that rebellion stems from inequalities bases on identity such as ethnicity, social class, gender or religion. Important contributions to the debate on the grievance side have been made by David Keen, who has also been an outspoken critic of the work ofCollier and Hoeffler. He accuses Collier of oversimplification and taking a numerical ap-proach, without paying too much attention to identities and ideas (Keen2014:765). Keentakes a more constructivist approach and explains rebellion as a means to a different goal: gaining political rights, secession, recognition of culture or religion or the abolition of a repressive caste system. However, both arguments have in common that they focus on a certain form of deprivation, whether it be economic (horizontal inequality) or based on ethnicity, race or religion (vertical inequality). Some causes of civil conflict can even be used to illustrate both arguments, such as the deprivation of education that can cause economic inequalities and drive uneducated young man to rebel to enrich themselves. Being denied access to education can be a key grievance to start fighting, as for example happened in Sierra Leone (Keen2014:761).

Important determinants of civil war are amongst others regime type (democracy/autocracy), poverty (GDP and GDP growth), population size, region and ethnic fractionalization, which will all be discussed in the next section. Other causes related to the onset of civil war are for example mountainous ter-rain, food supply, measured as the calories per capita, primary commodity exports and secondary schooling (Fearon and Laitin2003:85,Sobek and Payne2010:233,Collier and Hoeffler2004:574,Gurr

1968:1118). Rebels can look at conflict as a business opportunity, where they actually gain from the fighting (while the country suffers). This approach assumes that the benefits of conflict exceed the costs, making it profitable to continue the insurgency. In this case, export prices are important to the fighting parties, since they usually gain money from plundering natural resources or other commodi-ties. Lower prices would then shorten the duration of the conflict in that case, as the rebels’ resources will shrink (Collier et al. 2004:266 actually find significant correlation between lower export prices and shorter conflict duration, confirming this hypothesis). Collier and Hoeffler(2004:588-589) argue that population size and expected benefits of a rebellion are related, because a large population will probably contain different ethnic groups that might seek secession.

Another motivation for insurgents could be political, and in this case the pay-off will occur after the conflict while during the fighting costs are incurred, but typically only when rebels win. If the expected benefits of a victory are higher, the duration of a conflict will be linger. Lastly, conflict might persist because warring parties overestimate their chances of winning. Collier et al. (2004:254) call this the

(17)

2.2.Previous Work on Conflict Intensity 9

rebellion-as-mistake, and point out that a peace agreement can be hard to reach when both sides think they can get better when the conflict continues. A state-centric concept is put forward by DeRouen and Sobek (2004:305) and considers the state as a participant in the conflict. Rebellion as business claims insurgents fight because of the benefits fighting brings them, but why would they not have these benefits during peacetime? Fearon and Laitin(2003:84) show states with capacity are less likely to experience war and DeRouen and Sobek(2004:311) finds that state capacity is also important for the outcome of civil war, especially when states have a strong and efficient bureaucracy they are able to avert rebel victory. Balcells and Kalyvas(2014:1399) find that irregular conflicts, where lightly armed rebels fight against a more powerful state, last significantly longer than conventional conflict of symmetric nonconventional conflicts. Furthermore, they find ethnic fractionalization and per capita income to lead to longer conflicts (ibid). Conflict will also last longer when the number of veto players is higher, since it becomes significantly harder to reach a settlement (Cunningham2006:891).

2.1.3. Resolving Conflict & Aftermath

Conflict can be resolved through peace talks, peace treaties, peace accords, ceasefires, truce or slowly decrease in intensity until the actual fighting ceases. Power-sharing institutions are one means to resolve war, but conflict can also end by a victory of one of the warring parties, the partition of a state into newly formed countries or by establishing post-war institutions that favour controlled behaviour by politicians (Hartzell and Hoddie 2007:143). Research has revolved around the best measures to reduce violence and the durability of peace. Beardsley et al. (2018:3) finds that mediation by a third-party reduces the severity of conflict and decreases the number of battle deaths. Mediation in combination with peacekeeping can have large effects on the number of fatalities because the reinforce each other (ibid). International interventions in the form of peacekeepers helps to maintain peace (Fortna2004:288). Krause et al. (2018:1005) investigate the effects of woman participation in peace negotiations on the durability of peace, where they measure female participation as signatories on peace accords by women. They find a robust relationship between woman signatories and durable peace combined with a higher implementation rate. Including civil society representatives increases the chances the peace will last by bettering the legitimact of the agreements (Nilsson 2012:244). Negotiated settlements are more likely to endure when the state had a stable democratic regime before and the agreement includes a provision for threatened groups to have some territorial autonomy (Hartzell et al. 2001:202). Power-sharing arrangements increase the chances peace endures by giving each group a certain level of power and control and enables them to watch the other groups closely (Hartzell and Hoddie2007:143).

2.2. Previous Work on Conflict Intensity

Even thought conflict intensity has not been researched as well as the duration and onset of civil con-flict, the subject has not been neglected completely. Gurr (1968) already introduced the importance of conflict intensity, which he connected to the intensity of political and economic discrimination. Gurr

(1968:1107) measures conflict intensity as the number of battle deaths. Conflict intensity is one of the three aspects ofGurr’s (1968:1123) measure for the magnitude of civil strife and is found to be directly related to the intensity of relative deprivation, but the effect of the number of battle deaths alone cannot be determined from his results. Mesquida and Wiener(1999:187) find that a large number of young men is related to conflict intensity. The ratio of the number of men aged 15 to 29 compared

(18)

to the number of men aged 30 or older is found to account for a third of the variance in the number of battle related deaths (ibid). BethanyLacina(2006) uses Ordinary Least Squares (OLS) regression to find statistical evidence of a correlation between 10 explanatory variables and the severity of civil war, which she measures in battle related deaths using data from the Uppsala Conflict Data Program (UCDP). Lacina(2006:285-286) finds democracy to be a significant predictor of lower battle deaths, but while GDP per capita does have a negative influence, it is insignificant. Lujala(2009:50) also uses data from the UCDP/PRIO project and OLS regression to examine how natural resources (e.g. gems, oil and drugs) influence conflict intensity. and finds that gemstone mining and the production of oil or gas in the conflict zone exacerbate conflict.

Heger and Salehyan(2007:385) use the data provided by Lacina and Gleditsch (2005) on battle deaths to find a relation between the size of the ruling coalition and conflict intensity, which they find to be negative correlated (e.g. a smaller coalition leads to a greater number of deaths). They argue that this can be explained by the fact that leaders from a smaller coalition are less limited to use force and more likely to be followed by their supporters because they have prospects of larger benefits compared to larger coalitions (idem:386). Using the same data,Balcells and Kalyvas (2014) find that the type of civil war that is fought influences the severity of the conflict. Their hypothesis that conventional conflicts are more severe on the battlefield than irregular wars and SNC (symmetric nonconventional wars) is confirmed using negative binomial models (idem:1402, seeKalyvas and Balcells(2010) for the an explanation of the different types of war). Democracy appears to be related to lower levels of battle deaths, confirmingLacina’s (2006:285) findings (Balcells and Kalyvas2014:1405).

Also looking at the link between the type of conflict and conflict intensity isEck(2009), who is looking at ethnically mobilized conflicts. Firstly, she finds that a conflict with ethnic mobilization has a higher probability to escalate into war (idem:378). Using the data provided by Lacina and Gleditsch (2005), she then extends her analysis to conflict intensity measured as battle related deaths and finds that ethnic mobilization has a small but (significant) positive effect on the number of battle deaths, as does ethnic pluralism, which is not significant (idem:384).Medeiros(2017) looked at the effect of linguistic vitality on the intensity of conflict looking at territorialized linguistic minorities. No causal relationship is determined, which might be caused by the lack of linguistic vitality data (idem:640). Conflict intensity is measured combining the MAR variables protest and rebellion to create a continuum (idem:635). Combining these two variables is proposed by Regan and Norton(2005:327), even though they use the combined variable for protest and rebellion to dichotomize three variables, protest, rebellion and civil war, and use this as their measure for conflict intensity.

2.2.1. What determines the intensity of conflict?

Understanding what influences the severity of conflict starts with finding variables that can explain the variation in conflict intensity. Previous authors have chosen their variables based on the literature on civil conflict onset. Variables such as regime type, income and ethnic diversity that the literature suggests are related to the onset of conflict can possibly also predict conflict intensity. However, the question to be answered is what causes the variation in conflict intensity, considering a country is already experiencing some form of civil conflict and not to predict the possibility a country will suffer from a conflict with a low or high intensity in the future. Therefore, explaining variation in intensity adds

(19)

2.2.Previous Work on Conflict Intensity 11 to the current literature on conflict onset rather than substitutes it. Following both Lacina(2006:281) andLujala(2009:59), this thesis uses variables that explain civil conflict onset and duration in order to find out if they also influence the intensity of civil conflict.

Regime Type

The intensity of conflict might be related to the regime type. At first sight, it might be expected that democracy leads to less intense conflicts, since democratic regimes allow minorities to represent them-selves in the political systems, thereby reducing the need to rebel. Discrimination and repression are expected to occur less in democratic systems because all citizens possess some form of political power (voting rights) and reduce the risk of civil war (Fearon and Laitin 2003:79). However, no statistical significant result is found between democratic regimes and civil war onset, meaning that democracies are not less likely to experience civil conflict (idem:84). While the evidence for regime type in con-flict onset is small,Lacina(2006:282) finds that democratic regimes experience less intense conflicts and provides three reasons to explain this relation: democratic norms, selection effects and political adaptability of institutions. Democratic norms can cause political leaders to answer with restraint to insurgencies, using measures that will limit casualties to avoid backlash from the civilian population. Democracies tend to select the fights they can win and avoid unnecessary killing. Concessions might be granted to rebels under public pressure instead of using force when being faced with large insur-gencies , resulting in only minor insurinsur-gencies to be observed (Lacina 2006:282). Lastly, democratic regimes have institutions available that can be used by rebels and insurgents to be able to participate in policy formation and share in political power. Being able to better compromise with insurgents and work together could be a reason insurgencies do not escalate. BothLujala(2009:63) andBalcells and Kalyvas (2014:1405) are able to reproduce these findings and conclude that democracy reduces the level of battle deaths. Thus, democracy seems to be negatively correlated with conflict intensity:

Hypothesis 1a: Conflicts in democracies are less intense.

It is unsurprising though, thatLujala(2009),Lacina(2006) andBalcells and Kalyvas (2014) come to the same conclusions, since all three of them use the Lacina and Gleditsch (2005) combat deaths data in combination with data from the Polity project on regime type. Sambanis(2002:224) points out that according to empirical findings, not the most or least democratic regime suffers the highest risk of civil war, but states that are in the middle between autocracy and democracy. Autocratic regimes reduce their risk at civil war by repressing anti-regime voices and precluding every opportunity of civil war while democratic regimes prevent grievances and reduce the possibility people will rebel to voice their grievances (ibid). While civil war, protest and rebellion are not equal, the conclusions from Sambanis

on civil war might be extrapolated to protests and or rebellion. Medeiros(2017:637) also uses MAR data as his conflict intensity measure, and finds no significant relation between regime type and con-flict intensity, which could be because the regimes in the middle of autocracy-democracy spectrum experience the highest conflict intensity. Saideman et al. (2002:107-108) on the other hand, expected to find a negative relationship between democracy and the occurrence of ethnic protest or rebellion. Their hypothesis is based on the characteristics of democracies. Because of freedom of speech, the cost of protesting is less than in autocratic regimes, but the possible benefit is greater (Saideman et al.

2002:107). Protest can influence politics and lead to beneficial effects. The question is of course, if this relationship between democracy and protest will extend to violent protests and rebellion. Saideman

(20)

12 2.Theoretical Framework et al. (2002:118) find that democracies do suffer more from protest and rebellion, possibly because it is easier to organize as a group in a democratic system compared to an authoritarian system. This leads to the second hypothesis:

Hypothesis 1b: Conflicts in democracies are more intense.

Poverty

Poverty is thought to be an important factor in the outbreak of a civil war. A lower GDP per capita is assumed to lead to a higher risk of civil conflict (Sambanis2002:229). Collier and Hoeffler(2004:574) find higher per capita income to have a reducing effect on civil war risk andFearon and Laitin(2003:83) find a strong correlation between per capita income and civil war onset: $1,000 less income increases the chance of civil war onset with 41%. Poverty can be a grievance that drives rebels to fight, and

Collier and Hoeffler (2004:588) proxy income as earnings foregone in insurgency. Fearon and Laitin

(2003:80) see income per capita as an indicator of the state strength. Developed countries with good infrastructure and connections to rural areas will usually have a higher per capita income. A high GDP per capita indicates that a state possesses certain administrative, financial and military capabilities. The risk of insurgency decreases when economic prospects are good and potential insurgents have better alternatives than to fight. However, GDP per capita is not an indicator for an equal distribution of wealth. Similar relationships are confirmed between GDP per capita and the severity of civil conflict.

Balcells and Kalyvas(2014:1403) finds a lower GDP per capita to be related to a higher number of battle deaths, as doesLacina(2006:286), but her GDP variable is insignificant. Eck(2009:379) also finds per capita income to be insignificant when predicting the risk of conflict intensification, even though the coefficient is always negative.

Hypothesis 2a: A higher GDP per capita leads to a lower conflict intensity.

Regan and Norton (2005:330) find GDP per capita to be positively related to the onset of rebellion, meaning that when the income increases, there is a higher chance of rebellion. Contradictory findings are presented bySaideman et al. (2002:119), who find richer countries to engage less in violence. Interestingly, both previous authors use rebellion and protest data from the MAR-dataset; they also both find capita per income and protest to be unrelated. In our analysis, rebellion and protest variables will be combined, and it is expected to find that a higher GDP leads to a higher conflict intensity.

Hypothesis 2b: A higher GDP per capita leads to a higher conflict intensity.

Population Size

Population size is often included as a control variable but has repeatedly been found to be positively correlated with war onset. However, asSambanis(2004:822) points out, this may be a result of the 1,000 deaths criterion. Conflicts with a minimum of 1,000 fatalities only take place in states with a larger population. Sambanis (2004:822) describes the Greco-Turkish war in Cyprus (1963) that is usually omitted from data-sets (due to insufficient death count) even though the number of deaths is around 1,000 and all other civil war criteria were met. Considering Cyprus’ population size, an equally intense conflict in the Netherlands would have accounted for 17,000 deaths and would definitely have been included. Other authors believe that the correlation between population size and civil war is not

(21)

2.2.Previous Work on Conflict Intensity 13 artificial at all and can be explained. DeSoysa(2002:400) connects population size to level of openness to trade, stating that countries with a larger population are generally less open open to trade. States with more open economies tend to have larger governments, less corruption and a larger share of government to GDP which are all risk reducing factors. Her empirical analysis confirms her hypothesis that the effects of population size are related to governance factors and also contradicts Sambanis because she uses a threshold of 25 battle deaths instead of 1,000, thereby reducing the possibility the correlation is artificial.

A third possible explanation for the significance of population size in civil war onset analysis is the argument that larger countries have an increased ethnic fractionalization. Both grievances and oppor-tunities may increase with population size, for example because an increasing population results in a larger heterogeneity (Collier et al. 2004:588). On the other hand,Fearon and Laitin(2003:85) control for ethnic diversity and still find a correlation between the risk of civil war and population size. Previous literature on conflict intensity has only paid attention to population size sideways as a control variable. Lujala(2009:62) finds population size to be positively associated with the number of combat deaths. Conversely,Lacina(285) finds a large population does not lead to more battle deaths. Based on the evidence that population size is related to civil war onset and the first indications that it is related to conflict intensity as well the following hypothesis is formulated:

Hypothesis 3: A larger population leads to a higher conflict intensity.

Ethnic fractionalization

Ethnic polarization and fractionalization have been thought to be related to civil conflict onset by many researchers. Donald Horowitz’ (1985) book on ethnic groups in conflict has been of great influence to generations of political scientists. His main thesis contends that societies hosting a plurality of eth-nic groups are at a higher risk to experience civil conflict (idem:140). There is only mixed evidence supporting Horowitz’ argument: Fearon and Laitin (2003:88) conclude that indicators of ethnic and religious diversity do not predict a higher risk for civil war and even when insurgents are observed to mobilize their fighters along ethnic lines in poor countries, the causes of conflict are not ethnic diversity but structural problems as poverty or instability. Reynal-Querol(2002) finds religious polarization to be correlated to the incidence of civil war. An explanation for this correlation and the absence of any results supporting correlations for ethnicity or language could be that divisions along religion lines are sharper and more evident. Someone can be related to multiple ethnic groups or master multiple lan-guages, but will only adhere to one faith. These results are confirmed byRegan and Norton, who find ethnic and linguistic heterogeneity to significantly increase the risk of civil war and rebellion. Daxecker

(2011:46) also finds a positive relationship between ethnic fractionalization and civil war onset.

Sambanis(2002:230) blunty concludes that the literature finds ethnic diversity to not increase the risk of civil conflict, but that it in fact may decrease said risk. However, the contradictory findings in the literature suggest that idea of a correlation between civil conflict and ethnicity should not be com-pletely abandoned but should lead to a new research agenda. While the evidence for ethnicity as a cause of civil conflict is lacking, ethnic fragmentation and polarization might still be of importance for conflict intensity. Loyalty based on ethnic or religious identity instead of political identity is thought to be more lasting and less open to compromising with other groups, leading to conflicts that are more

(22)

severe (Lacina2006:284). The shared culture and language heighten the costs of conflict that group members are willing to bear (ibid). If conflicts between groups are based on historical hatreds that have been present for generations, coming to an agreement could be impossible and clashes will be more intense (idem:284). Lacinafinds religious diversity to not have an impact on conflict severity but does finds ethnic polarization to be negatively correlated to conflict intensity. This might be due to nonselective violence, resulting from the difficulty to determine who is on what side as a result of the ethnic homogeneity of the population (ibid). On the other hand,Lujala’s results show that democratic states with ethnically heterogeneous populations experience less intense conflicts and that ethnically diverse countries have fewer battle deaths (2009:52).

Contradictory findings are presented by Eck(2009:384) who finds evidence that ethnic mobilization gives a 92% higher chance that an existing conflict escalates into war compared to non-ethnically mobilized conflicts. Countries with a larger part of the population belonging to the biggest ethnic group have a higher risk of war, confirming that less pluralistic countries are more to prone to civil war (idem:379). The above presented literature uses similar measures for ethnic polarization, e.g. the share of the population belonging to the largest ethnic group or the size of minority and fractionaliza-tion as the possibility two randomly drawn citizens are sharing ethnicity, religion or language (Lacina

2006:285,Collier and Hoeffler2004:595 andEck2009:377,Daxecker2011:44). Combining above pre-sented results, the validity of these measures might be questioned (as doesSambanis2002:230-231). Using the number of groups involved in conflict in a certain state in a given year, this thesis will add another measurement for ethnic plurality to the literature (see3.1). Taking into account the previous literature the following hypotheses are presented:

Hypothesis 4a: A higher number of groups involved in conflict in a state in a certain year increases the conflict intensity.

Hypothesis 4b: A higher number of groups involved in conflict in a state in a certain year decreases the conflict intensity.

Region

All regions in the world have been affected by civil conflict but some more than others: Africa and in particular Sub-Saharan Africa has been affected the most with Asia following in second place ( Samba-nis2002:216). This raises the question whether certain regions share common characteristics making them more prone to conflict (Medeiros 2017:636). Regions can share a history, culture, language or economic traits, things that can be considered probable variables explaining the onset of civil war (Fearon and Laitin(2003:87). Major wars can have an impact on neighbouring countries and influence the security and stability of a whole region, possibly creating a more fertile environment for civil conflict to originate. War-torn countries will lose foreign investors who will invest their funds in other, more stable countries, draining the sources of the host countryMurdoch and Sandler(2002:92).

According toMurdoch and Sandler(2002:92), these consequences of war will probably not be confined to the warring countries, but spillover to neighbouring states and affect their investments and trade. They find larger civil wars, with more than 25,000 deaths, to have longer lasting effects, for example on the GDP per capita, in both the country that suffered from conflict as well as neighbours. Combining

(23)

2.2.Previous Work on Conflict Intensity 15 these findings with the evidence found by various authors that a lower GDP and a slow economic growth are causes of civil conflict, the assumption could be made that neighbours of a country experiencing a large civil war are also more prone to civil conflict in the future (Sambanis 2002:229). Since most wars over the past decades have taken place in Africa and Asia, other countries in this region might be at risk to experiences insurgencies as well. Regarding conflict intensity, it is expected that certain regions will experience conflict with a higher intensity because of the above described effects.

Hypothesis 5: Conflicts in Asia and Africa will have a higher conflict intensity.

Balcells and Kalyvas (2014:1401) confirm this hypothesis and find that both Asia and Sub-Saharan Africa variables are positively correlated with conflict intensity, while Medeiros(2017:637) only finds this result for Asia. However, both analysis include regions based on ’civilization’ aspects instead of geographic regions that will be used in this thesis.2

2_{These regions include: Asia, Eastern Europe and former Soviet Union, Latin America and Caribbean, North Africa and the Middle}

(24)

3

Research Design

3.1. Data

The empirical analysis will use time-series data, with each observation indicating a country-year data-point. Time-series analysis has been previously applied on civil war data for example byEck(2009),

Walter (2004) and Saideman et al. (2002). Eck(2009:375) studies conflict intensification and uses conflict/year data-points, which means that multiple conflicts per year can be coded for the same coun-try, on which she then employs the Cox proportional hazards model. The same model is applied by

Daxecker(2011:40), who uses leader/year observations to determine whether shocks to the distribu-tion of power lead to a higher risk of civil war. Country-year observadistribu-tions are used by both Walter

(2004) and Saideman et al. (2002), who study recurring civil war and ethnic conflict respectively.

Walter(2004:376) applies a logit regression on her binary dependent variables, whileSaideman et al.

(2002:117) perform Prais-Winsten regressions. In this thesis, OLS regression and Random Forest will be applied to the time-series data, which will consist of country-year observations. In order to obtain this form of data, multiple data-sets have to be recoded since they consist of event-year observations, as is explained in the following sections.

3.1.1. Conflict Intensity

Two measures for conflict intensity used in previous literature will be compared. One measure for conflict intensity as proposed among others byGurr(1968:1107) andLacina(2006:278) is battle related deaths. FollowingLacina(2006:278), the UCDP/PRIO battle deaths data will be used. Another measure for conflict intensity will be derived from the MAR project, followingRegan and Norton(2005:327-328) andMedeiros(2017:635) by combining protest and rebellion intensity variables. In order to make the comparison between the two conflict measures as accurate as possible, the data-sets will be organized in a similar way. The overlapping time-span between the two data-sets is 1989-2003, and the unit of observation will be country-year.

Protest and Rebellion

The MAR data-set is generated using the Minorities At Risk Data Generation and Management Pro-gram (MARGene2003) and includes all politically-active ethnic groups for every country in the world in every year. A minority at risk is defined as ”an ethnopolitical group” that:”collectively suffers, or benefits from, systematic discriminatory treatment vis- -vis other groups in a society; and/or collec-tively mobilizes in defense or promotion of its self-defined interests.” (MARGene2003). Every group in the world that meets this definition is included in the data, which means that some countries have multiple data-points per year representing different minorities. To create country-year cases, the data of the different minorities has to be aggregated per country per year. The MAR data contains a protest indicator, measured on a 5-point scale, and a rebellion indicator measured on a 7-point scale. Following

Regan and Norton(2005:327-328), these two measures are combined into one civil conflict measure that ranges from 0 to 12, where protest takes the values 1 to 5 and rebellion 6 to 12 (see Table3.1). If

(25)

3.1.Data 17 both the protest and rebellion measure take a value, the highest value is coded as the conflict intensity measure. Years without any data are added with a conflict intensity of zero.

To use the protest and rebellion value as the dependent variable, various possibilities for combining the intensities of the groups have to be considered. The first option is to take the aggregated intensity per year of all groups for a certain country that are present in the data-set. However, the drawback of this approach is that countries with a lot of groups end up with the highest intensity, even though the variable per group is very low.3 _{The second option is to aggregate for all minorities per country per}

year and then average over the number of minorities with a conflict variable higher than 0 (since battle deaths can only occur when there is conflict). The problem with this approach though, is that groups experiencing a high conflict intensity (which is most likely to cause battle deaths) are taken together with groups experiencing a lower conflict intensity which in turn decreases the average conflict intensity. This can lead to a very distorted image: for example if the MAR data-set codes a civil war (conflict intensity 12) for one group and symbolic resistance for another (2), the country-year observation will end up with conflict intensity 7 (campaigns of terrorism). To solve this problem, the maximum intensity coded for one of the groups in a country is taken as the conflict intensity for the country-year data-point. This also reflects the possibility battle deaths occur best, and since the results of the analysis will be compared with the battle deaths data this is viewed as being the best option.4

Table 3.1: Coding Conflict Intensity Measure Minorities at Risk

Protest variable Conflict intensity variable

1 Verbal opposition 1 2 Symbolic resistance 2 3 Small demonstrations 3 4 Medium demonstrations 4 5 Large demonstrations 5

Rebellion variable Conflict intensity variable

1 Political banditry, sporadic terrorism 6 2 Campaigns of terrorism 7 3 Local rebellions 8 4 Small-scale guerilla activity 9 5 Intermediate guerilla activity 10 6 Large-scale guerilla activity 11 7 Civil war 12

3_{In this case, India will almost every year have the highest conflict intensity because it has an enormous number of minorities.} 4_{The linear regression was executed for all three possibilities. However, when summing all groups, the linear analysis learns that}

the number of groups immediately becomes the most important predictor variable. Averaging over the number of groups in conflict mainly effects countries with multiple groups, where the conflict intensity differs a lot between them. The OLS results of averaging compared to taking the maximum are very similar but since the intensity is flattened out, the regression results are less significant.

(26)

18 3.Research Design Battle Deaths

The battle related deaths data-set from the Uppsala Conflict Data Program includes all conflicts in the world, classified into four different types of conflict: extrasystemic, interstate, internal and interna-tionalized internal (Uppsala Conflict Data Program2017,Pettersson and Eck2018). Only the internal conflict data-points will be used in the analysis since this is the most similar to the data from the MAR project. The data-points represent conflicts, and are included if they meet the criteria from the UCDP definition of armed conflict (see AppendixA.2). This means that only conflicts with at least 25 battle deaths in one year are included. The data-set is already in a country-year form, where the location of the conflict represents the country. However, for some years multiple conflicts are present in the data-set, and these will be aggregated. It is assumed that in missing years the number of battle deaths equated zero, these data-points are added. The UCDP data-set contains a best, high and low estima-tion of the number of battle deaths. In this analysis, the best estimaestima-tion will be used, to make our results as accurate as possible.

An overview of the countries with conflict is presented in the AppendixA.3for both data-sets and shows the number of country-year data-points that are available after the preparation of the data. Since the overlapping period of time is 1989-2003, a maximum number of 15 data-points per country can exist. In order to use time-series analysis, a minimum number of three years with conflict per country has to be present to be included in the analysis. This results in 600 country-year data-points for 40 unique countries for the UCDP data-set and 1,650 country-year data-points for 110 unique countries for the MAR data-set.5,6,7,8_{Examples of the structure of the data are included in section}_3.2_.

Comparability

When comparing two data-sets, it is important to pay attention to the differences and similarities between them. The MAR-data consists of minorities at risk all over the world and keeps track of their protests and rebellions. Large scale conflicts including minorities will also appear in the UCDP data-set when they have caused more than 25 battle deaths. For example the Front for the Liberation of the Enclave of Cabinda (FLEC), representing the Cabinda minority in Angola, has been fighting for a

5_{Yugoslavia and the Soviet Union both undergo changes in the years 1989-2003. The Soviet Union collapses in 1991 and with}

its dissolution 15 republics were created. In the analysis, Russia and the Soviet Union are treated as being the same country, since Russia comprises almost three quarters of the previous Soviet Union territory. Armenia, Azerbaijan, Belarus, Estonia, Georgia, Kazakhstan, Kyrgyzstan, Latvia, Lithuania, Moldova, Tajikistan, Turkmenistan, Ukraine and Uzbekistan are treated as independent countries. Because of this coding decision, these former Soviet states will not have any conflicts coded in 1989 and 1990. To do a robustness check, the analysis will be executed for the period 1991-2003.

6_{Different states split off Yugoslavia in the period 1991-2003. For consistency, a Yugoslavia variable will be constructed that}

consists of a weighted average of Yugoslavia and the former-Yugoslavia countries based on population size. This Yugoslavia variable will take into account Macedonia, Croatia and Slovenia and Bosnia and Herzegovina after they gained independence. Serbia and Montenegro formed the Federal Republic of Yugoslavia between 1992-2003 and are therefore still coded as Yugoslavia in the data-sets. For battle deaths, the numbers for different countries will be aggregated instead of averaged. In order to check for robustness, the model will be run without the Yugoslavia data-points. Note that some of the former Yugoslavia countries might have less than the threshold of three data-points through time (see AppendixA.3). These are included into the Yugoslavia variable to make it as accurate as possible.

7_{East and West Germany have different data-points in 1989 in the MAR data-set. These are recoded to Germany.}

8_{Czechoslovakia dissolved in 1992. In the MAR data-set, no conflicts are recorded for Czechoslovakia from 1989-1992. therefore}

the country Czechoslovakia is removed from our analysis. For Czech Republic and Slovakia, the years 1989-1991 are created using the data from Czechoslovakia without loss of information (e.g. income, regime). The UCDP data does not record any battle deaths in Czechoslovakia, Czech Republic or Slovakia in the period 1989-2003, so these states are not included in the data-set and analysis.

(27)

3.1.Data 19 separatist state since 1975 and has lead to almost 300 casualties between 1989 and 2003 according to the battle deaths data. Other minorities that have been engaged in violent conflicts are for example the Kashmiri in India, the Chechen in Russia, the Kurds (PKK) in Turkey and the Tamils (LTTE) in Sri Lanka. The SPLM/A, fighting for the southeners in Sudan, has been able to achieve secession and the creation of the new state of South-Sudan. While many of the larger conflicts will occur in both data-sets, protests and low-intensity rebellions will not be recorded by the battle-deaths data-set since it wields a threshold of 25 battle-deaths for any conflict to be included (see Appendix A.2). Some conflicts on the other hand, do not include minorities present in the MAR-data, such as the conflict in Colombia between the government and communist insurgency groups such as the FARC and ELN.

3.1.2. Regime Type

To code regime type, the Polity IV Annual Time Series data-set is used (Marshall et al. 2018). Previ-ous research used two variables from the data-set to create a regime type variable (Saideman et al.

2002:115, Lacina2006:285). The first variable, democracy, ranges from 0 to 10, where 10 indicates the most democratic regimes. Another variable exists for autocracy, which also ranges from 0 to 10, 10 being the most autocratic. Subtracting the autocracy variable from the democracy variable results in an indicator for regime type, ranging from -10 to 10. However, a new Polity data-set has been released since, in 2018, and includes this regime type indicator (the so called polity variable) as a recognition to its common usage (Marshall et al. 2018:16). The Polity IV data-set consists of data-points for each country and each year and can be easily added to the conflict intensity country-year data.

The Polity IV project version 2002 also included a variable polity2 in addition to the polity variable and recodes some data to make the variable suited for time series analysis. This variable has for example been used bySambanis(2004:836) but is not uncontested. Thepolity2variable adds new information to thepolityvariable but also assigns zeroes to previously missing data points. Cases previously coded as -66, -77 or -88 (missing cases, interregnum cases or transition cases) are now modified and changed to zero.9 _{However, this puts these countries in the middle of the autocracy-democracy spectrum while}

it is very questionable if they belong there. Plümper and Neumayer (2010) wrote a paper on the validity of the variable and conclude that this is not an appropriate coding for years of interregnum or transitions. In this thesis, the countries with apolity2score of 0 are treated as missing.10,11,12_Following

Lujala (2009:59),Balcells and Kalyvas(2014:1405) andFearon and Laitin(2003:84), de regime type variable is lagged one year.

9_{For example Afghanistan, which has}_polity2_{score -8 from 1989 to 1991 and -7 from 1996 to 1998, has been assigned a score}

of 0 for the years 1992 to 1995.

10_{Of course some countries really are in the middle between autocracy and democracy and these zeros are kept. Only cases}

with a previous value of -66,-77 or -88 are treated as missing.

11_{For Namibia, the 1989 data-point is not present because it only gained independence from South-Africa in 1990. The polity}

score for Namibia in 1989 is set to the polity score in South Africa in 1989, which is 4. The polity score for Namibia is 6 throughout the years 1990-2003.

12_{For Czech Republic and Slovakia the polity score for Czechoslovakia is used for the years 1989-1992. The polity score is -6 in}

1989 and 8 for 1990-1992. Czech Republic has polity score 10 throughout the years 1993-2003 while Slovakia scores 7 for the years 1993-1997 and moves to 9 for 1998-2003. Using the Czechoslovakia seems to match the other scores.