POLITICAL SURVIVAL THROUGH ARMED CONFLICTS

Transcription

1 POLITICAL SURVIVAL THROUGH ARMED CONFLICTS Kristopher Jordan UNIVERSITY OF OSLO Department of Political Science Master s Thesis in Political Science Spring 2016 Advisor: Håvard Mokleiv Nygård Senior Researcher - Peace Research Institute Oslo Word Count: 34153

2 Kristopher Jordan 18th May 2016 Political Survival Through Armed Conflicts Kristopher Jordan Print: Reprosentralen, University of Oslo i

3 ii

4 Abstract This thesis investigate to what degree political leaders are perceived as accountable for engaging in an armed conflicts, measured as perseverance in office. Previous research show that engaging in armed conflicts has a negative effect on political survival (Brody, 1991; Brody & Page, 1975; Bueno De Mesquita & Siverson, 1995; Croco, 2011; Kernell, 1978; Mueller, 1973; Norpoth, 1987), but has only addressed how engaging in an interstate conflict affect degree of survivability. By using Gleditsch et al. (2002); Pettersson & Wallensteen (2015) framework, I explore how different types of armed conflicts alter the leaders perseverance in office differently. Additionally, I use two theories on political accountability, the selectorate theory (Bueno De Mesquita et al., 2005) and Weeks (2014) index on autocratic regimes, to explain how a reduced degree of liability gives political leaders a greater leeway in their policy. Using data on political leaders and armed conflicts from 1946 to 2014, this thesis provides three main contributions. Firstly, I find that different types of conflicts affect political survival differently. Secondly, the degree of accountability the political leaders has to their citizens and political elites influence their perseverance in office after engaging in an armed conflict. Although same regimes have the same Polity- IV value (Marshall, 2002), the implementation of regime categories (Weeks, 2014) explains variations in survivability. Secondly, utilising penalised splines on some continuous variables gives an explanation of how pre-conflict tenure and starting age affect political survival. Thirdly, the discrepancy between the theoretical expectations and the statistical findings, questions the measurement validity (Adcock, 2001) of the selectorate theory. iii

5 iv

6 Acknowledgements There are many people to whom I am grateful for their contribution to finalize this thesis, both on an academic and personal level. Words do not suffice when describing the significance of your help and support through this process. I can unfortunate only name some. Firstly, I want to thank my advisor, Håvard Mokleiv Nygård. Your methodological knowledge, theoretical insight and thorough commenting have been of great importance. Secondly, I want to give a huge thanks to family, friends and fellow students. Magnus, your patience and help when I stumbled upon problems in R have been significant. Thank you mom and dad for proof reading and other forms of support these past years. Finally, I would like to thank my wonderful wife. Your encouragement and support has been imperative throughout this whole process. Any mistakes within this text are fully my own. The data and documentation is available at v

7 vi

8 To Eva, for her tremendous love and support, and to Viktoria, Wilhelm and Benjamin, for their awesomeness and ability to keep things in perspective. vii

9 viii

10 Contents Acknowledgements List of Figures List of Tables v xi xiii 1 Introduction Central Concepts Political Survival Armed Conflict Political Regimes The Plan of the Thesis Theory Degree of Accountability Authoritarian Regimes Literature review Political survival Political Survival and Armed Conflicts The Selectorate- and Weeks (2014)-Theory in Literature Situation of Thesis Summary - Theory and Literature Research Design Event History Analysis Non-Linear Effects Missing Data Data Dependent Variable Independent Variables The dataset Summary Analysis Descriptive Statistics ix

11 5.2 Results The Selectorate Theory on Political Survival Weeks (2014) Regimes on Political Survival Political Survival When Engaging War Conclusion Robustness Diagnostics Proportional Hazard Assumption Subsets of the Sample Residuals Deviance Residuals Score Residuals Limitations Summary Discussion and Conclusion Summary and discussion Conclusion References A Appendix vii ix

12 List of Figures 1 Welfare of coalition members Weeks (2014): Autocratic regime types Total Duration for Political Leaders, based on starting age and PolityIV 21 4 Total Duration for Political Leaders: Total Duration for Political Leaders, based on starting year Average Duration: Selectorate Theory World Map 2006: Size of Winning Coalition World Map 2006: Size of Selectorate Correlation matrix: Weeks and GWF regime categories UCDP/PRIO: Number of armed conflicts by type: Age when entering office for political leaders Size of Winning Coalition and Selectorate on Conflict Type Weeks(2014)-regimes on Conflict Type Hypotesis 1: The Selectorate Theory on Duration Model 6: GDP Per Capita and Starting Age Survival Curve: Winning Coalition Hypotesis 2: Weeks(2014) Regimes on Duration Model 3: Engaging War Model 3: Pre-Conflict Tenure Deviance Residuals Hypotesis 1: Score Resiuals Hypotesis 2: Score Resiuals Weeks(2014) Regime Index on Polity-IV Value ix 24 GWF Regime Index on Polity-IV Value ix 25 UCDP/PRIO Armed Conflicts: , Intensity level x 26 Pre-Conflict Tenure x xi

13 xii

14 List of Tables 1 Summary of Weeks (2014) assumptions on autocratic regimes Summary of Hypotheses Overview of the dataset Coefficients Model 1: S and W Matrix Subsets of Model 1 and Descriptive statistics xi 7 Imputated Modell xii 8 RESULTS: HYPOTHESIS xiii 9 RESULTS: HYPOTHESIS xiv 10 RESULTS: MODEL xv 11 Proportional Hazard Assumption Test xvi xiii

15 xiv

16 1 - Introduction When you ve spent half your political life dealing with humdrum issues like the environment, it s exciting to have a real crisis on your hands. Margaret Thatcher On the Falkland Islands war STATE leaders are perceived as their country s de facto leader. They have considerable influence on policy, both domestic and abroad. Questions of whether engaging in conflicts or pursing more peaceful solutions is one of central policy field, where their leadership is of great importance. Margaret Thatcher s quote above indicates how some policy areas are more important than others for state leaders. Manoeuvring their state through a crisis is one of their greatest concerns. If failing, a shrinking approval rate or losing office is probable consequences in the aftermath. The greater risk for losing office, the higher importance the situation has, for both citizens and leaders. This thesis will investigate to which degree citizens and political elites hold their state leaders accountable for policy crisis, measured as duration and perseverance in office. Engaging in an armed conflict is a potential crisis. High death tolls and prolonged wars could give the opposition support, increasing the probability of losing office for the incumbent government or leader. The next paragraph will investigate some cases of political leaders, whose commonalities are engaging in an armed conflict. In April 1982, General Leopoldo Galtieri of Argentina lead their military forces to occupy the British overseas territory Falkland Islands. The British forces struck back, resulting in a severe defeat to the Argentinian forces after 74 days. Likewise, in August 1990, Saddam Hussein invaded Kuwait. In this case a large Western coalition intervened, leading to Kuwait s renewed independence. These two cases have a lot of similarities, and some differences. Leopoldo Galtieri and Saddam Hussein where both dictators in the same time period; They ruled ruthlessly upon their citizens at large, and where only accountable to a small group of political elites; They initiated an armed conflict on a superior Western military power. The dissimilarities are striking. While Saddam Hussein continued his rule an additional thirteen years, Leopoldo Galtieri was replaced by another general four days after defeated. 1

17 What could explain the variation in duration after engaging in an armed conflict? In the same matter, George H. W. Bush intervened in Kuwait after aforesaid attack from Saddam Hussein and Margaret Thatcher dispatched the British Navy to recapture the Falkland Islands. While George Bush Sr. lost office after one period in office, Thatcher boosted her popularity significant during this war. Scholars argue that post-conflict support allowed her to regain office twice (Norpoth, 1987). The events have some similarities and differences. Both had a positive war outcome against dictatorships, both intervened as an answer to the other parts military aggression and both lead military superpowers in the same time period. The dissimilarities are mainly that Thatcher defended own territory, while Bush sr. intervened in a domestically area of interest. From these two examples, there seems to be some different factors affecting postconflict duration in office for political leaders. Saddam Hussein and Galtieri were leaders in autocratic regimes, with a low degree of accountability to their citizens. The former ruled as a personalistic dictator in a military regime, and the latter was elected from his comrades, with a higher degree of political constraints to domestic elites (Arquilla & Rasmussen, 2001, p. 762). This variation in accountability to elites may explain the difference in post-conflict tenure for these leaders, but not the differences between Thatcher and G. W. Bush. The similarities is striking; the seemingly similar political systems and ruling in the same period of time. The differences lie maybe in type of conflict? While Thatcher engaged in an interstate war, between two national states, Bush intervened in an ongoing conflict. Could this variation explain their post-conflict popularity? Horowitz et al. (2015) argue that the history of political events is made by people (p. xi). The theory assumes that political leaders are actors with influence, rather than just marionettes incentivized by international structures and domestic politics (ibid, p. 186). This thesis treats political leaders likewise, as an important actor in both international and domestic politics, whose influence on policy should not be underestimated. This thesis will mainly investigate to what degree the political leader is held accountable for their country s policy when engaging in an armed conflict. It assumes that political leaders are rational, and one of their main goals is to maximise their duration in office. An instrument to strengthen their own probability to maintain office, is relocation of the country s capabilities. Bueno De Mesquita & Smith (2010, p. 936) argue that political leaders want to maximise their political influence and design policies that shape the provision of private goods and public goods as the basis for 2

18 doing so. 1 How can engaging in an armed conflict be seen as rationale behaviour for political leaders? And to which degree are they identified as responsible for their policy to the electorate, both those entities supporting and not supporting them in the previous election. As in the example above, the electorate, both voters and/or political elites, will punish or reward their leaders depending on circumstances in the country (Lewis-Beck & Stegmaier, 2000). Bueno De Mesquita & Smith (2010, p. 936) points out that individual political leaders, as an unit of analysis, is a growing theme in political science. This may be, as Goemans (2000) argue, because the impact individual leaders have on the states policy, especially through crisis, should not be underestimated. This thesis will focus on perseverance in office for political leaders through these policy crises. The research question for this thesis is therefore: RESEARCH QUESTION: To what degree does engaging in an armed conflict affect the duration in office for political leaders in the post-war era? This thesis will seek to explore important factors affecting the duration for political leaders through armed conflicts. In the examples above, degree of accountability, autocratic regimes and conflict type could be prominent explanations for the variation in post-conflict tenure. Measuring accountability is an ambitious task without utilizing a theoretical framework. The analysis will therefore be conducted with the basis in two existing theories regarding political survival; the first on degree of accountability and the second concerning political leaders in autocratic regimes. In the Selectorate theory, Bueno De Mesquita et al. (2005) utilizes the degree of accountability the political elite have as the main explanation for the regimes latitude in policy. The different size ratios between the enfranchised entities and the subset of these needed to empower the political leaders gives them a divergent degree of leeway. In a typical democratic state, the political leadership is empowered by approximately one half of the enfranchised citizens. In autocratic regimes, due to, for example, rigged election systems, the proportion of enfranchised citizens needed to constitute political leaders is much smaller. Before presenting the theoretical framework utilised, some central concepts will be discussed. 1 This implies that political leaders behaviour is primarily based on rational choice theory. There are some weaknesses in regard to this theory. See Hindmoor (2011) for a brief discussion around some controversies concerning rational choice theory, and its appliance in political science. 3

19 1.1 Central Concepts To facilitate reading, I will first clarify the main terminology of the thesis. These concepts include: political survival, armed conflict and political regimes Political Survival The main concept for this thesis is the degree of perseverance in office for political leaders. Political survival is defined as the probability of holding office for the effective incumbent political leader. By effective leader, I utilize the same term as Goemans et al. (2009), which implies the person who de facto exercised power in a country. This thesis assumes, like Downs (1957, p.150), that political leaders chose policies to maximise their duration in office, through attaining the income, power, and prestige that go with office. These sets of policies affect to what degree the political leaders manage to hold office. The higher odds a political leader has of loosing office, the lower survival probability the leader have. The thesis will utilise the same assumptions regarding the behaviour of political leaders as Bueno De Mesquita & Siverson (1995). First, politics is competitive, even in autocratic regimes. The topics contested over and its regulations alters, both spatial and longitudinal. Second, it is a high priority for political leaders to maintain in office, hence power, and use the available sets of instruments at hand to accomplish this objective. Third, political leaders view the holding of office as a prize rather than as a burden. When maximising the probability of maintaining in office, the leaders utilize the set of applicable tools suitable in their polity. The next section will address armed conflict as a concept Armed Conflict An armed conflict is more than epic battles, fought between two or more rivalling states, like the napoleonic wars, the battle of the Somme or the Vietnam War. This thesis will utilize a broader definition of armed conflict. The definition will also encompass modern conflicts; whose battle related death tolls not necessary surpasses six digits and where the total duration could be counted in days. The Uppsala Conflict Data Program (UCDP) and Peace Research Institute, Oslo (PRIO) have collaborated on the UCDP/PRIO Armed Conflict Dataset. Here an armed conflict is defined as a contested incompatibility that concerns government and/or territory where the use of armed force between two parties, of which at least one is the government of a state, results in at least 25 battle-related deaths in one calendar year (Pettersson & Wallensteen, 4

20 2015). The research cited in this thesis has varying definitions, but the analysis utilize this aforesaid definition on armed conflicts Political Regimes Choosing between different democracy indexes should be a result of the research question Collier & Adcock (1999); Hadenius & Teorell (2005). This thesis focuses on how engaging in armed conflicts affect duration in office for political leaders, hence, the degree of democracy could be highly influential. The more autocratic a regime is, the less accountable the leader may be to their citizens at large. The reduced degree of accountability autocratism entail, could give the leader more leeway in his policies. Following Bratton & Van de Walle (1997, p. 38) political regimes is defined as the sets of procedures that determine the distribution of power. 2 Defining democracy and dictatorship is not straightforward. The differences that has to be considered is whether democracy is a continuous or dichotomous variable (Adcock, 2001; Munck & Verkuilen, 2002); whether the existence of particular institutions or some underlying principles is most important (Beetham, 1999; Cheibub et al., 2010); and thirdly, which sets of characteristics and traits that hallmark these regimes (Knutsen, 2011). The different approaches have their benefits and disadvantages. Every regime classification and index have problems with reliability or validity (Cheibub et al., 2010; Munck & Verkuilen, 2002). Therefore, as mentioned above, choosing how to measure democracy should be chosen as a means to answer the research question. The degree of accountability that the political leaders have, could be highly influential on their political survival. This thesis will therefore utilize two different measures of democracy, who focus primarily on the degree of accountability. The Selectorate Theory (Bueno De Mesquita et al., 2005) theorise that it is the size difference between the selectorate and the winning coalition, those given suffrage and the subset of these needed by a political leader to gain and hold office, that is most important. This index does not differ between political regimes, but measures political accountability as a continuous scale. Svolik (2012, p. 45) argue: the difference between democracies and autocracies is first one of kind and only then one of degree. Therefore Weeks (2014) autocratic regime index will also be utilized. This is because it classifies regimes in two dimensions: the degree of accountability the political leadership has and whether the leader has a military or civilian background. The inclusion of the accountability-aspect could be of great importance when analysing autocratic regimes. 2 "The regime" could be utilized referring to the incumbent leadership rather than regime type. 5

21 1.2 The Plan of the Thesis The thesis is organised as follows. Chapter 2 presents the theoretical framework, and derive a set of empirically testable hypotheses relating to political survival through armed conflicts. Chapter 3 review the scholarly debate on political survival, both concerning the utilized theories and political survival. Chapter 4 presents the quantitative research design utilized, based on event-history analysis. The findings are reported in Chapter 5, with its statistical robustness examined in Chapter 6. Chapter 7 will discuss the findings from Chapter 5, based on the theoretical framework utilised, preceding findings and the statistical limitations. Finally, some concluding remarks will be given. 6

22 2 - Theory I told all four that there are going to be some times where we don t agree with each other, but that s OK. If this were a dictatorship, it would be a heck of a lot easier, just so long as I m the dictator. Georg W. Bush, 18th December After meeting the congressional leaders, newly elected US. President Georg W. Bush stated the quotation above. Does he implicitly want to rule as a dictator, without the accompanying constraints imposed by democratic institutions? Did he think the reduced degree of accountability would give him a greater leeway in implementing his policy? In this chapter, I will present the theoretical framework, emphasising the political leaders and elites degree of accountability as the main explanation for why some leaders persevere in office after engaging in armed conflicts. Political leaders have a substantial impact on their country s policy, especially in foreign policy crisis. This chapter presents a theoretical framework for how engaging in an armed conflict may affect political survival, in light of his degree of accountability and the potential type of autocratic regime. If a state leader has a low degree of liability, the country ruled is most often an autocratic regime. On the other hand, this is not always the case. States with both autocratic and democratic institutions will not necessary meet those criteria required to be defined as an autocratic regime. Gates et al. (2006); Gurr (1974) argue that the most systemic unstable regimes is found in states where the political institutions have both autocratic and democratic characteristics. Thus, both Autocracies and Democracies exhibit institutional consistency that is self-enforcing (Gates et al., 2006, p. 907). This may affect duration after engaging in an armed conflict. When facing a severe policy crisis, the lack of self-enforcing political institutions may affect the ability to persevere in office. 3 Remarks by the President-Elect Following a Meeting with Congressional Leaders. 18th December

23 In this chapter I will present two theories on how the degree of accountability may affect political survival through an armed conflict: the selectorate theory (Bueno De Mesquita & Siverson, 1995) and Weeks (2014) theory on autocratic regimes. The selectorate theory emphasises the leaders degree of accountability to political elites and/or citizens, which is important for gaining or maintaining office or re-election. The key elements of the selectorate theory is the size ratio between those possessing the role as electors, and the size of the subset of these that is sufficient for the political leader to gain office. The smaller winning coalition, the subset of the electorate the leaders need to gain office, compared to the electorate, the lower degree of accountability. In Weeks (2014) theory, autocratic regimes are divided through two dimensions: the military - non-military axis and the degree of accountability the leader has upon political elites. This creates four different autocratic regimes, whose labels are collected from Slater (2003): Machine, Junta, Boss and Strongman. The theory emphasises how autocratic regimes, where the leaders have a military background, are more likely to utilize military measures in policy crisis, because of their knowledge of these. Likewise will political leaders with a low degree of accountability to political elites have a greater leeway in their foreign policy. These two theories emphasise two different aspects of consideration in political survival: the degree of accountability the leader endows and the potential military background. When measuring political accountability, several factors are seemingly important. First, the degree of support a leader needs for their day-to-day policy making. With a low degree of accountability a political leader have fewer restrains in their policy. Although accountability is important, regime type does also affect political survival. Different regimes could have a variety of underlying characteristics that could alter the total tenure for political leaders. Following Svolik (2012), both regime type and political accountability is included in the analysis. The reason for this is that different measurements could give complementary explanations on how engaging in an armed conflict could affect duration in office. The chapter will proceed as following. First I will present Bueno De Mesquita et al. (2005) Selectorate theory on political accountability. Second a presentation of Weeks (2014) regime classification. One hypothesis is derived from each of the theories, summarised in Table 2. 8

24 2.1 Degree of Accountability Bueno De Mesquita et al. (2005) initially investigates different characteristics autocratic regimes have that best could explain the perseverance of political leaders in office. They argue that it is not the type of regime in it self that affects their survivability, but the regime type is a result of how broad support the leader needs to obtain and preserve his position. The selectorate theory emphasises the percentage of citizens with suffrage the incumbent leader needs to accede to for maintaining office as the most important aspect of political survival. The selectorate theory categorises citizens in polities according to their importance for the political leaders survival. The residents (R) are the total amount of inhabitants in a polity. Nested within the residents, there is a selectorate (S), defined as the set of people whose endowments include the qualities of characteristics institutionally required to choose the government s leadership and necessary for gaining access to private benefits doled out by the government s leadership (ibid, s.42). This definition is different from an electorate, those citizens with suffrage. The definition of the selectorate implies that all residents have a desire to be in the selectorate, so that they could benefit from the private commodities granted by the leader. In post-modern democracies, the selectorate is usually citizens with suffrage, where there are few formal restrictions on voting in the election. The subset of the selectorate the leader needs to obtain or gain office is called the winning coalition (W). This group is defined as the subset of the selectorate of sufficient size such that the subset s support endows the leadership with political power over the remainder of the selectorate as well as over the disenfranchised members of the society (ibid, s. 51). The endowment of a leadership from a sufficient size of the selectorate leads to a winning coalition empowering the leader. The disenfranchised entities is residents not in the selectorate, N S, and does not have any political influence on a day-to-day policymaking. All relevant polities have a leader (L), all whom is facing one ore more potential challenger (C). 4 Leaders and challengers compete in support from a large enough share of the selectorate to gain or maintain office. This could be through allocating resources in their control, or pledging potential resources for the challengers. In an ideal world the political leadership needs support from S/2 + 1 members of the selectorate to obtain or acquire his post. If S N, there will be few disenfranchised entities because of the broad support the leader has throughout the polity. 4 The abbreviations used in the selectorate theory will be embedded throughout this thesis. Residents=R; Selectorate=S; Winning Coalition=W; Leader=L and Challenger=C. 9

25 Liberal democracies differ from countries where the political institutions have autocratic characteristics. These regimes create a set of required qualities for accessing membership in the winning coalition, creating an artificial scarcity of members in this group (ibid, p. 54). The different combinations of sizes of W and S give the leader different sets of appropriate policies. The selectorate theory utilises the size difference between the selectorate and the winning coalition as the main explanation for why some leaders have a greater degree of perseverance, especially through policy-crisis. This relationship W/S, which they call the loyalty norm, has a theoretically high explanatory power for the leaders degree of accountability (ibid, p.65f). One of the most important aspects in the selectorate theory is that individual members of the winning coalition are interchangeable with members in the selectorate. This is most apparent in policies with a small winning coalitions and a large selectorate. This combination is most common in autocratic regimes, because the demand of attaining a membership in the winning coalition far exceeds the supply. As a result of this created scarcity, the leader could create a coveted upper class, from which he could sustain his position by granting private goods, and other benefits, to the members of the winning coalition. In the unlikely event of an individual member redrawing from a small winning coalition, the individual concerned is easily replaced by someone in the large pool of selectorate-members. This creates a high cost for members in the winning coalition for opposing the current regime and gives the leader great latitude in their policy. In case a potential challenger or their supporters oppose the regime, leaders in autocratic regimes, with a small W and a large S, could engage in the most brutal and extensive oppression. The high loyalty from the small winning coalitions will make these leaders prevail through policy crisis. Bueno de Mesquita et al. theorises that, although there is a wide range of seemingly different regime types, such as democracy, monarchy, autocracy, junta, oligarchy et cetera, very few political systems are spatial or temporal identical. Bueno De Mesquita et al. (2005, p. 69) argue that the size of the winning coalition and the selectorate can be readily related to conventional labels for describing different political systems. In democratic states S N and W is typical a simple majority of S. In a single-party dictatorship, S could be less than 10 percent of N, and the leader could ignore the vast majority (N S) in a day-to-day policy making. In juntas the selectorate is also small, and the winning coalition could be a handful of colonels or generals. But even though regime categories exist, the selectorate theory emphasises the importance of estimating the values of W and S on a continuous scale. Regime categories cannot explain variations within, for example, juntas, but the selectorate theory 10

26 could find an explanation for these variations in small fluctuations in the size of W or S. The likelihood for a new leader in regimes with a small W is therefore low. Because of this, we find the most repressive political leaders in these regimes. The larger size of the winning coalition, the harder it is for the leader to maintain the aforesaid loyalty from W. The larger W, the more the leader has to rely on public goods to appease the winning coalition. This is due to the cost of private goods is lower when doled out to many, instead of private benefits. When W is large, all members of the winning coalition only grant public goods. Since public goods in nature is public, all residents get the approximately the same benefit. The loyalty to a particular leadership with a large winning coalition will therefore be low. The theoretical welfare of individual members of W is highest with a very small or a large W. This relationship is illustrated in Figure 1, a replication of Figure 3.2 in Bueno De Mesquita et al. (2005, p.97). Figure 1: Welfare of coalition members The main reason this theory is important when conducting research on political survival, is that the size of the selectorate and the winning coalition could set the premises for the leaders post-conflict survivability. In countries with a large winning coalition, the political leader needs a broad consensus from the large winning coalition before applying military force as a policy-instrument. Without this support the leader would suffer an increased odds for allocation. Because of the large W, the leader cannot compensate the winning coalition with private goods in an 11

27 event of a foreign policy failure. It follows that democratic leaders generally only attack when anticipating an easy victory and needs a broad political consensus for not loosing office. Democratic states are therefore highly selective on which conflicts to fight or negotiate. In those cases where democracies chose to fight, they are more likely to win (ibid, p.236). The size ratio between W/S is important when deciding whether to fight or to negotiate. States engage in an armed conflict when the expected utility exceeds that of negotiation: P A > (χ + Ψ + k)/(1 + Ψ) + R A /(1 + Ψ)W A. Where P A is country A s probability to win the armed conflict; χ is the military balance for country A, 1 χ for country B; Ψ is the leader s utility for remaining in office; k is the players per capita cost for engaging war, where the utility for victory equals to 1 k and defeat k; R A is the total pool of resources available for country A and W A is the size of the winning coalition in country A. Given the large winning coalition in a democratic state, the term R A /(1 + Ψ)W A will be small. The importance of the term (χ + Ψ + k)/(1 + Ψ) depends on the value of holding office. This indicates that political leaders in democratic states, with a large winning coalition, have to be almost certain of a favourable military outcome before engaging in an armed conflict, or else risk an increased odds of loosing office (Bueno De Mesquita et al., 2005, p.238ff). As illustrated in Figure 1 on the previous page, the members of a small winning coalition have incentives for maintaining the small coalitions, or else potentially lose their welfare. The selectorate theory discusses primarily small or large winning coalition, and gives little attention to medium-sized coalitions. In fact, the word medium is not used once in their book. From Figure 1 on the preceding page we see that the medium-sized winning coalition have a small amount of welfare. How will this affect political survival? There is a tipping point where the loyalty to their leader diminishes, because the probability of a new leadership surpasses a certain level. Because of the low level of welfare, combined with the raised possibility of a leadership change, political leaders in medium sized W will have little room for errors and policy-crisis. There is some uncertainty about the policies and degree of accountability for political leaders in medium sized coalitions. Bueno De Mesquita et al. (2005, p. 96) points out that an enlarged W has two effects: the individual reward of a membership in W decreases, because the overall number of members the leader has to allocate resources is increased. Second, because of the reduced benefits each member of the winning coalition receives, their loyalty decreases. The number of disenfranchised members will also potentially decrease with a 12

28 higher W. The motives to protest diminish when the society is dependent on a large coalition, because there are a low number of oppressed disenfranchised entities. Likewise, the willingness to protest is reduced when the costs is too high, because the probability of a new leader is low, due to a small W. The implication of this is that there may be a set of W and S, approximately corresponding to a medium sized winning coalition, which could enhance the probability of internal unrest. Based on these assumptions on states with medium-sized coalition, a testable hypothesis is deducted: H1: For political leaders in countries with a medium sized winning coalition, their duration in office shortens significantly in case of an armed conflict. Monarchies and juntas with possible the same size of W and S, could inherently have different properties, concerning to which sets of policy they most likely utilize in policy crisis. Chang & Golden (2010) find evidence that the size of the winning coalition is approximately the same across authoritarian regimes, but the selectorate differs in this theoretical order: military juntas and monarchies < single-party dictatorships < personalistic regimes. This gives support for that regime categories have some explanatory power because juntas and monarchies have approximately the same size combination of W and S. Juntas and monarchies have seemingly different characteristics, most apparent because of the differences in the political leaders military background. Weeks (2014, p.8) does not manage to find the same linear relationship between the winning coalition, the selectorate and their respective policy implications, as the selectorate theory assumes. Weeks makes the point that, in case of a change in political leadership, not all of the members in the winning coalition will lose their power and influence. This is because many of the members in small winning coalitions have their own independent bases of power, and the potential new leader, the challenger, needs support from these bases to legitimate their role as a leader (p. 8). Different autocratic regimes could therefore be important for explaining variations in duration for political leaders when engaging in an armed conflict. The next section will present Weeks (2014) theory on autocratic regimes. 13

29 2.2 Authoritarian Regimes In this study, I will analyse how engaging in an armed conflicts may subsequently affect duration in office for political leaders. Scholars often use one-dimensional measures of democracy and autocracy, although some exceptions exists, for example Gates et al. (2006). While this provides some insight on whether an increase or decrease in the aggregated level of democracy or autocracy is associated with different patterns of post-conflict duration for political leaders, it masks the variation between democracies and different types of autocratic regimes. Thus, including autocratic regimes in the analysis may give some explanations for variations, besides merely the degree of democracy. Weeks (2014) theorises that different types of autocratic regimes will have divergent incentives for engaging in war or to negotiate, when solving international affairs. This is apparent even though these regimes could have the same W/S-ratio. According to Weeks, it is the variance between different types of autocratic regimes that has the greatest explanatory power, when explaining their policy. The probability of using military force as a policy instrument, according to Weeks, differs across four aspects: the benefits of winning a war compared to maintaining status qou; the costs of engaging war; the cost of eventually losing the war and the likelihood of winning the war (Weeks, 2014, p.15). In Weeks theory, autocratic regimes are arranged in a two-dimensional index: whether the political leader is accountable or not to a strong domestic actor and whether the leader has a civilian or military background. This leads to four ideal types of autocratic regimes: nonpersonalistic civilian regimes, where the leader have a civilian background and a high degree of accountability to strong domestic civil elites; nonpersonalist military regimes, where the leader have military background and a high degree of accountability to strong domestic military elites; personalised non-military regime, where the leader have a civilian background and a low degree of accountability to a weak domestic civil elites; and a personalised military regime, where the leader have military background and a low degree of accountability to weak domestic military elites (ibid, p.17ff). Weeks adopt Slater (2003) labels on autocratic regimes: nonpersonalistic civilian regime, (machines); nonpersonalist military regimes, (juntas); personalistic regimes led by civilians, (bosses); and personalistic regimes led by military officers, (strongmen). Weeks utilise cut-points on the regime index. The countries surpassing a certain level on both the accountability axis and the background axis, will be classified as a regime. If the country does not exceed both of these criteria, the regime will 14

30 not be defined as a Weeks(2014)-regime, and be classified as inapplicable in the regime index. This is illustrated in Figure 2. The X-axis illustrates whether or not the political leader has a military background, and the Y-axis illustrates in which degree the leader is constrained from a influential domestic actor. The stapled line illustrates the theoretical cut-points and the boxes those combinations of X an Y that is applicable to be defined as a regime. Figure 2: Weeks (2014): Autocratic regime types A contemporary example of a civilian-led machine is China. These types of regimes are highly selective in which conflicts to escalate with military force. The theory explains this by the leaders high degree of accountability to domestic civil elites, in this example the Communist Party, and their willingness to attain a consensus among these. Additionally the leader and elites of a typical Machine do not have a military background, at least not to a great extent. They are therefore not in the same way acquainted to these policy-instruments compared to Junta and Strongmen regimes. This leads to a very low number of machines that have engaged in an armed conflict (Weeks, 2014, p. 36). In juntas the political leader, with a military background, is accountable to a domestic elite composed primary from within the armed forces. The leader could on the one hand be reluctant of using force in a policy crisis, because he is strongly accountable in case of failure. On the other hand, Juntas are more likely to engage in armed conflicts than machines because of their knowledge about military as a policy instrument. The military elites does not have the same confidence in diplomacy 15

31 as an effective and reliable policy instrument, and the costs of using the armed forces as a tool are perceived as lower. Juntas may also engage in an armed conflict more often because their military training implies the advantages of offensive action and first strikes. To possess the initiative in an armed conflict can reduce the fog of war (Clausewitz, 1976), the uncertainties arising during these crisis (Weeks, 2014, p.24ff). Personalistic dictators in the regime type bosses have a low degree of accountability upon their inhabitants and elites. They are therefore resilient for punishment from domestic actors in case of a policy failure. This could lead to enhancing the probability of utilising the armed forces. Personalistic dictators are less likely to perceive high costs for using military force or defeat at war, due to their biased cost analysis. While juntas and machines compute for their domestic audiences potential wrath in a polity crisis, Personalistic dictators only account for their own personal cost (Weeks, 2014, p.29ff). Because of the dictator s low degree of accountability to their political elites, the consequences of a policy-failure are low. On the one hand, they are typically only accountable to a few chosen Generals and/or politicians, who safely could control their subordinates with an iron fist. On the other hand, these leaders do not have military background, so they could therefore be more reluctant of using armed forces as a policy instrument. Political leaders in strongmen regimes, personalist leaders with a professional military background, are more likely to use military force as a policy instrument because of their knowledge and earlier experience with the armed forces. The leaders in strongmen regimes enjoy a small degree of accountability, and will therefore have decreased odds for domestic retaliation in case of a failure in foreign policy, compared to leaders in juntas who also have military background. This may lead to an increased probability for strongmen regimes to engage in an armed conflict, in comparison to juntas (Weeks, 2014, p.32f). As bosses, leaders in strongmenregimes have biased and self-centred cost analysis when considering different policies. Compared with bosses, a political leader in strongman-regimes has increased odds for utilising the armed forces, because of their personal military background. Table 1 summarises the political leaders different theorised likelihood for initiating an armed conflict, for defeat and for domestic retaliation, a replication of Table 1.3 in (Weeks, 2014, p.35). As described in Table 1, the lower degree of accountability a political leader has to domestic elites, the lower probability for retaliation from domestic elites or inhabitants in case of a severe failure in policy. When the odds of a domestic punishment decrease, the sets of policy instruments a leader could utilise without domestic re- 16

32 Table 1: Summary of Weeks (2014) assumptions on autocratic regimes Machines Juntas Bosses Strongman Armed conflict initiation Low Medium High Highest Probability of defeat Low Medium High Highest Probability of punishment High Medium Low Low Process Consultative, Consultative, but Little consultation; See juntas cautious, focus on military violence/military and bosses. incorporates rather than force seen as civilian and diplomatic or political business as usual military input aspects of decision (Saddam Hussein (North Vietnam (Argentina/Falklands and Joseph Stalin) and USSR after War and Japan Stalin) 1930s/1940s) taliation gets broadened. These leaders have, according to Weeks theory, higher odds for initiating an armed conflict. We should expect that strongman and bosses engage in significantly more armed conflicts, due to the leaders perceived costs of initiating them are low. They will also, because of their low accountability, in a greater degree endure through these policy crises. From this, a testable hypothesis is deducted: H2: For political leaders in Boss and Strongman regimes (Weeks, 2014), engaging in an armed conflict has no effect on their duration in office. While Weeks focus on Militarized Interstate Disputes (MID) 5, the proportion of civil wars has surpassed these in the post-war era (Themner & Wallensteen, 2014). As discussed in the introduction, there are several types of armed conflicts. How will different types of armed conflicts affect the duration for political leaders in the different regime types? From the degree of accountability the leader have, we may expect that leaders with low accountability with a military background will not suffer a decreased probability of survival through an armed conflicts, for example a civil war. On the one hand, their vast sets of policy instruments may give them an advantage when managing threats to their position. On the other hand, the probability of a civil war is heterogeneous, some countries and regions are at higher risk of a civil war (Collier & Hoeffler, 2004; Collier et al., 2008; Fearon & Laitin, 2003; Hegre et al., 2001). Inconsistent regimes are more prone to a civil war, because of 5 From the Correlates of War (Sarkees, 2010) 17

33 their weak institutions failure to address challenges to their authorities (Gates et al., 2006; Hegre et al., 2001). Although autocratic leaders prevail longer in office than their democratic colleagues (Bueno De Mesquita et al., 2005), democracy as a system far exceeds both autocratic and mixed regimes in duration (Gates et al., 2006, p.900). The causal relationship between how systemic factors affect civil war probability and duration of political leaders is uncertain. On the one hand, Gaddafi, Mubarak and Bashar al-assad prolonged tenure could have influenced their political survivability, because of their accumulated sets of instruments and means to maximise their current and future power and authority. 6 On the other hand, their prolonged tenure could stir up a rising level of discontent among their inhabitants and elites, which in turn lead to a domestic revolt, for example the so-called Arab spring. Political leaders in bosses and strongman have a broader set of applicable repressioninstruments because of their low degree of accountability, and revolts may therefore not affect their survivability. The two theories on political survival have different explanatory factors on the leaders degree of accountability. From the theories, two testable hypotheses have been deducted. These are summarised in Table 2. Following chapter is a brief summary of earlier research on different aspects affecting political survival. 6 The last part is collected from the main assumption of Gates et al. (2006). 18

34 3 - Literature review Warfare is a means and not an end. Warfare is a tool of revolutionaries. Fidel Castro, on Che Guevara 7 The literature review will look into what factors affect political survival in general, and more specifically how engaging in an armed conflict affects total tenure. While a prolonged tenure and avoidance of non-natural causes of death (Goemans, 2000), is one of the primary objects for political leaders, which factors influence political survival? This literature review will give a brief summary of important findings on this subject. Firstly, I present research relevant to political survival in general. Second, I present a more extensive discussion of the relationship between political survival and initiating an armed conflict. In the third part, I present key findings relating to the selectorate theory and Weeks (2014) theory. Finally, I will situate my own work as a continuation of existing work. 3.1 Political survival Political survival is, as mentioned in earlier, defined as the degree of perseverance a political leader has in office. Several factors alters the probability of losing office. Accumulated tenure is a good predictor of whether or not a leader will lose office. Bienen & Van de Walle (1989) find evidence showing that the tenure a political leader accumulates captures variables that we cannot measure independently, such as skill-level and the ability to build robust networks. Likewise Bienen & Van de Walle (1992) find that political leaders who have prevailed a decade, has an increased odds for retaining power. Pre-conflict tenure may be therefore be crucial for political leaders engaging in armed conflicts. Bueno De Mesquita & Siverson (1995) finds evidence that a one-unit increase in the length of their pre-war tenure [...] reduces the risk of postwar removal by 38 percent. 7 Castro (2006, p. 91) 19

35 The effect prolonged tenure have on leader survivability is theoretically declining through infirmity, because of high age. Both Bienen & Van de Walle (1992) and Burke (2012) find evidence supporting age on entry have a negative effect on duration. Although Bienen & Van de Walle (1992) lack significant results from the age-variable when adding the squared, this effect may be non-linear in nature. Horowitz et al. (2005) finds that, as the age of the political leader increases, the risk of engaging in an armed conflict increases. This effect is apparent in democratic states. In personalist regimes this effect is inverted. Older leaders may be driven of an interest in establishing a legacy fast, because of declining health Horowitz et al. (2005, p.46). This could drive them to take riskier decisions. (Svolik, 2012) argue that a dictator must solve two conflicts to gain power and enjoy a prolonged tenure: the problem of authoritarian control and power-sharing. The first threat is from subjects over whom they govern, and the second threat is from political elites the dictators share power with. Authoritarian states lack, compared to democracies, an independent authority, like a court of law, with the power to enforce agreements among key political actors. In these states, non-constitutional brutality, repression and violence are an likely alternative in a policy crisis. There are two main types of dictatorships: contested and established autocracy. In the latter, the ruler has effectively monopolised power, like in Weeks (2014) regimes Stongman and Boss. Svolik observe that those countries where the authoritarian ruler has an effective term limit, like PRI era Mexico and post Deng Xiaoping China, the power is not in the same way inheritable. Ambitious political clients could support the next generation of political leaders, reducing the vacuum of power normally created in dictatorships in an event of leadership successions. Another key to success is depersonalisation of political authority. This method of authoritarian power-sharing reduces the risk of a mortal post-office fate. The youngest state leader in the dataset is Jigme Singye Wangchuck of Bhutan and Hussein bin Talal of Jordan. They gained power at the age of 17 and ruled respectively 26 and 46 years. In this dataset, presented in chapter 4, 15 political leaders gained office before the age of They had an average duration in office of years, compared to 4.44 years on all leaders. Bienen & Van de Walle (1992) use a concave curve to explain the theoretical effect age has on political survival. The problem with this method is that the youngest and the oldest leaders are found in autocratic regimes. Figure 3 shows how degree of democracy, measured with the Polity-IV index 8 See Figure 11 on page 51 20

36 Figure 3: Total Duration for Political Leaders, based on starting age and PolityIV (Marshall, 2002), and starting age affects the leaders duration in office. The brighter colour, the longer average duration for the groups identified by the x- and y-axis. The Polity-IV index measures, in short, degree of democracy. The scale goes from - 10 to 10, where higher numbers indicates a more democratic state. The figure shows that the youngest leaders have the longest duration in those regimes with the most lack of democratic institutions. The most democratic countries do not have leaders with a starting age under 30. For political leaders in the middle age group, the effect of starting age and Polity-IV value have an almost linear effect on duration. Those leaders whose starting age surpassed 60 years have a reduced mean duration. This could also be because of their total potential duration is reduced due to high age. A countries state of economy has a large impact on leader survival. It is especially the elected leaders degree of survivability that will be highly influenced. Voters punish and reward leaders based on good economic achievements (Barro, 1973), which indicates that economic conditions are highly influential in developed countries. Voters reward governments in good times and cast them out in bad (Lewis-Beck & Stegmaier, 2000). This effect is unclear in developing countries, but apparent in states with an emerging market (Wilkin et al., 1997). Their findings indicate that economic growth in election-year influences the vote of the major party in office. Burke (2012, p.32) finds that GDP growth, as a result of increased prices for commodity exports, strengthens the prospects of a prolonged tenure for the leader. Likewise will economic recession depreciate the leaders support, especially when 21

37 utilising constitutional methods. Burlacu (2014) finds evidence that both good economy and governance are important for political survival. Good governance is particularly significant in poor democracies with a high degree of economic growth. The importance of economic growth on political survival may therefore be more necessary in weak economies (ibid, p.181). Tir & Singh (2013) investigates how foreign crisis affect the incumbents support. The results indicate that crisis-related popularity [...] cannot repair the damage caused by the economy that is faltering at a highly unusual rate (p.98). The phenomenon is entitled the rally round the flag-effect (Mueller, 1973). He argues that political leaders enhance their temporary probability for survival if they initiate an armed conflict. This demonstrates the interaction seemingly different policy areas have on political survival. Hence, the next section will examine literature relating to the research question more closely. 3.2 Political Survival and Armed Conflicts Democracies are better at signalling their intentions in international disputes, because of their high audience cost if they choose to escalate a policy crisis to the brink of an armed conflict (Fearon, 1994). Weeks (2008) questions the low audience cost in autocratic regimes, which Fearon models with. Electoral procedures are not the only level of accountability a political leader must adhere to. Elite coordination is equivalent with electoral procedures in autocratic regimes, Weeks argue. The political leader could utilise monitoring and punishment to restrain elite coordination or prevent foreign decision makers insight of such coordination, thus it is harder to observe for outsiders. The rally round the flag-effect (Mueller, 1973), mentioned in the last section, may increase the leaders support in armed conflicts. Oneal & Bryan (1995) investigate this effect in all 41 U.S. foreign policy crisis between 1950 and 1985, and find a small, but significant increase in support for the incumbent president post-crisis. Likewise Morgan & Bickers (1992) argue that political leaders could shift focus from their personal policy failure, and blame the recession on international conflicts. The problem with the research on post-conflict popularity is that the authors mainly address support from voters in democracies, and most studies focus on presidential support in the United States (Tir & Singh, 2013, p.83). Is this effect also apparent 22

38 in non-democratic countries? In democratic states, where the leader have support from a large selectorate (Bueno De Mesquita et al., 2005), the leader could initiate an armed conflict to boost, perhaps not successfully, their short-term popularity. Or like in the Falkland War, initiate an armed conflict to move the attention away from internal tumult (Bueno De Mesquita & Siverson, 1995, p. 841). Levy (1989) argues that political leaders more often engage in armed conflicts when facing an increased risk of losing office. Likewise Chiozza & Goemans (2003, p. 459) find evidence that more time in office increases the probability of crisis initiation, and more periods in office decrease the probability of crisis initiation. This implies that leaders with a prolonged tenure may have, due to their experience, more confidence. The authors also argue that, as the risk of losing office increases, leaders become less likely to initiate a crisis. In the same way, as the risk of an international crisis increases, leaders become more likely to lose office. The literature is divergent, and it seems to be some uncertainty on the causal direction. Is it the elevated risk of international crisis that affect political survival negatively, or is the extended tenure that elevates the risk of an armed conflict? Or are there a set of independent factors that cause both extended tenure and an elevated risk of an armed conflict? Horowitz et al. (2015) use a different perspective for why some leaders are more war-prone. They utilize a psychological perspective on leadership behaviour. The focus is more on internal factors regarding the leader than external aspects, on for example a systemic level. They investigate how events in adolescence and early adulthood affect policy, more specific future military behaviour of political leaders (ibid, p. 34f). The main findings indicate that those who served in the armed forces, but without real combat experience, have a heightened risk of engaging in a war. Those leaders who experienced real combat are more cautious in engaging an armed conflict. Horowitz et al. (2015) explain war probability with both systemic factors and leader properties. Some have a high level on one measure, but those who have a high value on both measures are very likely to engage in an armed conflict. Their main point is that Individual leaders do matter; heads of state are not simply interchangeable or continuously overwhelmed by exogenous factors. These executives can change the course of history in powerful ways (ibid, p. 179). Bueno De Mesquita & Siverson (1995) is the first study that approach political survival through an armed conflict with a multinational dataset. There had been some research on political survival and war, but their concern was limited to the United States (Brody, 1991; Brody & Page, 1975; Kernell, 1978; Mueller, 1973) and Great Britain (Norpoth, 1987) in the second half of the twentieth century. Bueno De Mesquita & Siverson (1995) article applies data from the Correlates of War project 23

39 (Sarkees, 2010; Singer & Small, 1972), consisting of 191 cases of interstate conflict participation between 1823 and The main findings indicate that, when initiating an armed conflict, non-democratic political leaders have longer duration in office than leaders who govern democratic states. These findings gives some support to the selectorate- and Weeks (2014)-theory. Because non-democratic leaders have a reduced degree of accountable to a broad public opinion among their citizens, it creates a greater leeway in their policymaking. Both Chiozza & Goemans (2003) and Bueno De Mesquita & Siverson (1995) use a threefold regime classification, based on dummy coding of the Polity index (Marshall, 2002). Goemans (2000) examines incentives individual leaders have to end or continue a war and the implications this may have for the duration and costs of war. He does not directly estimate the leaders probability for survival, but the findings are of considerable importance when analysing their probability of holding office after engaging in an armed conflict. Individual leaders have a significant influence on war-outcome. Leaders of dissimilar regimes can expect different post-office fate depending on the war-outcome. Leaders of mixed regimes can expect the same degree of punishment whether they lose moderately or disastrously. Therefore these leaders have incentives for continuing fighting, even when facing a certain loss. Chiozza & Goemans (2004) study supports this finding. Political leaders in autocratic and mixed-regimes suffer a much more severe punishment, measured in duration, than those in democracies, if their common denominator is losing a war. Political leaders facing a severe crisis experience a lower hazard of losing office, than those who engage in an armed conflict. This creates an incentive to continue fighting for these leaders, although the prospects of winning are limited. If the state wins the war, there is no significant positive effect on duration for political leaders, for either of the regime types. Likewise Croco (2011) examines how individual political leaders affect the waroutcome. The results indicate that the culpable leader, the incumbent leader who either presides at the beginning of the war or a political leader from the same party that inherits the war, in greater degree achieves favourable war outcomes. Domestic audiences will spare the non-culpable leader in the event of losing the war, but not the culpable leader. On the other hand, the voters will not have mercy on the nonculpable leader forever. They have a window of opportunity to disengage their countries military engagement. If not utilised, their risk of removal is increased. Engaging in an armed conflict is not equivalent to losing office for political leaders. It depends on factors like regime type, degree of accountability, state of economy and if the leader is culpable for the conflict. The next section will review the 24

40 literature in regard to the selectorate theory and Weeks (2014) regime classification. 3.3 The Selectorate- and Weeks (2014)-Theory in Literature The selectorate theory has, as Bausch (2014) points out, emerged to become one of the major theories for scholars in international relations, especially when attempting to explain the effect of domestic politics on issues of war and peace. Bueno de Mesquita and Siverson have used the theory for different policy aspects. Most apparent is Bueno De Mesquita & Smith (2010), using the selectorate theory to analyse the survivability for political leaders through revolutions and finance crisis. Although oil exports, GDP and population size lack a consistent statistical significant effect on leader survival, the size of the selectorate is highly significant. This implies that political survival is improved when the leader can select supporters from a large pool of inhabitants. The size of the winning coalition has no effect alone. Combined with variables like the political leaders age, W has an effect. The age variable could therefore be considered as important in non-democratic states. Bausch (2014) conducts a laboratory test to investigate if leaders of polities with larger winning coalitions invest more heavily in public goods. The results indicate that leaders in small coalitions keep more of the polity s wealth for them self. Clarke (2008) finds evidence to support the notion that the size of the winning coalition is a very good measure when predicting the provision of public goods and private benefits. They contradict some of the critics from Clarke & Stone (2008), that the use of residualisation in (Bueno De Mesquita & Siverson, 1995) analysis is questionable because of omitted variable bias. Gallagher and Hanson s (2015) main critic of the selectorate theory is it lack of regime categories. They support Svolik (2012, p. 45) statement, that the difference between democracies and autocracies is first one of kind and only then one of degree. The selectorate theory captures, like the PolityIV-index (Marshall, 2002), the degree of democratisation. On the other hand, regime categorisation versus degree of democracy as a continuous measure is a ongoing heated discussion (Knutsen & Nygaard, 2015). Bak et al. (2015) utilise Weeks (2014) regime classification to investigate which regime that is most likely to engage militarily against opponents. The results indi- 25

41 cate that all regime types, except for military regimes, are equal cautious selecting targets in militarized disputes. Military regimes are significant more war-prone than other regime types. Likewise personalistic regimes (Weeks, 2014) are more likely than others to pursue nuclear weapons (Way & Weeks, 2014). This gives evidence that these regimes pose a serious threat to international peace and security. Because Weeks regime classification is modelled along two dimensions, it seems to improve some aspects of research on autocratic regimes, capturing a broader understanding of authoritarianism. After first presenting two theories on international relations and political survival in chapter 2, then a brief summary of previous research in the first part of chapter 3, the following section will situate my thesis as a continuation of research on this field. 3.4 Situation of Thesis This section will situate the thesis as a continuation of earlier research. After presented the literature on political survival, it is apparent that there are some areas that need to be further scrutinised. Firstly, research on the duration of political leaders in office use mainly "old" data. For example Bueno De Mesquita & Siverson (1995) studies political leaders between 1823 and There have been some major upheavals since the nineteenth century. More citizens benefit from political leaders with a greater degree of accountability and the media coverage is more comprehensive in the post-modern era, compared to before. Burke (2012) uses data on political leaders from 1962 to 2006, but does not consider how engaging in an armed conflict affects leader survivability. This thesis first contribution is to also use contemporary political leaders, with data up to 31th December See Section 4.3 on page 36 for an overview. Secondly, previous research on how engaging in an armed conflict affects political survival (Brody, 1991; Brody & Page, 1975; Bueno De Mesquita & Siverson, 1995; Croco, 2011; Kernell, 1978; Mueller, 1973; Norpoth, 1987) has only addressed how engaging in an interstate conflict affect degree of survivability. Because there have been few interstate conflicts since 1980, implementing only these conflicts will lead to a small degree of statistical variation. Therefore other types of armed conflicts will be included, like civil wars. 26

42 Thirdly, the use of regime categorisation is underrepresented in previous research on political survival. Analyses of systemic duration find that countries with both democratic and non-democratic institutions have a negative affect on duration (Gates et al., 2006; Gurr, 1974). There is something about regimes with both democratic and non-democratic institutions that affect regime duration negative (Gates et al., 2006; Gurr, 1974), and the same factors could be apparent in political survival. Bueno De Mesquita & Siverson (1995); Chiozza & Goemans (2003) and Croco (2011) operationalize democracy and autocracy as a dichotomisation of the country s Polity value (Marshall, 2002). Implementing regime categories could potentially decrease omitted variable bias and give further understanding of these regimes. 27

43 3.5 Summary - Theory and Literature In this section I will highlight important elements from the theory and literature chapters. In chapter 2, I presented two theories that could explain the diversity in post-conflict tenure for political leaders; The selectorate theory and Weeks (2014) regime index. The selectorate theory (Bueno De Mesquita et al., 2005) emphasises the size difference between the selectorate and winning coalition as the single most important aspect for explaining a political leaders duration and policy. It is the ratio between these groups that explains the leaders degree of accountability and secondly his ability to persevere in office. States with a large selectorate and a small winning coalition will have leaders with the lowest level of accountability, leading to prolonged tenure and the highest odds for initiating an armed conflict. This is due to the high cost individual members of the winning coalition will suffer if opposing their leaders. The ratio between the selectorate and winning coalition is theorized as the loyalty norm. Figure 1 on page 11 show that those states with a high and low winning coalition have a high degree of welfare. In a medium sized coalition, where the winning coalition has least welfare, the loyalty to the leader is decreasing under a threshold, making leader changes possible outside of the established administration. Because of uncertainties relating to political leaders perseverance in office in medium-sized winning coalitions, a testable hypothesis is deducted. Weeks (2014) theory may be better than the selectorate theory when explaining variance between different autocratic regimes, potentially with the same loyalty norm. This is because the theory captures important aspects relevant to the different types of autocratic regimes, like the variation in policy between political leaders with a military or non-military background. At the same time the theory includes the degree of accountability the leader has upon their political elites, which could affect their degree of leeway in policy making. The selectorate theory may be better when explaining differences within a regime category. Not all regimes fit Weeks (2014) strict coding rules. The selectorate theory could better explain these uncategorised regimes and the variation between regimes measured by Weeks classification. If a domestic crisis encounters, the selectorate theory assumes that political leaders with a small winning coalition and a large selectorate will endure, because of the W/S loyalty norm. Political leaders in Weeks (2014) regimes bosses and strongman, because of their low level of accountability, will also endure. The hypothesis proposed, summarized in Table 2, is a result of how the selectorate theory and Weeks (2014) regime index theoretically will affect political sur- 28

44 vival through armed conflicts. The theories suggest that regime type and accountability are important factors when researching the duration of political leaders. The literature with respect to how engaging an armed conflict affect political survival is ambiguous. There seems to be some fundamental truths. Leaders whom enjoy a small degree of accountability to their citizens, have fewer restraints in their policy making. This is apparent because non-democratic leaders survive in office longer than their democratic colleges after initiating an armed conflict (Bueno De Mesquita & Siverson, 1995; Croco, 2011). Political leaders in autocratic and mixed-regimes suffer a much more severe punishment than those in democracies, if they lose a war Chiozza & Goemans (2004); Goemans (2000). Age entering office may have a negative effect (Bienen & Van de Walle, 1992; Burke, 2012) and accumulated tenure may strengthen the odds for survival (Bienen & Van de Walle, 1989; Burke, 2012). The risk of losing office and initiating an armed conflict has a negative covariance (Chiozza & Goemans, 2003; Levy, 1989). Lastly, economic factors are highly influential on political survival (Barro, 1973; Burke, 2012; Lewis-Beck & Stegmaier, 2000; Wilkin et al., 1997), but good governance is equally important (Burlacu, 2014). Next follows Table 2, a summary of the hypotheses embedded through the text in chapter 2. The next chapter presents the research design, and will discuss methods and variables for testing the hypotheses. Table 2: Summary of Hypotheses Hypothesis Description H1 For political leaders in countries with a medium sized winning coalition, their duration in office shortens significantly in case of an armed conflict. H2 For political leaders in Boss and Strongman regimes (Weeks, 2014), initiating an armed conflict has no effect on their duration in office. 29

45 30

46 4 - Research Design A recession is when your neighbour loses his job. A depression is when you lose yours. And recovery is when Jimmy Carter loses his. Ronald Reagan Although state leaders utilise different definitions of recession and depression, as the quote above indicates, the state of economy has, as mentioned in previous chapters, a large impact on political survival. How will different systemic and personalistic characteristics affect the political leaders survivability in office? This chapter will develop a quantitative research design for testing the hypotheses deducted from the research question. The purpose of the chapter is to bridge the gap between the theoretical framework and the empirical analysis. The chapter is organised as follows. First, a description of the statistical method utilised. Second, I discuss two methodological challenges, namely non-linear effects and missing values. Third, I present the data collected and the operationalization of both the dependent and independent variables. 4.1 Event History Analysis Investigating how engaging in an armed conflict affects political survival could either be measured as duration in office or as a binary indicator denoting dismissal. The findings will be affected by the analytical approach chosen. The literature regarding governmental survival considered both, but concluded that event history analysis takes both into consideration by estimating the time-to-event (Laver & Shepsle, 1998). An OLS approach is inadequate, because the assumption of Best Linear Unbiased Estimation does not hold. The residuals in a time-to-event analysis, e j are not normal distributed, but mostly non-symmetric or bimodal. OLS is not robust to these violations (Cleves, 2008, p.2). 31

47 Event history analysis models have a time-to-event approach. Time is in this case the time-differential between entering and losing office for political leaders. The event is a shift from one state of interest to another, in this case losing office. An additional reason OLS is deficient is because some leaders still prevail in office after 31th December 2014, their event has not occurred in the time span measured. Those units are treated as censored because subsequent history after last event is unobserved (Box-Steffensmeier & Jones, 2004, p.8). Political leaders entering office before the initial observation point, 1th January 1946, are left-truncated, which implies that they are omitted in the analysis. The risk of an event occurring in event history analysis is measured as hazard rates: h(t). The hazard rate h is the risk of whether the event will take place at time t, given that the unit have survived until (t) (Yamaguchi, 1991, p. 11), see equation 1. The hazard rate is therefore a conditional probability rate (Box-Steffensmeier & Jones, 2004, p. 13f). At a general level, positive coefficients indicate a decreasing survival time and negative coefficients the opposite (ibid, p.50). h(t) = f(t) s(t). (1) The baseline hazard is the underlying effect prolonged duration have on independent variables, when set to zero (Jonathan Golub in Box-Steffensmeier et al., 2010, p. 531). The nature of the baseline hazard could either be parametric or semiparametric. If firstsaid, the baseline hazard is specified by an underlying distribution, where the expected survival follows a chosen shape, and the variables effect is the deviation from this shape. Examples of distributions are Weibull, Gompertz, gamma or log logistic. In semi-parametric models, there is no assumptions on the shape of the baseline hazard, the survival rate does not have a parametric form but is instead estimated from the data. An example of a semi-parametric model is the Cox Proportional Hazard model (ibid). Box-Steffensmeier et al. (2010, p. 534) argues that [..] A particular shape should only be chosen on strong theoretical grounds, and never simply because it proves convenient. A strong theoretical expectation could be assumed when the basic behaviour of the objects studied is uncomplicated and strongly determined by, for example, the law of physics. The political survival literature has few commonalities; hence, political leaders duration in office is highly complicated. Restraining their expected perseverance in office by utilising an parametric baseline hazard may therefore be a crude measure. This is an argument for that a semi-parametric Cox model is appropriate when analysing the duration of political leaders, because the total tenure is most likely not a result of an underlying 32

48 distribution. In the Cox model, the hazard rate of the i th unit is: h i (t) = h 0 (t)exp(β x), (2) where h 0 (t) is the baseline hazard function and β x are the covariates and regression parameters (Box-Steffensmeier & Jones, 2004, p. 48). The hazard ratio between two hazards is demonstrated in equation 3: h i t h 0 t = exp(β (x i x j )) (3) A premise when utilising the Cox model in event history analysis is the proportional hazard assumption, which emphasizes that the hazard ratio between any two units is constant over their entire survival time. Equation 3 show that the relationship between h i t and h 0 t is fixed over time. Violating this assumption can result in bias, which could lead to incorrect estimates. When the proportional hazard assumption holds, h 0 t is assumed to be unknown and is left unparameterized (Box-Steffensmeier & Jones, 2004, p. 48ff). The Cox model uses partial likelihood when predicting survival time. It is not the time intervals between successive duration times that is important, but the ordered failure times. This creates a problem for ties, when two or more political leaders have the same duration in office. According to Figure 4 on page 40, most political leaders have the same duration in office as at least one other leader, when time is measured in years. When two or more units have the same duration, a coterminous event occurrences (Box- Steffensmeier & Jones, 2004, p. 53ff). Units are tied if they have identical event times or identical censored survival times. This is a problem for the Cox model, where a central assumption is that the hazard function is continuous, thus identical survival times are theoretical impossible. The problem with ties is minimised by a twofold measure. First, finegraining the time-unit, from years to days in power, following Gates & Strand (2004) suggestion when modelling civil-war duration. Second, the semi-parametric Cox model has some mathematical solutions concerning ties. See Box-Steffensmeier & Jones (2004, p ) for an overview. The Efron method is chosen for two reasons. First, because it accounts for how the risk set changes depending on the order of tied events (ibid, p. 55). Secondly, because the method comply with the Cox model analysis used in the statistical software package Survival in R. Some political leaders complicate this study by either engaging their country into more than one armed conflict, or by regaining office after losing to another op- 33

49 ponent. This aspect complicates our research. As Box-Steffensmeier & Jones (2004, p. 158) point out, if one assumes that the first event is no different from [...] the fifth event, then one may miss important and useful information regarding the timing of the repeated event, conditional on the occurrence of past events. This is addressed as recurrent or repeated events in the survival analysis literature. The variance is assumed independent across clusters, but not necessary within them. The units will therefore be clustered on political leaders. 9 Covariates could either be fixed over time or time-varying. Fixed covariates are constant through the whole study, like gender or country. Time-varying covariates (TVCs) could change values across the span of the observation period for each unit in the study (Box-Steffensmeier & Jones, 2004, p. 95). GDP per capita is an example of a TVC, which is set as yearly parameters. Inclusion of TVCs is a straightforward extension of the partial likelihood function used in the Cox model. Calculations of the hazard rates are only made at failure times, hence, the coefficient for TVCs can be interpreted as the change in the log-hazard ratio for units failing at time t, compared to the remaining observations in the risk set at time t (ibid, p 103f). The Cox model assumes that the covariates have linear effects on the dependent variable. This could be problematic when utilising continuous variables, as starting age for political leaders, where one unit increase of χ does not necessarily have the same effect on Y for all levels of χ. The result of this may be a heteroscedastic errorterm, which may lead to bias in the estimations. Including non-linear effects to some variables, whose theoretical effect is non-linear, may improve the explanatory power for the applicable variable. The next section will address the linearity assumption Non-Linear Effects Beck & Jackman (1998) argue that social scientists trust in to large degree on linear modelling, where the dependent variable is a global linear function of Y = βχ, despite theoretical evidences for the complexity of this relationship. If the researchers theory does not explicitly portray the interaction as linear, convex or a logarithmic function of Y, an alternative tool is the Generalized Additive Model (GAM). This is a generalised linear model with a linear predictor in- 9 Using the cluster(id) function in the Survival-package in R. 34

50 volving a sum of smooth functions of covariates. In Wood (2006), a general model have this structure: g(µ i ) = X i θ + f 1 (x 1i ) + f 2 (x 2i ) + f 3 (x 3i ) +... with µ i E(Y i ) and Y i some exponential family distribution. Y i is the response variable, X i is a row of the model matrix for any strictly parametric model components, θ is the corresponding parameter vector. The f i are smoothing functions for the covariates x k, the variables assumed a non-linear functional form. Researchers using GAMs must find a balance between flexibility and parsimony, by setting the right number degrees of freedom. If the degrees of freedom is the same as the number of observations, the regression line will go through every data-point, and regression as a data-reduction tool will be useless. Therefore it is important to chose the right smoothing parameter, λ (Nygård, 2013; Wood, 2006, p. 33ff; 128f). When implementing a GAM, penalised regression smoothers, penalised splines are utilised. There are several types of splines, see Wood (2006, Chapter 4) for an overview. Penalised splines will be used in this analysis on some continuous variables, because of its flexibility. It is important to set the right degrees of freedom, to few could suppress local effects, and too many could potential over-fit the variable. The analysis will fit the curves with 3 degrees of freedom, a cautious measure to prevent finding potential effects where there are none. In R, the Survival package, used to compute the survival model, cooperate with the package mgcv, by using the pspline function. This specifies a penalised spline basis for the predictor by fitting a comparatively small set of splines and penalising the integrated second derivative. This is similar to smoothing splines with a knot at each data point but is computationally simpler. 10 Next the methodical challenges from missing values will be discussed. 4.2 Missing Data Approximately 94 percent of the contemporary political science literature use listwise deletion, removing data rows with one or more missing values. King et al. 10 The Survival-package description. survival/survival.pdf 35

51 (2001) argue that, as a result of listwise deletion, researchers lose in average onethird of their data. The point estimation is therefore about one standard error from the truth, leading to over-confidence in inaccurate results (ibid, p.52). They argue for the benefits of imputation missing values. The software chosen to imputate the missing values is the Amelia II package in R (Honaker & King, 2010; Honaker et al., 2011). The software is good when utilising time-series-cross-section data, because it controls for cross section variables and longitudinal trends. I used countries and years to index the cross-section and time units. Imputation is conducted on the variable measuring the size of selectorate, the Legis03, the effectiveness of legislature, from Banks dataset (Banks & Wilson, 2015). More on this in Section For an overview over the percentage missing for each variable, see Table 6 in the appendix. It is apparent that most variables has a small degree of missing values, therefore will listwise deletion of missing values be utilised on the other variables in the analysis. Next follows a presentation of these variables, and lastly an overview of the dataset. 4.3 Data This section will present the data 11 used in the analysis. The section is divided in several parts. First, an overview over which datasets the analysis is based on. Second, a presentation of the dependent and independent variables and lastly an overview over how the dataset is built up. To conduct this research, a comprehensive database on political leaders will be acquired. Goemans et al. (2009) introduce the Archigos dataset, with data on political leaders from 1875 to 2004 in 188 countries. The Archigos identifies the effective leader, the person that de facto exercised power in a country (ibid, p. 271), as coded by Gleditsch & Ward (1999). The data describe dates of entry and exit from office for the principal leader of each nation. Through private exchange with Dr. H. E. Goemans and Dr. K. S. Gleditsch, I have been granted access to the updated Archigos version 4.0, with data on leaders entering office up to 31th December The Archigos dataset utilise a leader-year approach, where every row in the dataset 11 The data and documentation is available at MasterThesis 36

52 presents data on a specific political leader. 12 Data on armed conflict is collected from UCDP/PRIO Armed Conflict Database (From now on ACD) (Gleditsch et al., 2002; Pettersson & Wallensteen, 2015). ACD defines an armed conflict as a contested incompatibility that concerns government or territory or both, where the use of armed force between two parties results in at least 25 battle-related deaths in a calendar year (Pettersson & Wallensteen, 2015, p. 536). The dataset contains data on armed conflicts from 1946 and through The Polity-IV dataset (Marshall, 2002) and Banks dataset (Banks & Wilson, 2015) is used to collect data on regimes. The Polity-IV dataset contains data on all national states with more than inhabitants. Due to differences between Archigos, PolityIV and Banks dataset with regard to which countries included, some measures had to be taken. 13 In Archigos 4.0 there are 2219 leadership changes in the period between 1946 and 2014, where each row in the dataset provides information on that specific political leaders continuation in office. The problem with this method of collecting data is that it is not straightforward to merge with other political science datasets, which normally contains data on a country-year format. 14 The UCDP/PRIO dataset is based on conflict-dyads, where each row of data contains information on a particular year in a conflict, if the conflict surpassed 25 battle related deaths for that specific year. This thesis dataset is built up with the Archigos dataset as a base. Following Chiozza & Goemans (2003, p. 449), I Firstly transform the data into countryyear format by repeating each row equivalent to the number of years in office. To minimise a possible problem of endogenous regressors, where the explanatory variables is heavily correlated with the error-term, the row corresponding to the leaders starting year omitted. The problem, if included, is that country-specific variables will correlate perfect in country-years for those leaders in the year of the change. The result of this is initially that political leaders with a duration in office 12 An updated Archigos 4.1 is found at hgoemans/data.htm, with data up to 31th December This was released 29th February Because most R-codes with regard to the Archigos dataset was finished before this, and the UCD- P/PRIO Armed Conflict database have data through 2014, Archigos 4.0 is used. 13 Anjouan, Aruba, Bahamas, Barbados, Belize, Bophutswana, Ciskei, Federation of Malaya, Maldives, Netherlands Antilles, the Palestinian Autonomous Areas, Tanganyika, Transkei, Republic of Vietnam( ) and Venda is removed due to a small population size and/or political status. The manual modifications: Germany DR = German Democratic Republic, Germany FR = German Federal Republic, Serbia and Montenegro and Serbia = Yugoslavia, The United Arab Republic = Egypt, USSR=Russia, Democratic Republic of Vietnam = Vietnam, Rhodesia=Zimbabwe, Zaire=Democratic Republic of the Congo 14 For example PolityIV and Banksdata: One row of data for each country each year, e.g. United States:1946; United States:1947;[...] etc. 37

53 of less than two years are omitted. These are included; creating duplicated countrylevel explanatory variables for these county-years. The complete dataset have 9544 country-years, in the timespan between 1946 and Term Limits: Some political systems have legal restrictions on the consecutive terms the leader could hold office. This is known as term limitations. Political leaders whose duration ended as a result of term limitations are censored. This is because their event (losing office) did not occur as a result of dissatisfaction from political elites or voters, but regulations on their jurisdictions. Chiozza & Goemans (2003) writes that second-term American presidents [...] are coded as 0 in their last year in office; that is, these observations are considered censored. There are some uncertainties regarding how they treat other countries that have restrictions on the duration for their leader. There are two main types of term limits: consecutive and lifetime. The difference is that in consecutive systems, the leader could return to office after a set period of time, like in Russia. In systems with lifetime restrictions, the leaders do not have the opportunity to return holding office, like in the US. Collecting data on term limits was arduous, thus some indulgence measures with regard to sources was taken. I found the World Bank Database of Political Institutions 15 inadequate for this research, because it did not account for countries that have multi-term limitations. I used the dataset as a base and complemented the data with information found on The CIA World Factbook and Wikipedia. 16 I censor political leaders losing office because of term limitations, if these sets of conditions are met: Their duration where equal as the maximum duration set by term limitations 17 and they exited office regular, as defined by the Archigsos dataset. Hence, 162 political leaders is censored because of term limits, losing 800 country-years. Most apparent is political leaders in Switzerland, with one year term limit, and Mexico, with the Sexenio system. For leaders loosing office because of term limitations, their event did not occur, hence these are censored. Next I will present the variables utilized in the analysis html 16 Wikipedia: Term Limit, URL: term_limits 17 Some discretions are made: ± 2 months 38

54 4.3.1 Dependent Variable The dependent variable is measures as the difference in time between the start and end date in the Archigos dataset. This time is given in years. For example: Sardar Shan Mahmud Khan Ghazi of Afghanistan entered office and was overthrown , giving a total duration of years. The high number of digits is, as mentioned earlier, a measure done to reduce the number of tied duration times. Because this thesis use a country-year approach when merging systemic properties to the dataset, some measures with regard to the political leaders start and stop indicators had to be conducted. Arranging time in office to a country-year dataset, there has to be two time indicators; one defining time in office at the beginning and one at the end of the respectively country-year. Additionally, a censoring variable, indicating whether or not the leader lost office in that country-year, is included. The variables Time1 and Time2 specifies the duration at the respectively start and end of that countryyear. In the leaders final year of power, Time2 indicates their total duration. Since the leaders first country-year is omitted, due to a potential endogenity problem, the first Time1 indicator for every leader is always zero, and the first Time2 variable represents time in office at the end of the second year. Leaders with only one datarow, because their duration does not exceed two years, Time2 indicates their full duration. The event variable represents whether or not the leader loose office that specific country-year. Table 3 contains an sample of the dataset. The distribution of duration in office for political leaders is plotted in Figure 4 on the following page. Figure 4 illustrates that 654, of the total 2026 leaders, lose power the first year in office. The tendency flattens out from one to five years leaders have a duration 5 years. From 6 up to 30 years, the total duration is steadily decreasing, with some outliers over 30 years. Fidel Castro has the longest duration in the sample with years in office. Those who retain office on 31th December 2014 are right-censored, more on this in Section 4.1 on page 31. Dark blue indicates leaders not censored while light blue represent right-censored entries, because they still prevailed in office 31th December Orange indicates the leaders died of natural causes or committed suicide 18 when in office. Black indicates leaders who lost office because of term limitations. Figure 5 display the political leaders total duration (Yaxis), based on year entering office (X-axis). The colour-coding is equal to Only one in this time span: Guzman Fernandez, the 46th President of the Dominican Republic, from 1978 to Because he is the only case, suicide in office will be coded as natural death. Another questionable natural death (for some) is Harold Holt, prime minister in Australia from 1966 to I code him as a natural death (drowning), although some believe he was kidnapped by a Chinese submarine while surfing (Grey & Spiegel, 1986). 39

55 Figure 4: Total Duration for Political Leaders: Figure 5: Total Duration for Political Leaders, based on starting year 40

56 Figure 5 show total duration and a censoring indicator for political leaders. The orange line shows a weighted average of duration based on the starting year, using splines with four degrees of freedom. From this figure, there is evidence that the political leaders starting period could affect duration in office. Leaders entering office in the 1960 s have in average longer duration than others. In for example and the second half of the 1980 s, there seems to be fewer extreme-outliers, decreasing the mean duration. For additional information regarding the Archigos dataset, see Goemans et al. (2009) Independent Variables This section will present the independent variables. These are divided into three subgroups. First, those concerning systemic characteristics, like the operationalization of selectorate theory and Weeks (2014). Secondly, variables on armed conflicts and lastly personal characteristics with regard to the political leaders. Selectorate theory: Following Bueno De Mesquita et al. (2005), the size of the winning coalition and the selectorate is constructed from the POLITY-IV dataset (Marshall, 2002) and Arthur Bank s cross-national time-series data (Banks & Wilson, 2015). The operationalization of size of winning coalition builds on four variables: two considering executive recruitment; one on the degree of competitive party system and one for measuring whether it is a military regime or not. The size of selectorate is based on Banks (2015) legis03-variable. The specific coding of winning coalition is as following; One point is added to the index for each of the following conditions: if the Bank s regime type variable 20 is non-military; if XRCOMP is 2; if XROPEN is > 2; and if PARCOMP = 5. The variable is divided by four, creating an index between 0 and 1. The size of the selectorate is operationalized using Bank s legislative selection variable, the Legis03. It is coded zero if no legislature exists, 1 if selection is largely ineffective, 2 if the legislature is partly effective and 3 if the legislative is effective. The variable is divided by three, creating an index between 0 and 1. This variable has 1863 missing values, almost 20% of the cases. This creates a problem for the 19 The 876 pages long Archigos codebook is found here: college/faculty/hgoemans/archigos_4.1.pdf It gives an overview over variables and which leaders included. 20 The Polit02-variable 41

57 analysis using listwise deletion of missing values. If one data-row contains a single missing value, the row is omitted from the analysis. From Table 6 in the appendix, it is apparent how the variable defining the size of selectorate have considerable more missing values than the others. Imputating this variable will enlarge the basis for analysis significant. Utilizing the R package AmeliaII (Honaker & King, 2010; Honaker et al., 2011), the variable defining size of selectorate is imputated. The results from two different models, run with and without imputed values, are found in the Appendix, Table The coefficients do not change significantly when implementing imputed variables; the standard deviation decreases slightly, giving more precise estimations. It is obvious from the coding rules that the variables measuring the size of the winning coalition or the selectorate is not as continuous as the selectorate theory aspires. The variable is therefore dummy-coded appropriate to their value. Instead of utilising the loyalty norm, W/S, this thesis will use the combinations of W and S by interacting them. The winning coalition has five levels and the size of selectorate has four, which gives 20 combinations of W and S. This method approaches the size combinations between W and S better than W/S. The actual formula is:w/log((s + 1) 10/3), but as Gallagher & Hanson (2015) argues, this is an uninterpretable muddle given the meaning of the two component parts. The combination of value on W and S is compared to the political leaders respective average duration in office. This relationship is plotted in Figure 6. Those leaders with a small winning coalition have a higher probability of an extended tenure, especially when combined with a medium to large size of selectorate. The combinations W=1 and S=0 and S=0.33 are theoretically possible, but non-existing in this sample. From Figure 6, it apparent how different combinations of W and S affects total tenure for political leaders. The size of winning coalition is seemingly most influential on duration in office. Political leaders with the longest average continuation in office is found in the smallest winning coalition. This figure gives some preliminary evidence that the effect W and S has on political survival is ambiguous. Figure 7 gives an overview over countries respective size on winning coalition in It is apparent that most countries thought to be highly democratic have a large winning coalition. On the other hand, countries such as North Korea, Qatar and Saudi Arabia, which had a polity-value 22 on -10 in 2006, had a size of winning coalition estimated to Peru and South Africa had a polity value on 9, but has a size on winning coalition on Although the Polity-index and the size of the 21 The missing Legis03 values is estimated from polity2, xrreg, xrcomp, xropen and parcomp from the Polity dataset and the polit09 variable from Banks dataset. 22 Degree of democracy. Where -10 is least democratic and 10 is most democratic. (Marshall, 2002) 42

58 Figure 6: Average Duration: Selectorate Theory winning coalition are correlated, they are not equivalent to each other. Figure 8 shows the size of the selectorate for countries in 2006 on a world map. Myanmar (Burma), Thailand and Nepal have the smallest selectorate, while typical countries with a very high level of democracy have the largest selectorate, due to their effective legislature. In the selectorate theory, the most oppressive leaders are found in regimes with the largest selectorate and smallest winning coalition. This is a result of their low degree of accountability to political elites and citizens. From the world map, this combination of W and S is not apparent in In the dataset, there are seven country-years with this combination: Guatemala-1984, Algeria-2009 and 2013, Pakistan- 2007, Guinea-Bissau-2012 and Cambodia-2009 and The only one of these without an imputated size of selectorate is Guatemala-1984, under Mejia Victores. Weeks(2014) Regime classification: Coding Weeks (2014,p. 39) regime categories (from now on JW(2014)-regimes) is quite strait forward, utilising the dataset from Geddes (2003). Geddes dataset distinguishes between a military, single-party or personalist regime. Weeks also utilise the variable from Banks & Wilson (2015) which differs between a military-civilian regime. The problem is that Geddes dataset has renewed their data; from variables describing different properties the regime 43

59 Figure 7: World Map 2006: Size of Winning Coalition Figure 8: World Map 2006: Size of Selectorate 44

60 and their leaders have, to Geddes et al. (2014)-dataset (From now on GWF), which use predefined regime categories. The JW(2014)-regime dataset has data on political regimes from 1945 to The implication of this is, when using the JW(2014) regimes, the analysis will lose 14 years and 629 leaders due to listwise deletion of missing values. The GWF dataset has data from 1946 to 2010, losing 4 years and 310 political leaders. The elimination of these country-years could on the one hand result in less precise estimations of potential trends in the 21-century. On the other hand, most political leaders entering office after 2010 are right-censored, so this could be of minor importance. Because Geddes dataset is not renewed, Weeks (2014) replication dataset on regimes will be utilised in the analysis. A correlation matrix between Weeks (2014) and Geddes, Wright and Frantz regime categories is plotted in Figure 9. Figure 9: Correlation matrix: Weeks and GWF regime categories Figure 9 demonstrates that there is some overlap between Weeks and GWFs regime index. Orange indicates regime types that correlate strongly with each other. The Figure points out how some types of regimes are more tangible than others. This is apparent through the high correlation between some of the different authors regime classification. Most apparent is Weeks (2014) Machine regime and GWFs Party. The same applies to Weeks Junta and GWFs Military regime. It is worth mentioning the correlation JWs Strongman has to GWFs Party, Military and 45

61 Personal regimes. JW(2014) Boss correlates with both GWF Party and Personal. It seems that JW(2014) do not have a good indicator on monarchies, because GWFs Monarchy negative correlation with all JW(2014) regimes. Figure 23 in the Appendix illustrates the differences between Weeks (2014) and Geddes et al. (2014) regime classifications, based on their respective PolityIVvalue (Marshall, 2002). In these boxplots the strict coding rules of Weeks is apparent. The 0.5 cut-of point, se Figure 2 on page 15, leads to most regimes to be in the inapplicable category. This has two consequences: On the one hand, the typology has accurate defined regimes, which could lead to more precise estimations when utilised it in the analysis. On the other hand, the strict coding rule will exclude some obvious regimes. Their value of, for example, 0.45 on the degree of accountability dimension will exclude them as a regime. Economy: A positive economic development is highly influential for political leaders seeking a prolonged duration in office. Bueno De Mesquita & Smith (2010, p. 943) indicate that an increase in economic growth of 1 percent reduces the risk of deposition by about 4 percent for a small-coalition leader. Following Burlacu (2014), economy is a twofold measure: GDP per capita and the yearly percent change in real GDP per capita. Gross National Product Per Capita at market prices is collected from both Banks dataset and The World Bank GDP per capita dataset. GDP per capita is measured as gross domestic product divided by midyear population. The reason to include both Banks(2015) and The World Bank datasets is due to the degree of overlay. Banks data have longer time span, from 1946 to 2012, but The World Bank data have more recent data, from 1961 to The variable GDP per capita is calculated as the average between Banks economics03 variable and the World Banks GDP per capita variable. If one of these has a missing value, the other value will be used. Banks GDP per capita variable had initially 1127 missing and World Banks GDP per capita 2395 missing country-years. The combined variable have only 504 missing values, approximately 5 percent. The GDP data is reported in current prices for each year. This makes longitudinal comparisons difficult due to inflation. To overcome this, the Consumer Price Index Calculator from the Bureau of Labour Statistics is utilises to recode the economic variables into 2015-dollars. 23 This will not calculate country-specific variations. Since the world marked has an highly exogenous effect on countries economy, the variation in country-specific inflation based on currency is seen as less important. The variable is standardized; hence, 0 equals to mean GDP per capita

62 and an one unit change corresponds to an standard deviation change in GDP per capita. Economic growth is measured as yearly percent change in GDP from the preceding year. Population Size: is collected from Banks pop1 variable. Because the variable is highly skewed, log(population size) will be used in the analysis. Hence, the coefficients effect is corresponding with one percent increase in population size. Armed Conflicts: When analysing how engaging in an armed conflict affects the duration in office for political leaders, a comprehensive database in regard to armed conflicts is necessary. The Peace Research Institute Oslo (PRIO) and Uppsala University Department of Peace and Conflict Research (UCDP) have collaborated on the UCDP/PRIO Armed Conflict Dataset (ACD). The dataset contains information regarding armed conflicts from 1946 through 2014 (Pettersson & Wallensteen, 2015). The reasons for choosing the PRIO-UCDP conflict data over The Correlates of War (CoW) dataset (Sarkees, 2010), another dataset on armed conflict often used in political science with a much longer time span, are many. First, the ACD have a threshold on 25 battle-related deaths per year, while CoW have 1,000. An example why this is important is the Falklands War. With 907 battle related deaths for that given year 24, the conflict does not exceed COW s threshold. Nevertheless, the Falklands War is included in COW s dataset with an appropriate 1001 battle related deaths. COW has a variable defining the exact number of battle related casualties, while ACD distinguishes between two categories: over or under 1000 fatalities a given calendar-year. This is illustrated in Figure 25 in the Appendix, which shows a clear time-trend towards more conflicts not exceeding the 1000 battle related deaths threshold. Second, including more wars have some statistical advantages. With a high threshold there is simply "not enough" wars. An argument against the ACDdataset is that the CoW project has an extended time span. But, as Gleditsch et. al (p. 617) points out, "Are the theoretical explanations equally reasonable for the whole period?" The UCDP/PRIO conflict data differs between four conflict types: extrasystemic, a conflict that occurs between a state and a non-state group outside its own territory; interstate, an armed conflict between two or more states; intrastate, an armed conflict between the government of a state and an internal opposition group, and internationalized conflict, an armed conflict between the government of a state and internal opposition groups, with intervention from other states in the form of

63 troops. The number of armed conflicts from 1946 to 2014 is illustrated in Figure 10, a replication of Figure 1 Number of Armed Conflicts by Type, in Pettersson & Wallensteen (2015, p. 539). Figure 10: UCDP/PRIO: Number of armed conflicts by type: Figure 10 shows how intrastate conflicts have developed to become the major conflict type. The number of armed conflicts steadily increased in numbers up to the end of the cold war, then decreased. There is some evidence for an upward trend in the number of conflicts, especially with regard to internationalized armed conflicts. See Pettersson & Wallensteen (2015) for a comprehensive review of the newest UCDP-PRIO armed conflict dataset. Some countries have engaged in more than one conflict per country-year. This creates a problem when analysing, because these cases will have duplicated rows. The solution applied is "collapsing" the ACD-dataset, reducing the number of observations too one per country-year. 25 In the case of more than one conflict on a single country-year, the conflict with the highest intensity level is used, anticipating this conflict has the most impact on political survival. After removing these, 1548 conflict-dyads remained. Conflict Intensity: Studies indicate that battle related casualties have a negative effect on leader survival (Bueno De Mesquita & Siverson, 1995; Croco, 2011; Quiroz 25 The package doby was utilised, using the summaryby command. 48

64 Flores, 2012). The ACD-dataset does not contain a variable with the exact number of battle related causalities. Instead, the variable Int indicates conflict intensity. The UCDP Battle-Related Deaths Dataset could be applied instead (Sundberg, 2008). Because of the low time span ( ), leading to a low number of non-censored political leaders, the UCDP-PRIO armed database is chosen. The Int variable is a dummy variable, and differs between if there where 25 to 999 or 1000 or more casualties for given conflict-year. This variable may capture the different effect few and many battle-related deaths has on the leaders survivability. The variable Int is coded into a dummy variable:0 = no conflict, 1 =low intensity, 2 = high intensity. Figure 25 in the Appendix shows that the percentage of low intensity conflicts is steadily increasing. The number of high-intensity conflicts is rather stable, except for some more in the eighties. This is another reason for why the ACD is chosen rather than the Correlates of War-dataset. Pre-Conflict Tenure: Time in office prior to engaging in an armed conflict could, as Bienen & Van de Walle (1989); Chiozza & Goemans (2003) argue, capture independent variables that increase their odds for survival. More time in office gives leaders the ability to build larger networks and could also boost their skill set. A prolonged pre-conflict tenure could therefore have a positive effect on survivability. Some political leaders, like Shimon Peres in Israel and Alvaro Uribe Velez in Colombia, inherit conflicts from their predecessors. If this is the case, they will have a negative pre-conflict tenure. This could be fruitful to the analysis because it will encapsulate the difference between initiating an armed conflict and inheriting. Croco (2011) differs between leaders that are culpable or non-culpable for the conflict. The disadvantage of measuring culpability as pre-conflict tenure is that it does not distinguish between if the leader inheriting the conflict is from the same political party. The benefit is it differentiates between the duration in office when the conflict was initiated. If a leader inherits a conflict that started many years before he gained office, the ongoing conflict will not necessary alter his duration in office. The political elites and inhabitants could get accustomed to the conflict situation, and therefore this may not affect the leaders duration. The variable is coded as the date the conflict surpassed 25 battle-related deaths in the UCDP/PRIO dataset minus the start date for the political leader from the Archigos dataset. This duration is measured in years. Figure 26 in the Appendix shows the distribution of pre-conflict tenure for political leaders. Most political leaders have zero ± 1 year in pre-conflict tenure, with outliers from Bratusek in Slovenia too Gaddafi with respectively and 41.5 years of pre-conflict tenure. 49

65 Because the variable defining pre-conflict tenure both captures positive and "negative" tenure, the effect is most likely non-linear. A leader with prolonged preconflict tenure could have an heightened probability of survival in office, because of the seniority and experience already gained prior to the conflict. And for leaders who inherits a protracted conflict, the potential negative effect an armed conflict has on political survival could be mitigated. Because of the theoretical un-linear effect pre-conflict tenure has on duration in office, splines with 3 degrees of freedom will be applied. Domestic Conflict Event Data (Banks & Wilson, 2015): Arthur Bank s cross-national time-series data contains several variables with information on domestic conflicts. They are mostly collected from The New York Times, and contain variables whose definitions is collected from Tanter (1966). I will utilise one of these: Domestic2: General Strikes: Any strike of 1,000 or more industrial or service workers that involves more than one employer and that is aimed at national government policies or authority. The reason for not incorporating more than one Domestic-variable from Banks dataset is because the others, guerrilla warfare, riots and revolutions, are heavily correlated with ACD s conflict types and/or the Domestic2-variable. Because the data is primarily collected from The New York Times, there is some bias towards more information collected in the latter years. Additionally, there is more data on developed countries. The Domestic2-variable is included to test how domestic nonviolent upheavals may affect political survival. An effect from the variable may facilitate further for further research on how non-violent uprisings affect the survivability for political leaders in office. The variable indicates the number of incidents each country-year. Initially there where 700 missing country-years, these are coded to zero, thus no domestic conflicts that year. The variable Domestic2lag use the Domestic2-value for the preceding country-year, t 1. The inclusion of non-violent turmoil may also facilitate for further research on non-violent campaigns (Chenoweth & Stephan, 2011; Geddes et al., 2014; Shaykhutdinov, 2010; Svensson & Lindgren, 2011). Age entering office: Age when entering office could be a important non-theoretical independent variable. Bienen & Van de Walle (1989) argues that political leaders could have heterogeneous skill-levels when entering office. Skill-level is a latent variable and is difficult to measure statistically. Based on Oscar Wilde s statement 50

66 "With age comes wisdom", the variation in skill-level could be heavily correlated with age when entering office. Bienen & Van de Walle (1992); Burke (2012) find that age on entry has a negative effect on political survival. The effect of age could be important and will be included as an independent variable. Both Bienen & Van de Walle (1992) and Burke (2012) have modelled the effect from age as linear. The effect prolonged tenure has on political survival is theoretical declining through cumulative age related infirmity. Leaders at a high age may therefore be easier to replace, like Fidel Castro. On the other hand, those leaders entering office at young age have a potential longer duration, because their remaining estimated life-expectancy is long, but could also be shortened because lack of experience. The variable age entering office may therefore be non-linear in nature, and Splines with three degrees of freedom will be applied. The distribution of age entering office is plotted in Figure 11. The mean is and median 54 years old when entering office, with a standard deviation on Figure 11: Age when entering office for political leaders Non-constitutional entry: The Archigos differentiates between different types of entry to office. If the leader had a non-constitutional entry, the variable has the value 1, if regular entry 0. Previous times in office: Some political leaders have recurrent entries in office. These are dummy-coded: 1 if ruler has been in office before, and 0 if not. 51

67 Gender: is coded 1 if the political leader was a woman and 2 if man. There are 68 women in the dataset, distributed on 246 country-years. Female political leaders had a average duration on 3.68 years and men 4.57 years. Descriptive statistics regarding the variables utilised is found in Table 6 in the Appendix. The next section will present an overview over the dataset. 4.4 The dataset A sample of the dataset is shown under in Figure This gives the reader an outline of how the dataset is structured. Each country-year in the analysis have their own row, except from some cases where there are multiple leader changes in the same year. These are discussed in Section 4.3. The Time1, Time2 variables indicate the leaders duration at the start and end of the corresponding country-year. In the leaders first country-year, Time1 is zero. In rows not representing the entry or exit for the leader, for simplicity, the Time1 and Time2 indicators are measured in whole years for that country-year. This makes it easier when computing time-varying covariates. For the last country-year, Time2 indicates the leaders whole duration, measured in days/365.25, as showed under in Figure 3. Table 3: Overview of the dataset Country Ccode Year Leader Startdate Enddate Time1 Time2 Event Variables Colombia Ospina Perez X 1, X 2 Colombia Ospina Perez X 1, X 2 Colombia Ospina Perez X 1, X 2 Colombia Ospina Perez X 1, X 2 Colombia Laureanco Gomez X 1, X 2 Colombia Roberto Urdaneta Arbelaez X 1, X Fiji Bainimarama X 1, X 2 Fiji Bainimarama X 1, X 2 Fiji Bainimarama X 1, X 2 Fiji Bainimarama X 1, X 2 Fiji Bainimarama X 1, X 2 Fiji Bainimarama X 1, X 2 26 The data and documentation is available at MasterThesis 52

68 4.5 Summary In this chapter I have presented both the methods and data to test the hypotheses derived from the theoretical framework presented in Chapter 2. The purpose has been to bridge the gap between the theoretical framework and the empirical analysis. First I argued why Event History Analysis is suitable when analysing survivability in office for political leaders. The following chapter contains the analysis, which will be run on a dataset covering 2026 political leaders over 9490 countryyears, from 1946 to

69 54

70 5 - Analysis Somebody recently said to me: "Mr. President, you are so yesterday; Justin Trudeau has completely replaced you: he s so handsome, he s so charming, he s the future." And I said: "Justin, just give it a rest." I resented that.. Barack Obama White House Correspondents Dinner There seems to be different opinions regarding what characteristics are important for political leaders in office. Some believe relatively young state leaders, like Justin Trudeau, is the future. While others may resent this, arguing with the positive correlation between age and wisdom. How will different characteristics, systemic and personal, affect perseverance in office? This chapter will test the hypotheses, and control for different aspects, as age and state of economy. The empirical analysis will proceed in two main stages. First, a simple descriptive analysis of the data and then the analysis. When testing the hypotheses, the results will be separated into three sections; one on each hypothesis and one analysing a subset of the leaders, those who engaged in an armed conflict. 5.1 Descriptive Statistics Before commencing the main analyses, a preliminary look at the data will be conducted. Plots and descriptive tables are necessary to get a good understanding of the data, and may also help avoiding errors Achen (2005, p. 338). This section will give an overview over which regimes and sizes of winning coalitions and selectorate is most likely to engage in an armed conflict. Figure 12 show the different types of conflicts (Pettersson & Wallensteen, 2015) on a mosaic plot, based on the size of winning coalition and selectorate. This type of plotting is good when describing how two or more categorical variables distribute 27 remarks-president-white-house-correspondents-dinner 55

71 (a) Size of Winning Coalition (b) Size of Selectorate Figure 12: Size of Winning Coalition and Selectorate on Conflict Types (Pettersson & Wallensteen, 2015) Type of Conflict: 1=Extrasystemic; 2=Interstate; 3=Internal; 4=Internationalized on each other. In these figures the different sizes of the winning coalition and selectorate is plotted on the x-axis and the different conflict types are plotted on the y-axis. The size of the winning coalition seems to affect if the country have engaged in a extrasystemic conflict. On the one hand, in a large part, only the larger Winning coalitions have engaged in these conflicts. On the other hand, there have been few country-years with this type of conflict. It is apparent that those countries with a medium sized W in larger part have engaged in an interstate conflict. Internal an internationalized conflicts have an almost alike distribution over the size of winning coalition, except the largest coalition is underrepresented. The size of the selectorate does also seem to affect which types of conflicts their country most likely engage in. There are noticeable more extrasystemic conflicts in countries with the largest selectorate. Medium to low size of selectorate are dominant in inter-, intrastate and internationalized conflicts. This gives evidence for that the size of selectorate and winning coalition affects which type of conflict, if any, the country most likely will engage in. It shows also that some types of conflicts are more common than others, as Figure 10 illustrates. Internationalized conflict is more apparent in smaller winning coalitions and selectorates. These conflicts include the Cambodian governments conflict with Khmer Rouge, Government of Morocco against POLISARIO and the war against terror, al- Qaida, Taliban and IS. The Figure gives an overview for the basis of the analysis. The more cases, the more robust results. Figure 13 show the same as Figure 12, except with Weeks (2014) regime categories. Figure 13 show two things of importance. First, the distribution of how Weeks (2014) regimes engage in different conflicts. Non of Weeks (2014) regimes 56

72 Figure 13: Weeks(2014)-regimes on Conflict Types (Pettersson & Wallensteen, 2015) Type of Conflict: 1=Extrasystemic; 2=Interstate; 3=Internal; 4=Internationalized have engaged in a extrasystemic conflicts. For interstate conflicts, all regime types have engaged in a similar matter, except from Weeks (2014) Junta, which is underrepresented. In intrastate and internationalized conflicts, Weeks (2014) Boss and Strongman regimes have engaged in significantly more conflicts than Machines and Juntas. Is this because of their reduced level of accountability to political elites and citizens at large, compared to the others? The second thing of importance in Figure 13, compared to Figure 12, is, in the latter, approximately half of the cases had no conflicts, compared to Weeks (2014) regimes where two thirds of the cases have engaged in a conflict. Weeks (2014) regimes are more war-prone than non-regimes. On the other hand, there are few regimes in the dataset. For a further look at the distribution of variables, descriptive statistics is found in the appendix, Table 6. The next section presents the models used to analyse the hypotheses. 5.2 Results This section presents the results from the analysis, testing how engaging in an armed conflict affects duration in office for political leaders. The section will be divided in three sections, whereof the first concerns Hypothesis 1, testing the selectorate theory (Bueno De Mesquita et al., 2005). The second section tests the 57

73 hypothesis on Weeks (2014) and the third investigates a subset of political leaders, those who engaged in an armed conflict in their period of office. Similar to all sections is the non-theoretical independent variables and methods utilised. As explained in Chapter 4, all estimates are based on event history analysis on political leaders between 1946 and When utilising the Cox PH model, the estimations are relative; the main focus is to assess whether some characteristics affect the leaders probability of maintaining office another year, measured as hazard rates. The hazard rate is a conditional probability rate, which gives the rate at which units fail by a time t, given that the unit has survived until t (Box-Steffensmeier & Jones, 2004, p. 13f). Positive coefficients indicate decreasing survival time and negative coefficients increasing survival time (ibid, p. 50). All models are estimated in R version 3.2.4, using the Survival-package The Selectorate Theory on Political Survival In the selectorate theory, the different combinations of sizes of winning coalition and selectorate has a different degree of accountability, which again gives them a different sets of acceptable policy instruments. From Figure 7, it is apparent how countries thought to be highly democratic both have a large selectorate and winning coalition. In states with the largest selectorate and the smallest winning coalition, the leaders are least accountable. In this dataset, examples of this are Guatemala under Mejia Victores and Guinea-Bissau-2012 under Raimundo Pereira. This section will test hypothesis 1, how the size of winning coalition affect political survival through an armed conflict. When testing Hypothesis 1, For political leaders in countries with a medium sized winning coalition, their duration in office shortens significantly in case of an armed conflict, variables concerning the respective sizes on the winning coalition and selectorate will to be included. A medium sized winning coalition is defined as those values of W not corresponding to the largest or smallest coalition. Table 8 in the Appendix show the results from 6 models testing Hypothesis 1. The models has more variables included the higher model-number. At the bottom of the table, the PH-test indicator shows the result of the proportional hazard test for each model. As mentioned in chapter 4, the Cox model assumes proportional hazards over time. If violated, the analysis will overestimate the impact of variables whose associated hazards are systematically shifting over time (Box-Steffensmeier & Jones, 2004, p. 132). 28 R-homepage: 58

74 When the PH-test s p-value<0.05, the proportional hazard assumption is not held for the whole model, and the model has to be dismissed. Models 1 to 4 are highly significant, thus, these models are inadequate. Model 5, where a dummy variable indicating which decade the political leader entered office, has a PH value on.14. In Model 6, the variables measuring starting age, a dummy for irregular entry to office and gender for the political leader is included. This model has a PH value on.08, indicating that the assumption of non-proportionality is not violated for the full model. Additionally, the size of selectorate interacted is with time in office. In the process of analysing hypothesis 1, it was apparent how the effect different sizes of selectorate had on political survival covariated with accumulated tenure in office. Without this interaction, the full model is highly significant on the proportional hazard test. I will further analyse Model 6, with regard to hypothesis 1. For the proportional hazards test for all variables in model 6, see Table 11 in the appendix. Model 6 includes 15 variables, namely the variables defining size of selectorate and winning coalition, an interaction term between these, conflict type, intensity level for armed conflicts, standardized GDP per capita, %GDP per capita from t-1, log(population size), dummy variables indicating starting decade, Banks & Wilson (2015) domestic2 and lagged by one year domestic2lag, irregular entry to office, recurrence in office, gender male, and starting age with a penalized splines function with 3 degrees of freedom. The GDP per capita-variable is a time-varying covariate. The theoretical framework suggests that political leaders in medium sized winning coalitions in less degree will persist after engaging in an armed conflict. In Figure 14, the size of winning coalition has an effect on survival in office for political leaders, but the theoretical expectations finds only limited support in this model. Political leaders in countries with a size of the winning coalition on 0.5 and size of selectorate on 0.33 have a statistical significant coefficient through model 1 to 5. The standard deviation is too large in model 6 to make valid conclusions, due to large confidence intervals. With a point estimation on -0.31, the hazard - probability to lose office at t i+1 for the units survived until t i - is expected to decrease with 73.3%. 29 The interpretation of this is that aforesaid group has =26.7% higher probability to sustain office another year, compared to the reference category. Because of large standard deviations, the results have too high degree of uncertainty. Thus, it is impossible to make valid preliminary conclusions with regard to hypothesis 1. Table 4 summarises the hazard rates from the interaction between winning coalition 29 Formula for relative change in hazard: % h(t) = [ eβx 1 e βx 2 e βx 2 ] 100 The example: [ e e e ] 100 = 73.3 (Box-Steffensmeier & Jones, 2004, p. 60) 59

75 Figure 14: Hypotesis 1: The Selectorate Theory on Duration Notes: Cox Proportional Hazards model. Point estimates are hazard rates, with ± 1 and 2 standard error represented by the lines. Global proportional hazard test =

76 and size of selectorate. Table 4: Coefficients Model 1: S and W Matrix Size W (1.03) (0.83) (0.40) 0.18 (0.53) 1.02 (0.84) (0.25) 0.10 (0.43) 1.42 (0.78) (0.80) 0.52 (0.41) 0.92 (0.88) 0 S-size Model 1: Cox Proportional Hazards model. Coefficients are hazard rates. Standard deviations in parenthesis. Table 4 show that the different sets of combinations of S and W have different reference categories. exp( 0.31) gives these leaders 27 % higher probability to hold office another year, compared to their reference category. Bueno De Mesquita et al. (2005) theorises that political leader in countries with the largest selectorate, but the smallest winning coalition has the greatest degree of perseverance because of the low level of accountability. Regimes with a very small winning coalition and a large selectorate could choose followers from the large pool of supporters in the selectorate. If a member of the winning coalition withdraws from their position, they are easily replaced. This will make the political leadership more robust in polity crisis, not affecting the tenure of the leaders. Hypothesis 1 suggest that, because medium sized winning coalitions have a raised level of accountability, compared to the smallest W, there may be a threshold size of W affecting the leaders durability in office. Model 1 do not find evidence for this claim, due to the lack of statistical significant coefficients. For W-size 0.75 and S-size 1, a typical democracy, the coefficient shows a negative effect on political survival, exp(1.02) 177 percent lower probability to hold office another year. The coefficients W-size and S-size are included, without an interaction. The point estimate from W-size 0.75 and S-size 1 needs to be summarized with these for the true effect. W-size 0.75 has a point estimation on 1.02 and S-size The equation is therefore: exp( ( 1.58) ) = exp( 0.35) = 0.70 Political leaders in countries with a size of winning coalition on 0.75 and largest size of selectorate, have a 30 % reduced probability for another year in office. Is this finding statistical significant? The standard deviations for the three coefficients are 61

77 0.84, 0.76 and The variance 30 is = The standard error is therefore 1.41/3 = With an estimated hazard rate of -0.35, a standard error of 0.69 gives confidence intervals surpassing zero, too large for assessing reliable conclusions regarding the effect W-size 0.75 and S-size 1 have on political survival. This give some preliminary evidence for rejecting hypothesis 1, the interaction between W and S does not have a significant effect on political survival. Only the largest size of selectorate has a significant effect on political survival. Compared to the smallest, those leaders in countries with the largest selectorate have 79.5 % increased probability to hold office another year. The coefficient is statistical significant, with a p-value<0.05. In Model 1, the size of winning coalition does not have an effect on political survival. Engaging in an armed conflict has a negative effect on political survival. Extrastate conflicts have a point estimation on 0.99, 169 % higher risk of losing office. The variable is statistical significant, with a p-value< Interstate conflict has a point estimate on 0.45 and intrastate conflicts These findings are statistical significant, with a p-value<0.05. Internationalized conflict has a point estimate on 0.23, but is not statistical significant on any conventional levels. Is engaging in an armed conflict rational for political leaders? In Model 3, see Figure 18, a subset of the political leaders, whose common characteristics is that all have engaged in an armed conflict, is analysed. This model will, together with Model 1, provide the basis for a thorough discussion of the hypotheses in chapter 7. The size of selectorate is interacted with time, a measure done to not violate the assumption of proportional hazard rates. Compared to the smallest coalition, the largest have a point estimation of 0.05, statistical significant with a p-value<0.01. A S-size of 0.33 has a point estimate of -0.07, with a p-value< The interpretation of this is that, for each year in office, the largest selectorate have approximately exp(0.05) 5 percent lower probability to hold office. An S-size of 0.33 have, for each year in office, exp( 0.07) 7 percent higher probability to hold office. This effect is seemingly small, but for political leaders in countries with a S-size of 0.33, already accumulated 10 years in office, the effect is exp( ) 50 percent for another year in office. The time-varying covariate GDP per capita has seemingly no effect on the estimation, with a point estimate and standard deviation on zero. The coefficient for a TVC can be interpreted as the change in the log-hazard ratio for units failing at time t, compared to the remaining observations in the risk set at time t (Box- Steffensmeier & Jones, 2004, p. 103f). A change in GDP per capita, compared to 30 σ 2 62

78 those in the same risk set, does not have an effect on political survival. The effect yearly percent change in GDP has on political survival is plotted in Figure 15. (a) (b) Figure 15: Model 6: GDP Per Capita and Starting Age Yearly growth in GDP per capita has a positive effect on political survival. A point estimate on -.14 means each percent yearly positive change in GDP per capita raises the probability to hold office another year with 13.1%. The variable is highly significant, with a p-value < From Figure 15 (a) we see the regression estimation in the solid line and the dotted line marks ±2 standard deviations. The effect is clearly significant, because the line marking the standard deviations is different from zero. In this dataset, most political leaders had a GDP per capita growth on ±5% the year they lost office. Over and under this, the bases of analysis are limited, due to a small sample. The plot shows how a yearly positive change in GDP per capita increases the leaders risk of another year. Likewise Figure 15 (b) show how starting age affect political survival, using splines with 3 degrees of freedom. Starting age has a negative effect on holding office. The older the leader was on his onset, the more likely he is to lose office the preceding year. It is worth noting that there are few political leaders whose starting age is lower than 30, see Figure 11. The effect is nevertheless apparent, because the confidence intervals surpass zero. The dummy variables indicating starting decade for political leaders are highly significant, compared to those entering office between 1946 and to 1990 has a significant negative coefficient, which means that these leaders is estimated a longer duration, compared to those starting before Political leaders who entered office in the 1980 s, has an estimated 66.4 % higher risk of ruling another year, compared to the base category. 63

79 Irregular entrance and recurrence to office has, because of too large standard deviations, no significant effect on political survival. Male leaders have higher odds of holding office, but the coefficient is not significant at conventional levels. The level of uncertainty surrounding the estimates is simply too high to provide assurance that the effect is statistically different from zero. Banks & Wilson (2015) domestic2, and its lagged, are highly significant. For every general strikes with more than 1,000 workers, the risk of losing office increases with exp(0.14)= 15%. If the country had similar strikes preceding year, the effect is not statistical significant. The interpretation of this is, if a political leader manages to hold office the year after a domestic strike, this turmoil has no effect on his durability in office. Figure 16: Survival Curve: Winning Coalition Notes: The line marks estimated survival probability with 95 % confidence interval. Censoring times is marked with X. 64

80 Figure 16 show the effect size of winning coalition has on political survival. The most striking finding is the effect a winning coalition on 0.5 has on political survival. The 95 % confidence interval surpasses all other estimates, hence, the coefficient is statistical different than the others. A winning coalition on 0.5 has a clearly positive effect on political survival, even when controlled for engaging in an armed conflict. This gives evidence for rejecting Hypothesis 1, political leaders in countries with a medium size winning coalition does not have a higher risk of losing office than those with a small or large winning coalition. These findings will be discussed in chapter 7. The next section will present the models testing hypothesis 2, on Weeks (2014) regimes Weeks (2014) Regimes on Political Survival This section will investigate the second hypothesis, how Weeks (2014) regimes affect political survival through an armed conflict. The hypothesis is; for political leaders in Boss and Strongman regimes (Weeks, 2014), initiating an armed conflict has no effect on their duration in office. By creating an interaction between regime type and armed conflict, we can investigate the hypothesis directly. The model is clustered on political leaders, to account for repeated entries in office. Figure 17 shows Model 6 from Table 9 (Appendix), with the point estimates for the coefficients, and their respective ± 1 and 2 standard deviation. Weeks (2014) theory suggest that Strongman and Boss regimes, because of their low degree of accountability and military background, will persist through polity crisis. Their durability in office will therefore not be affected by engaging in armed conflicts. This is tested in hypothesis 2, plotted in Figure 17. Table 9 in the Appendix show the results from 6 models in regard to Hypothesis 2. Because Model 6 has the lowest Akaike information criterion, a non-significant PH-value and second most variables included, this model is our main model when testing hypothesis 2. Figure 17 show that both Weeks (2014) Strongman and Boss regimes have a highly positive effect on political survival. Strongman regimes has a point estimate on -1.79, hence 83% higher probability of another year in office, compared to a non-regime. Likewise Boss regimes have a high degree of perseverance, with a point estimation on -1.17, 69 % higher probability of another year in office. Both estimates is highly significant, with a p-value < Machine and Junta regimes 65

81 Figure 17: Hypotesis 2: Weeks(2014) Regimes on Duration Notes: Cox Proportional Hazards model. Point estimates are hazard rates, with ± 1 and 2 confidence intervals represented by the lines. Global proportional hazard test =

82 is omitted in Figure 17, the coefficients and standard deviations is found in Table 9. Leaders in these regimes also have a positive significant probability to hold office another year: Machines with 45 percent and Juntas with 62 percent. The results is statistical significant, with a p-value< Engaging in an armed conflict raises the probability of loosing office the following year. The extrastate and intrastate conflict coefficients are statistical significant, with a point estimation on 0.75 and 0.37, respectively. Engaging in these types of conflicts raises the probability to lose office with % and 44.7 %. The coefficient from interstate conflicts is 0.49, with a standard deviation on 0.27, hence, engaging in an interstate conflict has a negative effect on duration. However, the uncertainty connected to the high p-value makes it impossible to draw valid conclusions from engaging in interstate conflict. The different regimes and conflict types is interacted to analyse if there is any statistical interference between these. Only strongman- and junta-regimes interacted with internationalized conflicts is statistical significant. This has two interpretations. First, engaging in an armed conflict does not, in large, have an effect on duration for political leaders in strongman and boss regimes. Second, the interaction between strongman regimes and engaging in an internationalized conflict has a point estimation on 1.49, hence, these political leaders has over 300 % higher probability of losing office. Leaders in Junta regimes have a point estimation on , hence a near infinite high probability of holding office. This estimation may be a result of a small sample. The only case is Paul Kagame of Rwanda, still holding office 31th December Hypothesis 2 is therefore, in large part, verified. Engaging an armed conflict has no effect on the duration in office for political leaders in Strongman and Boss regimes (Weeks, 2014). These regimes prevail through most domestic policy crisis, except from foreign interventions. These conflicts reduce the probability for another year in office for leaders in strongmen regimes severely. This is discussed further in chapter 7. Yearly change in GDP per capita has a highly significant effect on duration in office. For each percent yearly growth in GDP per capita, the leader enjoys an 11.3 % higher probability of holding office. The coefficient is highly significant, with a p-value< The time varying GDP per capita has no significant effect on duration. The variable Log(Population size), has a significant negative effect on political survival. For each percent growth in population size, the leader has exp(0.03) = 3% lower probability to hold office another year. Peaceful domestic demonstrations are highly influential. Both domestic 2 (Banks & Wilson, 2015), and its lagged, are highly statistical significant with a p-value<

83 Both point estimations is 0.15, thus, a domestic anti-governmental strike with more than protesters reduce the leaders probability to hold office another year with 16 %. Gender has an effect on survivability, but only statistic with a p-value<0.1. If this degree of uncertainty is accepted, being a male political leader reduces the probability slightly for another year in office. The next section will use a subset of those political leaders in Model 1 and 2, namely only those engaged in an armed conflict. This is to make a more exhaustive investigation of potential factors that affect political survival when engaged in an armed conflict Political Survival When Engaging War Following Bueno De Mesquita & Siverson (1995), this section will only analyse political leaders who engaged an armed conflict. I use the same two models from the preceding sections, but including the pre-conflict tenure variable. While the original models analysed how engaging in an armed conflict affects political survival, compared to those who did not, this model will investigate how systemic factors and properties regarding the political leaders affect political survival when engaging in an armed conflict. This is included to investigate if political leaders who engage in armed conflicts have distinguished sets of unobserved heterogeneity that may affect the findings from the preceding models. Following is the results in Figure 18, also listed in Table 10 in the Appendix. The variable defining extrastate conflict is omitted as a reference conflict type. Figure 18 plot two analyses, based on the models testing Hypothesis 1 and 2. The models includes the variable pre-conflict tenure, thus, all political leaders missing pre-conflict tenure is omitted from the analysis. Standard deviations exceeding ±10 is left out from the figure. The full model is found in the appendix, Table 10. The most apparent change from the full models is neither of the variables defining starting decade is statistical significant. When engaging an armed conflict, no period of time, since 1946, has altered the leaders degree of survival. Engaging in an armed conflict has a negative point estimate, thus a longer estimated duration in office. In the first model, compared to extrastate conflicts, engaging in an interstate conflict has a point estimate on -0.96, intrastate and internationalized All coefficients are statistical significant, with a p-value<0.05. In the second model, the conflict types have a non-statistical different effect from engaging in an extrastate conflict. The reason for this may be because of the interaction between conflict type and Weeks (2014) regimes. The intensity level has a large impact on survivability for political leaders. For level 1, less than 1000 bat- 68

84 Figure 18: Model 3: Engaging War Notes: Cox Proportional Hazards model. Point estimates are hazard rates, with ± 1 and 2 confidence intervals represented by the lines. 69

85 tle related death for given year, the point estimate is 1.59 and 1.42 for the models. Level 2, if battle related deaths exceed 1000 for given year, has a point estimate on 1.93 and A raising hazard the more battle related deaths. All coefficients have a p-value< This gives evidence for that, if engaging in an armed conflict, reducing the death toll is a good measure for holding office. This finding underpin Bueno De Mesquita & Siverson (1995) findings, battle related deaths have an negative effect on political survival. Like in model 1 and 2, economic growth has a positive effect on holding office. Each percent yearly increase in GDP per capita raises the probability to hold office another year with approximately 8 and 12 percent. The findings have a p- value<0.05. Economic growth could therefore be crucial for leaders to maximize their probability of holding office. The time transformed GDP per capita variable have no effect on the models. Size of winning coalition and the selectorate has an effect on holding office when engaging in an armed conflict. Political leaders in countries with the largest winning coalition and size of selectorate on 0.67 have a point estimate on -2.99, hence almost 100 percent increased probability of survival another year. This combination is theoretically found in democracies. Other statistical significant findings are the largest selectorate combined with a medium sized winning coalition on both 0.5 and Political leaders in these regimes have a point estimate on 4.14 and 4.53, hence a large negative effect on survival in office. This finding gives some evidence for not rejecting hypothesis 1; political leaders in countries with a medium sized winning coalition have a lower probability of holding office when engaging in an armed conflict. Weeks (2014) Boss and Strongman regimes have a highly significant negative coefficient, like in model 2. These leaders enjoy a greater leeway, because of their reduced degree of accountability to their citizens at large, and will in greater extent hold office if engaging in armed conflicts. With this finding, hypothesis 2 is not rejected. Engaging in an armed conflict does not have an effect on political leaders in Weeks(2014) regimes. This will be further discusses in Chapter 7. Next, the effect pre-conflict tenure has on political leaders engaging in armed conflicts is more thorough investigated. Because the variable pre-conflict tenure has splines with 3 degrees of freedom, the point estimate of the coefficient is ambiguous. Therefore the predicted in-sample risk for the coefficient is plotted in Figure 19. The variable measures years from entry in office to engaging in an armed conflict. If the political leader did not initiate, but inherited a latent conflict, he has negative pre-conflict duration. The effect is 70

86 (a) (b) Figure 19: Model 3: Pre-Conflict Tenure seemingly non-linear. In model 1 the line marking the 95 % confidence interval does not surpass zero except for conflict initiated more than 50 years before entering office. From Figure 26 in the appendix, there are few cases of leaders with more than 50 years of negative pre-conflict tenure. Therefore the significant effect from Model 1 plotted may be because of a small sample. In model 2, pre-conflict tenure has a different effect on survivability. In large, conflicts initiated from 20 years before to a couple of years after entering office have a negative effect on political survival. While conflicts initiated approximately 10 years after entering office has a positive effect. This gives evidence for that some pre-conflict tenure may boost the leaders chance of holding office another year. Negative pre-conflict tenure does not measure how long into their reign the conflict surpassed the needed 25 yearly battle-related deaths. But it assumes that the conflict was latent the preceding years before again flaring up. The findings are not clear, because of the different estimations between model 1 and Conclusion This chapter have analysed how engaging in an armed conflict affects the probability of holding office for political leaders. The findings are summarised in this chapter, and will be more thoroughly discussed in chapter 7. In large, armed conflicts reduce the leaders likelihood of another year in office. Political leaders in Weeks (2014) regimes has a higher odds of holding office. Engaging in armed conflicts does not have a statistical significant effect on the duration for these leaders. 71

87 How valid are these findings? The next chapter will investigate how statistical robust the results is by investigating the proportional hazard assumption, residuals and potential extreme outliers based on the leaders duration. 72

88 6 - Robustness Diagnostics We re going to be looking at a lot of different things Donald Trump 31 In this chapter I will test the statistical robustness of the results have from the preceding chapter. First, I will test if the assumption of proportional hazard rates holds. Second, I will investigate if some subsets of leaders, based on their total duration, affect the coefficients in the preceding chapter. Thirdly, the residuals from Model 1 and 2 will be analysed, checking how statistical outliers affect the models. Lastly, I utilize Adcock s(2001) framework for assessing the validity of the findings. The conclusions are made in the next chapter, after discussing the findings from chapter 5 and 6 in light of the theoretical framework presented in chapter Proportional Hazard Assumption Box-Steffensmeier & Jones (2004, p. 131) argues whether the proportional hazard assumption holds is arguably the primary concern when fitting a Cox model. When testing this assumption, I use cox.zph function in the R-package Survival. This function tests both the whole model and each variable for the proportional hazard assumption. The results is listed in Table 11 in the Appendix. On the one hand, the global test for model 1 shows that the model holds, with a p-value on 0.08, hence the assumption of proportional hazard is held. On the other hand, this is quite close to the chosen threshold level of In large part, most variables have a high p- value, except from the interaction between S-size 0.33, S-size 0.67 and time in office. Without this interaction, the model is highly significant, and fails the PH-test. One solution could be to remove the variable measuring the size of selectorate from the analysis. But since it is an important theoretical variable, it is not omitted. The consequences is that, on the one hand, the time in office interacted with S-size is overestimated. On the other hand, by not controlling for time, the S-size variable will violate the PH-assumption. 31 Recurring statement during the 2016 Republican Party presidential primary elections 73

89 The model used for testing hypothesis 2 has a non-significant global test for proportional hazard rates, with a p-value on The results are listed in Table 11 in the Appendix. The variables are mainly non-significant on the PH-test, except from those variables measuring Weeks (2014) Boss, Strongman and Machine regimes. Two measures were tested: Interaction with time and making a time dependent covariate. Neither solved the problem. The variables are nevertheless included, because of their theoretical importance. Because the variables violation on the proportional hazard assumption, some of the effect the coefficient have may be due to this. The conclusion from this section is that the models in large part do not violate the assumption of proportional hazard rates, except from some variables. These are not omitted because of their theoretical importance, but their point estimation may be inaccurate. This will be further discussed in chapter 7. Next section will investigate subsets of the model, based on their total duration. 6.2 Subsets of the Sample This section will test subsets of the full model, based on the political leaders total duration. Other subdivisions that could be interesting may be dividing the models with basis in the leaders gender, which continent their country is located in or by economic conditions, like developed/developing country. The division based on the leaders total duration is done because leaders with a prolonged duration in office may skew the results in the original models. Six models will be tested, 3 on each model. Political leaders with a total duration less than 5 years is in the first subset, the second between 5 and 10 years, and third those leaders with more than 10 years. This is to investigate if the findings are robust for all leaders, independent of their total duration. Table 5 shows the findings from all six models. Model 1 represents hypothesis 1 and model 2 hypothesis 2. In the original Model 1, S-size 1, intrastate-, interstate and extrasystemic conflicts, GDP per capita growth, domestic2, S-size interacted with time in office and most starting decades had coefficients that where statistical significant. In Table 5, there are no single variable that is statistical significant through all subsets. This gives evidence for that the findings from the preceding chapter may be influenced by subset of political leaders based on their total duration. Engaging an armed conflict has a negative effect on political survival, hence the positive coefficients. The findings are not apparent in all subsets. Engaging in an extrasystemic conflict affects strongly in model 1 and 2 for those leaders with a 74

90 Table 5: Subsets of Model 1 and 2 Dependent variable: Duration in Office for Political Leaders Model1.1 Model1.2 Model1.3 Model2.1 Model2.2 Model2.3 W-size 1.07 (.11).34 (.26).82 (.49) W-size 2.13 (.21).31 (.42) 2.15 (1.42) W-size 3.01 (.35) 2.59 (1.09) 7.83 (2.11) W-size 4.38 (.38) 2.41 (1.10) 2.42 (2.32) S-size (.16).42 (.68) 3.05 (.96) S-size (.30) 1.96 (1.34) 3.68 (1.19) S-size 1.31 (.39) 1.89 (1.31) 3.78 (2.71) W-size 0.25:S-size (.20).13 (.41) 1.00 (.70) W-size 0.5:S-size (.27).37 (.51) 2.60 (1.51) W-size 0.75:S-size (.39) 1.98 (1.14) 5.98 (2.25) W-size 1:S-size (.45) W-size 0.25:S-size (.31) 1.69 (1.13) 1.59 (1.06) W-size 0.5:S-size (.36) 1.85 (1.21) 2.68 (1.64) W-size 0.75:S-size (.46).14 (.41) 5.98 (2.30) W-size 1:S-size (.49) 1.19 (2.51) W-size 0.25:S-size 1.97 (.57) 8.21 (2.52) W-size 0.5:S-size 1.45 (.44) 2.16 (1.18) 3.28 (3.48) W-size 0.75:S-size 1.45 (.52) 5.82 (3.18) W-size 1:S-size 1 S-size 0.33:Time2.01 (.05).10 (.09).05 (.04) S-size 0.67:Time2.00 (.05).03 (.10).06 (.05) S-size 1:Time2.05 (.04).03 (.09).05 (.11) Time (.37) (.69) (.61) ConflictEx.43 (.12).21 (.55).71 (.17) 1.41 (.64) ConflictInter.31 (.24).06 (.50) 1.40 (1.01).10 (.44) 2.39 (.55) 1.06 (.67) ConflictIntra.03 (.10).16 (.29) 1.11 (.44).17 (.15).51 (.25).44 (.43) ConflictInt..ized.19 (.14).12 (.31) 1.92 (.74).92 (.21).74 (.43).97 (.64) Intensity Level 1.00 (.10).07 (.29) 1.11 (.50).07 (.15).23 (.29).73 (.32) Intensity Level 2 GDP per capita.00 (.00).00 (.00).00 (.00).00 (.00).00 (.00).00 (.00) GDP % Growth.02 (.02).01 (.04).12 (.07).07 (.03).24 (.06).22 (.05) log(population).01 (.02).01 (.03).01 (.07).01 (.02).00 (.04).12 (.05) 1950-Dummy.06 (.10).20 (.23).32 (.47).15 (.14).29 (.29).53 (.32) 1950-Dummy.11 (.10).08 (.22).28 (.46).18 (.15).26 (.27).55 (.30) 1970-Dummy.19 (.10).16 (.21).32 (.48).43 (.15).04 (.25).46 (.29) 1980-Dummy.13 (.10).20 (.22).24 (.49).56 (.15).12 (.28).42 (.31) 1990-Dummy.02 (.09).05 (.20).64 (.57).35 (.14).31 (.27).19 (.35) 2000-Dummy.19 (.10).23 (.22).99 (1.27).47 (.15).58 (.28) 1.04 (1.50) 2010-Dummy.63 (.19).87 (.24) Domestic2.02 (.03).06 (.11).56 (.27).08 (.03).03 (.17).54 (.13) Domestic2 lag.01 (.03).10 (.15).10 (.36).08 (.03).05 (.15).47 (.15) Gender Male.02 (.12).01 (.28) 3.32 (1.16).17 (.20).29 (.34) 1.29 (.54) Irregular Entry.00 (.06).19 (.17).18 (.23) Recurrence to Office.09 (.05).16 (.16).33 (.41) Weeks(2014) Boss 1.04 (.08).83 (.64) 1.50 (.31) Weeks(2014) SM.45 (.32).59 (.45).99 (.31) Weeks(2014) Machine.14 (.34).18 (.38).28 (.24) Weeks(2014) Junta.52 (.29).66 (.33).06 (.36) Boss:ConflictEx Boss:Conflictinter.19 (1.16) Boss:Conflictintra.32 (.70) 1.24 (.70) Boss:Conflictint..lized.90 (.64) (.92) SM:ConflictEx SM:Conflictinter (1.35) 1.78 (.74) SM:Conflictintra.27 (.45).59 (.63).40 (.56) SM:Confint..lized (1.07).68 (.59) 1.43 (.94) Machine:ConflictEx Machine:Conflictinter.48 (.54) (1.14) Machine:Conflictintra.46 (.86).29 (.48) Machine:Confint..lized Junta:ConflictEx Junta:Conflictinter 1.50 (.92) Junta:Conflictintra 1.23 (.72).84 (.44).23 (.64) Junta:Confint..lized (.69) AIC R Max. R Num. events Num. obs Missings PH test p < 0.001, p < 0.01, p <

91 total duration less than 5 years. Both models lack results from the leaders with the longest duration, because no leader with a very long duration have engaged in the conflict type. Engaging in an interstate conflict has no effect for all subsets except from model 2.2. The confidence intervals are too large for making valid conclusions. Engaging in an intrastate conflict affects some subsets. Both models with leaders in office less than five years, engaging in an intrastate conflict has no effect on their perseverance. Internationalized conflicts has a large impact on political survival in some subsets. In model 1, for those leaders with a total duration exceeding 10 years, engaging in this conflict type reduces the probability of another year in office dramatically. The reason for why this effect is not apparent in model 2 may be because the interaction between regime and conflict type. For leaders in Boss-regimes, internationalized conflicts boost their perseverance in office, for those with a total duration over 10 years. Strongman-regimes have a highly positive effect for leaders with less than 5 years of total duration. Both coefficients is most likely a result of little variation in the sample, because of few "events". Economic growth has a positive effect on political survival in all model 2 subsets, but not in model 1. This could be because inclusion of time interaction with S-size in model 1. In large part, most political leaders enjoy a positive growth in GDP per capita. Domestic turmoil has in large parts a negative effect, but more apparent in the full models. Domestic strikes lagged with one year have less impact than the nonlagged, as in the full model. Although male political leaders have a slightly reduced probability of holding office than their female colleagues in the full model, the effect is more ambiguous in the subsets. After 10 years, men have a greater degree of perseverance than women, in both models. For most variables, the effect is similar in the subsets and the full models. Some interesting remarks is that, even though some variables are highly statistical significant in the full model, the effect alters according to the leaders total duration in office. There may therefore be some outliers that heavily affect the model. The next section will address this issue, by investigating the residuals from Model 1 and Residuals In this section I compare predicted and observed duration. In survival analysis, defining residuals is more difficult than in OLS, where e = y i ŷ i. There are several types of residual tests in survival analysis, with different features of a models fit 76

92 and specification (Box-Steffensmeier & Jones, 2004, p. 119). Examples are Cox-Snell, Schoenfeld, Martingale, deviance and score residuals. Two of these will be tested, namely deviance and score residuals, because of their differences. While deviance residuals are used to calculate the residual for the given subject, with respect to given time, score residuals are calculated with respect to a given subject on each covariate Deviance Residuals Deviance residuals should be symmetric round zero when the correct model is estimated. It is based on normalisation of Martingale residuals. In general, negative residuals imply that the predicted failure time is less than the observed failure time (Box-Steffensmeier & Jones, 2004, p. 130). For these observations, the estimated hazard rates are overestimated, thus the political leader holds office longer than the model could predict. (a) (b) Figure 20: Deviance Residuals From Figure 20 (a) we see that, on the one hand, the residuals distribute in large parts between ± 1. On the other hand, the residuals are not identical distributed around zero. There is a higher density between 0 to -1, than 0 to 1. This indicates that the model in large parts overestimates the duration for many leaders. The two leaders with the most positive deviance residual are Laureanco Gomez and Borusewicz, whose events occurred the same day as entering office. The exclusion of these does not alter coefficients in the full model. Figure 20 (b) show the deviance residuals from the model used to analyse hy- 77

93 pothesis 2. It seems to be a bimodal distribution: one around zero, and another around 3. On the one hand, this indicates that there is a large part of the observations that the model underestimates, their total duration is longer than predicted. There are no residuals greater than -1. On the other hand there are many observations with a deviance residual exceeding 3. Figure 20 illustrates that the model analysing hypothesis 1 have a seemingly better fit than model 2. The residuals have a more uniformly distribution around 0, and there are fewer outliers greater than ± 1. Further, to investigate how each observation influences the point estimate for the variable, score residuals are utilised Score Residuals Score residuals are a good tool for investigating if there are any influential units in the analysis. The score residual is scaled by multiplying it with the variancecovariance matrix from the Cox model. This gives us an influence matrix that shows how a given observation is altering the parameter estimates from the Cox Model. The scaled score residual, referred to as the dfbeta, can be interpreted as standard deviation changes to an estimate for the i-th observation (Box-Steffensmeier & Jones, 2004, p. 128). Following is 10 variables plotted with their respective dfbeta value, namely all sizes of winning coalition and selectorate, and annual change in GDP from preceding year. A horizontal line marks the zero line, where the observation has no impact on the coefficient. The y-axis denotes the value of the scaled change. Figure 21 shows that there are few observations with a big impact. The range of the y-axis changes according to the variables extreme values with regard to the dfbeta-value. This has two benefits. First, it is visibly easy to detect large outliers and second when the zero-line is close to the top or bottom it indicates if there is many high og low outliers. The variables W-size 0.75 and S-size 0.33 have each one observation that have a dfbeta-value higher than 0.8. This is Raimundo Pereira, prime minister in Guinea-Bissau from 9th January 2012 to 12th April Based on his large dfbeta-value, it seems that he may influence these covariates heavily. Removing him from the analysis alters the W-size 0.75 coefficient with and its standard deviation decreases with The S-size coefficient alters and standard deviation decreases with Removing Raimundo Pereira from the analysis has virtually no effect on the model. The variable size of selectorate on 0.67 has an skewed distribution of score residuals. It is apparent that the model overestimates this coefficient. On the one hand, this could have an impact on the point estimation and its standard deviation. On the other hand, the deviation is small, with a dfbeta-value on

94 Figure 21: Hypotesis 1: Score Resiuals The score residuals for the model corresponding to hypothesis 2 is plotted in Figure 22. In Figure 22, I have plotted the theoretically most important variables and their respective score residual value for model 2. For these variables there are no extreme outliers, no variables has a score-residual value exceeding ± 0.1. The model investigating hypothesis 2 have therefore few outliers that could alter the point estimates and standard deviations. 6.4 Limitations Some notable limitations with the study need attention. As mentioned earlier, the operationalization of the selectorate theory may be inadequate. In this section I will use Adcock and Colliers (2001) framework for assessing the validity of the research. They use a threefold measure of measurement validity; whether the operationalization reflects the concept the researcher seeks to measure Adcock (2001, p. 539). Firstly, content validity the degree to which an indicator represents the universe of content entailed in the systematised concept being measured (Adcock and Collier 2001, 537) is low for some aspects of the analysis. The size of selectorate is merely operationalized on Banks & Wilson (2015) Legis03 variable, the effectiveness of legislature. As Bueno De Mesquita et al. (2005, p. 134) points out: the selectorate theory is not focused on legislative selection per se, but the selection mechanism for the legislature 79

95 Figure 22: Hypotesis 2: Score Resiuals seems to be a reasonable, albeit crude, indicator of the inclusiveness of the polity s selectorate. The problem with this measure is twofold. The first issue is the degree of missing values, almost 20 % of the full dataset. This is partial solved by imputation of missing values (Honaker et al., 2011). The second problem is more substantial. Norway, Germany, France, United States of America and Kuwait have, among other, the most effective legislative selection. Iran under Mohammad Reza, Syria under Bashar al-assad and Djibouti under Gouled Aptidon had the second most effective legislative selection. The variable should measure the proportion of the population involved in selecting the ruler, not the effect of an institutionalised selection processes. A new and more nuanced operationalization of S is in the works (Bueno de Mesquita et al., 2002), until then, researchers testing and utilising the selectorate theory must be cautious about drawing too firm conclusions in regard to its content validity. The operationalization of the size of winning coalition also has a questionable degree of content validity. The full coding rules of W is given in Section Gallagher & Hanson (2015, p. 376) argues that the main problem with the operationalization of the winning coalition is the PolityIV-variable PARCOMP. When the PARCOMP-variable defines the country-year as a competitive political system, W=1 in about 97 % of the cases. On the one hand, this gives evidence for that the operationalization of the winning coalition is not directly an estimation of the leaders authority, based on the percentage of enfranchised needed to gain (hold) office, 80

96 but political contestation. On the other hand, the degree of political contestation is heightened the larger winning coalition. Hence, although not exactly measuring what claimed, the operationalization of the size of winning coalition is more adequate than that of the size of selectorate. Secondly, convergent validation - whether the scores produced by an indicator are empirically associated with scores for other variables, called criterion variables, which are considered direct measures of the phenomenon of concern (Adcock and Collier 2001, 540) - is difficult to determine in this analysis, because of the lack of alternative indicators. The degree of convergent validation relies on Bueno De Mesqita et al s (2005) and Weeks (2014) operationalization of the selectorate theory and regime category. Thirdly, construct validity - whether a given indicator is empirically associated with other indicators in a way that conforms to theoretical expectations about their interrelationship (Adcock and Collier 2001, 542) - is low for the variables defining the selectorate theory. In countries with the largest selectorate and smallest winning coalition, their perseverance in office is high, due to their low degree of accountability to political elites and citizens. This is not apparent in the models tested. Therefore it is evidence for that it is a discrepancy between the theoretical expectation and the statistical findings. When analysing Figure 16, we see how a medium sized winning coalition has a higher probability of holding office. This indicates that the operationalization of the selectorate theory has a low degree of convergent validation, because the theoretical likelihood of a prolonged duration in office for leaders in these regimes are low. For political leaders in Weeks (2014) regimes, their perseverance in office comply with the theoretical expectations, thus, a higher degree of convergent validity. The limitations regarding the operationalization of the selectorate theory, and its relationship between theoretical expectations and findings, are striking. As discussed above, the inaccurate measure of the respective sizes of winning coalition and selectorate casts doubt over the findings for hypothesis 1. Weeks (2014) regime index has a higher degree of validity, leading to more trustworthy findings. 6.5 Summary This chapter set out to look at a lot of different things. We tested the Cox models assumption of proportional hazards, if the results from the analysis are affected by subsets of political leaders, based on their total duration in office, and utilized Adcock and Colliers (2001) framework for assessing the degree of validity. The findings from this chapter are: First, the assumption of proportional hazard is largely 81

97 holds for both models. Second, some subsets of the full analysis could affect the results in the analysis. There is also some evidence for this when investigating the deviance residuals. Model 1 overestimates and model 2 underestimates to a degree the duration for many political leaders. The Score residuals have few large outliers. Excluding these does not alter the coefficients notably. Lastly, the analyses measurement validity is discussed. While Weeks (2014) regime index has an acceptable degree, the selectorate theory s lack of theoretical assumptions and statistical findings is striking. This may undermine the findings from Model 1. The next chapter will discuss the findings from chapter 5 and 6, in light of chapter 2 and 3, and try to give an answer to the research question. 82

98 7 - Discussion and Conclusion Despite the challenging economic conditions, our goal is to achieve a growth rate no lower than the world Dmitry Medvedev on Twitter 21th April 2016 After analysing factors that influence political survival, the logic behind Medvedev s tweet is apparent. Economic growth will increase the probability of sustaining their current political status-quo, with Putin as president. This chapter is divided in two sections; First, I will discuss the findings from chapter 5 and 6, in light of the theories applied and former literature on the subject. I will address some implications for the political leaders duration in office when engaging in an armed conflict. Secondly I will make some concluding remarks. 7.1 Summary and discussion This paragraph will discuss my findings on how engaging in an armed conflict affects the duration in office for political leaders, in light of the theories utilised and former literature. First, the findings related to the selectorate theory (Bueno De Mesquita et al., 2005) and Weeks (2014) regime index will be discussed. Secondly, how engaging in armed conflicts affect the leaders probability of holding office.... FYLL UT HER.. The selectorate theory (Bueno De Mesquita et al., 2005) emphasises, on the one hand, how countries with a large selectorate and small winning coalition have political leaders with a greater degree of perseverance. This is because individual members in the winning coalition are easy replaceable due to the large selectorate. The rewards of maintaining in the winning coalition far surpasses the costs of withdrawing, incentivizing them to support the incumbent leadership, even through polity crises. The leaders in these countries have a small degree of accountability to most citizens, because they could depend on the support from the small winning coalition. On the other hand will political leaders in states with both a large selec- 83

99 torate and a large winning coalition have a high degree of accountability to their citizens and political elites. These countries will only engage in armed conflicts if there is a broad consensus supporting this policy. If engaging, the leader will not suffer a higher risk of losing office, because of the pre-conflict consensus. These political leaders have to expect that their policy will be scrutinised by those who elected them. If failing to please these, the winning coalition will, due to the large selectorate, be replaced. Hypothesis 1 emerged from these theoretical assumptions. If states with a small winning coalition have a small degree of accountability and states with a large winning coalitions is based on consensus, states with a medium sized coalition will have an unclear effect on political survival. Hypothesis 1 tests if leaders in medium sized coalitions has a reduced probability of holding office if engaging an armed conflict, because of their theoretical degree of uncertainty surrounding their survivability in office. The results from Model 1 show that no size-combinations of winning coalition and selectorate have a statistical significant effect on durability in office for political leaders. In Model 3, when only analysing political leaders who engaged in an armed conflict, the leaders in democracies, with the largest winning coalition and a medium sized selectorate, have a higher probability of a prolonged duration in office. The explanation of this may lie in the consensus-seeking political leaders, who in larger part only engage in armed conflicts if there is a pre-conflict political consensus supporting it. From model 3 we see that, for leaders in countries with the largest selectorate and medium sized winning coalition, engaging in an armed conflict reduces the probability of another year in office. In model 1, from Figure 16, we see that, in states with a medium-sized winning coalitions, political leaders have a higher probability of holding office, compared to the other sizes. On the one hand, this does not reject hypothesis 1, it gives evidence for the fact that political leaders in countries with a medium sized winning coalitions may have a reduced likelihood of holding office if engaging in an armed conflict, but not otherwise. On the other hand, some of the reason for not finding a clear relationship between the sizes of S, W and political survival may lay in the operationalization of the selectorate theory, as discussed in Section 6.4. The size of the selectorate is measured as the effectiveness of the legislature, assuming that the smaller size of selectorate, the less effective legislature. Section 6.4 discuss how this may be a crude measure of the size of selectorate. For leaders in the smallest winning coalition and largest selectorate, their theoretical degree of accountability is very low. When measuring the size of the selectorate as legislative 84

100 effectiveness, the lack of compliance between the theoretical assumptions and statistical findings are obvious. This combination is found in the dataset, but almost exclusively in cases where the selectorate-variable is imputated because of missing data. A high degree of legislative effectiveness is an indicator on highly democratic institutions, which does not comply with the selectorate theory. Although the selectorate theory was compelling to utilize when analysing political survival, the lack of conformity between the theoretical framework and statistical findings mitigated the bases of making valid conclusions. The effectiveness of the legislature has an impact over time for political leaders. Although the Legis03-variable may be an insufficient operationalization of the size of selectorate, it gives some evidence in regards to the importance of political institutions for political survival. As shown in Models 1 and 3, time interacted with the size of selectorate is highly significant. For leaders in countries with the most effective legislature, their probability of holding office reduces by 5 percent each year. The cumulative effect this gives after 10 years in office is severe. In the same manner, a largely ineffective legislature selection increases the leaders likelihood of holding office with 7 percent each year. The interpretation of this may be that political leaders are more easily replaceable the higher degree of political contestation. The variable s effect is compared to countries where there is no existence of a legislature. Having a selection process that is largely ineffective makes is difficult for political challengers to gain power; hence, the probability of holding office increases. The operationalization of the size of winning coalition is more theoretically compliant than the size of selectorate. But nevertheless, the effect the size of winning coalition has on political survival is only to a certain degree evident in models 1 and 3. As discussed in Section 6.4, there is a discrepancy between the theoretical expectations and the analytical findings, indicating a low construct validity (Adcock and Collier 2001, 542). Because of the inconsistent statistical findings on medium sized winning coalitions and the low degree of measurement validity for the parameters, making valid conclusions in regard to hypothesis 1 is difficult. Further research on the topic and a better operationalization of the selectorate theory are needed before drawing more reliable conclusions. Hypothesis 2 is more palpable, Weeks (2014) dictators have a more profound unambiguous impact on political survival. In Section 6.4, I advocated for the incorporation of regime indexes not based on a dichotomisation of the Polity-IV scale (Marshall, 2002) in political survival. To increase the findings convergent validation (Adcock, 2001), further research could compare Weeks (2014) regime index with the Polity-IV-scale. 85

101 Political leaders in Weeks (2014) regimes have a great degree of perseverance in office. Political leaders in Strongman and Boss-regimes have a more indisputable relationship between the theoretical expectations and statistical findings. In both Model 2 and 3, their probability for a prolonged tenure is significantly higher than other leaders. In Model 2, while engaging in armed conflicts have a negative effect on other leaders, Weeks dictators have a non-significant coefficient of engaging in an armed conflict. This is also apparent in Model 3, when analysing the subset of leaders who engaged in armed conflicts. Head of states in Junta and Machines are more liable to political elites, their political party or military officers, than leaders in Boss and Strongman regimes. This relationship is not hypothesised, but the analysis gives some support to the theoretical expectations. The degree of perseverance in office for these leaders is not as distinct as for leaders in Boss and Strongmanregimes. Hypothesis 2 is therefore not rejected; engaging in armed conflicts does not affect the perseverance in office for political leaders in Strongman and Bossregimes (Weeks, 2014). Different types of conflicts affects political survival different. I pointed out in Section 3.4 that different types of conflicts have yet to be analysed in the political survival literature. The findings indicate that extrasystemic conflicts, occurring between a state and a non-state group outside its own territory, has the most negative point estimation in Models 1 and 2. When militarily intervening in other countries internal affairs, the political leader will have a reduced probability of holding office another year, compared to other conflict types. The explanation may be that foreign policy may be less important than domestic policy for the electorate appointing their leadership. Because of few cases of extrasystemic conflicts, I am cautious drawing conclusions on this conflict type. While intrastate conflict has approximately the same negative effect as interstate conflicts, internationalized conflicts has a non-significant effect on leadership duration in Model 1 and 2. On the one hand, engaging in an internal conflict with a foreign intervention has an unclear effect for most leaders. On the other hand, internationalized conflicts is the only conflict type with an significant effect on leaders in Weeks (2014) regimes. Internationalized conflicts have a significant negative effect on leader duration in Strongman regimes. These conflicts are most often a continuation of intrastate conflicts, with an eventual foreign intervention. Contemporary examples is Muammar al-gaddafi in Libya; overthrown due to a foreign intervention after persecuting domestic insurgents. In Weeks (2014), these leaders are perceived as ruthless oppressors, and are the most likely to engage in armed conflicts. The implications for dictators in Strongman-regimes is, when engaging in 86

102 intrastate conflicts or barbaric indigenous persecuting, minimise the number of victims and refugees and maximise own military capabilities (Binder, 2015), reducing the probability of a Security Council-resolution supported intervention. Indicating whether the leader is culpable for the conflict (Croco, 2011) turned out to be difficult in the ACD-dataset. Instead, pre-conflict tenure was included. Negative pre-conflict tenure indicates that the leader was not directly responsibly for the conflict, and positive that the leader is culpable. In Section 3.4, I argued for the use of non-linear models in political survival. The use of penalised splines gave some substantial findings with regard to post-conflict tenure. In Model 1, preconflict tenure does not have an effect on political survival. Model 2 show how leaders with more than 7 years of pre-conflict tenure have an elevated probability of holding office. Conflicts starting up to 20 years before entering office has negative effect on political survival. Is it then rational for political contestants to seek office in countries with a latent conflict? On the one hand, the findings gives some discouraging evidence for prospective leaders in these countries. Their total tenure may be affected by preceding leaders conflict initiation. On the other hand does these findings need more scrutinising. How will regime type interacted with pre-conflict tenure affect political survival? While engaging in armed conflicts affects both leaders in office and possible forthcoming leaders, there are other circumstances that may affect their perseverance in office positive, like economical growth. A positive growth in GDP per capita is highly robust in all models. In Figure 15a it is apparent that a political leaders needs at least some positive economic growth before it increases their probability of holding office. The same is apparent when utilising penalised splines. Economic growth surpassing ±5 % is non-significant, and the leaders need at least one percent yearly growth in GDP per capita to ensure a positive effect on political survival. This findings supports earlier literature on political survival, economic growth is highly important for maximising the leaders durability in office. In the same matter will economic recession decimate the leaders support. This finding is robust for all regimes and subsets in the analysis. Likewise will peaceful domestic turmoil, measured as general strikes, mitigate the leaders support. Each strike reduces the likelihood of another year in office with about 15 percent. This finding is robust through all models. Because the analysis use data up to 31th December 2014, the findings could be used to investigate the risk of losing office for contemporary political leaders. I will assess Model 3 to estimate Vladimir Putins s risk of another year in office at the end of the dataset. Because he has engaged in armed conflicts, I will use Model 3 to 87

103 determine his probability of another year in office. Although Vladimir Putin s duration in office is affected by term limitations, he is defined in the Archigos-dataset as the de facto leader in Russia since his first entrance in office. At the end of the dataset he had accumulated 15 years in office. His calculation is as following: Size of winning coalition = Size of selectorate = W-size 0.5*S-size S- size*15 years + type of conflict = intrastate + intensity of conflict = 1 + change in GDP from preceding year = % + entered office in gender = male + population size = log(141025). The risk of another year in office for Putin, using the point estimations from model 3, is: V ladimirp utin 2014 = ( 1.24) ( ) + ( 1.05) ( ) + ( 0.02) ( 0.03 log(141025)) = 0.83 The probability for Putin holding office another year in 2014 where exp( 0.83)= 43.6 percent higher than the baseline hazard. Hence, the model have anticipated his survival in office, despite the negative economic growth from the preceding year. On the one hand, this give evidence for that political survival is more than the economic condition and accumulated tenure. The inclusion of several measures, like variables concerning starting age, systemic features and conflict situation, have enhanced the model. On the other hand, how will an inclusion of personal data on the individual political leader affect the model, utilizing theoretical frameworks like Horowitz et al. (2015)? The framework could be used to give a more profound understanding of political survival. Theoretical degree of accountability seems to be the most important factor for holding office for political leaders through armed conflicts. This is apparent for Weeks (2014) regimes in Model 2 and 3, and to some degree by utilising the selectorate theory in Model 3. A low degree of accountability for political leaders give them a broader set of acceptable policy instruments, with leads to a smaller degree of retaliation from their political elites and citizens. In the preamble I asked what could explain the difference between the leaders post-conflict tenure. I have tried to explain the variation between Galtieri s and Saddam Houssein s survivability in office using Weeks (2014) regime index. The difference between the survivability for leaders in Strongman and Junta regimes could explain their different post-conflict tenure. For explaining the difference between Thatcher s and G. W. Bush s post-conflict tenure, I have included several types of armed conflicts as an 88

104 explanatory variable. Different types of armed conflicts have dissimilar effects on political survival, but this does not explain Thatcher s and G. W. Bush s pre-conflict perseverance in office. Although the American military intervention in Kuwait has some resembles to an internationalized conflict, it is not defined as one in the ACDdataset; The American military presence, and many resembling cases, is excluded. Incorporating these types of military interventions may give some valuable additions of understanding the accountability for political leaders after engaging in armed conflicts. 7.2 Conclusion The research question is: To what degree does engaging in an armed conflict affect the duration in office for political leaders in the post-war era? To answer this question, two testable hypothesis where made. The first in regard to how degree of accountability could affect the perseverance in office, utilising Bueno De Mesquita et al. (2005) theoretical framework. The second hypothesis investigates how being a leader of an autocratic regime affect their duration in office, using Weeks (2014) regime index. Both theories emphasise the political leaders lack of accountability to elites and citizens as the main explanation for why some leaders persevere through crisis. Because of the unambiguous relationship between the theoretical framework and the statistical correlation, there is not enough evidence for neither rejecting nor keeping hypothesis 1. The inadequate operationalization of the selectorate theory leads to a lack of conformity between the theoretical expectations and the statistical findings. The uncertainness in regard to hypothesis 1 is therefore too large to draw valid conclusions. Further research on the relationship between the selectorate theory and political survival is needed to further comprehend and assess its utility. Hypothesis 2 tests whether engaging in an armed conflict has an impact on the perseverance in office for political leaders in Boss and Strongman-regimes (Weeks, 2014). With a correspondence between the theoretical expectations and the statistical significant effects, hypothesis 2 is not rejected. While engaging in an armed conflict has a negative effect on survivability in office for political leaders, it has no effect on political leaders in Weeks (2014) Strongman and Boss regimes. The only uncertainty with regard to the robustness of this findings lies on the significant coefficients concerning the proportional hazard assumption. Thus, the effect from the coefficients measuring these regimes could be overestimated. The short answer on the research question is therefore: engaging in an armed conflict has mostly a negative effect on the perseverance in office for political lead- 89

105 ers, but the degree of political accountability affect their perseverance. For political leaders in the most brutal and repressive regimes, as defined by Weeks (2014), engaging in an armed conflict does not have an effect on their survivability in office. The only deviation from this is for leaders in Strongman-regimes (Weeks, 2014), whose political survival is highly affected by military interventions from foreign troops. For a more nuanced comprehension of political survival, several dynamics are worth further exploration. First, the inclusion of a regime index has given a further understanding of political survival. Its compliance between the theoretical framework utilised (Weeks, 2014), and the findings are striking. I have not compared it with other measures of democracy or regimes, like dichotomizing the polity-index (Marshall, 2002) or Geddes et al. (2014) regime index. Further research could investigate this, and clarify its convergent validation (Adcock, 2001). Second, the findings indicate that non-violent turmoil, measured as general strikes, reduces the political leaders likelihood of another year in office. The research on how non-violent campaigns affect regime duration is starting to receive more attention (Chenoweth & Stephan, 2011; Geddes et al., 2014; Shaykhutdinov, 2010; Svensson & Lindgren, 2011). How will these same factors affect political survival? Lastly, how will individual characteristics concerning political leaders affect the model? Research in the borderlines between political science and psychology, like Horowitz et al. (2015), could enhance the statistical accuracy, making more valid predictions in regard to political survival. 90

106 References Achen, Christopher H Let s Put Garbage-Can Regressions and Garbage-Can Probits Where They Belong. Conflict Management and Peace Science, 22(4), Adcock, Robert; David Collier Measurment Validity: A Shared Standard for Qualitative and Quantitative Research. American Political Science review, 95, Arquilla, John, & Rasmussen, Maria Moyano The Origins of the South Atlantic War. Journal of Latin American Studies, 33(04), Bak, Daehee, Kenwick, Michael R, & Palmer, Glenn Who is Careful: Regime Type and Target Selection. European Journal of International Relations, Banks, Arthur S., & Wilson, Kenneth A Archive. Databanks International. Jerusalem, Israel. Cross-National Time-Series Data Barro, Robert J The Control of Politicians: An Economic Model. Public choice, 14(1), Bausch, Andrew W An Experimental Test of Selectorate Theory. International Interactions. Beck, Nathaniel, & Jackman, Simon Beyond Linearity by Default: Generalized Additive Models. American Journal of Political Science, Beetham, David Democracy and Human Rights. Polity. Bienen, Henry, & Van de Walle, Nicolas Time and Power in Africa. American Political Science Review, 83(01), Bienen, Henry, & Van de Walle, Nicolas A Proportional Hazard Model of Leadership Duration. The Journal of Politics, 54(03), Binder, Martin Paths to Intervention What explains the UN s Selective Response to Humanitarian Crises? Journal of Peace Research, 52(6), Box-Steffensmeier, Janet M, & Jones, Bradford S Event history modeling: A guide for social scientists. Cambridge University Press. Box-Steffensmeier, Janet M., Brady, Henry E., & Collier, David The Oxford Handbook of Political Methodology. Oxford University Press. i

107 Bratton, Michael, & Van de Walle, Nicholas Democratic Experiments in Africa: Regime Transitions in Comparative Perspective. Cambridge University Press. Brody, Richard Assessing the president: The media, elite opinion, and public support. Stanford University Press. Brody, Richard A, & Page, Benjamin I The Impact of Events on Presidential Popularity: The Johnson and Nixon Administrations. Perspectives on the Presidency, 140. Bueno De Mesquita, Bruce, & Siverson, Randolph M War and the Survival of Political Leaders: A Comparative Study of Regime Types and Political Accountability. American Political Science Review, 89(04), Bueno De Mesquita, Bruce, & Smith, Alastair Leader Survival, Revolutions, and the Nature of Government Finance. American Journal of Political Science, 54(4), Bueno de Mesquita, Bruce, Morrow, James D, Siverson, Randolph M, & Smith, Alastair Political Institutions, Policy Choice and the Survival of Leaders. British Journal of Political Science, 32(04), Bueno De Mesquita, Bruce, Smith, Alistar, & Siverson, Randolph M The logic of Political Survival. Cambridge, MA: The MIT Press. Burke, Paul J Economic Growth and Political Survival. The BE Journal of Macroeconomics, 12(1). Burlacu, Diana Its Not Only the Economy, Stupid! Good Governance Matters in Elections. Electoral Studies, 36, Castro, Fidel Che: A Memoir. Ocean Press. Chang, Eric, & Golden, Miriam A Sources of Corruption in Authoritarian Regimes. Social Science Quarterly, 91(1), Cheibub, Jose Antonio, Gandhi, Jennifer, & Vreeland, James Raymond Democracy and Dictatorship Revisited. Vol Chenoweth, Erica, & Stephan, Maria J Why Civil Resistance Works: The Strategic Logic of Nonviolent Conflict. Columbia University Press. ii

108 Chiozza, Giacomo, & Goemans, H. E International Conflict Is War and the Tenure of Leaders : Stil Ex Post Inefficient? American Journal of Political Science, 48(3), Chiozza, Giacomo, & Goemans, Hein E Peace Through Insecurity Tenure and International Conflict. Journal of Conflict Resolution, 47(4), Clarke, Kevin Retesting Selectorate Theory: Separating the Effects of W from other Elements of Democracy. American Political Science Review, 102(3). Clarke, Kevin A, & Stone, Randall W Democracy and the Logic of Political Survival. American Political Science Review, 102(03), Clausewitz, Carl V On War. Princeton, NJ: Princeton University Press. Cleves, Mario An Introduction to Survival Analysis Using Stata. Stata Press. Collier, David, & Adcock, Robert Democracy and Dichotomies: A Pragmatic Approach to Choices About Concepts. Annual Review of Political Science, 2, Collier, Paul, & Hoeffler, Anke Greed and Grievance in Civil War. Oxford University Press, 56(4), Collier, Paul, Hoeffler, Anke, & Rohner, Dominic Beyond Greed and Grievance: Feasibility and Civil War. oxford Economic papers, gpn029. Croco, Sarah E The Decider s Dilemma: Leader Culpability, War Outcomes, and Domestic Punishment. American Political Science Review, 105(03), Downs, Anthony An Economic Theory of Political Action in a Democracy. The journal of political economy, Fearon, James D, & Laitin, David D Ethnicity, Insurgency, and Civil War. American political science review, 97(01), Fearon, Jd Domestic Political Audiences and the Escalation of International Disputes. American Political Science Review, 88(3), Gallagher, M, & Hanson, J Power Tool or Dull Blade? Resilient Autocracy and the Selectorate Theory. Comparative Political Studies. iii

109 Gates, Scott, & Strand, Håvard Modeling the Duration of Civil Wars: Measurement and Estimation Issues. In: presentation at the Joint Session of Workshops of the ECPR, Uppsala. Gates, Scott, Hegre, Håvard, Jones, Mark P., & Strand, Håvard Institutional Inconsistency and Political Instability: Polity Duration, American Journal of Political Science, 50(4), Geddes, Barbara Paradigms and Sand Castles Theory Building and Research Design in Comparative Politics. University of Michigan Press. Geddes, Barbara, Wright, Joseph, & Frantz, Erica Autocratic Breakdown and Regime Transitions: A New Data Set. Perspectives on Politics, 12(2), Gleditsch, Kristian S, & Ward, Michael D A revised List of Independent States Since the Congress of Vienna. International Interactions, 25(4), Gleditsch, Nils Petter, Wallensteen, Peter, Eriksson, Mikael, Sollenberg, Margareta, Eriksson, Mikael, & Strand, Havard SPECIAL DATA Armed Conflict : A New Dataset *. 39(5), Goemans, H. E., Gleditsch, K. S., & Chiozza, G Introducing Archigos: A Dataset of Political Leaders. Journal of Peace Research, 46(2), Goemans, Hein E Fighting for Survival: The Fate of Leaders and the Duration of War. The Journal of Conflict Resolution, 44(5), Grey, Anthony, & Spiegel, Eleanor The Prime Minister was a Spy. London: Weidenfeld and Nicolson. Gurr, Ted Robert Persistence and Change in Political Systems, American Political Science Review, 68(04), Hadenius, Axel, & Teorell, Jan Cultural and Economic Prerequisites of Democracy: Reassessing Recent Evidence. Studies in Comparative International Development, 39(4), Hegre, Havard, Ellingsen, Tanja, Gleditsch, Nils Petter, & Gates, Scott Towards a Democratic Civil Peace? Opportunity, Grievance, and Civil War, American Political Science Review, 95(1), Hindmoor, Andrew Review Article:Major Combat Operations Have Ended? Arguing about Rational Choice. British Journal of Political Science, 41(01), iv

110 Honaker, James, & King, Gary What to do about Missing Values in Time- Series Cross-Section Data. American Journal of Political Science, 54(2), Honaker, James, King, Gary, Blackwell, Matthew, et al Amelia II: A Program for Missing Data. Journal of Statistical Software, 45(7), Horowitz, Michael, McDermott, Rose, & Stam, Allan C Leader Age, Regime Type, and Violent International Relations. Journal of Conflict Resolution, 49(5), Horowitz, Michael C, Stam, Allan C, & Ellis, Cali M Why Leaders Fight. Cambridge University Press. Kernell, Samuel Explaining Presidential Popularity: How Ad Hoc Theorizing, Misplaced Emphasis, and Insufficient Care in Measuring one s Variables Refuted Common Sense and led Conventional Wisdom Down the Path of Anomalies. American Political Science Review, 72(02), King, Gary, Honaker, James, Joseph, Anne, & Scheve, Kenneth Analyzing Incomplete Political Science Data: An Alternative Algorithm for Multiple Imputation. American Political Science Review, 95(01), Knutsen, Carl Henrik The Economic Effects of Democracy and Dictatorship. Oslo: Unipub. PhD. Thesis. Knutsen, Carl Henrik, & Nygaard, Haavard Mokleiv Institutional Characteristics and Regime Survival: Why Are Semi-Democracies Less Durable Than Autocracies and Democracies? American Journal of Political Science, 59(3), Laver, Michael, & Shepsle, Kenneth A Events, Equilibria, and Government Survival. American Journal of Political Science, Levy, Jack S The diversionary theory of war: A critique. Handbook of war studies, Lewis-Beck, Michael S, & Stegmaier, Mary Economic Determinants of Electoral Outcomes. Annual Review of Political Science, 3(1), Marshall, Monty G; Keith Jaggers Polity IV Project: Political Regime Characteristics and Transitions, The Polity Dataset. v

111 Morgan, T Clifton, & Bickers, Kenneth N Domestic Discontent and the External Use of Force. Journal of Conflict Resolution, 36(1), Mueller, John E War, Presidents, and Public Opinion. John Wiley & Sons. Munck, Gerardo L, & Verkuilen, Jay Conceptualizing and Measuring Democracy Evaluating Alternative Indices. Comparative political studies, 35(1), Norpoth, Helmut Guns and Butter and Government Popularity in Britain. The American Political Science Review, 81(3), Nygård, Håvard Mokleiv Five Essays on the Political Economy of Regime Survival. Ph.D. thesis, University of Oslo. Oneal, John R, & Bryan, Anna Lillian The rally round the flag effect in US foreign policy crises, Political Behavior, 17(4), Pettersson, T., & Wallensteen, P Armed Conflicts, Journal of Peace Research, 52(4), Quiroz Flores, Alejandro A Competing Risks Model of War Termination and Leader Change. International Studies Quarterly, 56(4), Sarkees, Meredith The COW Typology of War: Definition and Categorizing Wars. Shaykhutdinov, Renat Give Peace a Chance: Nonviolent Protest and the Creation of Territorial Autonomy Arrangements. Journal of Peace Research, 47(2), Singer, David J., & Small, Melvin The Wages of War, : A Statistical Handboo. John Wiley & Sons. Slater, Dan Iron Cage in an Iron Fist: Authoritarian Institutions and the Personalization of Power in Malaysia. Comparative Politics, Sundberg, Ralph Collective Violence : Global and Regional Trends. States in Armed Conflict Svensson, Isak, & Lindgren, Mathilda Community and Consent Unarmed Insurrections in Non-Democracies. European Journal of International Relations, 17(1), Svolik, Milan W The Politics of Authoritarian Rule. Cambridge University Press. vi

112 Tanter, Raymond Dimensions of Conflict Behavior Within and Between Nations, Journal of Conflict Resolution, Themner, L., & Wallensteen, P Armed Conflicts, Journal of Peace Research, 51(4), Tir, Jaroslav, & Singh, Shane P Is It the Economy or Foreign Policy, Stupid? The Impact of Foreign Crises on Leader Support. Comparative Politics, 46(1), Way, Christopher, & Weeks, Jessica LP Making It Personal: Regime Type and Nuclear Proliferation. American Journal of Political Science, 58(3), Weeks, Jessica L Autocratic Audience Costs: Regime Type and Signaling Resolve. International Organization, 62(01), Weeks, Jessica L P Dictators at War and Peace. Cornell University Press. Wilkin, Sam, Haller, Brandon, & Norpoth, Helmut From Argentina to Zambia: A World-Wide Test of Economic Voting. Electoral Studies, 16(3), Wood, Simon Generalized Additive Models: An Introduction with R. CRC press. Yamaguchi, Kazuo Event History Analysis. Vol. 28. Sage Publications. vii

113 viii

114 A - Appendix Figure 23: Weeks(2014) Regime Index on Polity-IV Value Figure 24: GWF Regime Index on Polity-IV Value ix