Dictators, Personalized Security Forces, and Coups

Transcription

1 Dictators, Personalized Security Forces, and Coups Wonjun Song August 10, 2018 Abstract Dictators rely on coercive forces to remain in office, as violence is the ultimate arbiter of power in these regimes. However, coercive forces can also remove the dictator from office in a coup. This presents the dictator with a dilemma. One way to address this dilemma is to personalize the security forces. This paper argues that personalizing the security forces decreases coup risk by: (a) linking the security elites fate more closely to the leader s and (b) increasing the informational advantage the leader has over security elites. Using a new measure of the personalization of the security apparatus, I show that personalization decreases coup risk in dictatorships, but this stabilizing effect of personalization disappears after the dictator s exit from office. This study documents how dictators transform the security apparatus to stabilize their rule, with implications for how dictatorships survive and collapse. Keywords: authoritarian regime, coups, security forces, personalism

2 1 Importance of Security Apparatus in Autocratic Regimes Coups are the most common way by which dictators lose power. About two thirds of dictators lost power by coups and when we include the reasons indirectly related to loyalty of the security forces such as popular uprising or assassination its significance only increases (Svolik 2009, 487). It is therefore unsurprising that coups garner significant attention from scholars in comparative politics and international relations. There are two strands of research linking the structural conditions to coup risk. A first group emphasizes factors that alter the propensity of military leaders to attempt forceful overthrow of leaders. For example, Huntington (1991) argues that bribing military officers through increased military spending or personnel can reduce coup risk. This logic is supported by the findings that military size is negatively associated with democracy (Albertus & Menaldo 2012), and greater military spending per soldier decreases the probability of coup attempts (Powell 2012). A second body of research posits that social and economic conditions shape the military s opportunity for intervention in domestic politics. This research argues that negative economic growth (Nordlinger 1977), economic shocks induced by weather (Kim 2016), social unrests (Finer 1975), economic insulation (Powell & Chacha 2016), Cold War superpower competition (Marinov & Goemans 2013), inequality as a proxy for the regime s reliance on the military for repression (Svolik 2013), and domestic instability events such as successful terrorist attacks (Aksoy, Carter, & Wright 2015) increase the likelihood of a (successful) coup. When government s performance is viewed unfavorably, the risk of a coup increases because coup attempts are more likely to enjoy popular support (Johnson 1962, 260). Fearing coup attempts, dictators take a series of actions known as institutional coupproofing to reduce the military s ability to successfully organize their ouster by reshaping the institutional characteristics of security forces and their relationship with political leaders. These schemes include creation of rival parallel organizations as coordination obstacles and counterweights (Quinlivan 1999, ; Black 2000; Belkin & Schofer 2003; Powell 2012; De Bruin 2017), recruiting military personnel along ethnic-tribal or family lines (Enloe 1980; 1

3 Horowitz 1985; Harkness 2016), monitoring the military with political commissars (Fainsod 1967), and favoring incompetent officers while purging competent, more threatening officers (Ezrow & Frantz 2011, 157). Institutional coup-proofing, these theories posit, reduce coup risk. However, to this point, the civil-military relations has not directly examined security forces personalization as a type of coup-proofing strategy. This paper builds on prior research by analyzing how increased personalization reduces the likelihood of coup attempts. 1 Personalization is the expansion of a dictator s power over her allies and rival elites in the security forces. 2 I argue that aggrandizement of dictator s personal control over the security forces relative to the security officials decreases the likelihood of coup attempts. 3 Personalization differs from the institutional or organizational coup-proofing strategies in two ways. First, it aims not only to create an information gap and coordination obstacles among the elite members of the security forces, but also reshapes the institutions and their leadership compositions. Second, personalization policies recruit and encourage the growth of the security elites whose primary and foremost allegiance is to the dictator at the expense of the elites who have to stronger social attachments to and thus are more loyal to the state, regime, or military organizations. This creates a subgroup of security elites who serve as Personal loyalists for the dictator and her close family members. Unlike the larger group of apolitical career officers and bureaucrats or other elites without personal connections, these Personal loyalists are tied to the dictator and less likely to survive should the leader lose power: Their career advancement and positions of power within the security apparatus result from social, ethnic, or familial relationships with the regime leader and not from meritocratic criteria or bureaucratic service. Thus, the personalization enables a dictator to align her and the security elites utility more easily by more closely linking their careers to the regime leader and her grip on power. The leader in these regimes also has better information about the security elites preferences as the size of the winning coalition is smaller and she can more easily detect and address their grievances. Moreover, the existence of praetorian guard units that are personally loyal to the dictator increases the probability that troops will fight to protect the leader until the end, serve as an additional deterrent against threats from disenfranchised elites. Creating (possibly multi- 2

4 ple) security organizations with leaders tied closely to the regime leader and whose promotion depends on personal relationships with the regime leader not only increases coordination costs for security elites planning a coup attempt, but does so by creating competition among security elites jockeying for a share of power distributed from the regime leader. To test the claim that security force personalization reduces coup risk, I use newly-collected data on different features of autocratic rule by Geddes, Wright, & Frantz (forthcoming). I create a continuous measure of security force personalization. I validate this measure by examining how well over-time variation in the measure corresponds to case study evidence from a handful of autocratic regimes. Employing this measure, I show that personalization decreases coup risk. The temporal effect, however, varies in expected ways. Initial attempts at personalization increase coup risk as these attempts often yield backlash from the security forces; but after successful personalization coup risk drops quickly. To demonstrate that personalization distinct from institutional counterweights against the regular military, I test whether the effect of personalization endures after a dictator s exit from office. I find that it does not, suggesting that it is the personal link between the dictator and her security forces that reduce coups. 2 Importance of Security Forces Personalization 2.1 Remedying Security Forces Dilemma All dictators need support groups to rule. Since the most fundamental difference between democracies and non-democracies is the lack of institutionalized consent by a popular majority (Svolik 2012, 3), this makes repression of the excluded population with the help of security forces crucial for retaining power. Historically, security forces have been the primary threat to a dictator: coups are the most frequent way dictators lose power and those who exit office via irregular methods such as coups are the most at risk for a nasty, even deadly, post-exit fate (Svolik 2009; Debs & Goemans 2010, 435). Moreover, executive leaders of the security apparatus, such as military generals, internal troop commanders, or intelligence agents, are willing to let autocratic regime collapse or replace the dictator if they are able keep their posts 3

5 under a subsequent regime or leadership (Nordlinger 1977; Sutter 1995). 4 One way to partially remedy this dilemma is to build a security apparatus that is personally loyal to the dictator. This can be done if the leader exploits qualitative difference among winning coalition members and not just the size of the winning coalition The Setting According to the selectorate theory, a political leader has the highest chance of survival if the selectorate is large while the winning coalition is small. Under this setting, it is easier for the leader to replace winning coalition members by other members who do not belong to the winning coalition; on the other hand, the winning coalition members are heavily dependent on the leader to maintain the access to private goods (Bueno de Mesquita et al. 1999; Bueno de Mesquita et al. 2013). The critics of the theory argue that at least for authoritarian regimes, it is also important to consider the composition and preferences of the winning coalition and their changes over time, not just the size of it (Gallagher & Hanson 2015, 371). For example, Deng Xiaopinig could implement his economic reform policies because he could bring in new provincial elites to the Central Committee while purging the old guards that were tied to the Cultural Revolution (Shirk 1993). On the other hand, Mawdsley & White (1990) point to the lack of turn-overs and entrenched interests among the Party elites to explain the stagnation of the Brezhnev era. Following this logic, security forces personalization is a constant effort by dictators to alter the security apparatus by replacing security elites with other elite members 6 who are more willing to serve as Personal loyalists. I posit that personalization of security apparatus does not end in one-time event. First, attempts to replace a large number of support coalition members at once increases the risk of coup attempt and rebellion. Second, dictators maximize their safety by reshuffling the winning coalition members. There is no guarantee that the security elites personally appointed by the dictator will not develop their own agenda or grievances during their tenure. 7 For example, Stalin continuously purged his secret police (NKVD) chiefs, from Genrikh Yagoda to Nikolai Yezhov. Kim Jong-un also repeatedly promoted, demoted, and re- 4

6 educated prominent security chiefs, such as Kim Won-hong (State Security Department) and Hwang Pyong-so (General Political Bureau) (Berlinger & Lee 2017). Such reshuffling of the inner-group security elites who are relatively more cohesive, hierarchical, and armed compared to civilian counterparts 8 is possible because they hold heterogeneous identities and preferences. In other words, we should view a security institution as a sum of qualitatively different elites, not as a monolithic organization with all members striving to achieve certain objective corporate interests. Related works on civil-military relations elucidate this point further. First, multiple factions within the military (or in any other security organization) may be divided by different preferences. In his study of military intervention in politics, White (2017, 582) names them Professionals and Politicians-in-uniform. Professionals prefer to concentrate military s resources and attention to the main task of national defense and avoid being bogged down in political tasks such as regime preservation or leadership succession (White 2017, 583). On the other hand, Politicians-in-uniform favor military intervention in domestic politics and are more willing to support a particular leader in exchange for greater military institutional privileges or involvement in national politics (ibid). Second, there is no a set definition of what constitutes the overall military s corporate interest or institutional preference. Although military organizations are hierarchical, their institutional preferences are a product of intra-politics among senior officer corps, not exogenously given or imposed on every single military officer through a top-down processing (Stepan 1971, 40; Debs 2016, 74; White 2017, ). The process of security forces personalization is analogous to that of Svolik s (2012, 57-63) sequential personalization game, where a dictator concealing her intention of consolidating personal power at the cost of the ruling coalition, initiates the game. This personalization process succeeds or fails depending on the support group s ability to interpret imperfect signal of whether the dictator is attempting to personalize and the probability of rebellion s failure (Svolik 2012, 59). 9 The study finds that the probability of leader exit by a coup attempt decreases as the dictator s time in office increases (Svolik 2012, 75-77). Combining these logics enables us to theorize the relationship among (1) a dictator who 5

7 wants to personalize her security apparatus, (2) pre-existing Personal loyalists, (3) would-be Personal loyalists, and (4) Un-enfranchised security elites, where the third group is a subset of the fourth group. 10 I posit that a dictator personalizes her security apparatus which is a short-term risky option to ensure long-term safety. Therefore, security forces personalization should be more likely to occur during the beginning of a dictator s tenure and decrease over time as the leader consolidates power. The personalization process marginalizes Un-enfranchised security elites or even pre-existing Personal loyalists, depending on needs, and replaces them with would-be Personal loyalists. 11 The degree of personalization is an accumulative one (speed of the personalization varies depending on the circumstances) and we cannot pinpoint to a single event that we can use to strictly classify when a regime has become a perfectly personalist or non-personalist regime. In other words, the relationship between the dictator and her security elites should be described as a continuum rather than a binary concept. Nor does it stop completely at a certain point; long-living dictators periodically engage in purges, demotions, and frequent post transfers of important security commanders (Svolik 2012, 79). 2.3 Effects of Increased Personalization Personalization helps a dictator mitigate the guardianship dilemma in a number of ways. First, security apparatus personalization aligns security elites preferences with the dictator s preference. Unlike apolitical appointees, if the members have personal or ethnic connections to the leader, their career fates are more closely link to the leader s tenure. In other words, because they are personal, and not meritocratic or bureaucratic appointees, they are more easily identified as strong supporters of the leader and therefore more likely to face retribution should the leader lose power. 12 Knowing this, security forces composed of Personal loyalists will be less likely to oust the leader via a coup, and are more likely to protect the leader against threats such as a popular uprising or insurgency. The studies on post-autocratic regime collapse find that that the greater the personal control a dictator wields to political elites, the less likely these regimes transition to democracy (Escribà-Folch 2013; Geddes, Wright, & Frantz forthcoming). Anticipating that their post-exit fates are unlikely to be safe, both the dictator and her personal- 6

8 ized security forces cling to power in the face of mass uprising and are willing to use violence in a large and protracted scale (Escribà-Folch & Wright 2015). Second, dictators who surround themselves with Personal loyalists as the winning coalition members have better information on the elites preferences than when they are surrounded by a larger number of career military generals and intelligence officers. To reiterate, the efforts by a dictator to amass personal power at the expense of regime elites succeed or fail depending on the elites ability to observe dictator s intentions and the probability of a elite rebellion s success (Svolik 2012, 59). Altering support coalition by staffing it with more loyal members decreases the cost of monitoring and detecting potential dissidents, leading to higher chances of preventing the grievances from growing to a dangerous level (Ostrom 1990, 222). Given that there is no guarantee of security elites who are personally appointed by the dictator are not going to develop their own agenda or grievances during their tenure, having a smaller but more loyal winning coalition helps to curb such a probability. Finally, because fearful Personal loyalists will fight to the end in order to prevent their leader from losing power, it will decrease the chances of a successful coup attempt. 13 Given the nature of coups, a successful coup attempt has to be conducted swiftly with a minimal number of plotters. Coup success is not determined through the clash of physical forces that coup plotters and government troops mobilize; it is determined by the ability of coup plotters to persuade the rest of military and security forces that the outcome has already been determined and there is nothing they can do to revert it (Singh 2014, 21). Geddes (2009, 6) posits that other military officers have three choices: Rush to defend the government, mobilize troops to support coup plotters, and do nothing. Most will choose the third option and support whoever turns out to be the winner because it is the least risky option (Geddes 2009, 6). Appointing Personal loyalists to senior security apparatus posts increases the chances of important military commanders coming to the rescue for the dictator or even launch a counter-coup attempt if coup plotters succeed as the linked fate with the leader changes the risk and reward calculation. Expecting that this would happen, potential coup plotters either give up the idea of ousting the leader by a coup or seek to recruit additional officers to the conspiracy which in turn, increases 7

9 the chance of coup plan being detected by the regime. 14 To summarize, personalization reduces the probability of a coup attempt in dictatorship by altering the composition of the supporting coalition in the security sector. Newly-minted security elites are more likely to defend the leader against potential usurpers as they share the fate and preferences with the leader. On the other hand, it is easier for the dictator to monitor her core security elites with closely aligned preferences and family or ethnic ties. Hypothesis 1 (personalization effect): Personalization of security forces in autocratic regimes decreases coup risk. Even if such personalization is effective for reducing coup risks, we also observe dictators who fail to personalize the security forces and even face a backlash from disgruntled entrenched elites. Although a high level of security forces personalization reduces the chances of coup attempt once it is in place, it is possible that the personalization policies such as purging military officers or creating a personally loyal paramilitary forces may increase the chances of coup plots in the short term. Nordlinger (1977) identifies the direct challenges to the corporate interest of military as the primary reason that can build up the disposition to overthrow the government. Following this argument, De Bruin (2017, 7) finds that counterbalancing adding security forces outside of the regular military increases the risks of coup attempts immediately after its implementation. Thus, political leaders anticipating such resistance from the military tend to implement coup-proofing strategies when the coup risk is low (Sudduth 2017a, 3). Since the personalization process is a gradual and accumulative one, reduction in coup risks does not come instantaneously; rather, I posit that it should decrease over time. Hypothesis 2 (grievance effect): Personalization of the security apparatus should first increase coup risk but then decrease coup risk. 8

10 3 Measuring the Personalization of Security Forces 3.1 Identifying Personalization Policies To operationalize security apparatus personalization, we need to identify what policies dictators implement to attain personalized security forces. I posit that the following five acts by a dictator are the observable indicators that we can use to predict a prelude of security forces personalization. 15 The first of such a sign is the creation of a new paramilitary or security service outside of the regular military that is personally loyal to the dictator. These organizations are not the same as party-controlled armed groups such as the Nazi Party s SS or Iran s Islamic Revolutionary Guard Corps that are established to defend the ruling party and the regime against the regular military, respectively (Wehrey et al. 2009, 9-12). One example of such personal paramilitary organization is the Tonton Macoute created by Francois Papa Doc Duvalier of Haiti in It was composed of illiterate young men and the organization answered to Duvalier only, committing numerous acts of political terror (Council on Hemispheric Affairs 2010; Kendall-Taylor, Frantz,& Wright 2017, 15). Another question that needs to be asked is whether the dictator or other groups of elites such as the ruling party or senior military officers control the security forces. The ability to mobilize, command, and control the armed forces is crucial when determining the power balance between the leader and her support group. The support group can deter the dictator s possible move toward the personalization as long as they can credibly threaten to oust the leader through a coup attempt (Geddes 2009, 10; Svolik 2012, 61). However, if the dictator can directly command authority without sharing it with other elites from regime institutions such as the ruling party s military commission (as in Hu Jintao s first term) or a consultive organization among high ranking officers (junta in military regimes), it decreases the probability of a successful rebellion by the ruling coalition members (Geddes 2009, 12; Svolik 2012). In an attempt to build a highly personalist regime at the expense of ruling coalition members, North Korea s first regime leader Kim Il-sung first effectively removed the military and internal security troop commanders who did not belong to his faction by passing the blame for failed offensive against 9

11 South Korea during the Korean War before moving on to personalize the ruling party (Song & Wright 2018). The last type of pattern that we can observe during the personalization process is the ability to appoint Personal loyalists to important posts of the security apparatus (or vice-versa, the ability to remove anyone from these posts) at will, not just in high-level military command but also in intelligence agencies, the (secret) police, and internal security troops (Kendall-Taylor, Frantz, & Wright 2017, 14-15). In an effort to further consolidate his personal control over the security apparatus, Putin merged the interior ministry troops, special police crack units (OMON and SOBR), the Federal Migration Service, and the Federal Drug Control Service to create the National Guard of Russia (Panfilov 2016). This 340,000-strong internal security unit not only exceeds the number of the regular Russian ground forces of 290,000 (Luzin 2017), but also may subordinate the regular armed forces under its command during a domestic crisis (Golts 2017). Putin selected Viktor Zolotov, a personal loyalist who has been working for him since the 1990s as the chief bodyguard, to be the director of the National Guard. Dictators may also choose to promote family members. Saddam Hussein appointed multiple family members to crucial security posts: his son Qusay to the Revolutionary Guard divisions; his son Uday to Saddam s personal paramilitary organization known as the Fedayeen Saddam; his cousin Barzan Abd al-ghafur to the Special Republican Guard; and his half-brothers Barzan and Sabawi Ibrahim al-tikritis to the Iraqi Intelligence Service and the Directorate of General Security, respectively (Woods, Lacey, & Murray 2006; Macintyre 2003; The Independent 2013). 3.2 Conceptual Structure of Security Apparatus Personalization Before moving on to the data and measurement, I address the following questions on the conceptual structure of security forces personalization. First, the five observable policies used in the construction of the latent measure are not mutually exclusive. In other words, I assume that dictators who face direct threats to their physical security should they lose power would make attempts to maximize their personal powers with whatever means necessary. Although the exact sequential order of these personalization policies (or whether one policy is neces- 10

12 sary for another) is not the focus of this paper, from the IRT model we can learn that certain policies are easier for dictators to implement than others because they are observed more frequently. For example, Table 1 and Figure 2 reveal that promoting military officers based on personal loyalty (Personal promotion) is the most prevalent among the five policies, followed by Personal appointment, Personal control, Personal paramilitary, and Personal purge. 16 The goal of using an IRT model to measure the level of security forces personalization is to estimate the latent power of dictators vis-à-vis their security forces using multiple observable indicators that capture behaviors we would expect to see if the dictator has more relative power over security organizations than the de facto leaders of these organizations. The personal power of the dictator vis-à-vis the security apparatus is inherently unobservable, so we use observable indicators gleaned from systematic reading of the case study literature on different authoritarian regimes. To do so, I have selected the manifest variables based on the criterion of what dictators are able to do. In other words, dictators are considered to be more powerful the more we observe them engaging in personalization. Measuring a dictator s personal power based on what she can actually do (de facto capabilities) is important because, in many cases, constitutions in autocracies do not clearly specify command and control authorities or are intentionally written ambiguously in order to confuse and disrupt coordination among potential coup plotters (e.g., Black 2000, 10; Gause 2006, 5). In these cases, de jure capabilities of dictators do not correctly reflect the level of true personal powers they hold. Relatedly, I refrain from including variables that measure structural features of security apparatus such as the ethnic composition of military and the number of counterbalancing organizations because they do not belong to the concept of security apparatus personalization, both conceptually and empirically. Conceptually, I consider security forces personalization as a dictator s observed effort to reshuffle the composition of her security apparatus with Personal loyalists. The number of counterweight organizations or having a co-ethnic military might overlap with personalization process (and also correlate with lower coup risks); however, the key here is distinguishing whether a dictator has created (para-)military organizations that are designed to be personally loyal to the dictator (Personal paramilitary) or filled military posts 11

13 with co-ethnic officers based on personal connections (Personal promotion). Moreover, variables that measure ethnic stacking do not fully distinguish if a dictator chose to build her armed forces around personal-ethnic loyalty or (1) she merely inherited an ethnic army structure from the colonial period (e.g., Syria) or (2) ethnicity is not politically relevant (e.g., China or North Korea). Thus, personalization policies should be limited to observed acts by dictators to increase personal loyalty among the security elites. Without such a criterion, we would also include variables such as the level of military spending. Such monetary compensation could be a policy of recruiting additional Personal loyalists and may also reduce the coup risk; however, we are not able to distinguish if the expansion or reduction in military budget resulted from dictator s effort to create a more loyal security apparatus. On the other hand, the empirical pattern also supports the decision not to include the structural features of the security apparatus as observable indicators. Utilizing the IRT model is advantageous in this situation because the latent variable models offer a principled way to collect, aggregate, and structure different pieces of information. Alternatives to using a latent variable model such as the IRT would be an additive index of the observable indicators or relying on arbitrary decisions by a human researcher without knowing how the indicators covary or how much information they contribute to a latent measure. Appendix Figures 5, 6, 10, and 11 consistently show that incorporating the structural variables of ethnicity and counterbalancing does not reveal much information to the unidimensional concept of security forces personalization. Moreover, Appendix Figure 9 posits that adding the structural variables decreases the fitness of the models. Setting principled theoretical and empirical standards for which observable indicators to be included in the latent variable model is important because they enable us to analyze how observed events systematically contribute to the latent concept that we are trying to measure. Last, the fact that only Personal control has passed the 95% significance threshold in relations to the coup risk (Appendix Table 9) does not mean that the other four indicators should not be included in the latent model. In other words, testing whether a personalization policy reduces coup risk is not the same as using it to measure latent level of dictators personal powers. 12

14 All items contribute information to the latent variable, as evidenced by the item discrimination parameters reported in Table 1. If the four indicators did not contribute information to the latent variable, the item characteristic curves of these indicators would be flat lines in Figures 2. Including the other four indicators not only yields greater within-dictator variation of personal powers but also more information on which personalization policies are easier or more difficult to be implemented compared to the personal command and control of security apparatus, as evidenced by the difficulty parameters. 3.3 Data on Security Apparatus Personalization I use newly-collected GWF data on different dimensions of autocratic rule (Geddes, Wright, & Frantz forthcoming) to gather observable indicators of the security forces personalization. The data includes 4,457 regime-year observations and 280 autocratic regime cases from 1950 to While the previous GWF regime typology data allows cross-regime variations, it does not capture changes within the same regime over time. For example, both China and the Soviet Union are categorized as dominant party regimes (i.e., no variation during regime duration), but different Chinese and Soviet leaders developed various levels of personalist power. The data collection involves extensive reading of case studies, biographies, and news reports in order to attain information about the relationships between the dictator, the ruling political party (provided that there is one), and the military. The coders recorded information for January 1 of each calendar year to allow variations over time within autocratic regimes (Wright 2017, 8-9). They are also asked to identify the first regime year which a specific personalization policy occurs and the rest of regime years are coded as positives (1s) as long as the leader who started it is in power or until no longer able to do so. For example, the indicator for whether the dictator is able to purge high-ranking military officers at her will is coded as positives during the Soviet regime years under Stalin, but it is coded as negatives (0s) when Khrushchev takes power in The constructed dataset contains five observable variables that are related to security forces personalization. Their names and coding rules are: 13

15 Personal paramilitary: Does the regime leader create paramilitary forces, a president s guard, or new security force loyal to herself? Personal control: Is the security apparatus controlled personally by the regime leader? Personal promotion: Does the regime leader promote military officers loyal to herself or from her ethnic, tribal, regional groups? Personal purge: Does the regime leader imprison or kill military officers from other groups without fair trial? Personal appointment: Does the regime leader have discretion over appointments to high office or appoint relatives in these positions? Personal paramilitary variable captures whether the leader builds paramilitary or praetorian guards that are personally loyal to her and that are independent from the regular military leadership. Pro-government militias and death squads that act as informal state agents to carry out repressive actions (Mitchell, Carey, & Butler 2014) and paramilitary forces that are merely serving as a counterweight against the regular military and intelligence agencies are not included. Secret service and presidential palace guards also do not qualify if they are loyal to the office of the president, but not the specific leader herself. 17 Personal control identifies whether the regime leader has command and control authority over the security agencies or the authority belongs to or shared with the members of dominant party or a group of senior military leaders. 18 Personal promotion, Personal purge, and Personal appointment are coded positively if a regime leader is able to promote or purge military officers, cabinet members, or other high-level officials including intelligence heads or interior ministers based on personal loyalties or family and ethnic ties, and not through institutionalized formal promotion and punishment processes. To my knowledge, the earliest work that tests this argument with a large-n analysis is Belkin & Schofer s (2003) article on structural coup risks. They measure counterbalancing by counting the number of military and paramilitary organizations and comparing the relative size of paramilitary to the total number of armed forces (Belkin & Schofer 2003, 613). Similarly, Pilster & Böhmelt (2011; 2012) count effective number of ground combat organizations and each organization s personnel. While Powell s (2012, 1031) work finds a negative relationship 14

16 between the paramilitary-to-military ratio and the coup attempts but not for effective organization numbers, Frantz & Stein (2016, 18) find largely a positive association between effective organization numbers and the probability of coup attempt in their two-step models. Measures that count numbers of organizations or their personnel do not capture loyalty schemes that dictators may employ to alter the structure of the military and security apparatus. Often security apparatus organizational charts and their roles can be deceiving in autocracies, and this problem is more acute in personalist regimes such as Libya and North Korea. Qaddafi created an intentionally confusing military chain of command to prevent military elites from amassing power independent of the regime leader (Black 2000, 10). In North Korea, the regime leaders have frequently overrode official chains of command and the flow of power and information depends more on personal connections between the generals and the regime leader (Gause 2006, 5). Moreover, numerical counts of organizations do not contain information about each security organization s role and independence of operational control. The creation of an additional military unit reflects the leader s attempt to counterbalance the armed forces as in Iraq s Republican Guard divisions. It can also illustrate an increased investment in national defense, adding specialized units for different mission environment (e.g., marine corps or naval infantry divisions) and creating internal security troops to lift the burden of domestic security from the regular armed forces. 19 Harkness (2016, 587) work, however, is one attempt to overcome this issue. She constructs original data on the initial leader s choice of ethnic composition of the officer corps in Africa; her analysis finds that attempts to build co-ethnic armies despite the pre-existing ethnic diversity will lead to greater chances of violent resistance from the military (Harkness 2016, 610). Nevertheless, the ethnic composition data does not capture the post-decolonization changes in African security institutions. We better understand the relationship between institutional characteristics of the security forces and the coup risks by observing both qualitative differences between security organizations and how they transform over time. In other words, not only accounting for the number and respective strength of security apparatus, we need to trace the continuous process of how regime leaders shape the security institutions to ensure their loy- 15

17 alty. For example, Greiten (2016, 12-13) contends that autocratic security organizations are institutionally malleable by describing how the regime leaders in Taiwan, the Philippines, and South Korea shaped the design of coercive apparatus over the several decades based on their dominant perceived threat. 3.4 Measuring the Latent Concept With these five observable indicators, I construct a latent measure of the personalization of the security apparatus. The simplest way of combining multiple indicators would be additive indexing. However, this approach assumes that each personalization policy carries a same weight toward the latent concept. This is theoretically illogical as imprisoning or killing military officers is very different from creating a new paramilitary force. Therefore, I estimate an Item-Response Theory (IRT) model to construct the index. The IRT model is used to determine the relationship between the latent variable (degree of the security forces personalization) and the items that are intended to measure the latent variable (the five observable indicators) (Reise & Waller 2009, 29). 20 Since the items used in this paper are all binary responses (i.e., dichotomous indicators), I use the two-parameter logistic model to assume that each item varies in their discrimination or weight (Birnbaum 1968). 21 Table 1: Two-Parameter Logistic Model Estimates Indicator Discrimination Difficulty Personal control Personal paramilitary Personal appointment Personal promotion Personal purge Notes: Variables are in descending order of discrimination. In Table 1, I present the discrimination and difficulty parameters from the IRT model. It shows that Personal control has the highest level of discrimination the ability to distinguish better between low and high levels of personalization (Baker 1985; 2001, 34). Other indicators 16

18 % 64.5% 39.7% 23.7% 60.7% 35.3% % Percent % 64.7% % 39.3% % 0 Categorical Regime Paramilitary Control Promotion Purge Appointment Positive (1) observations Non-positive (0) observations Notes: This figure provides a graphical illustration of the regime-year distributions for the five binary indicators used to construct Personalization index and the GWF personalist regime classification. The sample has 4,457 observations with 117 country and 280 autocratic regime cases, with years from 1950 to This illustration shows that specific personalization policies do not equate with being a personalist regime leader. Figure 1: Descriptive Statistics for Personalism Indicators show lower levels of discrimination. This makes sense since the regime leader has to have the ability to control and command the security apparatus before promoting or purging officers based on personal loyalty. The second parameter of difficulty is related to the location of an item on the latent variable scale. Personal promotion variable has the lowest difficulty of 1.111; the probability of being able to promote military officers based on the level of personal loyalty is higher than the probabilities of being able to execute other types of personalization policies. The item characteristic curves in Figure 2 is the visual representation of the discrimination and difficulty parameters. The steeper curves indicate a higher level of information (discrimination). Personal control curve is the most steep, revealing the greatest amount of information. The level of difficulty for each item is shown on the x-axis at which the curve crosses the 0.5 probability value on the y-axis. Their positions along the x-axis are in the order of difficulty parameter, from Personal purge (most difficult at.55) to Personal promotion (least difficult at 17

19 1 Item Characteristic Curves.75 Probability Difficulty Personal paramilitary Personal control Personal promotion Personal purge Personal appointment Figure 2: Two-Parameter IRT Item Characteristic Curves -1.11). The data is plausible as purges and creation of a personal paramilitary forces are direct threats to physical security and corporate interests of military officers. Using these estimates, the IRT model generates a continuous index that numerically represents the latent trait measuring the extent of security apparatus personalization. 22 The correlation matrix in Appendix Table 1 illustrates that all observable indicators are positively correlated to each other and also to the generated index. This indicates that the manifest indicators are measuring the same latent concept of security apparatus personalization. Roughly 20 percent of the variation in the measure is from within-regime case and 80 percent is across regimes (Appendix Figure 1). I also run tests to incorporate the uncertainty of the personalization measure into model estimation (Appendix Figure 3) and the co-ethnic military variable as an additional item (Appendix Figure 5). 18

20 3.5 Verification Through Example Regime Cases In Figure 3, I plot the generated Personalization index time series graphs to examine face validity of the index by observing actual histories between regime leaders and security forces. I select the regimes that have relatively well-known histories and leaders generally assumed to have highly personalized powers. The plots shown in Figure 3 illustrate temporal changes in the standardized values (from 0 to 1) of Personalization index and Pilster & Böhmelt s (2011; 2012) count of effective number of ground combat organizations in each autocracy. Max Soviet Union/Russia Max China Min Year Min Year Max Syria Max Ghana Min Year Min Year Personalization index Effective number Figure 3: Changes in Security Apparatus Personalization in Selected Regimes Both the Soviet Union/Russia and China have dictators with highly personalized powers early in the regime spells. The index for the Soviet Union declines rapidly after Stalin s death in 1953, increases marginally during Khrushchev s tenure when he was able to fend off rival elites in the Presidium in 1957 with the support of Defense Minister Marshal Zhukov and other 19

21 security departments (Tompson 1995, ). 23 Fearing Zhukov s power and popularity, Khrushchev dismissed Zhukov from the defense minister position in the same year while he was on leave to visit the Albanian counterparts (Chaney 1996, ). The index decreases again as Brezhnev reinstated the collective leadership (Service 2009, 378) but starts to increase under Yeltsin and Putin s tenures. Personalization index for China spikes when Mao starts the Cultural Revolution in The index reaches the highest point and declines from 1968 when the Red Guards are violently suppressed by the PLA (Meisner 1986, 361). The trend starts to flatten as Mao s personal powers are weakened during the later years of his tenure ( ) and the power-sharing arrangement within the Chinese Communist Party is instituted (Lieberthal 2004, 132). The Syrian and Ghanaian cases contain multiple successful coup attempts. The Syrian leaders did not exercise full personal control over the security forces before Hafez al-assad came to power in 1970 through a military coup. The level of security apparatus personalization increased rapidly under Hafez because of an event in the 1960s. Although then-syrian army s top posts were filled by the Sunnis, the Alawites that al-assad s clan belong to, were numerically superior. When President Amin al-hafiz a former Sunni general came to power through a military coup, he purged many high-ranking Sunni officers throughout the mid-1960s which opened a door for the Alawite officers to fill the top posts (Bhalla 2011). Enjoying co-ethnic dominance in the military, Hafez al-assad created independent paramilitary and special forces that are personally loyal to him, including the Defense Companies in 1971 which later became the 4th Armored Division (Sinai 1988, 269), the Republican Guard divisions in 1976 (Paul 1990, 50), the General Intelligence Directorate in 1971 (MEIB 2000), and turning the Air Force Intelligence Directorate as a powerful personal agency due to the fact that Hafez al-assad was once an air force commander (MEIB 2000). The personalization level drops during the early tenure of Basher al-assad because the old guards that were personally loyal to his father, wielded moral weight within the regime (Bar 2006, 311). He gained complete personal control after the Baath Party s 10th Congress in 2005 when Basher replaced the old guards with younger regime members that are close to him (Bar 2006, 385). 20

22 Ghana also experienced multiple successful coups. Ghana s level of security forces personalization increased during the late tenure of Nkrumah when he expanded his presidential guard unit from a company to a full regiment. He removed the unit from the regular military command chain and granted its operational independence in 1964, making the unit only answerable to him. Because Nkrumah angered military officers by favoring his personal forces over the regular military, he lost power by the coup of 1966 (Biney 2011, 156). Years between Nkrumah s departure and Rawlings coming to power in 1981, Ghana experienced several democratic and military dictators that shared power through juntas (e.g., National Liberation Council, National Redemption Council, and Supreme Military Council). Rawlings built a highly personalized regime by 1984, when his Provisional National Defense Council (PNDC) government created a countrywide network of the Civil Defense Organizations (CDOs) and the Forces Defense Committees (FDCs). The CDOs were paramilitary forces with a mission of assisting the government during national emergencies; the FDCs were the military and police force s counterpart to the Committees for the Defense of the Revolution (CDRs), mass revolutionary organizations that were installed in workplaces, cities, and villages (Owusu 1995, 199). Utilizing more detailed data based on the relationship between each dictator and military not just institutional makeup can show how the regime leader s power relative to the security apparatus varies within same regime case over longer temporal range. 24 In the next section, I use Personalization index to test if the security apparatus that are personally loyal to the dictator are effective deterrent against coup attempts in autocracies. 4 Estimation Strategy 4.1 Data The unit of analysis is autocratic regime-year. The regime cases come from the Geddes-Wright- Frantz (2014; forthcoming) data, with years ranging from 1950 to 2010, depending on the data availability. The merged dataset includes 4,457 regime-year observations and 280 autocratic regime cases. The dependent variable used to test the the hypotheses is the binary indicator of 21

23 whether a coup attempt occurred in a given regime-year. The coup attempt variable comes from Powell & Thyne s (2011) data where they code 1 if one or more coup attempts occur in a given country-year, and as 0 if otherwise. They define a coup as extra-judicial and overt attempts by military or other elites to remove the incumbent executive leader (Powell & Thyne 2011, 252). I also include a number of state-level predictors of coups in the models. Following Powell (2012), I include the dollar-amount of military spending and the number of military personnel in thousands from the Correlates of War project s National Material Capabilities data (Singer, Bremer, & Stuckey 2007). To account for state s general level of economic development, I use the logged dollar amount of GDP per capita data from the World Development Indicator (2012) and the Penn World Table (2012). I also use the logged amount of oil production per capita in metric tons from Wimmer, Cederman, & Min (2009) and Wimmer & Min (2006) to capture the regime leader s ability to bribe security forces without the need to increase taxation. As Roessler (2011) and Harkness (2016) show, the practice of stacking security institutions with the dictator s co-ethnic group has been in place for a long time in order to ensure loyalty from the armed forces. A binary variable that indicates whether the high ranking positions in the armed forces are dominated by the group that the regime leader belongs is included. 25 This variable is coded 0 for autocracies where ethnicity is not politically salient, such as in North Korea. 26 This comes from the updated Geddes-Wright-Frantz dataset (2017). To account for regime instability, I include the UCDP s three-level civil conflict data and code conflict event-year with at least 25 but less than 1,000 battle deaths as low-intensity and 1,000 or greater number of battle deaths as high-intensity civil conflicts. Following the findings on the positive relationship between protests and coup attempts (Johnson & Thyne 2016), I account for the number of violent and nonviolent protest campaign movements in a given autocratic regime-year. The data come from the Nonviolent and Violent Campaigns and Outcomes (NAVCO ver. 2) project (Chenoweth & Lewis 2013). I include decade dummies to account for changing international environment of both Cold war and post-cold War and decolonization period and post-decolonization period (Marinov & Goemans 2014; Berger et al. 2013). 27 The last variable I use involves the information on the dictator s mode of entry, which is 22