Going, going, gone? Varieties of dissent and leader exit

We examine how popular dissent affects the likelihood that political leaders lose power, distinguishing between types of dissent in terms of nonviolent/violent primary tactics as well as the level of individual participation. We posit that protests threaten leaders both directly through the governance costs of citizen non-compliance, and indirectly through the increased risk of elite defections in the ruling coalition. In a series of propositions we detail how the type of dissent and the magnitude of participation influence the odds of leaders surviving in office. We argue that mass nonviolent challenges tend to be more threatening to a leader’s rule than violent dissent, given the characteristics of movements likely to choose nonviolent tactics. Moreover, the effectiveness of the challenge increases in the scale and size of the dissident campaign, and movements that can mobilize larger numbers have a comparative advantage in nonviolent tactics. Employing data on political leaders’ tenure and dissident campaign characteristics, we provide evidence consistent with our expectations.


Introduction
Now, I've just got to cut loose Before it gets late So I'm going I'm going I'm gone (Bob Dylan, Going, Going, Gone, Planet Waves, 1974) We develop a theory of how, and under what conditions, violent and nonviolent dissent influence a leader's political survival. We posit that dissent threatens leaders both directly through the governance costs of citizen non-compliance, and indirectly through the increased risk of elite defections in the ruling coalition. In short, leaders are substantially more likely to leave office when active mass mobilization makes it difficult to hang onto power. However, mass nonviolent challenges are likely to be more threatening to a leader's rule than violent dissent, given the characteristics of movements likely to choose nonviolent tactics.
Maintaining power is often taken as a leader's key motivation, shaping decisions and political outcomes (Bueno de Mesquita et al., 2003). In democracies, political power is ultimately allotted based on competitive elections with popular participation. In non-democracies, political power is typically seen as emanating primarily from other elites, who decide to support or replace a leader based on the leader's ability to offer rents or policy concessions (Gandhi, 2008;Svolik, 2012). The existing literature on leaders' political survival has emphasized the risk of intra-elite threats such as coups, and paid less attention to the role of large-scale popular challenges from outside the ruling coalition. Mass challenges can be either violent or nonviolent. The conventional assumption is that violent challenges must be more serious -as reflected in Mao's famous quote that 'power comes from the barrel of a gun'. However, some argue that nonviolent tactics can be more effective (Chenoweth & Stephan, 2011).
By contrast, we argue that the characteristics of dissident groups, which make the choice of a specific tactic more likely in the first place, are key in explaining why 'nonviolence works'. We broaden measures of dissent impact by focusing on whether leaders lose office. Existing studies examine mainly cases where leaders facing dissent are ousted or irregularly removed from power (Casper & Tyson, 2014;Aksoy, Carter & Wright, 2015;Beger, Dorff & Ward, 2016;Johnson & Thyne, 2018;Chenoweth & Belgioioso, 2019;Kim & Kroeger, 2019). This entails a far too limited understanding of the possible impact of dissent as many leaders resign in the face of challenges, precisely because they anticipate that their rule will become untenable. In the perceptive words of Bob Dylan, a leader may go before they are likely to be 'gone'.
While all dissent is damaging to leaders and increases the odds of exit from office, we expect nonviolent direct action to be generally more damaging than violent direct action. Compared to violent conflict, nonviolent mobilization can generate more significant governance costs by threatening a state's economic base and fueling elite divisions. Violent conflict is often confined to distinct ethnic groups in the periphery where the center's reach is weaker, which in turn generates fewer incentives for elite opportunism than turmoil at the center. By contrast, nonviolent uprisings typically reflect universal/nonsectarian goals and take place in urban areas where the state seems stronger by traditional measures of state capacity. Yet the governance costs of large-scale nonviolent mobilization in urban areas (which tend to be the fiscal backbone of a state) typically exceed those arising from a limited insurgency in the periphery.
Thus, governments are more sensitive to widespread non-compliance. Since the effectiveness of dissent increases in the scale and size of the campaign, we propose that nonviolent challenges have a comparative advantage in unseating the leader because they can generate higher participation. This is because nonviolent movements have a broader popular appeal and lower barriers for entry. We test our theoretical expectations by examining all leader exits and mass mobilizations between 1945 and 2006, using data from the Archigos project (Goemans, Gleditsch & Chiozza, 2009) and the Nonviolent and Violent Campaigns and Outcomes (NAVCO v2.0) (Chenoweth & Lewis, 2013). The results provide strong evidence in support of our propositions, and are robust to alternative explanations affecting the risk of exit and likelihood of dissent.
This article advances our understanding of political survival and social resistance movements in several ways. First, we show that popular compliance is the ultimate requirement for being able to rule (Sharp, 1973;Levi, 1988), rather than the monopoly on the use of force (Huntington, 1968;North, 1981;Weber, 1918Weber, /2004. Practically, large-scale nonviolent mobilization and popular non-compliance pose a greater threat to rulers than violent dissent. Second, we emphasize the role of agency in challenges (e.g. groups strategically choosing the dissent tactic likely to be most effective based on their underlying characteristics and resources), in contrast to work that emphasizes the direct consequences of nonviolent tactics (e.g. Chenoweth & Stephan, 2011). Finally, we conduct a comprehensive empirical analysis of dissent and political survival, explicitly comparing campaign types against nomobilization cases, with a series of robustness tests, including matched data and strategies to consider possible implications of strategic selection in mobilization.

Political survival and popular dissent
Much of the literature on political leaders and regimes emphasizes the importance of governments' resource superiority in maintaining political control (Svolik, 2012). This perspective tends to focus on elite challenges to leader tenure, since non-state actors cannot match the state's resources. In line with this, rebels rarely win outright in a civil war (Bapat, 2005;Cunningham, Gleditsch & Salehyan, 2009), revolutions fail more often than they succeed (Beissinger, 2013;Weyland, 2009), and popular dissent increases the likelihood of an elite coup attempt but not success (Aksoy, Carter & Wright, 2015).
Other research highlights several challenges to regime stability apart from regime insiders. Some emphasize how losing interstate wars undermines a leader's rule (Bueno de Mesquita et al., 2003;Chiozza & Goemans, 2011), and there is also evidence that civil wars pose a more severe threat to leaders who have recently gained power than established leaders (Chiozza & Goemans, 2004;Debs & Goemans, 2010). In a forecasting model of irregular leader removal, Beger, Dorff & Ward (2016: 5) find the largest weight for 'contagion and internal conflict', underscoring how domestic dissent poses a serious challenge.
There is also a literature on the role of popular pressure and dissent in promoting the end of autocratic regimes and transitions to democracy. Many studies extending the modernization literature propose that popular pressure promotes regime change. However, empirical studies have often inferred its effects from structural factors (e.g. inequality), not considered mobilization explicitly, or not discriminated between distinct types of dissent (Acemoglu & Robinson, 2006;Boix, 2003;Svolik, 2013). Rivera Celestino & Gleditsch (2013) find that nonviolent protests increase the likelihood of transition to democracy, while violent civil conflicts primarily increase the risk of irregular leader removal leading to new autocracies (see also Ackerman & Karatnycky, 2005;Teorell, 2010;Chenoweth & Stephan, 2011;Dahlum, Knutsen & Wig, 2019;Kim & Kroeger, 2019).
Existing research has offered compelling insights, but there are several important limitations in previous work on the consequences of dissent on political survival and regime change. In particular, (1) what is deemed to be 'successful' is often based on subjective, narrow, or restrictive criteria; (2) 'effectiveness' overlooks dissent impact short of full democracy; and (3) measures of 'success' are contingent on the timing of assessment.
First, attempts to code dissent 'success' inevitably have to rely on subjective criteria, as dissent can be influential at different time points, and outcomes may satisfy some actors but not others. The Orange revolution in Ukraine illustrates the lack of consensus, with some observers characterizing it as a successful revolution or democratization, others labelling it authoritarian turnover, or even elite consolidation after a failed movement (Beissinger, 2013: 579). Mass dissent typically entails broad coalitions, unified only in the opposition to a sitting leader, but individual factions often have very different longer-term aims. Movements often fall apart after the initial success in unseating the leader, as disagreements come to the front and factions turn against each other. In the 1979 Iranian Revolution, for example, factions seeking democratic reform were sidelined and repressed by the Islamists after the fall of the Shah (Gheissari, 2009).
Second, while some studies use measures of success with a clear delineation such as regime change (Chenoweth & Stephan, 2011;Rivera Celestino & Gleditsch, 2013;della Porta, 2014), this overlooks any outcomes short of transitions (Aksoy, Carter & Wright, 2015;Kim & Kroeger, 2019). More disaggregated measures of degree of democracy do not solve the core problem, as major changes in coalitions holding the power may not be reflected in changes in the degree of democracy. For example, the fall of the Shah is clearly a major political change spurred by dissent, but Iran's Polity score does not change significantly after the Islamic Revolution (Jaggers & Gurr, 1995).
Finally, attempts to classify whether a campaign is successful can vary dramatically with the timing of assessment. For example, the 2011 Egyptian revolution spurred efforts to unseat Mubarak, with the vice president transferring power to the Armed Forces Supreme Council. However, limited political competition prevented open democratic elections, and renewed protest in June 2013 was followed by a direct military coup. In this case, dissent clearly had important short-term effects, even though we do not see changes in regime or institutions. Therefore, it is essential to assess the short-term effectiveness of dissent (e.g. unseating leaders or coalitions) without conflating this with long-term consequences (e.g. promoting transition to democracy). 1

Conceptualizing dissent effectiveness
We use the ability of political leaders to stay in office as a simple and clear observable outcome of dissent effectiveness, avoiding the problems discussed above for alternative understandings of campaign success. This is linked to a specific actor whose (assumed) preference to remain in power is uncontroversial (Bueno de Mesquita et al., 2003). Any unplanned exit unmistakably represents a failure for the individual leader. Even when dissidents do not demand outright that the leader step down, a change in leadership at a time of unrest is likely the result of ruling elites responding to the challenge. Focusing on leader exit does not rely on subjective judgments, and can capture consequences below regime transition. It also extends existing research on dissent and irregular leader exit (Casper & Tyson, 2014;Aksoy, Carter & Wright, 2015;Beger, Dorff & Ward, 2016;Johnson & Thyne, 2018;Chenoweth & Belgioioso, 2019;Kim & Kroeger, 2019), since it covers cases where exit is regular and putatively voluntary, but arguably would not have happened in the absence of mobilization.
Many examples attest to the relevance of voluntary resignations prompted by mobilization. Honecker in East Germany and Milosevic in Yugoslavia/Serbia, for example, resigned in the face of mass protests. In the final days of the GDR, several Politburo members started to plan Honecker's removal from power following concerns over his hardline approach against the growing protests and the increasing alienation from the USSR this generated. When the vote for his dismissal was tabled, Honecker himself voted in favor of the proposal in line with the tradition of the Politburo of seeking a unanimous vote. 2 Similarly, Milosevic tendered his resignation since the Yugoslav electoral committee had 'reconsidered its previous ruling on the election outcome', and the revised results indicated that his rival Kostunica had won. 3 He made no explicit reference to the protests, which probably spurred the electoral committee to reconsider its initial decision. Neither of these exit cases constitutes an irregular leader change by the Archigos criteria.
What is more, irregular leader exits have become increasingly less common over time. Figure 1 plots the relative share of irregular to all exits per year. These have fallen from about one-third of all leader exits in the 1970s to a low of less than one-tenth after the 1990s. This reflects in part the decline in the frequency of military coup attempts, possibly because coup-makers are likely to face sanctions and thus find it harder to hold onto power (Powell & Thyne, 2011;Marinov & Goemans, 2014;Thyne et al., 2018). But whereas military coups have declined, popular dissent has increased (Chenoweth & Lewis, 2013;Chenoweth & Ulfelder, 2015). This also contrasts with the trend in conventional civil wars, which have declined after the Cold War, largely due to a decrease in secessionist ethnic conflicts (Cederman, Gleditsch & Wucherpfennig, 2017).
The end of Mugabe's rule in Zimbabwe illustrates the ambiguities in classifying the causes of leader exits and how potential coup makers have incentives to pressure for voluntary resignation. The military officers assuming power went out of their way to emphasize that they were not carrying out a coup, and saw it as imperative that Mugabe resigned voluntarily, to the point of offering reassurances of a safe retirement if doing so, and threatening impeachment if he failed to comply. 4 The resignation of Algerian President Bouteflika follows a similar pattern. Bouteflika's announcement that he would seek a fifth consecutive term in February 2019 was met by massive street protests, dubbed the Smile Revolution. On 11 March Bouteflika withdrew his candidacy, and on 1 April he promised to resign by the end of the month. When the protests continued, the military weighed in and called for the president to be declared unfit to rule. The military invoked article 102 of the constitution, which states that the presidency can be vacated if the incumbent is too ill to exercise his functions, revealing a strong preference to have the president removed through legal means rather than a coup. The very next day, on 2 April, Bouteflika announced that he would relinquish power immediately. 5 The events in Algeria also highlight the practical challenge in pinpointing the 'ultimate' cause of a leader's departure. Some may argue that the military was decisive in wielding the knife, while others may emphasize the 2 Honecker officially resigned due to his poor health, see http:// www.independent.co.uk/news/world/world-history/fall-of-theberlin-wall-history-catches-up-with-erich-honecker-the-east-ger man-leader-who-praised-the-iron-curtain-and-claimed-it-preventeda-third-world-war-9826715.html/. 3 For further details, see http://www.nytimes.com/2000/10/07/ world/showdown-yugoslavia-overview-milosevic-concedes-hisdefeat-yugoslavs-celebrate.html. mass protest as the ultimate cause since the military probably would not have acted in the absence of this. Efforts to code data on the specific factor leading leaders to stand down would ultimately rely on subjective judgments and counterfactuals that can be highly controversial. There is an instructive analogy from medicine in the well-known problem of low agreement in efforts to classify individuals' underlying 'cause of death' (Mieno et al., 2016). Examining how active dissent increases the risk of leader exit generally, be it voluntary or forced, provides a more straightforward empirical approach.
Focusing on leader exit also draws attention to some paradoxes in the relationship between state capacity and vulnerability due to challenges from dissent. State strength is often defined as the monopoly on the use of force (Weber, 1918(Weber, /2004, and much of the literature on civil wars argues that state weakness encourages violent challenges (Fearon & Laitin, 2003;Gleditsch & Ruggeri, 2010). However, many violent conflicts do not seem particularly threatening to rulers, and countries with longstanding civil wars have seen considerable stability in political leadership. President Ne Win of Burma (Myanmar), for example, ruled for 19 years despite a large number of violent rebellions in the periphery. Although not noticeably affected by these civil conflicts, Ne Win resigned when faced with the largely nonviolent Four Eights uprising in 1988. The vulnerability of Ne Win to the latter type of dissent challenges the folk wisdom that violent mobilization is the greater threat. In sum, a leader's ability to stay in power is a clear observable indicator of political capacity, and nonviolent and violent challenges are likely to have divergent implications.

Varieties of dissent and leader exit
A leader's power is typically assumed to come out of the purse or the barrel of a gun. Rulers can prevent dissent by providing selective benefits to key individuals and constituencies (Bueno de Mesquita et al., 2003;Gandhi, 2008;Svolik, 2012;Frantz & Kendall-Taylor, 2014). Alternatively, their ability to monitor the population and enact pre-emptive repression exacerbates the dissidents' collective action problem, and may deter mobilization altogether (Lichbach, 1995;Ritter & Conrad, 2016;Sullivan, 2016). If deterrence should fail, the overwhelming military superiority leaves rulers well placed to win in any confrontation (Geddes, Wright & Frantz, 2014).
We build on an alternative perspective on power, often associated with Sharp (1973), which stresses the need for popular compliance with state authority (see also Levi, 1988). Any government ultimately relies on its subjects' (tacit) compliance, and simply cannot rule if they refuse to follow orders (e.g. pay taxes), or they carry out acts the government tries to prevent (e.g. illegal strikes, demonstrations). Popular dissent signals to observers, be they ordinary citizens or elites, that there is a core of individuals unwilling to comply with the government's policies. If dissent mobilization acquires enough momentum, the governance costs imposed through non-compliance can make leaders unable to rule effectively and ultimately force them out of office (Lake, 1999;Braithwaite, Kucik & Maves, 2014;Chenoweth & Belgioioso, 2019). While dissidents rarely seize control of the state, popular dissent can make rulers chose to withdraw if high governance costs make their situation untenable.
Dissent can also increase pressure from within the ruling coalition, or the risk of elite defection and opportunism (Casper & Tyson, 2014;Kim & Kroeger, 2019). As in the aforementioned cases of Algeria, East Germany, and Egypt, rulers are often removed by regime insiders in the context of protest. Popular dissent can trigger defections against embattled leaders from otherwise loyal elites. 6 The direct governance costs created by dissent 'from below', and the indirect elite defection costs of revolt 'from within' are conceptually distinct. However, in practice the two are often intertwined and their relative weight is difficult to assess. Higher governance costs imposed from below can encourage elite opportunism and defection, as in the case of Egypt, and perceived elite splits may encourage mobilization and dissent from below, as appears to have happened in the Tiananmen protest in China.
Violent campaigns generally seek to coerce by threats to kill or the political ramifications of violence. But the dominance in military capabilities leaves governments better positioned to resist a violent movement, and the strategic advantage of violent campaigns in the periphery is often inversely proportional to its economic impact on 6 Elite defections manifest 'primarily via coups, which nonviolent anti-regime protests encourage in several ways' (Kim & Kroeger, 2019: 654). Most coup attempts are bloodless, especially when successful. As such, they are conceptually distinct from NAVCO violent campaigns, where armed conflict between a non-state actor and the government must claim more than 1,000 battle-deaths. Only under exceptional circumstances can a botched coup attempt set off armed confrontations at the level of a civil conflict. In that case, coup makers or allies of the previous incumbent would be considered a non-state actor even if they emerged from the old regime. the center. By contrast, nonviolent dissent seeks to coerce through popular non-compliance, by generating political crises via elite defection or imposing unsustainable governance costs. This happens because nonviolence often takes place in the urban core where the state is most sensitive to the economic disruption from non-compliance (Chenoweth & Ulfelder, 2015;Johnson & Thyne, 2018;Kim & Kroeger, 2019;Dahl et al., 2021). Thus, while all dissent is potentially damaging to leaders and increases the odds of exit, we expect nonviolent direct action to be more detrimental than violent conflict.
The Campaign Type hypotheses: (1): Leader exit is more likely under dissent.
(2): Leader exit is more likely under nonviolent than violent dissent.
In our view, many of the practical differences between nonviolent and violent mobilization are not direct consequences of the tactic per se, but stem from differences in the dissident groups' constituency and mobilization potential. These characteristics, in fact, make a specific tactic more likely to be selected in the first place. Groups strategically adopt tactics based on the best expected response given their resources and characteristics (Dahl et al., 2021). Nonviolent claims for democratization or regime change are non-sectarian, with a wide crossover appeal to many government opponents and plausibly defecting elites (Dahlum, Knutsen & Wig, 2019). A minority ethnic group has an exclusive audience or constituency, with little crossover appeal to elites or individuals in the dominant group. In practice, many civil wars reflect ethnic groups seeking secession in rural insurgencies (Buhaug & Gates, 2002), while nonviolent mobilization takes place when groups can mobilize large numbers in the capital and key cities (Dahl et al., 2021). 7 In practice, motive and resources also shape the plausible minimum and maximum individual participation in campaigns. Violent intrastate conflict often relies on covert planning and operation, and does not require large numbers to be feasible. The large opportunity costs and high requirements for equipment and training make maximum feasible participation low -at least at the outset, as converting novice recruits to competent soldiers takes time. Nonviolent dissent needs a high number of participants to exert significant pressure, but the low opportunity costs to participants make this relatively easier to achieve as the mobilization potential is high. The risk to individuals also decreases rapidly with growing participation given higher logistical and political costs for mass repression (Francisco, 2004;Siegel, 2011;Sutton, Butcher & Svensson, 2014).
In sum, nonviolent direct action is an attractive option for large and resourceful movements that can quickly mobilize many participants, whereas violent tactics may be the only feasible option for movements with a limited mobilization potential. The data confirm that nonviolent campaigns tend to involve more active participants than violent conflicts. Figures 2a and 2b, respectively, show the individual dissident participation in nonviolent campaigns from NAVCO, and the level of rebel forces from the Uppsala Armed Conflict dataset (Gleditsch et al., 2002;Cunningham, Gleditsch & Salehyan, 2013). The values are displayed on a log scale to better capture the long tail, with medians superimposed. Median participation in violent conflicts is only 4,000 troops, while the median participation estimate for nonviolent campaigns is 100,000.
We expect large participation in dissent to increase the threat to leaders. However, the risk of exit should increase more steeply for nonviolent dissent, since this type of campaign has a greater potential for further mobilization and is harder to disrupt once it gains momentum.
The Campaign Size hypotheses: (1): Leader exit is more likely with greater participation in dissent.
(2): Leader exit is proportionately more likely under greater participation in nonviolent than violent dissent.

Empirical analysis
We test these hypotheses using data from the Archigos v4.0 project (Goemans, Gleditsch & Chiozza, 2009), which provides information on the de facto head of state for all countries in the Gleditsch & Ward (1999) state list, with start and end dates. The dependent variable Leader exit is coded 1 for the time period when the leader loses power (regardless of whether it was via regular or irregular means), and 0 while the leader is in office. The indicators for violent and nonviolent campaigns are taken from the NAVCO v2.0 data, which provide information on the occurrence as well as the level of individual participation for the 1945-2006period (Chenoweth & Lewis, 2013. Our unit of analysis is leaderyear, since the NAVCO data are reported annually. The violent campaigns in NAVCO are organized contests between a non-state actor and the government over either control of the government or territory (a.k.a. maximalist goals), which must claim more than 1,000 battledeaths. The nonviolent campaigns are direct action movements over similar maximalist goals that mobilize at least 1,000 participants. Note that the focus on maximalist direct action excludes nonviolent dissent by regular means (e.g. legal petitions and political parties), or direct action over claims that are not maximalist (e.g. protests over environmental protection). As a result, the nonviolent and violent campaigns we examine are comparable in terms of their ultimate goal, even if the mode of resistance is different.
Besides our main independent variables, we control for a number of factors that can affect the popularity of leaders and the risk of exit. Specifically, we consider whether leaders entered in a regular manner (i.e. not in contravention of existing rules), since leaders who have entered irregularly are more likely to be removed in an irregular manner (Londregan & Poole, 1990;Goemans, Gleditsch & Chiozza, 2009). We also add leader age, based on Archigos, since older leaders are more likely to see challenges (Chenoweth & Ulfelder, 2015).
For economic indicators, we include the logged GDP per capita, since poor countries tend to be more unstable, and the growth rate, since leaders are challenged less often when the economy is doing well (Gleditsch, 2002). We also control for total population and percentage of urban population using data from the Correlates of War project (v5) (Singer, Bremer & Stuckey, 1972).
To account for institutional differences, we control for regime type by coding a state as a democracy if it has a score of 6 or above on the Polity IV scale (Jaggers & Gurr, 1995). 8 We expect leaders in democracies to be more likely to see exit because of shorter tenure and higher turnover. We also consider whether leaders who step down face legal term limits and thus are required to relinquish power. When effective, terms limits should accurately predict exit. This could encourage dissent if leaders mandated to step down are perceived as lame ducks late in their term. Our main source to identify term limits is the National Elections Across Democracy and Autocracy Dataset (NELDA v4.0) from Hyde & Marinov (2012). 9 We add information on term limits in authoritarian regimes, where elections may not be held or are irrelevant, from Svolik's (2012) Leadership Change in Dictatorships dataset.
To control for the government's ability to fend off challenges we include the composite indicator of national capabilities index from the Correlates of War, which encompasses military expenditure and personnel. For more specific indicators of coercive capacity we consider the government's willingness to curtail basic civil rights or engage in extrajudicial killings and torture, based on data from the V-Dem (v10) project (Coppedge et al., 2020). CSOs protection indicates whether civil society organizations are free to organize and criticize the government without fear of sanctions, while Physical  Figure 2b shows the level of participation in violent armed civil conflicts from the Uppsala Armed Conflict dataset. Medians are superimposed as red vertical lines. 8 We use a dichotomous measure of regime type to facilitate interpretation. Using the full [À10, 10] Polity scale does not change the results substantively, and a sample limited to nondemocracies (i.e. Polity < 6) yields substantively similar findings (see Online appendix A, hereafter OA). 9 We depart from NELDA in a small number of cases. First, we add some term limits not linked to elections (e.g. leader rotating mechanisms). Second, where Archigos and NELDA list different leaders, we use additional sources to determine if the leader in Archigos reached their term limit. integrity captures freedom from political killings and torture. Although our models control for observed repression, we recognize that they do not fully consider the potential deterrent effects on the outbreak of dissent.

Results and discussion
To examine how active dissent affects the risk of losing office, we estimate various Cox proportional hazard models of exit. The Cox model framework is useful as it allows considering how factors affect the risk of exit without needing to specify the correct hazard function, as the baseline hazard cancels out in the partial likelihood. Table I reports a series of Cox regressions on the hazard of leader exit. Model Ia includes no covariates other than campaign type. Compared to the base category of no active dissent, both violent and nonviolent campaigns increase the risk of exit. But nonviolent campaigns have an estimated coefficient about twice as large as that on violent campaigns, and the two coefficients are statistically different. Thus, nonviolent dissent is on average more detrimental to a leader's survival than both nonresistance and violent conflict. These findings provide initial empirical support for our Campaign Type hypotheses. We illustrate the relationship between leader survival and mode of dissent, or lack thereof, in Figure 3. Specifically, the plot shows the respective cumulative probabilities of surviving for the first four years in office (1,460 days). It is easy to see that the surviving rate associated with nonviolent campaigns is consistently lower than the one for violent conflict, which in turn is lower than the surviving rate for the no active dissent scenario.
Notably, we do not assume these consequences follow from the choice of tactic alone, which is itself a manifestation of underlying campaign characteristics. We argue that nonviolent mobilization is a bigger threat to leaders because popular dissent imposes governance costs and creates opportunities for elite opportunism. However, the primary tactic indicator we employ is not refined enough to disentangle the different pathways. But we specifically control for the magnitude of individual participation, which is a distinct campaign characteristic. Specifically, Model Ib examines the size of dissident  participation in either nonviolent or violent campaigns. The positive coefficient indicates that the likelihood of exit increases with higher participation. 10 Model Ic distinguishes between participation in nonviolent and violent campaigns. Higher participation increases the risk of exit for both types of campaign, but the increase is proportionately larger for nonviolent dissent. This is consistent with our Campaign Size (2) proposition, which posits that greater participation in nonviolent campaigns entails a bigger risk of loss of office for leaders than high participation in violent campaigns.
The risk of leader exit can obviously be influenced by a number of things other than dissent. To account for such factors, in Table II we add control variables to our initial models. We get similar results when adding control variables, and consistently find a higher coefficient on nonviolent dissent than on violent conflict, with a 2:1 ratio between the respective coefficients (see Model  IIa and IIc). Therefore, the previous results cannot be dismissed as arising due to these other factors influencing tenure and the risk of dissent. 11 In additional robustness tests we also find no evidence that these results are driven by the use of a dichotomous measurement for regime type (i.e. democracy and dictatorship), or by leaders in democracies. Specifically, re-estimating Model IIa using the full [À10, 10] Polity scale does not alter the main results (see Model A1b). Similarly, replicating Model IIa on a sample limited to non-democracies (i.e. Polity < 6) yields substantively similar findings, including the 2:1 coefficient ratio for the nonviolent versus violent campaigns (see Model A1c). Thus, our findings are not affected by including democracies where elections play a larger role in loss of office.
In terms of the control variables, we find a lower risk of exit for leaders that have entered in a regular manner, those in power during times of high growth rates, and rulers of wealthy countries generally. Evidence for this comes from the negative and statistically significant coefficient on Regular entry, Growth and ln(GDP/capita), respectively. We also find that older leaders and those who face term limits are more likely to see exit. Furthermore, the positive and statically significant coefficient on CSOs protection indicates a higher hazard rate when civil society can organize and mobilize without hindrance, which is particularly relevant for nonviolent campaigns. The coefficients on the remaining controls do not reach the conventional levels of statistical significance.
In sum, we find strong evidence that nonviolent movements are more destabilizing to leader tenure. Yet, we may still undercount the effectiveness of nonviolent mobilization since we examine only short-term effects. Research suggests that the threat to leaders from violent conflict actually decreases over time, and leaders are relatively more secure if they manage to stay in power after the initial outbreak (Debs & Goemans, 2010;Gleditsch & Ruggeri, 2010). Nonviolent mobilization is difficult to sustain over time as low barriers to entry also means low barriers to exit, and movements have less ability to control individuals than in violent conflict (Gates, 2002;Dahl et al., 2021). But political mobilization can have a more enduring impact in the aftermath of a nonviolent campaign, as it generates changes in the civil society and increases the incentives for elites to offer reforms (Dahl, 2016;Kim & Kroeger, 2019). According to della Porta (2014: 32), protest campaigns 'produce new relations and resources that favor mobilization' which do not totally die out when street unrest fades. Thus, there are reasons to expect that nonviolent dissent increases the risk of leader exit beyond the immediate campaign, due to its mobilizing effects on the civil society and increased incentives for elite opportunism. We see such differences in delayed effects of dissent tactics on leader survival as a natural consequence of movements' underlying characteristics. If leader exit would arise due to the specific tactic itself, then one should not expect any differences after the tactics are no longer in use (i.e. the campaign stops being active).
In Model Id we consider the effects of violent and nonviolent tactics over time, using a decay function where half time is set to 2. Substantively this means that, in the post-campaign period, the effect of dissent decreases to half every two years. Specifically, y¼ 2 À=a where is the time as a fraction of a year since the end of campaign, and a is the half time parameter (Burt, 2000). 12 The results indicate that violent campaigns have a negative long-term effect, meaning that leader exit becomes notably less likely in the aftermath of a campaign. Conversely, nonviolent campaigns have a positive long-term effect, as they continue to make exit more likely even in their aftermath. Stated differently, leaders become relatively more secure following a violent conflict they managed to survive. In contrast, the risk of loss of office remains higher even after an active nonviolent dissent fades. This evidence suggests that violent and nonviolent campaigns have different long-term effects, an aspect that future work should examine more systematically. Table III considers a number of robustness tests. First, the Cox model assumes that the hazard of the response (leader exit) is proportional in the sense that it does not depend on time at risk (leader tenure). One way to relax the proportional hazard assumption is to interact covariates with logged cumulative time (Allison, 2010;Keele, 2010). Model IIIa introduces interaction terms between campaign type and cumulative tenure. In non-linear models it is hard to assess conditional effects from the constitutive terms alone. We thus outline these quantities of interest in Figure 4a. Compared to the additive effects, these results suggest a slightly more complex relationship between the mode of dissent and the risk of losing office. Violent campaigns now have a higher risk of exit when tenure is low, but the risk declines more rapidly with time in office. Specifically, the estimated effect of violent campaigns becomes lower than the one for nonviolent dissent with merely 124 days in office, and essentially reaches zero when tenure is about four years. In contrast, the effect of nonviolent campaigns on the hazard rate is still positive at the four-year mark. Since most challenges (roughly 70%) occur to leaders who have been in office longer than 124 days, the risk from nonviolent dissent exceeds the one from violent campaigns for most observed cases.
Our results may also be influenced by imbalances or systematic differences in the distributions of covariates between observations with and without dissent. Ideally, we want to compare cases without active campaign with very similar counterparts that experience dissent. Yet, having balanced groups (i.e. groups with similar covariate distributions) is not guaranteed when using observational data. When there is little overlap between groups, we may be making inferences from implausible counterfactuals. To alleviate this concern, we preprocess the data through nearest neighbor matching to limit the comparison to observations with and without dissent that are as similar as possible. We match separately on nonviolent and violent dissent. Although there is some evidence of imbalances in the full data, reduced by matching, reestimating the model on the matched data yields results similar to those from the full data. Specifically, while both types of campaign increase the risk of exit, the coefficient on nonviolent campaigns is more than twice the size of the coefficient on violent conflict (see Model IIIb and IIIc). Hence, our findings are not sensitive to differences in the observed cases due to distinct covariate distributions.
Another possible concern is that the comparative advantage of nonviolent campaigns in unseating the ruler may be driven by strategic calculations. If nonviolent campaigns target leaders who are weak to a greater extent than violent challenges, then that could explain the higher success rate. Theoretically, this assumes that nonviolent and violent tactics are substitute strategies that can be selected at will, which is not necessarily the case.
For example, a minority ethnic group may lack the numbers for a viable nonviolent challenge, regardless of how weak or strong the leader is. However, we address this concern in supplementary empirical analyses where we use estimates of the probability of a leader being in office at the end of the year as a proxy for leader strength. The predicted probability is arguably an informative composite indicator since it aggregates the effect of multiple factors that influence a leader's job security.
The survival probabilities for all leaders were computed using the estimates from a logistic regression where the dependent variable equals 1 if the leader is still in power at the end of a given year. All variables from Model IIa are included in the analysis except, for theoretical reasons, the two campaign types. 13 The resultant probabilities range from 0.57 to 0.98. This means that a weak leader has close to a 50/50 chance of being ousted whereas a leader of a stable state with resilient institutions appears almost immune to outside challenges. Since these are computed and not observed values, we have to account for the uncertainty of our estimates. To do so, we randomly draw 100 values from the sampling distribution of each individual predicted probability (i.e. for each leader-year observation), where the sampling distribution's mean and standard deviation are equal to the probability estimate and its standard  error, respectively. The results in Models IIId and IIIe are the pooled coefficients from 100 Cox survival models, one for each set of predicted probability estimates. 14 Model IIId confirms that nonviolent dissent has a large comparative advantage in unseating the leader even when controlling for leader strength. Also, as expected, strong leaders are notably less likely to see exit. In Model IIIe we interact leader strength with campaign type and the results are illustrated in Figure 4b. More specifically, the graph shows the difference between the effect of nonviolent and violent campaigns, with positive values indicating that nonviolent dissent is relatively more effective. Notably, nonviolent campaigns are statistically more likely to unseat leaders who are strong by traditional measures of state capacity. In contrast, violent dissent appears more effective when leader strength is relatively low, but the estimate is not statistically significant (i.e. the 95% CI crosses the zero line).
These results also provide evidence that the effectiveness of nonviolent campaigns is not driven by strategic selection. Note the distribution of nonviolent and violent observations as outlined by the two rug plots in Figure 4b. The bulk of nonviolent and violent conflicts cover cases with similar leader strength, and the median values for the two campaign types are remarkably similar (i.e. 0.841 for nonviolent and 0.838 for violent campaigns). In other words, unlike what we would expect in case of strategic selection, the two sets of observations are not clustered at opposite ends of the strength scale. Thus, the finding that nonviolent campaigns are more detrimental to the leader's survival is robust and not driven by inferences from implausible counterfactuals. 15

Conclusions
We have argued that active dissent poses a clear threat to a leader's rule, and that the magnitude and mode of dissent matter. Against the conventional wisdom that violent challenges are the most severe, we posit that nonviolent claims are likely to have characteristics that make them a greater threat than violent challenges. We find considerable evidence for our arguments. All maximalist dissent can undermine leader tenure, but nonviolent dissent more so than violent dissent. Moreover, the effect of higher participation is larger for nonviolent than for violent challenges.
Whereas some studies on the effectiveness of nonviolence have emphasized the direct consequences of specific tactics (Chenoweth & Stephan, 2011), we argue that the observed differences between nonviolent and violent campaigns are likely to stem from the variation in the dissident groups' constituency and ability to mobilize. These characteristics, in fact, make a specific tactic more likely in the first place. Stated differently, groups that have prospects for mobilizing large numbers are more likely to choose nonviolent tactics, while resort to violence might be an adaptation to low mobilization potential. Since large and resourceful groups have a comparative advantage in coercion through non-compliance, they would likely be worse off by choosing violence. By contrast, small groups have poor prospects to exert influence through nonviolence, and may be better off mounting a violent dissent. Collective action requires motivation and resources, and it does not follow that all actors can oust the leader via any tactic. While violent conflict is more likely in weak states and rarely happens in more developed societies, our results indicate that even leaders in apparently high-capacity states are vulnerable to nonviolent dissent, provided sufficient motivation and successful collective action.
We conclude that dissent seems to clearly affect a leader's ability to retain political office. Moreover, large-scale nonviolent dissent is more likely to drive a leader over the edge, contrary to the folk wisdom that violent challenges entail a greater political impact or shock. There are a number of possible extensions to the research presented here. First, differentiating more between immediate and delayed effects could provide a more comprehensive assessment of the overall impact of violent and nonviolent tactics. Second, it would be useful to examine in more detail the factors that encourage dissent and mobilize individual participation, and to explore how these could be related to events that reveal possible vulnerabilities for a regime. Finally, it would be helpful to examine the impact of dissent through governance costs more directly, for example by looking at the economic impact of dissent actions and the process of defection. These extensions could help validate the causal mechanism we propose, and help us better understand the specific tactical choice and level of mobilization. This could also elucidate the countermeasures that governments may employ to stem dissent, as 14 Pooling is a technique to aggregate or combine estimates across multiple models, for example when estimating effects from multiple imputed datasets (King et al., 2001;Kropko et al., 2017). 15 We recognize that our robustness tests cannot conclusively evaluate the possible impact of unobserved factors. But in order for these to influence the comparison between violent and nonviolent dissent they would need to have different influence on each type of mobilization, and it is difficult to see how an unobserved factor that reflects leader weakness would mobilize only nonviolent protest but not violent dissent. well as how outsiders could influence the incentives of either the government or dissidents (or both) to promote a specific outcome, such as peaceful resolution.

Replication data
Replication data for the empirical analysis in this article, along with the Online appendix, can be found at http:// www.prio.org/jpr/datasets.