Divided We Fall: Ethnic Cleavages, Movement Cohesion, and the Risk of Escalation to Civil War in Non-Violent Uprisings

ABSTRACT Why do some non-violent uprisings escalate into armed violence while others do not? We suggest that horizontal polarisation contributes to the escalation of non-violent campaigns. We examine the effect of ethnic cleavages between the campaign and its opponent and movement cohesion as explanatory factors for escalation into civil war and non-state violence. Statistical analysis of all major non-violent campaigns (1970–2014) shows that non-violent conflicts with ethnic cleavages have a higher risk of escalating into armed violence in particular, when the conflict takes place over governmental aims. The results also indicate that movement cohesion alleviates the risk of armed escalation.


Introduction
Aside from the research demonstrating the power of non-violent resistance in challenging authoritarian regimes and bringing about major societal changes (e.g., Schock 2005, Chenoweth and Stephan 2011, Nepstad 2011, a growing line of research on non-violent uprisings explores their risks of escalation (Pearlman 2011, della Porta et al. 2017. 1 This line of research is motivated by the need for a deeper understanding of how movement characteristics shape the societal implications of non-violent resistance and the conditions under which non-violence can pave the way for armed escalation of societal conflict. 2 Cases where popular non-violent uprisings were followed by intergroup polarisation, various forms of foreign involvement and escalation of the conflicts into civil wars include Ukraine in 2014, Libya in 2011, and Myanmar in 2021.
Three sub-fields of empirical research have made particularly significant contributions in this regard. First, a number of studies have modelled the tactical choice of resistance (violent versus non-violent) as a non-sequential choice between two independent means of struggle (e.g., Chenoweth and Ulfelder 2016, Schaftenaar 2017, Karakaya 2018). Yet, such a comparative approach between non-violent and violent forms of conflict is not suited to study variations within non-violent conflicts and the sequential process of escalation. Second, there are studies that examine escalation as a sequential process, yet these tend to be limited in their scope: either they have explored the question only theoretically (Lichbach 1987); only regarding ethnic conflicts or self-determination movements (Gurr 2000, Öberg 2002, Germann and Sambanis 2021, Vogt et al. 2021; or only concerning governmental conflicts (Beardsley et al. 2015). Third, several studies have explored the question of escalation on the protest-level (Gustafson 2019, Ives and Lewis 2019, Ryckman 2019, rather than on overall campaign or country level. In this study, we set out to contribute to this growing research field and fill its research lacunas by exploring the conditions under which major maximalist non-violent campaigns risk catalysing large-scale armed violence, either in the form of state-based internal conflict or non-state armed violence. We argue that while non-violent uprisings invariably contribute to polarisationa process where individuals and actors in a society are increasingly forced to position themselves with one or the other side in a conflict -non-violent conflicts where mobilisation forms an ethnic cleavage between the regime and the campaign do not only polarise the relationship between the challenged regime and its citizens (vertical polarisation), but also catalyse polarisation between different social groups (horizontal polarisation). Whereas vertical polarisation is an essential component of non-violence, horizontal polarisation, we suggest, weakens non-violent campaigns and paves the way for a reconsideration of the tactical choice of contention in favour of violent means. Thus, collective identity-based cleavages in the campaign-regime relationship increase the risk of conflict escalation.
Specifically, we argue that the presence of ethnic identity cleavages induces major obstacles to the effectiveness of the non-violent campaign by decreasing the pool of potential participants, weakening the mechanism for gaining loyalty shifts, and lowering the prospects that government repression backfires. Moreover, polarisation of horizontal relations exacerbates strategic commitment problems between the opposition and the regime in power, as both parties fear for the future safety of one's (dominant) ethnic group. We expect this commitment problem to be particularly severe in nonviolent uprisings fought over governmental power rather than territorial autonomy. Moreover, we contend that movements can seek to mitigate horizontal polarisation effects by creating broad-based cooperative civil coalitions that transcend ethnic identity lines within the opposition.
The implication of our argument is that non-violent conflicts with an ethnic identity cleavage between the non-violent campaign and the regime have an increased probability of escalation into armed violence. The role of ethnicity, albeit predominant in the study of civil wars (e.g., Denny and Walter 2014), has only recently started gaining explicit focus in the study of the emergence and success of non-violent campaigns (Svensson and Lindgren 2011a, Thurber 2018, Hendrix and Salehyan 2019, Abbs 2020, Pischedda 2020. Our study contributes to this research frontier by broadening the focus from how societal ethnic cleavages facilitate or complicate non-violent struggle against the vertical opponent (the state) to how the presence of an ethnic cleavage makes non-violent conflicts catalysts to exacerbating existing intergroup relations within a society.
We test the argument through a statistical analysis of all major non-violent campaigns and their immediate aftermaths between 1970 and 2014, merging data on non-violent campaigns with data on ethnic identity cleavages in nonviolent conflicts, using and converging data from NAVCO, UCDP, and EPR. In line with our argument, we find that ethnic cleavages in non-violent conflicts increase the risk of escalation, particularly when ethnic identity cleavages in non-violent struggles are combined with demands for regime change. Moreover, movement cohesion -the degree of unity of the non-violent opposition -decreases the risk of armed escalation of non-violent uprisings.
It is important to study the escalation propensity of non-violent conflicts to get an empirically grounded understanding of the risks related to nonviolence. Recent research comparing non-violent and violent means of conflict has found that non-violent uprisings are substantially less likely to encounter mass killings (Perkoski and Chenoweth 2018). In our study, we find that armed escalation of non-violent conflicts rarely occurs (as the analysis below will clarify, between 4 and 9 per cent of the campaign years, depending on the type of violent aftermath examined). This aligns with the notion that transformations from unarmed into armed conflicts are generally rare (Chenoweth 2021, p. 151). Bearing in mind this generally low risk of escalation, particular research and policy focus should be given to those nonviolent conflicts that portray characteristics that make them vulnerable to escalation, including those where mobilisation reinforces ethnic identity cleavages.

Previous Research on Escalation in Nonviolent Uprisings
Why do some non-violent campaigns escalate and pave the way for violence, armed conflicts and civil wars, while others do not? Prominent explanations to armed violence would direct our attention to social movements' resources and the nature of their grievances (Gurr 1970) as well as the general challenge to overcome the information failures (Öberg 2002). More recently, scholars have answered the question from the perspective of the movement's characteristics as well as the role of the state opponent. For example, attention has been given to the implications of the level of movement organisation (Ives and Lewis 2019). Moreover, studies have examined the effects of internal fragmentation of the opposition and radicalisation of demands (Vogt et al. 2021), movements' organisational capacity for violence, especially the presence of radical flanks, as well as movements' lack of leverage (Ryckman 2019), and participating protesters' impatience (Gustafson 2019).
Other scholars have emphasised the role of the challenged regimes in particular, regime repression (Ives andLewis 2019, Geelmuyden Rød et al. 2021), state authority as well as coercive capability (Sullivan 2018), political system (Cunningham et al. 2017) and exclusion of groups from state power and withdrawal of autonomy and self-determination (Germann and Sambanis 2021). Still, others have examined the role of external actors and the international community in mitigating against escalation (Oberg et al. 2009, Beardsley et al. 2015.
These studies have contributed significantly to the cumulated knowledge of social movements and the prospects of mass anti-regime civil resistance campaigns, yet they have thus far paid less attention to the relational aspects in a non-violent conflict dyad. Non-violent uprisings do not only vary in terms of their internal characteristics or the characteristics of the state that they face, but they also vary in terms of the similarity versus difference in the composition of the two main parties forming the non-violent conflict: the non-violent campaign and the state regime. We argue that the structural characteristics of these relations, particularly the ethnic similarity versus cleavage between the movement and the regime, influence the dynamics of a non-violent conflict and the probability that it turns violent. Below we clarify our theoretical expectations.

Polarisation and Escalation
All non-violent uprisings seek to create societal polarisation to achieve sociopolitical change. Polarisation can be argued to be 'the prerequisite of collective actions because it creates the national unity, the attitudinal alignment and the emotional foundation which are required for the success of actions such as strikes, boycotts and general non-cooperation' (Boserup and Mack 1974, p. 33). Non-violent campaigns strive to increase overt manifestations of underlying structural conflicts that exist in a society to win mass support and gain loyalty shifts. Specifically, maximalist, non-violent uprisings seek to defy the 'vertical legitimacy' of a state, i.e., the rightfulness of the state institutions, the form of rule, and the subsequent capacity of the state to command loyalty from the masses. At the heart of vertical legitimacy lies the consent of the governed to what is perceived as rightful authority (Holsti 1996). Non-violent resistance aims at challenging this vertical legitimacy by aspiring to withdraw the consent of the key pillars of support underlying the current power holders (Sharp 1973).
While challenging the state's vertical legitimacy is part of all non-violent campaigns, non-violent uprisings can also invoke questions concerning the horizontal legitimacy of the state. Horizontal legitimacy concerns the identityquestion of the state: it is about the community/communities over which rule is exercised. Horizontal legitimacy is considered high when different communities tolerate and accept each other and the state as their representative. In contrast, when the state appears to reflect the aspirations of specific communities and to under-represent others, and when there is fundamental disagreement of the identity of the state (to whom it belongs), horizontal legitimacy is considered low (Holsti 1996). The underlying argument of this study is that non-violent conflicts in which mobilisation occurs along ethnic identity cleavages do not only challenge the state's form of rule but may also (not necessarily intentionally) induce collective insecurity concerns within ethnic communities concerning the future political power and safety. In other words, non-violent conflicts in which ethnic features serve as identitymarkers between the regime and the non-violent opposition invoke competition not only at the vertical level between the regime and the constituents but also at the horizontal level between society's different ethnic groups.
We adopt a broad understanding of ethnicity-formation, referring to collective identity groups primarily defined by descent-based attributes, such as language, religion, tribe, sect, race or region, or a combination of these, and focus on the implications that its politicisation has on conflict dynamics. While these identities are not static and their relevance to individuals and the wider society varies across time and space, they are found to be constrained in terms of change and more visible than other (often non-descent based) social identities, such as socioeconomic class (Horowitz 1985, Chandra 2012. The social nature of ethnic identities means that an individual does not have full control over her/his ethnic identity, but she/he is also constantly ascribed to specific groups within existing politically salient group boundaries. In line with Lake and Rothchild (1998) we assume that ethnic identities do not cause conflict -most often ethnic groups interact peacefully -but that the constrained and social nature of ethnic identities make tensions between ethnic groups particularly escalatory in times of insecurity concerning the future of specific groups.
Specifically, there are two interrelated mechanisms through which horizontal polarisation through ethnic identity cleavage in non-violent conflicts may increase the risk of large-scale violence. First, ethnic identity cleavage risks decreasing the effectiveness of non-violence by alienating the nonviolent campaign from the wider society and the key pillars of support of the regime. Second, heightened collective fears about the future of political power distribution among ethnic groups exacerbate commitment problems between the non-violent movement and the supporters of a challenged regime, which can be particularly acute in situations where non-violent uprisings are successful. Escalation, it should be noted, can be initiated through a shift in strategy by the movement, or by other actors taking up arms as a response to the non-violent opposition's increasing hold of power.

Ethnic Identity Cleavage and the Effectiveness of Nonviolent Uprisings
Withdrawing the vertical pillars of support (e.g., the crucial economic, political, and security institutions supporting the state) from the regime and increasing the general support for the opposition movement are at the heart of non-violent resistance (Sharp 1973). The presence of ethnic cleavage can decrease the campaign's ability to achieve loyalty shifts in the core state institutions, such as the security forces, particularly when these are deployed along the ethnic identity lines of the regime (Nepstad 2011(Nepstad , 2013. Inducing loyalty shifts in security forces is pivotal for a non-violent campaign as the state will find it increasingly difficult to continue opposing the demands of a non-violent campaign if its security forces disobey it and refuse to repress the opposition (Chenoweth andStephan 2011, Lee 2014). Yet, as Nepstad (Nepstad 2011, p. 129) observes, defections usually occur 'because troops shared a collective identity with civil resisters'. The presence of an ethnic division between the regime and the non-violent campaign can decrease the odds of such identification and subsequently shifting loyalties.
Furthermore, an ethnic identity cleavage can decrease the prospects of broader 'backfire' in the wake of violent repression. A condition that increases the chances of backfiring effects -transforming passive bystanders to active participants in mobilisation -is that these bystanders identify themselves with the side of the protesters as constituents facing an oppressive regime (Martin 2007, Sutton et al. 2014. The presence of an ethnic identity cleavage makes the situation more complicated for bystanders of different ethnic identities than the dominant campaign groups, as they will have to weigh between identifying with the vertical grievances against the state and their horizontal insecurity concerning the future inter-group power distribution.
The regime in power has incentives to emphasise the emerged ethnic cleavage and induce horizontal polarisation in society as a counter-measure against the non-violent uprising. By framing the civil resistance campaign as strife of a certain group, the regime can de-legitimise the demands of the resistance and draw a distinction between the campaign and the interests of the wider society. Moreover, by utilising ethnically divided repression and accommodation measures during the early stages of the resistance campaign the regime can consolidate intra-ethnic identification and solidarity within the regime's side, prevent defections from the regime's ethnic constituency and support base, and strengthen the distinct ethnic identity assigned to the non-violent movement. In short, the regime can use ethnic identity cleavage for its strategic benefit.
Subsequently, ethnic cleavage between the campaign and the regime can decrease active support towards the non-violent campaign by members of ethnic groups that are not among the dominant campaign groups. Increasing ethnic tensions are likely to discourage people who identify themselves with other than the predominant uprising group from actively participating in the protests. Once ethnic identities become entangled in a non-violent conflict and the conflict becomes horizontally polarised, further mobilisation and building of support networks will likely start following ethnic lines (see Cederman and Girardin 2007). Previous research has demonstrated that ethnic cleavage in a non-violent campaign lowers the likelihood of successful outcomes of a non-violent uprising (Svensson and Lindgren 2011a). This weakened efficacy is likely to frustrate dissidents and increase support for alternative dissident strategies (e.g., violence) and for tactical shifts in opposition strategy to military means. The perceived ineffectiveness of non-violent means is a vital reason for why opposition groups sometimes turn to armed resistance (Asal et al. 2013, Chenoweth and Lewis 2013, Cunningham 2013, Rørbaek 2016.

Ethnic Identity Cleavage and Commitment Problems
Aside from making non-violent resistance less effective, ethnic identity cleavage can create bargaining obstacles, particularly credibility problems for conflict parties. Ethnic cleavage in non-violent conflict reinforces the strategic obstacle of credible commitment in the bargaining process. Trustworthy commitments over time can be argued to become even harder to make when the substantive issues of incompatibilities are associated with intergroup power distribution. The regime side may promise reforms or a share of its hold on power but will have incentives to disengage from such promises once the pressure from the mass uprising has faded. The opposition, on the other hand, may promise to replace an authoritarian system with increased democracy, but will have incentives to secure the campaign's dominant ethnic group's power position as well. The possibility of any future political order and its benefits being defined along ethnic group lines deepens the problem of making credible commitments.
Ethnic identity cleavage makes bargaining for the civil resisters' particularly difficult as achieving their aims becomes associated with benefits for specific ethnic groups -even when the underlying grievances and aims of the movement would cut across ethnic groups in the society (McLauchlin 2018).
As ethnic identities are difficult to change and instrumental in fostering patronage networks (Chandra 2012:55), a (hypothetical) future political order with a specific group in power leaves little room of manoeuvre to people ascribed to other ethnic groups. Even if both sides benefit from a settlement short of armed conflict, the credibility problems in making long-term commitments will increase ethnic tensions and the risk of armed escalation.
Thus, even when aiming to reach out to the wide segments of the society and initially succeeding in diminishing the popular support of the regime (by appealing to shared vertical grievances and needs across ethnic groups), ethnic cleavage easily becomes an overriding element that locks the fate of one's future with that of one's ethnic group (see Horowitz 1985). Ethnic identities become the dominant cleavage along which decisions over allegiance and participation are made. For people who do not identify themselves with the dominating ethnic identity in the uprising, the campaign can seem alarming and risky even if they support the cause of the uprising. They may be incentivised to take up arms, leading to increased risk of conflict escalation into civil war.
The uprising in Syria in 2011 illustrates these processes. The distinct ethnic identity in the non-violent campaign -even when not deliberate -gave the regime a tool in its manipulation and repression tactics and induced intercommunal tensions between the collective identity groups prevalent in the conflict (Hilu Pinto 2017). Ethnicisation increasingly alienated the uprising from other ethnic groups and favoured radicalised voices on each side (Phillips 2015). The regime carried the main responsibility for fuelling sectarian tensions by portraying the non-violent uprising as a Sunni rebellion, repressing it through ethnic militias and releasing Sunni militants from prisons (Hilu Pinto 2017:136). However, structural aspects in the non-violent campaign itself facilitated this process: the campaign's predominately Sunni Muslim identity (different from that of the regime) mobilisation contributed to collective fears concerning the future power distribution across ethnic groups, which were present from the start of the protests. While the sectarian and tribal networks helped mobilise the non-violent uprising against the Baathist party, they also contributed to horizontal polarisation and escalation of the conflict into an armed rebellion (Leenders 2014).
A counterexample comes from Sudan's non-violent resistance movement of 2018-2019(/21), which catalysed the toppling of the country's long-term authoritarian leader Omar al-Bashir in the spring of 2019 and forced the Transitional Military Council to share power with civilian authorities (and that continues its struggle until today) (Zunes 2021). Alike Syria, Sudan is an ethnically heterogenous country that is no stranger to ethnicization of politics and political violence. In a similar fashion to the Syrian regime, al-Bashir tried to portray the non-violent resistance as a fringe movement of the marginalised (ethnic) segments of the society (Bishai and Elshami 2019). However, in contrast to the developments in Syria, these manipulation efforts were unsuccessful and the non-violent resistance in Sudan came to be (and to be seen as) a heterogenous and predominantly multi-ethnic challenge against the regime. Whilst the movement-level organisation and cohesion played a role in this (as will be discussed below), the dominant make-up of the resistance also protected it: the movement was from the outset predominantly ethnically heterogenous, drawing support from both Arab and African communities and encouraging wide segments of the society and its multiple ethnic groups across the different regions within the country to join (Zunes 2021).
Based on the theoretical arguments developed above, we derive three testable hypotheses. The first hypothesis concerns the general effects of ethnic identity cleavage on escalation:

Hypothesis 1: Non-violent conflicts with an ethnic identity cleavage between the regime and the opposition have a higher risk of escalation to armed violence than non-violent conflicts without an ethnic cleavage.
We further expect ethnic identity cleavage to be particularly polarising in campaigns aimed at challenging and overthrowing the regime, in comparison to territorially defined non-violent campaigns. Although territorial conflicts are basically invariably ethnically defined, some non-violent conflicts with ethnic cleavage are fought over the control over government. In fact, previous research shows that almost half of all ethnic armed conflicts in the post-World War II-period have been fought over the government control (Buhaug 2006:694). Previous research has shown that non-violent campaigns over territorial challenges are generally less effective than campaigns for regime change (Chenoweth andStephan 2011, Svensson andLindgren 2011a) and that territorial self-determination campaigns may choose other tactical choices to pursue their goals (Cunningham 2013).
We expect the second tier of our argument to be particularly acute in governmental conflicts. When a non-violent campaign aims at achieving autonomy or independence from the current state, the regime's continuity as the central authority of the state is questioned in somewhat (territorially) constrained way. In contrast, in governmental non-violent conflicts with an ethnic identity cleavage any result in the bargaining process will not only change the nature of the state rule but also the central political power relations between horizontal societal groups. Even when the aims of the uprising would emphasise the creation of a more democratic and inclusive government, the presence of an ethnic cleavage raises the issue of to which group the state belongs and what happens to the political power distribution among ethnic groups in the aftermath of the uprising.
The prospects of any change in the status quo of ethnic power relationships will create commitment problems for all parties to the conflict and incentivise the challenged ethnic groups to maintain intra-ethnic solidarity and consolidate their access to power-structures even by military means. The polarising effect of governmental non-violent conflicts with ethnic identity cleavage goes beyond the main ethnic identities involved and will likely produce fears among third party ethnic groups. Thus, we derive a second hypothesis, in which we expect the polarising and escalating effect of ethnic cleavage to be particularly evident in non-violent conflicts concerning regime power: Hypothesis 2: Non-violent conflicts with an ethnic identity cleavage and which are fought over governmental, rather than territorial, aims have a higher risk of escalation to armed violence than other governmentally aimed nonviolent conflicts.
Finally, opposition movements can try to strategically adapt to the structural context in which the conflict takes place. While the literature on ethnic cleavages and ethnic power relations has shown that ethnic exclusion increases the risk of civil war (e.g., Cederman et al. 2013), it has not focused specifically on the movement-based strategies that can be used by nonviolent movements to build bridges across divides. Previous research on civil resistance, on the other hand, has pointed to the importance of creating broad civic coalitions that draw from different segments of the society as a condition for effective non-violent action (Ackerman and Karatnycky 2005). The level of movement cohesion among the groups involved in the nonviolent struggle can be expected to condition the effect of ethnic cleavage between the regime and the movement. Creating social ties across societal divisions and enhancing broad-based movement cohesion is an essential element of successful civil resistance (Thurber 2018(Thurber , 2019. Pearlman (2011) shows that movement cohesion is one of the most important pre-conditions for maintaining the non-violent character of an insurgency over time. For example, in the recent case of Sudan, a high level of movement cohesionbrought by clear coordination and organisation -further safeguarded the movement from the efforts of the regime to portray the movement as that of instigated by marginalised ethnic African groups (Zunes 2021).
The level of movement cohesion can be expected to play a particularly important role in non-violent struggles involving an ethnic identity cleavage. Movement cohesion creates possibilities for coordination of action and communication, which helps to maintain focus on the vertical polarisation and undermine the horizontal fears. It binds different groups involved in the campaign closely together and can prevent a specific ethnic identity from becoming the front of the campaign. Hence, movement cohesion can serve as an antidote against the government's efforts to undermine trans-ethnic appeal of the opposition movement, even if such trans-ethnic appeals can become non-credible once the conflict has turned violent (McLauchlin 2018). In turn, the lack of movement cohesion risks to lead to outbidding between factions, creating space for the most radical ethnic voices and enabling undue external interference in the struggle. From this reasoning, we derive our third and last hypothesis about the mitigating effect of movement cohesion.
Hypothesis 3: Non-violent conflicts with an ethnic cleavage and lack of movement cohesion have a higher risk of escalation to armed violence than non-violent conflicts with an ethnic cleavage and cohesive non-violent movement.

Dependent Variable
To examine the violent escalation of non-violent uprisings, we need data on major non-violent campaigns and their potential escalation into armed violence. The data on the former is based on the Non-violent and Violent Campaigns and Outcomes (NAVCO) dataset, version 2 (Chenoweth and Lewis 2013), which includes all major armed and unarmed campaigns taken place in 1960-2006. As the focus here is on non-violent insurrections, we have only included campaigns that have started out as non-violent (primary method being non-violent in the first year of the campaign). We have extended the dataset to cover all major non-violent campaigns from 2006 to 2014 to include the most recent cases of major non-violent uprisings. The additional data have been collected using data by Sutton et al. (2014), which includes non-violent uprisings against authoritarian regimes until 2013, and through our own data collection efforts to cover the remaining major nonviolent campaigns during the period 2006-2014. 3 As a result, we have a dataset of 322 non-violent conflict years.
With regard to armed escalation, two sets of variables are constructed. Violent escalation of non-violent conflicts is ffirst understood as a direct escalation of an unarmed conflict into armed conflict that takes place between the government and at least one of the groups involved in the non-violent campaign. Armed conflict is measured using data from the Uppsala Conflict Data Program's (UCDP) monadic dataset, which covers all armed conflict onsets -defined as stated incompatibility between a government and a non-state actor that result in a minimum of 25 battle-related deaths in a calendar year (Gleditsch et al. 2002, Allansson et al. 2017. The variable takes the value 1 if the non-violent conflict has escalated into armedconflict during a non-violent campaign-year and 0 otherwise. This variable measures the immediate escalation of non-violent conflict into an internal armed conflict. Moreover, for a non-violent insurrection to be coded as escalated armed conflict, two conditions need to be fulfiled (besides the condition of 25 battle-related deaths): 1) the armed conflict has to be about the same incompatibility as the unarmed insurrection and 2) the non-violent conflict cannot at the time of its inclusion in the dataset already have escalated into armed violence according to the first-mentioned criteria. 4 It is important to note that we do not require consistency in terms of actors: escalation may or may not occur by the same actors that took to the streets. Based on these criteria, there are 14 escalated non-violent insurrection-years in the 322 non-violent campaign-years. Clearly, escalation in armed conflict is rare, as this occurs in approximately 4 per cent of the insurrection-years studied (Syria, Ukraine, Iraq, Libya, Kosovo Albanians, East Timor, West Papua, El Salvador, Panama, Romania, and Algeria).
Furthermore, non-violent conflicts may escalate between the main antagonist groups in a non-violent struggle (opposition versus the regime), but they may also lead to the souring of horizontal group-relationships beyond the strict government-opposition relationship. Therefore, aside from the variable measuring immediate intrastate conflict escalation of non-violent conflicts, a more flexible definition of violent escalation is adopted. This operationalisation includes all cases of immediate escalation operationalised above but it also includes cases of communal conflicts (nonstate conflicts) that take place between at least one group represented in the non-violent campaign and some other non-state communal group during the specific non-violent campaign-year or no later than 1 year after the last recoded non-violent campaign-year. 5 The data for the communal conflict escalation is taken from the UCDP's 'Non-state conflict dataset' that defines non-state conflict as armed violence between two informally organised nonstate groups that are constructed along shared identities (Sundberg et al. 2012). Again, for a conflict to be included in the UCDP data, it needs to exceed the threshold of 25 battle-related deaths in one calendar year.
For a non-violent campaign-year to be coded as having escalated at the horizontal level, the non-state conflict needs to be at least partially motivated by the issue of incompatibility driving the non-violent campaing, in addition of the conflict involving at least one of the groups also present in the nonviolent campaign. In Kyrgyzstan, for example, ethnic tensions between Kyrgyz and Uzbek in the south of the country escalated in the immediate aftermath of the 2010 non-violent uprising that ousted president Bakiyev and consequently changed the political power dynamics in the country.
Hence, the outcome that we label as 'general violent escalation' takes the value 1 if the non-violent conflict has escalated into armed intrastate violence within a specific campaign-year or there has been a violent escalation of communal relations (involving groups represented in the non-violent conflict) during the ongoing non-violent conflict-year or -in the case of last coded non-violent year -no later than 1 year after the last campaign-year. As the UCDP non-state data only cover years from 1989 onwards, the general violent escalation variable is only tested upon a subset of all non-violent conflict-years in the post-1988 period. Based on the broad definition of general violent escalation there are 21 escalated unarmed conflict-years out of which 10 include escalation of communal violence. Again, as this clearly shows, escalation is rare, manifested in the fact that only 4 per cent of nonviolent uprising years have been followed by escalation of communal relations.

Explanatory Variables
Our main explanatory variable, ethnic identity cleavage, is a binary variable that receives the value 1 if the non-violent campaign is separable from the regime on the basis of ethnic identity-markers, such as language, religion, race, region, sect or tribe in a campaign-year. 6 Information on the ethnic structure of the non-violent campaign has been gathered through encyclopaedic entries, newspaper articles and reports using the Factiva search engine. Similar to other studies on the ethnic structure of non-violent campaigns Lindgren 2011b, Thurber 2018), the variable captures and relies upon public information on the structure of the non-violent campaign in relation to that of the regime, e.g., news and other reporting on the campaign structure. Ethnic identity cleavage is coded to be present if the main groups participating in the non-violent campaign represent different ethnic groups than the dominant regime groups, based on the Ethnic Power Relations (EPR) dataset (Cederman et al. 2010). Specifically, if the ethnic groups coded as senior, dominant, or monopoly groups in the regime are not among the dominant non-violent campaign groups, an ethnic identity cleavage is coded to be present. The dataset overlaps with Thurber's (2018) data on the ethnic structure of non-violent campaigns for the period 1970-2006. Albeit imprecise with regard to the exact number of protesters identifying themselves with particular ethnic groups, the variable measures whether an ethnic cleavage has been manifested in a non-violent conflict during a specific conflict-year. This variable does not measure ethnic mobilisation per se, i.e., the campaign does not need to mobilise on the basis of ethnic grievances and ethnic group symbols, rhetoric, demands, etc. Out of the 322 campaign years, 108 are coded to include ethnic identity cleavage.
It is important to point out that this operationalisation differs from that of the NAVCO data project's ethnic diversity variable, which measures whether the campaign 'embraces ethnic diversity'. A campaign can embrace ethnic diversity as an aspiration and, yet, the conflict may nevertheless manifest an ethnic identity cleavage. Take the non-violent campaign in Syria 2011 for example. The campaign initially sought to have a trans-sectarian appeal and not to be seen as a Sunni-based uprising. However, a majority of the protesters were nevertheless Sunnis and the non-violent campaign never overcame the ethnic identity cleavage between the uprising and the regime with its basis in the Alawite-community. Hence, the coding is based on the presence of ethnic cleavage in the structure of the non-violent conflict dyad, rather than in the stated (ethnic) aims of the non-violent campaign.
The second part of the analysis furthermore considers the impact of different types of campaign goals. As argued in the theoretical section, campaigns that have governmental goals can be argued to be more vulnerable towards the escalating effects of ethnic identity cleavage, as the issue of incompatibility is directly linked to access to governmental power. The variable goals -coded on the basis of the NAVCO 2.0 campaign goals variabletakes the value 1 if the campaign aims at a territorial aim (independence, autonomy) and 0 if the campaign aims at regime change or major policy changes at the regime level. 7 There are 234 governmentally aimed campaignyears against 91 territorially aimed campaign-years in the dataset. Twentyfour campaign-years are governmentally aimed and have a salient ethnic cleavage, whilst there are 83 territorially aimed years with ethnic cleavage. 8 Finally, in order to test hypothesis 3, we include movement cohesion into the third part of the analysis. The variable comes from the NAVCO 2.0 data and measures the level of competition between the different groups participating in the campaign, ranging from 0 (seemingly united campaigns) to 3 (violent competition among the groups participating in the campaign). We interpret higher levels of competition to proxy lack of movement cohesion. 9 Most campaign-years can be defined as cohesive (n = 119), whilst 45 campaign-years have experienced considerable competition among the participating groups (either active non-violent or violent competition). There are 18 campaign-years with both ethnic cleavage and lack of movement cohesion present (with intra-movement competition being active) and 60 campaignyears with ethnic cleavage present and movement cohesion (seemingly united campaigns and cooperative campaigns with moderate disunity).
In addition to our main explanatory variables of interests, we include several control variables that can be theoretically expected to influence the relationship between ethnic identity cleavage and the onset of armed escalation of a non-violent conflict. First, we consider the size and the duration of the campaign. The size of the campaign is one of the conditions found to influence the outcomes of non-violent campaigns. It can also be theorised to relate to the presence of our main explanatory variable, as more popular campaigns might mitigate the visibility of ethnic identity cleavages. The variable size originates from the NAVCO 2.0 data, where campaigns are divided into 6 ordinal categories based on their size. 10 The duration of the campaign (in years) can also be argued to influence the outcome variable in various alternative manners. On the one hand, longer lasting uprisings can become more consolidated in their type of resistance and thus prevent armed escalation of the situation. On the other hand, long-lasting campaigns might also lead to frustrations in the campaigners, contribute to more pronounced ethnic cleavage, and give room for voices inducing a violent turn. Moreover, to control the potential differences between the conditions of nonviolent uprisings that took place during the Cold War period and the more recent non-violent uprisings we include a binary variable that receives the value 1 in the case of Cold War-period years and 0 otherwise. Furthermore, we control for the level of repression during a non-violent campaign. Violent repression of a non-violent campaign can backfire on the government and increase support for the campaign, but it can also escalate tensions and justify radicalisation of the campaign. We use NAVCO 2.0 to capture the repression a campaign-year faces. 11 We also control for the general ethnic power relations at the country-level. Existing inter-ethnic relations are important to consider as they can influence the emergence of non-violent conflicts with an ethnic identity cleavage and the risk of armed escalation of non-violent campaigns in general. The variable ethnic exclusion comes from the Ethnic Power Relations (EPR) data and captures the share of the population excluded from political power based on their ethnicity (Cederman et al. 2010). Aside controlling for the political discrimination of ethnic groups, we take into account the countries' ethnic demographics by including the ethnic fractionalisation index and its squared term (Fearon and Laitin 2003). Finally, we consider time dependencies in terms of armed conflict risk by including a measure of 'peace years' as well as its polynomials, following Beck et al. (1998). Table 1 presents a covariate tabular comparison of armed escalation of nonviolent campaign-years with an ethnic identity cleavage and non-violent campaign-years without this characteristic. 12 Two general notions arise. First, armed escalation of non-violent conflicts is rare. Only 14 (4 per cent) out of 322 campaign-years have experienced an escalation into intrastate conflict between the regime and an armed movement. When expanding the outcome variable to include violent escalation between non-state groups, there are 21 (≈9 per cent) escalated campaign-years in 229 campaign-years. 13 Second, non-violent campaigns with an ethnic identity cleavage have escalated into armed conflicts considerably more often than non-violent campaign-years without ethnic cleavages. Around 8 per cent of the ethnically mobilised non-violent campaign-years have seen armed escalation in comparison to around 2 per cent of the non-violent conflicts without an ethnic component. 14 To move beyond bivariate tabulations, Table 2 presents four multivariate logit-regression models of armed escalation of non-violent conflicts. We estimate armed escalation using the logit model as the dependent variable is a binary choice variable. We also report skewed dependent variable and rare-events estimation models in the Appendix due to the rarity of escalation events in the data. 15 Model 1 and 2 examine the years 1970-2014 and treat armed escalation as an escalation of intrastate armed escalation. Model 3 and Standard errors, clustered by country, in parentheses *p < 0.10, **p < 0.05, ***p < 0.01. 4 restrict the timescale, due to data-limitations, to post-1988 years and treat escalation as both intrastate and communal armed escalation. Model 1 and 2 support our first theoretical proposition (Hypothesis 1). Ethnic identity cleavage has a positive and statistically significant coefficient (on the 0.05 level) regarding the likelihood of intrastate armed escalation. This is the case for a parsimonious model including only the main independent variable and the variable measuring the intensity of repression as well as for the model including the other control variables. Based on Model 2, the risk of armed escalation increases from 0.004 to 0.026 when there is an ethnic identity cleavage between the regime and the non-violent campaign. Figure 1 illustrates the substantive effects of the main explanatory variables on the probability of armed escalation. 16 Model 3 and 4, which include armed escalation both at the intrastate level and between communities, show a similar positive and significant coefficient for the main independent variable. However, the estimated substantive effect of ethnic identity cleavage is considerably larger in these models. A nonviolent conflict with an ethnic identity cleavage is estimated to have 14 per cent risk of escalating into political violence, in comparison to the baseline of around 3 per cent. The difference in the substantive effects is mostly driven by the model components, such as increased variation in the dependent variable that is now operationalised in this broader sense. Yet, one can also argue that operationalising the dependent variable both as state-based and non-state violence can better capture the full range of observable implications of the theoretical mechanism. Horizontal polarisation and rising inter- group tensions in non-violent conflicts can give rise to violence that may or may not directly involve the state.

Non-violent Campaigns with Ethnic Identity Cleavages
In terms of the control variables, violent state repression of non-violent conflicts correlates positively with armed escalation of conflicts, pointing to the importance of the regime's actions in escalating a non-violent conflict. However, it should be noted that the coefficient of regime repression is not statistically significant at the 95 per cent confidence level in Model 2 and its substantive effects are not clearly separable from zero. Furthermore, longer campaigns appear to escalate more rarely than shorter ones, yet this association is only significant when restricting the focus to more contemporary campaigns and including both intrastate and communal escalations. Notably, ethnic demographics and ethnic exclusion do not seem to explain the risk of escalation of large-scale non-violent campaigns. The size of the campaign does not have any clear influence on the risk of armed escalation.

Non-violent Campaigns Over Government Power
To test the second hypothesis concerning the particular risk of situations where a non-violent conflict with an ethnic identity cleavage coincides with incompatibilities over the nature of the regime, the conflict issue will be included in the analysis. This is done in three ways portrayed in Table 3: model 5 includes a binary variable of governmental versus territorial aims in the model on escalation, model 6 examines intrastate escalation of a subset of governmentally aimed campaigns, and model 7 examines all escalations (including communal escalation) in the subset of governmentally aimed campaign.
In model 5, the analysis correlates a binary variable of governmental versus territorial aims (Territorial aims = 1) with the risk of escalation. Considering the campaigns' aims improves the fit of the intrastate conflict escalation model and shows that non-violent campaigns with governmental aims are significantly more likely to escalate than territorial conflicts. Crucially, controlling for the type of aims also increases the substantive effect of ethnic identity cleavage, implying a connection between the two explanatory variables. Figure 2 illustrates the estimated conditional marginal effects of ethnic identity cleavage in two scenarios based on model 5: non-violent conflicts concerning governmental power and non-violent conflicts with territorial claims. Notably, while the estimated marginal effects of ethnic cleavages with governmental claims are higher than ethnic cleavages with territorial aims, the overlapping confidence intervals impede rejecting the null hypothesis on the difference in the effects.
To investigate the matter further, Model 6 and 7 restrict the focus to governmentally aimed non-violent campaigns, hence excluding all territorial non-violent conflicts from the analysis. 17 This sub-setting is done deliberately to exclude those campaigns where ethnic identity is almost inseparable from the aims of the campaign (autonomy, independence for a specific ethnic group) and to focus solely on non-violent campaigns aiming at changing the political regime of the country. As the theoretical framework outlined, it is in these conflicts where the ethnic identity cleavage is expected to be particularly polarising.
Both model 6 and 7 allude to a significantly heightened risk of violent escalation of non-violent governmental conflicts when ethnic identity cleavage is present. Examining Model 7 with logit-regression of governmental conflicts' violent escalation, the likelihood of conflict escalation is 0.08 when all the explanatory variables are kept mat their median values. However, with ethnic identity cleavage (and maintaining everything else in their median values) the likelihood of conflict increases to approximately 0.40. Thus, nonviolent conflicts with governmental aims and an ethnic identity cleavage appear to be particularly vulnerable towards escalation.
While the results imply that the combination of governmental non-violent conflict and the presence of an ethnic identity cleavage is particularly vulnerable towards armed escalation, the findings should be approached with Standard errors, clustered by country, in parentheses *p < 0.10, **p < 0.05, ***p < 0.01.
caution. The number of observations in this subset is small in general, and the occurrence of non-violent conflicts with ethnic identity cleavage is particularly low in governmental conflicts. In fact, there are only 24 observations of non-violent campaigns with ethnic identity cleavage in the subset of governmentally aimed non-violent campaign-years. Out of these 24, seven have experienced violent escalation of either inter-group or intrastate conflict. 18 Crucially, however, the proportion of conflicts escalated in this sub-group is considerably higher than in the remaining governmentally aimed non-violent conflicts without an ethnic cleavage. Hence, while governmental non-violent campaigns with clear ethnic cleavage are themselves rare, the results suggest that their presence increases vulnerability to armed escalation.

Non-violent Campaigns and Movement Cohesion
Finally, our theoretical argument suggests that ethnic identity cleavage might be particularly conflict-inducing in campaigns that are less cohesive in their organisation. In contrast, activists may offset the potential structural risks of ethnic divisions by creating cohesive, broad-based coalitions that bridge and unite ethnic cleavages. To test this proposition and the robustness of our explanatory variable, we include a measure of the movement cohesion of non-violent campaigns. The descriptive statistics and regression models are reported in the Appendix. In line with our theoretical expectations, the interaction models imply an interdependent and escalation-inducing effect of non-violent conflicts with an ethnic identity cleavage and a campaign structure that lacks cohesion. Descriptively, 10 out of the 18 campaign-years with ethnic cleavage and lack of movement cohesion have experienced violent escalation, whereas only five out of the 60 campaign-years with ethnic cleavage and movement cohesion have escalated. 19 In multivariate estimation models, the interaction terms are statistically significant and positive and indicate a substantial increase in the likelihood of armed escalation when the campaign is both ethnically mobilised and lacks movement cohesion. Notably, including the interaction terms erases the significance of the coefficients for both ethnic cleavage and movement cohesion versus fragmentation. Figure 3 demonstrates this conditional effect of non-violent conflicts with ethnic identity cleavage and increased internal competition within the campaign. Given the small number of events containing both conflict escalation and the two predictors (ethnic cleavage and lack of movement cohesion), the findings here ought to be read as suggestive and highly uncertain, pointing to the need to further study the alleviating effects of movement cohesion in contexts where the opposition movement's front differs from that of the regime in regard to ethnic identity.

Additional Tests
To account for the rareness of the escalation cases, we re-run the key models using model estimation techniques better adjusted for the skewed distribution of the dependent variable. These are reported in Appendix A.3.2 and A.3.4. The results remain unaltered by these model specifications. Furthermore, we include several additional control variables that capture the states' capacities to quell a rebellion. Specifically, Table A.3.4 in the Appendix reports models that capture variations in regime type, proximity to regime instability and GDP per capita.

Conclusions
This study has aimed at advancing the literature on non-violent resistance by focusing on the risk of escalation of large-scale non-violent conflicts into intrastate armed conflicts or violence between non-state groups. The empirical analysis overall supports our theoretical framework and suggests that mass non-violent campaigns that differ from the regime in their ethnic group build-up have a higher risk of paving the way to armed escalation than those without such ethnic cleavage. Nevertheless, we want to underline that the relatively small sample used in this study and particularly the small absolute number of events with escalation pose methodological challenges and decrease the generalisability of the results. Still, a few key insights and their implications should be underlined.
We have put forward an argument explaining the risk of violent escalation based on the social mobilisation structure of non-violent conflicts. While we have focused on ethnic cleavages in this study, the implications of the escalatory potential of horizontal polarisation travel further. Whether other types of existing societal cleavages that become activated in a non-violent struggle have similar polarising and escalating effects is an interesting question for future research. Contemporary examples from Venezuela (2017), Thailand (2013/14) and the Philippines (2001) would seem to indicate that when a non-violent uprising becomes mobilised disproportionately around a certain political block or socioeconomic class it becomes more vulnerable towards the opponent's delegitimizing efforts and risks to polarise rather than unite societal groups.
Our study also suggests a sub-category of high-risk cases -non-violent opposition movements that are governmental in their aims and ethnic in their structure of mobilisation -that pose significant risks for escalation. In fact, as many as half of those cases in our sample have escalated into violence. More in-depth case study focus and micro-level empirical attention would be needed to study the mechanisms through which governmentally aimed non-violent uprisings with distinct ethnic-identity markers from that of the regime escalate horizontal fears and tensions.
Our study add to the discussion about that the use of state repression against non-violent campaigns can increase the risk of escalation into civil wars. We theorised that state repression is less likely to backfire against the regime when mobilisation follow ethnic fault lines. It should be said, however, that the relationship between repression and dissent is multifaceted and there is a possibility of reversed causality, in state repression against particular ethnic/societal groups can influence the onset and non-onset of certain types of non-violent challengers. We therefore encourage that future research should continue to disentangle the relationship between ethnically charged non-violent mobilisation, state repression, and conflict escalation.
Finally, the findings here call for more research and policy attention to how crosscutting bridges can be created among opposition groups sharing vertical grievances against a regime in ethnically divided societies in the wake of social change and political transformation. So far, insights from conflict resolution have largely escaped the attention of the study of strategic nonviolence. Yet, this is slowly starting to change (Dudouet 2017, Wanis-St John and Rosen 2017, Svensson and Lundgren 2018. While polarising vertical relations between the regime and its constituents is an essential part of non-violent uprisings, there is a need for a better understanding of conflict resolution and transformation needed to maintain intergroup peace during these upheavals.

Notes
1. Non-violent resistance can be understood as 'the application of unarmed civilian power using non-violent methods such as protests, strikes, boycotts, and demonstrations, without using or threatening physical harm against the opponent' (Chenoweth and Cunningham 2013:271). In this study, we use the terms 'non-violent resistance' and 'non-violent uprisings' interchangeably. 2. Non-violent uprisings are clearly less risky than violent uprisings. In fact, comparing violent and non-violent uprisings, Chenoweth and Stephan (2011) find that violent campaigns are more likely to be followed by post-conflict civil wars. The probability of post-conflict civil war is 42 per cent for violent insurgencies, against 28 per cent for civil resistance campaigns. 3. The data search has relied on the Global Nonviolent Action Database, Factiva search engine, and additional secondary literature. Only non-violent resistance that fills in the criteria set out in NAVCO 2.0 (maximalist goals and>1000 participants) have been included in the data. 4. If an unarmed insurrection indeed escalates while it is still active, succeeding years of non-violent action have been excluded. Furthermore, when armed conflicts are inactive or have temporarily ended and the unarmed insurrection continues, we have included these years into the dataset (as unarmed conflicts). East Timor is a case in point: the conflict escalates 2 times in our dataset: first in 1992 after the violent response of the police and armed forces to demonstrations, known as the Dili massacre, and a second time in 1997. 5. To clarify the coding scheme, with the onset and ongoing years of non-violent insurrection, the violent escalation of communal relations is coded on the basis of the specific campaign-year. However, to grasp potential lag of escalation in the aftermath of non-violent conflict activity, we also include communal escalation in one-year after the last non-violent conflict-year. If there has been an escalation of related communal conflict in this year, the last non-violent campaign-year is coded as 1. 6. The data on ethnic cleavages have been updated from Svensson and Lindgren (2011b). In the few cases where the coding differed from Thurber's, we have instead followed Thurber's coding. 7. For NAVCO 2.0, see Chenoweth and Lewis (2013). 8. See the Appendix for descriptive statistics tables concerning the main variables and their interactions. 9. As the movement cohesion variable is taken from the NAVCO 2.0 data, including it restricts the analysis to the years 1970-2006. 10. 0 = 1-999, 1 = 1000-9999, 2 = 10 000-99999, 3 = 100 000-499 999. 4 = 500 000-1 million, 5 => 1 million. 11. 0 = none, 1 = mild repression, 2 = moderate repression (physical, no apparent intention to kill, 3 = extreme repression (physical repression with the intent to silence and even kill the protesters). 12. The Appendix A.1 presents descriptive statistics for all the key variables. 13. See Appendix A.1.1 for a cross-tabulation of violent escalation and ethnic identity cleavage when escalation includes communal escalation. 14. Concerning the subset of cases after 1988, 14 out of 83 campaigns coded as having an ethnic identity cleavage have experienced escalation in comparison to 7 out of 146 non-ethnic cases. 15. Specifically, we use the scobit model to relax the assumption that the effects of the regressors are the strongest at 0.5 probability. This is because both the data structure -highly skewed dependent variable -and the theoretical considerations -the interplay between ethnic cleavage and repression, for examplesomewhat question the assumption in this case. Moreover, we run the models using the re-logit estimation method suggested to better suit data with a low absolute number of positive events (as in our case). See the Appendix for the model specifications. For discussion, see Nagler (1994) and King and Zeng (2001). 16. The predicted probabilities and first difference effects are estimated with the Clarify software program (see King et al. 2000). 17. Campaign size and repression variables have been left out of the model specifications that are restricted to governmental conflicts for high level of collinearity between the variables and governmentally aimed conflicts. Including the variables in the multivariate models impedes estimation of the parameter coefficients. 18. Countries with intrastate armed escalation in these contexts are Syria (2011), Iraq (2013) and Algeria (1992). South Africa (1990Africa ( -1992
*Regime type is captured by a binary variable measuring whether the regime can be characterised as inconsistent. Regimes that are neither consolidated democracies nor authoritarian systems -i.e., regimes that fall in the middle within democracy-autocracy scales -are generally seen as more vulnerable to instability and fluctuation than highly developed democracies or clearly autocratic states. The variable is constructed using the polity IV scale so that all regimes coded as −5-5 are coded as inconsistent whilst others are coded as consistent.