Shaping Electoral Outcomes: Intra- and Anti-systemic Violence in Indian Assembly Elections

Electoral violence is perpetrated by anti-systemic actors opposed to the democratic system, as well as by those vying for power through the electoral process. Even though the motivations for violent tactics are distinct, we do not know whether intra- and anti-systemic violence differ in their effects. Focusing on state-level elections in India – a country that combines nationwide elections with persistent political violence – we demonstrate that the distinction is crucial for understanding spatial patterns of electoral violence and effects on election outcomes. Based on an original dataset of violence in legislative assembly elections between 1985 and 2008, we show that both tactics depress turnout overall but that the effect is larger for anti-systemic violence. Intra-systemic violence not only appears to be more selectively targeted, as it is more likely to occur in constituencies where the incumbent belongs to the state-level opposition, but also generates electoral benefits for the party in control of state government.

systemic violence influence election outcomes. Evidence from India has played a crucial role in revealing the electoral motivations of political elites in instigating intra-systemic violence (Brass 2003;Varshney 2003;Wilkinson 2006). The involvement of politicians and elected officials in fanning communal violence in the run-up to elections, in unleashing violent goons on political opponents and in 'capturing' polling places on election day is well documented (see, for example, Berenschot 2011;Daxecker 2020). Elections in India have also occurred in the context of armed challenges to the existing political order. Separatist movements have turned violent in Punjab, Jammu and Kashmir, and Assam, while Maoist insurgents have operated across such states as Andhra Pradesh, Bihar, Chhattisgarh, Jharkhand and West Bengal (Staniland 2016). The specific strategies of insurgent movements towards the state vary across time and space, but such groups have regularly called for wholesale electoral boycotts in their areas of influence. To enforce boycotts, insurgents have assassinated poll workers and candidates, planted bombs at polling stations, and cut off voters' fingers marked by indelible ink (Kennedy 2014;Singh and Roy 2019: 338-40). A boycott call by an armed group thus unleashes a campaign of coercion and intimidation against voters and those involved in conducting the election in the affected area. In this article, we focus on such boycott calls by armed groups as instances of anti-systemic violence. While intrasystemic and anti-systemic violence disrupt the electoral process, the strategic motivations of actors engaged in these tactics shape who is targeted and how violence plays out. We analyse the effects of intra-and anti-systemic violence by focusing on three dependent variables: the location of violence; turnout; and incumbent vote share.
Our results demonstrate that intra-and anti-systemic violence influence the outcomes of elections in distinct ways. Both types of violence should be understood as disruptive and exclusionary tactics, as they often discourage participation in electoral contests and depress voter turnout overall. Yet, whereas anti-systemic violence aims to depress participation as much as possible in order to undermine the legitimacy of the election, intra-systemic violence seeks to suppress or drown out the voices of political opponents. We argue that this distinction is crucial for understanding the spatial patterns of electoral violence and its effect on election outcomes. Intra-systemic violence is shaped more directly by the institutional framework within which the election takes place, as its purpose is to make specific election outcomes more likely. In the context of India's first-past-the-post electoral system, this requires careful targeting across single-member electoral constituencies, as well as within them. Our analysis establishes party alignment between the statelevel government and the constituency-level incumbent as an important factor for understanding where intra-systemic violence takes place. Assembly constituencies are the units within which members of state legislative assemblies are elected, and constituency-level incumbents may or may not be co-partisans of the state government. Intra-systemic violence is less likely in aligned constituencies than in those where the local incumbent belongs to the opposition. Moreover, intra-systemic violence benefits state governments electorally. It decreases the vote share of constituency-level incumbents in non-aligned constituencies but does not negatively impact incumbent vote share in aligned constituencies. This pattern illustrates how the selective targeting of intra-systemic electoral violence benefits the party in control of the state government. Anti-systemic violence, by contrast, is not tied as closely to the institutional geography of elections, as boycotts aim to systematically disrupt elections across constituencies within the group's area of influence. Compared to intra-systemic violence, the effect of anti-systemic violence on turnout is substantively larger and less localized. Violence in the context of a boycott call not only depresses turnout in the constituency where it occurs, but also signals to voters nearby that insurgents are serious about enforcing the boycott. Anti-systemic violence is not systematically related to alignment, and the decrease in turnout does not systemically benefit state-or constituency-level incumbents. The two types of electoral violence thus follow distinct logics, which we need to take seriously to understand their effects.
To examine the impact of intra-and anti-systemic violence, we leverage an original dataset of electoral violence in legislative assembly elections across Indian states between 1985 and 2008.
Our data, which are based on news reports in the Times of India (ToI), are novel in two respects. First, we coded whether violence was intra-systemic or occurred in the context of a boycott call by armed groups, making it possible to distinguish more clearly between these two types of violence than in previous studies. Secondly, our data capture electoral violence in subnational elections. Existing geo-coded datasets of electoral violence, such as the National Elections across Democracy and Autocracy (NELDA) dataset (Hyde and Marinov 2015) or Electoral Contention and Violence (ECAV) dataset (Daxecker, Amicarelli and Jung 2019), are restricted to events related to national-level elections. Despite their substantive importanceespecially in large federal democracies like Indiawe thus know fairly little about violence in subnational elections.
Our subnational comparative approach allows us to analyse within-country variation while holding nationwide factors, such as the electoral system and the quality of election management, constant (Giraudy, Moncada and Snyder 2019). We focus on India for substantive and methodological reasons. Despite its strong democratic track record and professional election management, India has struggled to curb political violence. In cross-national analyses of electoral violence, India consistently emerges as having the highest number of violent events for a country not experiencing civil war (see, for example, Daxecker and Jung 2018). Moreover, as indicated earlier, there is variation in the types of electoral violence, and our dataset captures both intra-and antisystemic violent events. Further, India's federal system allows us to systematically study the effect of incumbency. Whereas previous work has focused on national-level incumbents in unitary systems (Hafner-Burton, Hyde and Jablonski 2014; Taylor, Pevehouse and Straus 2017), our subnational analysis leverages variation in incumbent governments across states to examine the role of alignment. In the following section, we discuss our concepts and theoretical expectations. We then briefly introduce our data before turning to the analysis. The final section discusses our findings and their implications for the study of electoral violence.

Concepts and Theory: Analysing Intra-and Anti-systemic Violence across Indian States
We conceptualize electoral violence as coercion directed towards actors and/or objects during the electoral cycle (Birch and Muchlinski 2020, 2;Höglund 2009, 415-16). Violence is part of a menu of electoral manipulation, which includes threats and coercion targeting voters, candidates and officials involved in the process (Fjelde 2020;Staniland 2014). Even though electoral violence often manifests in contexts in which other types of political or criminal violence are prevalent, it is distinct in the sense that it 'would not have occurred or would at least have manifested itself differently in the absence of an electoral contest' (Fjelde and Höglund 2016b, 8). Violence can take place in the run-up to elections during campaigns, on voting day and after the closure of polls when votes are counted and results announced. In this article, our empirical focus is on violence taking place before the polls close, as we are interested in effects on election outcomes.
Building on the literature that conceives of pre-election violence as strategic, we distinguish between intra-and anti-systemic violence (Capoccia 2002;Staniland 2014). In the former, an actor seeks 'to win or maintain power within the context of the democratic political system' (Staniland 2014, 108). The intention of violent electoral manipulation, in other words, is to tilt the playing field towards one particular candidate or party, while accepting elections as the appropriate means for determining who governs. Intra-systemic violence and electoral politics can be viewed as 'strategic complements' (Dunning 2011), in that these acts of violence support electoral aims. Intra-systemic violence comes in many forms, including riots or group clashes between party supporters and violent attacks on candidates, politicians or voters. In India, intra-systemic violence may also be 'outsourced' to specialists in violence, such as criminal gangs associated with politicians or parties (Berenschot 2011). This includes instances of 'booth capture', where armed goons take over polling places (Berenschot 2011;Verma 2005). The following quote about the 1990 election in Bihar illustrates this range: In Dongri, polling stations nos 87 and 91 were raided by about 100 men who came armed. They snatched the ballot papers stamped them and put them into the ballot boxes. According to the police nobody was arrested.… As news of this spread in the areas, large-scale violence erupted, with party workers freely flinging soda water bottles, bricks and stones at each other. A number of cars were damaged and some people were injured. (ToI, 27 February 1990) Anti-systemic violence, by contrast, seeks to disrupt the electoral process as such. This type of violence is 'aimed at destroying the rules of the game, not winning within them' (Staniland 2014, 108). Anti-systemic violence and electoral politics are 'strategic substitutes' (Dunning 2011), as these violent tactics constitute an alternative to electoral participation, as well as an attempt to undermine it. Anti-systemic violencealso referred to as 'insurgent electoral violence' (Condra et al. 2018)often takes place in the context of conflicts over territorial control between the state and insurgent groups. It involves the coercion of or threats against candidates, voters, poll workers or security forces, as well as the destruction of election infrastructure.
Strategies pursued by insurgent groups towards elections differ across time and space, and vary from seeking incorporation by allying with existing electoral alternatives or forming political parties, to armed opposition to the electoral process (Kennedy 2014;Matanock and Staniland 2018). Similarly, state strategies towards insurgents vary from co-optation to attempts at annihilation (Staniland 2015). In India, anti-systemic electoral violence often takes the form of violently enforced boycott calls by insurgent groups (see, for example, Spokesperson 2009). In such instances, insurgents not only state their opposition to the election, but also stake their reputation on their ability to disrupt its orderly conduct. The boycott call becomes a means for the group to demonstrate territorial control and popular support.
Key for the distinction between intra-and anti-systemic violence is the strategy of the actors towards the election. Whereas intra-systemic violence seeks to increase the chance that one particular candidate or party emerges victorious, anti-systemic violence seeks to delegitimize the winner by casting doubt on the state's ability to administer elections and by highlighting the importance of non-electoral alternatives. In making this distinction, it is important to acknowledge that identifying collective or individual motivations for engaging in violent tactics is notoriously difficult. Violence is generally considered illegitimate, and electoral violence in particular is unpopular with voters, creating incentives for actors to hide their involvement or distort their motivations (Birch, Daxecker and Höglund 2020, 5;Brass 2003, 14;Gutiérrez-Romero and LeBas 2020). Moreover, actors who instigate violence are unable to fully anticipate the consequences and may miscalculate. Against this background, we limit our analysis to short-term strategies and effects related to specific elections.
Intra-and anti-systemic violence undermine democracy. Intra-systemic violence respects the principle of elections but not the underlying norms of free and fair competition. Anti-systemic violence indicates that the principle of elections is not accepted uniformly by relevant actors. In the long run, both types of violence may feed into each other. Political groups that do not have a chance of winning within the system because there is no level playing field may turn against the system, for instance. With regard to the effects of violence, we focus on turnout and incumbent vote share in the election during which violence occurred. The long-run consequences of violence, either at the group or individual level, are beyond the scope of our analysis.
To understand the effects of intra-and anti-systemic electoral violence, we have to consider why actors engage in violent tactics. These electoral strategies, we argue, influence where we are likely to see violence and how violence shapes election results, specifically turnout and incumbent vote share. Focusing on the Indian context, we start by laying out the conditions under which actors develop their electoral strategies and decide whether or not to turn to violence.
The organization of free and fair elections depends on professional election management, as well as on impartial security forces that are able and willing to guarantee the safety of candidates, voters and poll workers. In India's federal system, elections require cooperation between the Election Commission of India (ECI)an autonomous constitutional bodyand state governments (Singh and Roy 2019;Sridharan and Vaishnav 2017). The architects of India's constitution believed that centralizing election management in one nationwide commission would provide better protections against electoral malpractice and political interference than state-level commissions (Gilmartin and Moog 2012, 138). Placing election administrators at arm's length from political influence is generally easier than insulating security forces from political pressure, however. Indian state governments have a constitutional mandate to uphold public order during the election, while they or their parties participate in the process and thus have a direct and immediate stake in the outcome. This dual position has often generated tension, and state governments have misused control over security forces to instigate violence or failed to effectively curb it. While attributing direct responsibility for specific violent incidents is difficult, there is substantial evidence that violence has the potential to generate electoral benefits and that political actors are aware of this potential (see, for example, Berenschot 2020;Brass 2003;Daxecker 2020;Verma 2005;Wilkinson 2006).
Elections not only influence actors trying to win at the ballot box, but also provide a flashpoint for insurgent groups, as campaigns and polling seek to engage the population in the political process under certain rules and procedures. If these groups are not aligned with any of the options on the ballot, they have a strong incentive to discourage participation, rather than to just let the election proceed unfettered. Campaigning in the run-up to the election provides opportunities for electoral forces to make inroads with the group's actual or claimed constituency. Candidates and party workers who enter the group's territory to disseminate campaign messages, even if only moderately successful, involve citizens in democratic politics and therefore present the threat of erosion of group support. In India, such boycott calls by armed groups go hand in hand with the intimidation or even assassination of candidates and election officials, which often requires them to enter the affected territory with armed escorts, if at all. Electoral participation is also symbolically important, as high turnout indicates broad support for selecting leaders through the electoral process. High levels of participation thus confer legitimacy on the eventual winner, which undermines the group's claim that existing electoral alternatives leave out important voices in society. Even though insurgent groups are not directly on the ballot, elections signify a referendum on non-electoral alternatives. Newspaper articles on elections in areas affected by boycott calls tend to prominently report the percentage of turnout and to relate this to group strength (see, for example, Banerjee 2009). Depressing turnout is thus crucial to demonstrate that non-electoral alternatives need to be engaged with. An article published in the run-up to assembly elections in Jharkhand illustrates this strategy: Handwritten posters have been put up by Maoists urging people to boycott Assembly Polls. The Red rebels have also challenged the police administrations and threatened leaders of political parties, besides creating panic in the tribal dominated villages of Bhandaria block of Garhwa district in the last 24 h. The Maoists, in fact, have threatened people and political leaders with dire consequences if they oppose the poll boycott call. 'No vote for Capitalists. No vote for a government that allows police to kill innocent people in fake encounters', read the posters pasted at Bargarh, Marda, Tengari, Jumnagarh, Ugra, and Karchali villages at Bhandaria bloc. The posters have spread panic among voters as well as leaders of various political parties. (ToI, 8 December 2009) Since insurgent groups cannot afford to ignore elections, the extent to which they succeed in disrupting elections allows them to demonstrate strength. Their long-term goal may be to negotiate with the government and to win concessions. Governments only have incentives to negotiate, however, if the groups are relevant potential partners.
Based on these motivations for engaging in intra-and anti-systemic violence, we now develop our hypotheses, starting with how violence influences turnout. Electoral violence is a disruptive and often exclusionary tactic, which implies that both types of violence should be associated with lower turnout (Bratton 2008;Collier and Vicente 2012;Gutiérrez-Romero and LeBas 2020;Trelles and Carreras 2012). In the case of anti-systemic violence, insurgent groups aim to obstruct or delegitimize the election. We therefore expect them to seek to decrease turnout as much as possible, ideally to the point where the election is cancelled or postponed, or where polling is disrupted to such an extent that the eventual winner lacks legitimacy. Violent tactics signal to everybody in the group's area of influence that there are severe repercussions for supporting, facilitating or participating in the electoral process. By making good on their threat in some locations, insurgents amplify the boycott call in neighbouring constituencies.
The goals of partisan elites are distinct from those of insurgent groups. To benefit parties who aim to win the election, polling must take place, and it must be credible enough to ensure that the winner has some legitimacy. If irregularities are systematic or widespread, the ECI will order a re-poll in affected constituencies, thus rendering the initial election result moot. Rather than disrupt the election wholesale, the purpose of intra-systemic violence is to suppress support for political opponents, while boosting the electoral fortunes of allies. This can be achieved through threats aimed at opposition candidates or supporters. It can also be achieved through 'booth capture', as in the first newspaper quote earlier. In such instances, armed goons block off access to a polling station and stuff ballot boxes. Overall, it is easier to keep voters away from the polls than to make them vote a certain way once they are in the polling booth. While the presence of armed goons at polling stations could galvanize participation as a mode of protest (Aytaç and Stokes 2019) and the stuffing of ballot boxes could also make it appear as if more votes have been cast, we expect that some groups of voters will be dissuaded from participating in the election to avoid the (threat of) violence. In the case of intra-systemic violence, we therefore also expect to see a decrease in turnout, though not as dramatic as in the case of anti-systemic violence: Hypothesis 1: Intra-and anti-systemic violence are associated with lower voter turnout. The decrease in turnout is larger for anti-systemic than for intra-systemic violence.
The location of violence is endogenous to the goals actors aim to achieve by turning to violent tactics. As indicated earlier, anti-systemic violence tends to be focused on the areas of influence of insurgent groups in order to demonstrate and maintain territorial control. Boycott calls generally affect contiguous areas encompassing multiple administrative districts, which, in turn, contain multiple assembly constituencies. A violent incident anywhere in the area affected by the boycott signals that the threat of coercion is credible. Insurgents therefore have quite a lot of flexibility in choosing their specific targets within their area of influence. The institutional framework of the election and the characteristics of constituency-level incumbents therefore matter less for targeting. The location of intra-systemic violence, by contrast, is influenced by partisan electoral goals. In the case of India's first-past-the-post electoral system, the single-member constituencies in which members of the state legislative assembly are elected are the focal point of competition. We argue that intra-systemic violence is more likely to take place in constituencies where the local incumbent belongs to the state-level opposition. Aligned constituenciesthat is, those where the incumbent member of the legislative assembly belongs to the party in control of the state governmentare therefore comparatively more likely to be peaceful. There are several reasons to expect this pattern. Incumbents are more frequently responsible for electoral violence, and opposition supporters are more likely to be targeted (Burchard 2020). Overall, as outlined earlier, incumbent elites in control of the state government have opportunities to deploy security forces strategically. Existing literature demonstrates that intra-systemic violence is likely to occur in locations where it generates electoral benefits to incumbents (see, for example, Daxecker 2020; Wahman and Goldring 2020; Wilkinson 2006). Incumbent elites may use their control of security forces to broaden their territorial base. The strategic deployment of security forces allows their co-partisans to campaign beyond the party's core constituencies. This is a crucial resource not available to the opposition, as only incumbent elites can draw on security forces to muscle their way into nonaligned constituencies. In plurality systems, pre-election violence is therefore particularly common in areas controlled by the opposition (Wahman and Goldring 2020). Fjelde (2020) shows that parties are more likely to resort to violence when they lack other means to appeal to voters, such as an infrastructure for organizational outreach and voter mobilization. This again points to opposition-controlled areas as likely sites of violence. While Fjelde's analysis is cross-national and looks at party strength in the aggregate, we expect the same logic to hold at the subnational level: Hypothesis 2a: Intra-systemic violence is more likely in non-aligned constituencies than in aligned constituencies. Hypothesis 2b: Anti-systemic violence is not systematically related to alignment.
Finally, if intra-systemic violence is more likely in non-aligned constituencies, then within these constituencies, it should be targeted at supporters of the state-level opposition. The effect of electoral violence would then be to strategically depress turnout among supporters of the local incumbent. These are the voters who would be on the receiving end of coercion and intimidation, and they may respond by withdrawing from the electoral process: Hypothesis 3: Intra-systemic violence decreases incumbent vote share in non-aligned constituencies but does not negatively impact the incumbent vote share in aligned constituencies.
In sum, we expect the distinct intentions of intra-and anti-systemic violence to influence where violence occurs and how it influences election outcomes. We now turn to the empirical test of our hypotheses.

Measuring Intra-and Anti-systemic Electoral Violence
Our analysis draws on an original dataset that includes electoral violence during legislative assembly elections between 1985 and the implementation of the new delimitation order in (Electoral Commission of India (ECI) 2008). This timespan ensures that the borders of assembly constituencies, which are our units of analysis, are consistent over time. As indicated earlier, previous efforts to collect data on electoral violence have focused on events related to national elections (see, for example, Daxecker, Amicarelli and Jung 2019; Hyde and Marinov 2015). Existing violence data that would cover subnational elections are related to either armed conflict (see, for example, Sundberg and Melander 2013) or communal violence (see, for example, Varshney and Wilkinson 2006), and therefore do not systematically capture intra-and anti-systemic violence in the same dataset. Methodologically, the scope of our project implies the need to select data that are available consistently between 1985 and 2008. We rely on election coverage in a national newspaper, namely, the ToI, the English-language daily with the largest circulation in the country. Since it is regarded as a quality (or 'broadsheet') newspaper, the ToI has been popular as a reliable source for event data on India (see, for example, Kennedy and King 2013;Varshney and Wilkinson 2006). It is the best source of information for our purposes for two reasons. First, the events we are aiming to identify are newsworthy, as they involve violence or the threat of violence. While newspapers filter incoming information, 'violent events, particularly when involving death, are generally more likely to make news' (Franzosi 2004, 168). Secondly, since the ToI is a national newspaper, it is more likely to offer comprehensive coverage of state assembly elections than international newspapers or newswires. This alleviates concerns about under-reporting and geographically biased coverage (von Borzyskowski and Wahman 2019), though it does not fully resolve them. Media-based data always remain accounts of 'reported violence', rather than capturing all violent events. Nevertheless, elections are central events in democratic societies, meaning that even subnational elections tend to draw considerable media attention. The articles we code include reporting by local journalists, who may be more likely to travel to and report from urban areas and expected violence hotspots (von Borzyskowski and Wahman 2019). Violence in rural areas and in constituencies with no previous history of violence may not be picked up. In addition, however, the articles also draw on updates provided by the ECI, which organizes the elections and staffs polling places, investigates whether incidents were reported, and, in severe cases, orders a re-poll. By referencing this information, ToI articles offer more comprehensive geographic coverage than traditional news reporting.
The collection of the data proceeded in two steps. First, we extracted articles covering elections from the ToI. For each election, we searched for articles published one month prior to the beginning of polling up until one month after results were announced. 1 The majority of articles are clustered around polling days. We extracted articles that contain the name of the state and the word 'election'. In addition, we specified nine search terms, at least one of which had to be present in the article: 'repoll', 're-poll', 'irregularit*', 'violen*', 'unrest', 'attack', 'intimidat*', 'booth captur*' or 'boycott*'. 2 For the 88 legislative assembly elections covered in our analysis, we coded 2,107 articles and, on average, we extracted 24 articles per election, with some searches yielding more than 100 articles (for example, in Punjab in 1992). In a second step, we coded incidents reported in the extracted articles based on three criteria: (1) election related; (2) taking place in the state holding assembly elections; and (3) substantiated, that is, either the event is presented as a fact by the ToI or the ECI recognized the event as substantiated. For each event, we coded the constituency name and a brief event description. 3 Where the information about the event was incomplete, we complemented information about extracted events with other news sources (for example, The Hindu or India Today magazine). If the ToI mentioned only the name of the village, not the constituency, we matched the village to a constituency.
In our analysis, we include Andhra Pradesh, Assam, Bihar, Chhattisgarh, Gujarat, Haryana, Jammu and Kashmir, Jharkhand, Karnataka, Kerala, Madhya Pradesh, Maharashtra, Odisha, Punjab, Rajasthan, Tamil Nadu, Uttar Pradesh, and West Bengal. These are the states for which census data are available at the district level. Combined, they accounted for 95 per cent of the Indian population in 2001. 4 In addition to our violence data, we draw on three datasets: district-level census data collected by Iyer and Shrivastava (2018); constituency-level electoral data compiled by Jensenius and Verniers (2017); and data on the composition of state governments provided by Harbers, Bartman and Van Wingerden (2019). 1 This time period is based on the de facto polling and result dates. If elections or the announcement of results were postponed, we extended the time period. As indicated earlier, our analysis is restricted to violence that took place before the election but includes incidents of pre-election violence reported after the election. 2 We accessed the ToI archives digitally through news aggregation databases and used loose search terms to identify articles (* in ProQuest, ? in Factiva and ! in Lexis Nexis). 3 Our data were subsequently cross-validated with Uppsala Conflict Data Program (UCDP) data, the only available alternative (Sundberg and Melander 2013). The datasets differ in three critical ways: (1) our events are related to the electoral process, whereas UCDP records all violent events; (2) UCDP records only events with at least one fatality, whereas we include non-lethal events; and (3) UCDP records events at the district level, that is, the administrative unit above the assembly constituency. Consequently, we could cross-validate only with regard to the number of deaths recorded and district affected. Cross-validation shows that, over the same time period, UCDP records 1,150 deaths compared to 438 in our data, which fits with the more restrictive election-related coding. However, since we include non-lethal events, we record 177 districts affected, whereas UCDP records only 114. 4 Of the three states created in November 2000 (Jharkhand from Bihar, Chhattisgarh from Madhya Pradesh and Uttarakhand from Uttar Pradesh), we only managed to track which constituencies 'changed' states and how they were renumbered for Jharkhand and Chhattisgarh.

Intra-and anti-systemic violence
The distinction between intra-and anti-systemic violence captures 'whether an actor in question is seeking to win or maintain power within the context of the democratic political systemic (intrasystemic) or whether instead the actor seeks to overthrow the status quo order (anti-systemic)' (Staniland 2014, 107). As highlighted earlier, identifying collective or individual motivations for engaging in violent tactics is difficult. In order to operationalize types of violence, we therefore have to render motivations observable. For anti-systemic violence, we rely on stated intentions by armed groups to disrupt the election and code whether such groups have declared an election boycott in a specific area. 5 In order for the boycott to be effective, it must be publicized. Armed groups generally put up announcements calling on residents to abstain from voting and implicitly or explicitly threatening those who do not comply, as in the newspaper article earlier. They may also issue statements to the press (for example, Spokesperson 2009). By announcing boycott calls before an election, armed groups essentially try to turn these constituencies into anticipated violence hotspots, which draws media coverage (von Borzyskowsi and Wahman 2019). Boycotts generally affect whole districts and often several districts. In addition to identifying whether a boycott call is in place in a constituency in a given election, we also code whether violent events are reported in the area affected by the boycott call and, if so, whether these events were lethal or non-lethal.
While boycott calls by armed groups have to be made explicit in order to be effective, violent tactics by actors trying to win the election are more covert. Parties and candidates generally cannot openly declare their intention to use violence against opponents. They at least have to maintain a semblance of commitment to the principle of free and fair elections. Our data are therefore asymmetric, in the sense that actor involvement is more observable in boycott calls than intrasystemic violence. We deal with this asymmetry in two ways. First, we code violent incidents that did not occur in an area affected by a boycott call and that involved actors affiliated with parties on the ballot, such as party workers, party supporters and voters, as intra-systemic violence. Through their connections with political parties, we thus assume that these actors are trying to influence the outcome of the election, and we use this operationalization of intra-systemic violence for our main analyses. In addition, however, we also create a second variablewhich we call 'non-boycott violence'that captures all violent events in our database that did not occur in an area affected by a boycott call, even if the actors involved are not clearly identified as situated within the political system. 6 We expect non-boycott violence to demonstrate greater similarity to intra-systemic than to anti-systemic violence in our analyses. We report results for non-boycott violence in the Online Appendix and, indeed, find that they are broadly consistent with patterns observed for intra-systemic violence.
Our unit of analysis is the constituency-year. We aggregate events at the constituency level into binary and categorical variables. The binary variables capture: (1) whether a boycott call was in place and the constituency was thus affected by anti-systemic violence (0 = no; 1 = yes); and (2) whether intra-systemic violence occurred (0 = no; 1 = yes). This dummy does not reflect the severity of violence, but it is useful for our analysis, as we are primarily interested in where 5 In operationalizing anti-systemic violence, we focus strictly on boycott calls by armed groups, which we distinguish from peaceful protests by citizens. In our coding, we came across several instances of villages or community organizations stating that they would not vote to protest issues like poor infrastructure or a lack of drinking water. Such citizen boycotts are not instances of anti-systemic violence according to our definition. 6 For instance, if an article reports 'Incidents of shots fired' (ToI, 18-06-1987, 'Heavy polling in Haryana') the event is coded as non-boycott violence, whereas if an article reports 'Clashes between supporters during voting' (ToI, 15-10-2009, 'Minor skirmishes at 70 per cent turn up in Haryana'), the event is coded as intra-systemic violence. The difference is whether actors are identified and if they are affiliated with parties on the ballot. For the reasons outlined earlier, we anticipate that nonboycott violence, which is a residual operationalization, contains more instances of intra-systemic than of anti-systemic violence.
violence occurs and how its occurrence influences the outcome of the election. Further, based on news coverage, it is difficult to accurately determine whether separate incidents occurred at multiple polling places or whether one large-scale event affected multiple polling places. Especially in elections where irregularities were widespread, articles tend to mention the constituencies where clashes between party supporters or attacks by armed groups occurred, though without fully disentangling each event or explaining the relationship between different events. 7 While it is therefore not possible to count events to measure the severity of the violence, we are able to distinguish between lethal and non-lethal events. To develop a more nuanced measure of violence, we also created categorical variables that capture: (1) the type of intra-systemic violence (0 = none; 1 = non-lethal violence; 2 = lethal violence); and (2) the type of anti-systemic violence (0 = no boycott call; 1 = boycott call in place; 2 = boycott call with non-lethal violence; 3 = boycott call with lethal violence).
Over our entire time period, we have 18,982 observations, and our variables are highly skewed towards zero (3.35 per cent of our units are affected by intra-systemic violence, and 1.63 per cent of our units are affected by anti-systemic violence). Figure 1 visualizes the constituencies that were affected by each type of violence. The maps confirm that, while there is some spatial overlap in Bihar and Punjab, intra-and anti-systemic violence generally take place in different areas, suggesting different dynamics.

Electoral results
We draw on data collected by Jensenius and Verniers (2017), which aggregates results from the ECI election reports. 8 For each assembly constituency, we identified the winner, the runner-up, their respective parties, the margin of victory and turnout in the election. To capture constituency-level Incumbency we draw on the results of the previous election. Margin of Victory measures the difference in vote share between the winner and the runner-up. Incumbent Vote Share captures the percentage of votes for the incumbent party. 9 Turnout reflects the percentage of eligible voters who cast their votes. To capture Alignment, we create a binary variable indicating whether the assembly constituency incumbent belongs to the party in power at the state level (or to the main party in the coalition). For this, we draw on data about governing coalitions collected by Harbers, Bartman and Van Wingerden (2019). Finally, we include dummy variables capturing three distinct types of assembly constituencies: general, reserved for scheduled castes (SC) and reserved for scheduled tribes (ST). Goel (2018) shows that reserved constituencies present distinct patterns of voting compared to general constituencies.

Demographic and geographic controls
We use Iyer and Shrivastava's (2018) census data to build our control variables. Census data are not available at the level of assembly constituencies, only at the district level. We use the (Electoral Commission of India (ECI), 1976) Delimitation Order to match each assembly constituency to its district. 10 For each variable, the same percentage is applied to all constituencies within a district. Linear interpolation was applied between the census years by Iyer and Shrivastava (2018) from 7 For example, 'In Chhattisgarh, [Deputy Election Commissioner Ajay Narayan] Jha said Left-wing extremist groups did not allow polling parties to reach eight polling stations in Dantewara. "We have received reports of snatching and damage to EVMs," he said' (ToI, 2 December 2003). We code this as non-lethal anti-systemic violence taking place in Dantewara constituency. 8 This dataset has later been expanded (see Agarwal et al. 2021) and is available at: https://lokdhaba.ashoka.edu.in/ 9 When the incumbent did not compete in election t because the incumbent party entered an electoral alliance, for instance, this variable was coded as missing. All analyses are replicated in the Online Appendix with consistent samples. Our temporal identifier throughout is electoral sequence, rather than year, as elections are staggered. 10 A similar procedure was used by Iyer and Shrivastava (2018) to match districts to states. 1985 to 2000, and we extend the period to 2008. From the census data, we extract three variables: Electrification, that is, the percentage of households with electricity; Literacy Rate, that is, the percentage of the population that is literate; and Urbanization, that is, the percentage of the population that lives in urban areas. In the turnout models, we also include a Spatial Lag of intra-systemic violence. 11 We also include Ruggedness, which measures the standard deviation of altitude within the assembly constituency (Fearon and Laitin 2003).

Analysis
Following our hypotheses, we analyse three different relationships. First, we model the effect of intra-systemic and anti-systemic violence on turnout (H1). Next, we analyse the effect of alignment on the likelihood of intra-systemic violence (H2a) and on the likelihood of anti-systemic violence (H2b). Finally, we model the interaction effect of intra-systemic violence and alignment on incumbent vote share (H3). We report descriptive statistics and alternative specifications, including those using different operationalizations of the violence variables, in the Online Appendix. Table 1 presents the results of our analysis of the effect of intra-systemic and anti-systemic violence on turnout. We use fixed effects to control for unobserved heterogeneity between

11
We used the pre-delimitation shapefile created by Sandip Sukhtankar and shared by Jensenius (2017), but we edited the file to reflect delimitation in Jammu and Kashmir in order to ensure the alignment of constituency and state borders, as well as the creation of new states. Since elections are staggered, the lag is calculated separately for each state on the basis of an inverse distance weights matrix so that spillover effects do not extend beyond state borders. We do not include a spatial lag for anti-systemic violence, as boycott calls generally affect entire districts. constituencies and do not include the time-invariant control variables in these models (that is, ruggedness, constituency type and state fixed effects).
Our results are in line with our first hypothesis: intra-systemic and anti-systemic violence are negatively associated with turnout, both in separate models and in a combined model. For intrasystemic violence, the negative effect emerges only for non-lethal events, which constitute 88.2 per cent of all intra-systemic violence in our data. The coefficient for lethal events is not significant. The effect of anti-systemic violence on turnout is stronger than the effect of intra-systemic violence (see Model 3 in Table 1), which makes sense considering that anti-systemic violence aims at depressing overall turnout and intra-systemic violence aims at depressing turnout selectively to favour one candidate over others. 12 While intra-systemic violence could lead to higher turnout, for instance, in the case of ballot-box stuffing, the aggregate effect is thus still negative. For anti-systemic violence, the size of the negative effect increases with the severity of violence. A boycott call, even without additional events reported, depresses turnout significantly, but the negative effect becomes more pronounced if there are specific violent incidents reported and increases further if there are fatalities. This contrasts with the pattern for intra-systemic violence, where the coefficient for lethal events is not significant, suggesting that both types of violence are qualitatively different. Further, the spatial lag for intra-systemic violence is positive, suggesting that while intra-systemic violence decreases turnout locally, it boosts turnout in neighbouring constituencies. Literacy and electrificationtwo variables that tap into the provision of public goodshave positive effects on turnout. Since this is a fixed-effects model, this indicates that an increase in the provision of public goods is associated with an increase in electoral participation.
In Table 2, we explore the effect of alignment on the incidence of intra-systemic violence (H2a) and on anti-systemic violence (H2b). Since we are not trying to account for the severity of violence in a particular place, we draw on the binary operationalization of intra-and anti-systemic violence for these models. 13 Our hypothesis is that alignment has a negative effect on the likelihood of intra-systemic violence but that it is not systematically related to anti-systemic violence. Notes: Robust standard errors in parentheses (cluster: district); * p < 0.05; ** p < 0.01; *** p < 0.001: fixed effects at the constituency level.

12
In models based on the binary operationalizations of intra-systemic and anti-systemic violence, the effect of anti-systemic violence on turnout is also stronger (see Table 1 in Online Appendix B). 13 Results for the categorical variables (see Table 2 in Online Appendix C) are substantively similar. Yet, while the effect of alignment on lethal intra-systemic violence is also negative, it is not significant.
For these models, we use zero-inflated Poisson regression with robust standard errors to account for the highly skewed nature of our dependent variables.
The results of this second analysis are also in line with our theoretical expectations: compared to non-aligned assembly constituencies, the probability of intra-systemic violence is lower in aligned constituencies (β = −0.339** [see Model 1 in Table 2]). In other words, constituencies where the incumbent belongs to the party in power are less likely to experience intra-systemic violence. As expected, alignment is not systematically related to anti-systemic violence. 14 Interestingly, we find that margin of victory is positively associated with both types of violence. This means that the less competitive the race, the higher the likelihood of violence. Especially with regard to intra-systemic violence, this finding is remarkable, as previous research on India has highlighted that close electoral races lead to riots (Wilkinson 2006). One possible interpretation is that the violence that appears in our dataset, which is generally limited in scale, is qualitatively different from full-blown communal riots. In this sense, the positive relationship we find is in line with research on intra-systemic electoral violence in other countries, where incumbents have been found to use violence to muscle their way into opposition strongholds (Fjelde 2020;Wahman and Goldring 2020). In the case of anti-systemic violence, it is possible that insurgent activity is endogenous to the lack of vibrant electoral competition, as these groups operate in areas where democratic institutions function less well and their presence further erodes such institutions.
Finally, in Table 3, we analyse how intra-systemic violence influences election results and whether the decline in turnout benefits parties in control of the state government. If so, we expect the vote share of incumbents in non-aligned constituencies to decrease in the case of violence. Once again, we use a fixed-effects model to control for unobserved heterogeneity. To model the combined effect of intra-systemic violence and alignment on incumbent vote share, we use Notes: Robust standard errors in parentheses (cluster: assembly constituency); * p < 0.05; ** p < 0.01; *** p < 0.001.
14 Boycott calls often affect entire districts, whereas alignment is measured at the constituency level. To account for this, we replicate the analysis with an alternative operationalization of anti-systemic violence, for which only specific incidents of violence related to boycott calls are considered (see Model 3 in Table 2 in Online Appendix C). This ensures that the level of analysis is consistent. The coefficient is not significant, indicating that alignment is not systematically related to violent incidents that occur in the context of a boycott call, just as it is not related to the overall boycott call. an interaction term. We again rely on the binary operationalization of intra-systemic violence here and report specifications based on categorical variables in the Online Appendix. 15 Results are in line with our hypotheses: we find that intra-systemic violence has a negative effect on incumbent vote share, which indicates that violence is associated with fewer votes for the constituency-level incumbent party. Further, the positive coefficient of alignment indicates that incumbents who belong to a party in power at the state level have comparatively higher vote shares than incumbents belonging to an opposition party. The positive interaction term (β = 7.993***) indicates that the negative effect of intra-systemic violence on incumbent vote share is positively mitigated by alignment, that is, that the negative effect of violence is diminished if the incumbent is aligned with the state government. This is consistent with our theoretical expectations: vote shares of incumbents belonging to opposition parties suffer more due to intrasystemic violence than vote shares of incumbents belonging to the party that governs the state. This suggests that within non-aligned constituencies, intra-systemic violence is targeted at supporters of the state-level opposition and thus strategically depresses turnout among supporters of the local incumbent. Combined, our analyses lend support to the idea that intra-systemic violence benefits state-level incumbents.

Conclusion
The strategic motivations for violent tactics are well documented in the literature on electoral violence. We also know that violence can serve as a strategic complement or as a strategic substitute to electoral politics (Dunning 2011), and that not all electoral violence is aimed at winning within the existing institutional framework. Yet, previous work has generally not considered whether intra-and anti-systemic violence differ in their effects on election outcomes, even though the distinct motivations for turning to violence would suggest that they do. In this article, we made the case for distinguishing between intra-and anti-systemic violence conceptually and empirically. We analysed the effects of both types of violence on the basis of an original dataset of violence in legislative assembly elections in India. Our results show that violence decreases turnout but that the effect is larger for anti-systemic violence. Intra-systemic violence appears intended to selectively depress turnout among opposition supporters, as it is more likely to occur in constituencies where the local incumbent belongs to the state-level opposition and it decreases the vote Note: Robust standard errors in parentheses (cluster: district); * p < 0.05; ** p < 0.01; *** p < 0.001; fixed effects at the constituency level.

15
Results in Table 1 in Online Appendix D are similar in direction, strength and significance, except for the category of lethal intra-systemic violence. As in the turnout model (see Table 1), the coefficient is positive but non-significant, with a very large standard error.
share of local incumbents within such non-aligned constituencies. For anti-systemic violence, the incidence of violence is not systematically related to the institutional geography of elections. However, the negative effect on turnout is larger than for intra-systemic violence due to the compounding effects of the boycott call and violence (both lethal and non-lethal), indicating that this type of violence is extremely effective in terms of keeping voters away from the polls and thus discrediting the electoral results due to low participation rates.
Overall, our results show that intra-systemic violence benefits parties in control of state governments and therefore the actors who are responsible for maintaining law and order. This corroborates previous work which has indicated that electoral violence can generate electoral benefits to incumbents, as governing parties are able to deploy security forces strategically to broaden their territorial base (for example, Daxecker 2020; Wahman and Goldring 2020;Wilkinson 2006). We find that anti-systemic violence, by contrast, has distinct spatial patterns and is not systemically tied to alignment. This is consistent with studies which show that insurgent strategies are aimed at disrupting the electoral process in order to undermine the legitimacy of the winner and to demonstrate the relevance of non-electoral political alternatives (Condra et al. 2018;Matanock and Staniland 2018). We add to previous scholarship by showing that distinct motivations for turning to violent tactics translate into distinct effects on election outcomes.
Our empirical focus on India allows us to observe intra-and anti-systemic violence and their effects within one country. India differs from many other democracies in its commitment to holding elections throughout the territory, rather than suspending polling in areas with limited state control, such as those in which insurgent groups operate. Further, despite its track record of democratic governance, India also contains actors that treat violence as a strategic complement to electoral politics, as well as those that engage in violence as a strategic substitute to electoral politics. This makes it possible to identify the effects of both strategies while holding a host of other national-level factors constant.
While the operationalization of our violence variables, such as the treatment of boycott calls by insurgent groups as an instance of anti-systemic violence, is context specific, our findings and analysis have implications beyond India. For cross-national studies of the effects of electoral violence, our results suggest the need to look more closely at the predominant type of violence per country. In the absence of data capturing intra-and anti-systemic violence, it may be necessary to exclude countries with active insurgencies in order to isolate the effects of intra-systemic violence, for instance. Otherwise, estimates of the effect of electoral violence on turnout or incumbent vote share are likely imprecise and also influenced by the relative balance of intra-and anti-systemic violence included in the analysis. Similarly, for subnational analyses of violence, scholars should consider which type of violence is present and whether there is variation in types of violence within the country. If data distinguishing between the two types of violence are not available, the best strategy for researchers may be to narrow the frame of comparison so that the type of violence is homogeneous across analysed cases. While intra-and anti-systemic violence are both disruptive and exclusionary tactics that threaten the core values and existence of democratic competition, our analysis cautions scholars to systematically consider the strategic motivations for engaging in violent tactics in the first place in order to better understand their effects.
Supplementary Material. Online appendices are available at: https://doi.org/10.1017/S0007123422000345 Data Availability Statement. Replication data for this article can be found at: https://doi.org/10.7910/DVN/ACRF5X Financial Support. Data collection for this project was supported by H2020 European Research Council (ERC Advanced Grant #323899) and H2020 Marie Skłodowska-Curie Actions (Marie Curie Fellowship #656361).