Education, Disinformation And Electoral Violence

(1)

1

Education, Disinformation And Electoral Violence

A Quantitative Study on the Association between Education and Violent Elections

Jan Rustemeyer Master's Thesis

Spring 2021

Department of Peace and Conflict Research, Uppsala University Supervisor: Emma Elfversson

Word count: 18 000

(2)

2

Abstract

The institutional, electoral and ethnic factors contributing to electoral violence have well been documented through both quantitative and qualitative research, while the mobilization process for electoral violence has been examined qualitatively. This study aims to contribute to quantitative research on factors that explain why citizens turn into perpetrators of electoral violence by examining how education can contribute to a decrease in electoral violence through the question:

How does the level of education influence the occurrence of electoral violence? Given the presence of disinformation about elections during the electoral cycle, this research asserts that education can contribute to a decrease in electoral violence by decreasing the acceptance of disinformation about elections. The hypothesis is tested through a large-N study on sub-national data of elections organized between 2004 and 2012 worldwide. The study’s results identify no support for the hypothesized association between education and election violence.

Key Words: electoral violence, education, elections, disinformation, mobilization

(3)

3

Acknowledgments

This thesis owes a great deal to those around me. I thank my supervisor Emma Elfversson, who served as a great guide and encouragement throughout the project. Thank you, Bryan, Jack and Maurice for your never-ending banter and (occasional) words of advice. I am very thankful for the support from my family in Amsterdam, Maastricht and Germany during my studies. And thank you Charlotte, for being there for me, listening to my ramblings and turning them into somethings.

(4)

4

List of figures and tables

Figure 1: Map on the global variation in the average years in education. ... 15

Figure 2: Distribution of the amount of national election rounds per country ... 21

Figure 3: Distribution of the frequency of events of pre-election during election day(s) violence22 Figure 4: Education and electoral violence plot ... 31

Figure 5: Electoral violence events and years in education per sub-national unit per election cycle ... 36

Figure 6: Numbers of mobilization for events of electoral violence plotted against the average years in education ... 39

Figure 7: Test for overdispersion of the dependent variable ... 51

Table 1: Descriptive statistics of the data ... 30

Table 2: Negative binomial regression table 1 ... 32

Table 3: VIF test for multicollinearity ... 34

Table 4: OLS regression table... 35

Table 5: Binary logistic regression table ... 37

Table 6: Negative binomial regression table 2 ... 38

Table 7: Negative binomial regression table excluding the control variable ‘Urbanization Rate’ .. 51

(6)

6

Introduction

The Nigerian general election of 2019 was marred by violence. Civil society organisations estimated that 626 people lost their lives due to election-related violence from the start of the election campaign until the commencement of the voting (Human Rights Watch 2019). Compared to the rest of Africa, Nigerian elections give rise to a large number of violent incidents (Burchard and Simati 2019), and the election of 2019 was no exception to this. In the run-up to election day, armed attacks and abductions took place against government actors, including election officials (Deutsche Welle 2019). The electoral violence was concentrated in a couple of Nigerian states, most prominently the Northern regions, the Southwest and the Delta (Council on Foreign Relations 2019). The election occurred in a highly contested context, where tensions between the two frontrunners for presidential office and their supporters ran high (Bamigbade and Dalha 2020).

Violence-inciting disinformation that negatively depicted the other side and its role in the election was spread across the country during the pre-election phase (CNN 2019). Incumbent president Buhari had to publicly refute claims that he had died and a clone had taken over his office (CNN 2019).

Following the relatively peaceful election of 2015, which also saw Nigeria’s first democratic transition of power since the end of the military dictatorship (Council on Foreign Relations 2019), the return to higher numbers of election violence might have come as a surprise. Research into factors that contribute to the occurrence of electoral violence allows for insights into why election violence resurfaced during the Nigerian general election of 2019. Given the potentially destructive impact on the stability of a democracy (Christensen and Utas 2008), understanding and preventing electoral violence is crucial in a world where autocratization is on the rise again (Lührmann and Lindberg 2019). The institutional, electoral and ethnic factors which contribute to the occurrence of electoral violence have well been documented through both quantitative and qualitative research (see Brosché, Fjelde, and Höglund 2020; Fjelde 2020; Fjelde and Höglund 2016; Hafner-Burton, Hyde, and Jablonski 2014; Wilkinson 2004). Qualitative studies, focussing on land grievances, have also examined the mobilization process through which electoral violence might occur (Klaus and Mitchell 2015).

Although global concern has increased in the recent years, disinformation has a long history (Lazer et al. 2018). The presence of disinformation campaigns during the electoral cycle has been detected in numerous cases and has been linked to the outbreak of election violence (Smidt 2020). Previous research has demonstrated the preventive effect of education campaigns on the occurrence of election violence through combatting the influence of disinformation within the violent

(7)

7

mobilization process during the electoral cycle (Smidt 2020). However, the general effect of variation in education levels on the occurrence of electoral violence has not been established. Given the potential role of disinformation in the mobilization process for election violence, as highlighted by the Nigerian general elections of 2019, research into this mobilization trajectory is necessary.

Combining findings in the fields of electoral violence, cognitive psychology and grievances literature, this study aims to examine the association between education and the mobilization and occurrence of electoral violence through the research question: How does the level of education influence the occurrence of electoral violence?

The study hypothesizes that education is negatively associated with electoral violence. Based on prior research, this paper argues that a decrease in average years in education decreases the average cognitive performance (Opdebeeck, Martyr, and Clare 2016). In turn, this increases the acceptance rate of disinformation (Pennycook and Rand 2019). Assuming that during the electoral cycle, anti- election disinformation is distributed (Judge and Korhani 2020), grievances against the election are more prominent if this disinformation is more likely to be accepted. Anti-election grievances give rise to mobilization against the election, which can spiral into electoral violence. Therefore, it is argued that education negatively impacts the occurrence of electoral violence; higher education levels covary with lower numbers of electoral violence, and vice versa.

Using sub-national data on education levels and the occurrence of electoral violence during election cycles worldwide from 2004 until 2012, the study quantitatively examines the hypothesized association. A negative binomial regression model including control variables and country fixed effects serves as the study’s main statistical test, while an OLS regression and a binary logistic regression function as robustness tests.

The results suggest that the hypothesis is not supported: the main negative binomial regression model, including control variables and country fixed effects, shows a significant positive association between education and election violence. Due to the hypothesis this study tests, no conclusions can be drawn from the positive association between education and election violence. It can only be concluded that no support was found for the hypothesized negative association between education and election violence. Data limitations, such as the systematic measurement error of occurrences of electoral violence in rural areas (von Borzyskowski and Wahman 2019), challenge the empirical validity of this finding. Future research could employ expert judgements on electoral violence occurrence, as suggested by von Borzyskowski and Wahman (2019), and mapping techniques based on sub-national variation to overcome the methodological limitations.

(8)

8

Literature review

Conceptualizing electoral violence

In order to examine the causes and consequences of the phenomenon at play, one has to define it first. Defining electoral violence entails differentiating it from other forms of violence. Höglund (2009) provides a characterization based on the motive, the actor and the timing of the violence, arguing that it is the motive and the timing which differentiate electoral violence. A similar argument is posited by Birch et al. (2020) who highlight the instrumental use of the violence. Both studies emphasize that electoral violence can be employed as a tool to (re)direct the trajectory of the elections by actors dissatisfied with the expected outcome; influencing the electoral cycle as a motive. Timing-wise, electoral violence occurs pre-election, during the election day(s) or post-election (Höglund 2009). The actors conducting the violence range from paramilitary groups, rebel groups, political parties and protestors to security forces (Höglund 2009). Election violence can come about in different activities such as riots, clashes, attacks and kidnappings, aimed at those involved in the electoral process (Daxecker et al. 2019). In conclusion, research distinguishes electoral violence from other forms of violence by its motive and timing; aiming to alter the course of the elections by employing violence pre-election, during election day and thereafter. Differentiating electoral violence from criminal activities and other forms of conflict is critical, given the role of elections for post-conflict stability (Brancati and Snyder 2013). Other implications of election violence include a (possible) decreased willingness among citizens to cast their vote (Condra et al. 2018;

Burchard 2015), even though violence is condemned by voters (Gutiérrez-Romero and LeBas 2020). Fear for election violence can in turn negatively affect general political knowledge, detrimental for the functioning of a democracy (Söderstrom 2018). The violence can threaten democratic institutions and question their legitimacy, paving the way for (renewed) modes of violent conflict in post-conflict societies (Höglund 2009). As such, understanding the determinants of election violence is critical to contribute to avenues of further democratization and stability in post-conflict areas.

On motives for electoral violence

Understanding the reasons for why actors move to commit electoral violence instead of (solely) contesting elections in a peaceful manner, requires distinguishing between distinct actors.

Incumbents apply electoral violence in order to stay in power after elections (Hafner-Burton et al.

2014). Intimidation and targeting of voters pre-election and during election day by security forces serves as a weapon for incumbents unwilling to accepts the norms of democracy (Hafner-Burton

(9)

9

et al. 2014). The employment of electoral violence by opposition groups, insurgents and other actors on the other hand, can stem from a strategy of targeting the symbolic institution of democratic elections, aiming to demonstrate the government’s inability to organize elections in an adequate manner (Condra et al. 2018). Post-election violence can be employed by both protestors and security forces to contest the results of the election (Birch et al. 2020). The election-related motive connects the acts of violence to the elections. Research on electoral violence often emphasizes the strategic element (Birch et al. 2020), in order to differentiate it from other forms of violence. Such an outlook aims to understand the occurrence of electoral violence primarily through a political elite-level perspective, where the move from a motive of electoral violence to the actual occurrence of the violence is influenced by factors which incentivise or discourage political elites, such as the strength of a political party, electoral designs and the strength of institutions.

Determinants of electoral violence

Research on the determinants of electoral violence has strongly focussed on structural elements through large N-studies (Birch et al. 2020). Explanations on why electoral violence occurs have been guided by rationalist, institutionalist and identity-based understandings (Klaus and Mitchell 2015). Rationalist underpinnings have established that violence is more likely to occur during competitive elections (Wilkinson 2004; Saleyhan and Linebarger 2015). In this environment, politicians could employ violence as a way to coerce, supress or frighten possible voters in order to secure their electoral mandate (Saleyhan and Linebarger 2015). Saleyhan and Linebarger (2015) conduct a large-N study on the outbreak of electoral violence in Africa and find quantitative support for the correlation between competitive elections and election violence.

Yet, evidence on the independently significant impact of competitive elections on the outbreak of violence during the elections is limited and other factors allow for more explanatory power on the occurrence of electoral violence (Birch 2020). These include institutionalist explanations on electoral violence variation (Birch et al. 2020). Fjelde and Höglund (2016) demonstrate in a quantitative study how majoritarian electoral designs, where the larger parties obtain a disproportionate number of seats, increase the possibility for electoral violence, compared to electoral designs following the proportional representation system. Majoritarian systems increase the stakes at play during the elections and might as such, incentivise elite actors to use political violence with the aim to prevent the loss of power (Fjelde and Höglund 2016). Further research has linked the strength of political parties to violent elections; the stronger a party’s organization the more it can rely on persuasive election mobilization instead of violence (Fjelde 2020). Weaker

(10)

10

organized political parties cannot do this effectively and might be incentivised to use violence when trying to reach their electoral aims (Fjelde 2020). Another possible determinant for the outbreak of election violence is the legacy of the country’s authoritarian past. The more exclusionary the former authoritarian regime was, the more the country is institutionally divided along ethnic lines, increasing the potential for violent political mobilization along these lines (Brosché, Fjelde, and Höglund 2020).

Identity-based theories explaining why electoral violence occurs have also been prevalent.

Wilkinson (2004) argues how ethnic polarization contributes to the strategic use of violence during the election cycle. Ethnic party elites aim to supress other ethnicities from voting and demonstrate the need to vote along ethnic lines (Wilkinson 2004). In ethnically polarized societies, the (ethnic) losers risk to lose any potential political power, and hence electoral violence becomes viable (Wilkinson 2004). A legacy of ethnic divisions in authoritarian regimes where no cross-ethnic coalitions existed can further contribute to violent mobilization along ethnic lines in the post- authoritarian age (Brosché, Fjelde, and Höglund 2020), highlighting the historical contexts within which structures of electoral violence might take place.

Mobilization for electoral violence

While examining the different theoretical strands that explain the occurrence of violent elections, it is often presented as a strategic instrument, yielded by elite actors with enough influence to initiate the outbreak of the violence. If they are incentivised to do so depends on the electoral, institutional and ethnic context. Theories explaining the occurrence of electoral violence through the agency of political elites provide an important understanding into the contexts that facilitate or hinder its occurrence by assuming elites are able to mobilize those willing to commit violence. Yet, this assumption can be challenged and the factors influencing the mobilization process, necessary for electoral violence to take place, can be scrutinized in order to understand why citizens turn into perpetrators of violence (Klaus and Mitchell 2015). Given the distinct nature of electoral violence (Höglund 2009), the mobilization process requires a particular explanation. Previously, Klaus and Mitchell (2015), employing a comparative case study, have demonstrated how grievances surrounding the distribution of land operate as a tool to mobilize people into committing electoral violence; - if these land grievances are constructed upon on a security threat. Elections serve as a mechanism for the distribution of power and their outcome can be framed as a threat to the land rights of particular social groups (Klaus and Mitchell 2015). This narrative does not have to stem from factual evidence in order to function as a mobilization mechanism against the process and outcome of elections (Klaus and Mitchell 2015). Lies and rumours about the prosperity or troubles

(11)

11

one will face after the election, depending on the outcome, can also operate as a mechanism to incentivize anti-election mobilization.

Continuing the study of mobilization for electoral violence, Smidt (2020) examines how education campaigns about the electoral process by the UN peacekeeping mission in Cote d’Ivoire influence the sub-national variation of the occurrence of electoral violence in the country. The study’s findings establish that election-education-events can reduce violence during elections by limiting mobilization (Smidt 2020). Theoretically, Smidt’s argument is built up on the relation between the spread of disinformation during the electoral cycle and the use of violence. Disinformation campaigns can serve as a tool for those mobilizing citizens into violence, by providing narratives of unfairness, exclusion and ineffectiveness which question electoral legitimacy and exacerbate societal divisions (International IDEA 2021). Ultimately, disinformation could push grievances (Wilkinson 2004). As such, Smidt (2020) argues that education campaigns can help citizens to resist the disinformation which could prompt them into violence. The employment of education campaigns challenges disinformation narratives that portray elections as unfair, fraudulent and ineffective (Smidt 2020). In conclusion, various studies have examined the factors which might contribute to violent mobilization during elections. Quantitative studies on these factors are limited though. Specifically, the association between disinformation and election violence has not been examined extensively.

Contrasting Smidt’s argument (2020), a variety of authors have hypothesized on a positive relation between education-related factors and the occurrence of electoral of violence (von Borzyskowski and Kuhn 2020; Söderberg Kovacs and Bjarnesen 2018 pp.87-113). While Smidt’s causal mechanism (2020) considers education as an impediment to electoral violence mobilization, both Von Borzyskowski and Kuhn (2020), and Söderberg Kovacs and Bjarnesen (2018) posit that more informed citizens canresult in an increase in electoral violence. Differentiating the actors and the timing of the violence allows for an understanding into this discrepancy; Von Boryzyskowski and Kuhn (2020) demonstrate that the positive relation between education and electoral violence occurrence is due the fact that higher educated voters are more likely to be targeted by security forces.

Söderberg Kovacs and Bjarnesen (2018) establish how urban areas are more likely to experience post-election violence, as more informed voters are more likely to protest after elections which were perceived as fraudulent. Aiming to explain post-election participation against perceived fraudulent elections, Young (2020) asserts that one’s general self-efficacy can increase the willingness to attend anti-government protests which could spiral into violence. If one assumes that education contributes to increased feelings of self-efficacy, it is possible that higher-educated citizens are more likely to engage in election violence (Young 2020), as established by Söderberg Kovacs and

(12)

12

Bjarnesen (2018). These empirical findings challenges assertions of a negative association between education and electoral violence.

Yet, Smidt’s causal argument (2020 - better educated voters results in a decrease in electoral violence - can be strengthened by identifying the scope conditions of her argument, being violence committed by non-state forces before/during the election day(s), as disinformation primarily functions as a mobilization mechanism for non-state actors before and during election day(s). The extent to which education could prevent electoral violence committed by state actors (e.g. security forces) seems limited, given that these actors are less likely to be mobilized into violence due to anti-election disinformation campaigns (Smidt 2020). The temporal scope condition can be introduced due to the empirical finding of how during the post-election phase, higher-educated citizens are more likely to protest possible fraudulent elections (Söderberg Kovacs and Bjarnesen 2018 pp.87-113).

The missing link: disinformation?

Research on the determinants of electoral violence has concentrated on factors which incentivise or discourage elites from opting for violence. Complementing this research, a couple of studies have examined the mobilization process leading to electoral violence (Klaus and Mitchell 2015;

Smidt 2020). Quantitative research on the factors conducive to (un)successfully mobilizing citizens into perpetrators of election violence has been scarce however. In order to prevent violence during the electoral cycle, and the potential breakdown of democracy subsequently, it is crucial to understand the entire range of factors influencing occurrence of electoral violence, including those impacting the mobilization. Building upon Smidt’s causal story on the role of disinformation on electoral violence mobilization (2020), this study aims to contribute to research on the determinants of the violence by examining how education influences the occurrence of electoral violence on a sub-national level. While Smidt (2020) specifically examines how UN education campaigns help to prevent electoral violence following from disinformation, this research focusses on the relation between the general education level and electoral violence, and hence complements her study. In order to provide a comprehensive overview of the role of education level on the occurrence of electoral violence, this thesis applies a sub-national cross-country quantitative study, answering the research question How does the level of education influence the occurrence of electoral violence? Examining a possible determinant of election violence could pave the way for policy recommendations that decrease the possibility for violent elections while emphasizing the government’s responsibility to provide quality basic education for all.

(13)

13

Theory

In order to elaborate on the theoretical underpinnings of this research, the separate elements within the theoretical framework are defined first. These are the concepts of electoral violence, education, cognitive performance, disinformation and grievances. Thereafter, the causal argument, the hypotheses stemming from the causal chain and the scope conditions are presented. Combining findings in the fields of electoral violence, cognitive psychology and grievances literature, this study aims to construct an interdisciplinary theory on the association between education, disinformation and election violence.

Electoral violence

As conceptualized in the literature review section, this study follows an understanding of electoral violence defined by timing, actor and motive (Höglund 2009; Birch et al. 2020). Arguably evident by its name, electoral violence distinguishes itself from other types of violence by its direct relation to the holding of elections; when no elections are organized, it does not occur (Birch et al. 2020).

Electoral violence can occur pre-election, during election day and post-election day. Actors able to commit the violence range from security forces to protesters, while it can entail actions such as attacks, kidnappings, intimidations and violent riots. The motives of the violence are often portrayed as changing the perceived course of the election and/or showcasing disproval with the election. This research zooms in on the occurrence of violence before election day and at the election day(s), excluding post-election violence. Additionally, electoral violence committed by state forces will be excluded from the hypothesis. The motivations for these delimitations are laid out below. The study uses the following defintion of electoral violence: violent actions before election day or on election day(s), undertaken by actors who are not affiliated to the state, with the motive to express dissatisfaction with the electoral process and/or change the outcome of the election.

Education

Ever since Seymour Martin Lipset studied the social requisites of democracy (1959), scholars have examined the level of education in society as a possible factor influencing the functioning of democratic institutions. In Latin America, educational improvements are considered to have been decisive for the furthering of democracy in the region (Valverde 1999). The independent effect of education on the presence of democracy has been contested; some studies find a significant impact of education on democracy (Glaeser et al. 2004) while others identify such an impact only in interaction with economic variables (Boix and Stokes 2003). After controlling for country fixed effects, Acemoglu et al. (2008) establish that variation in education levels across-countries does not

(14)

14

correlate with democracy. Nevertheless, education has been connected to furthering a wide array of socio-economic and health factors, such as gender-equity, child mortality and social mobility (Graetz et al. 2018).

Comparing the level of education globally is often conducted through the average years of education per spatial unit (Smits and Permanyer 2019). The average years of education contains the average number of completed years of education of a country’s population aged 25 years and older, excluding years where grades were repeated (UNESCO 2008). Data also covers the average expected years of education for children per country. Comparing education by means of years of attended education might not incorporate the influence of other insightful factors, such as the quality of the attended education and the social context of the schooling (Lövdén et al. 2020). Due to data scarcity however, research usually employs average numbers of years in education when comparing education statistics globally (Lövdén et al. 2020).

Average years of education rates vary strongly on a global cross-country scale; generally, states with higher levels of socio-economic development have a higher number of average years of education compared to states with lower levels of socio-economic development (UNDP 2018). In 2004, the Swedish population for instance, had an average of 16 years of completed education, while the Bangladeshi population averaged 8,2 years of completed education (UNDP 2018). Examining national years of education rates more closely, sub-national variation looms around the corner.

Research on Sub-Saharan Africa points out that populations in rural areas are likely to have lower numbers of average years in education, compared to those living in urban areas (Graetz et al. 2018).

The same study finds that in Sub-Saharan Africa, gender inequalities regarding educational attainment are frequent; the average years a woman has attended educational institutions is lower than the average years a man has followed education (Graetz et al. 2018).

(15)

15

Figure 1: Map on the global variation in the average years in education.

Figure 1 (Smits and Permayer 2020) provides an adequate overview on both national and sub- national variation of the average years in education in 2004 on a global scale. Red colours indicate relatively low numbers of average years in education, while blue colours indicate relatively high numbers of average years in education.

Cognitive abilities

The attainment of education has been associated with an increase in the performance of one’s cognitive abilities (Opdebeeck, Martyr, and Clare 2016). Cognitive abilities are here defined as processing aspects of cognition, including reasoning, vocabulary, literacy, general and specialized domain knowledge, while cognitive performance is defined as the functioning of these abilities (Lövdén et al. 2020).

Opdebeeck, Martyr and Clare (2016) establish through a meta-analysis of 109 studies that an increase in years of education is associated with an increase in cognitive performance. The association between education years and cognitive performance is robust when controlled for by variables such as race, gender and country (Opdebeeck, Martyr, and Clare 2016). Hence, one can

(16)

16

assert that an increase in years in education entails an increase in the performance of one’s cognitive abilities.

Disinformation and the perception of disinformation

Although global concern has increased in the recent years, disinformation has a long history (Lazer et al. 2018). Often used in overlap with the term ‘fake news’, this study defines disinformation as

“false information that is purposely spread to deceive people” (Lazer et al. 2018). This definition does not specify how the false information is spread; the term ‘fake news’ might be strongly connected to the online realm, but this study adopts a perspective on the spread of disinformation through all possible means, ranging from for example radio, TV, written media outlets as well as social interactions.

The use of disinformation campaigns during elections cycles has been detected in a significant number of elections around the world (Mutahi and Kimari 2020; Judge and Korhani 2020). For a functioning democracy, the free flow of political discourse is critical, and hence, regulating the spread of disinformation can be a challenge for incumbent governments; responding through repression might endanger the preservation of rights embedded in the constitution while a passive response could allow the disinformation flow to roam freely (Judge and Korhani 2020).

Substantively, electoral disinformation often questions the integrity and fairness of the elections (Judge and Korhani 2020), and highlights the ineffectiveness of democratic institutions (International IDEA 2021). It has also been associated with the spread of fear and the violent intensification of in-out group tensions (Smidt 2020). If the elections are fraudulent because of the manipulations of the out-group, then, the electoral disinformation aims to tell its reader, the in- group has to contest these elections.

Examining the cognitive process of how disinformation is accepted as valid information, two mechanisms can be detected (Pennycook and Rand 2019). Kahan et al. (2017) argue that the (political) bias of the one processing the disinformation significantly influences the likelihood of the disinformation being regarded as valid, as political and ideological preferences reign over our cognitive abilities. However, research conducted by Pennycook and Rand (2019) points out that it is not the bias of the disinformation processor which significantly impacts the likelihood of the disinformation being perceived as valid, but it is the cognitive reasoning employed by the one processing the disinformation. Strong cognitive reasoning can counter political biases in order to accurately identify disinformation as false (Pennycook and Rand 2019). The association between cognitive reasoning and disinformation holds in the opposite way as well; if cognitive reasoning

(17)

17

abilities are lower, then one can expect to be more susceptible to accept the claims of disinformation (Pennycook and Rand 2019).

Grievances

Accepting the validity of violence-inciting disinformation which questions the integrity of the elections can spark feelings of grievance against for instance, the elections as an institution, the group that is portrayed as in control of the elections and other societal actors depicted in a harmful way in the disinformation. Research on grievances and violence has a long history: in 1970, Gurr explained the outbreak of violence through the theory of relative deprivation, in which grievances regarding unfilled expectations are mobilizing people to take action against their situation. From there on, a myriad of researchers has examined the effect of individual and collective grievances, often measured through economic indicators, on the outbreak of violence, with mixed results (for instance Brush 1996; Østby 2008; Cederman et al. 2011). Besides economic-based inequalities, Wood (2003) highlights the contextual importance of grievances by showing how collective moral outrage can contribute to violent mobilization. It is not in the scope of this study to further assess how grievances might explain conflict. Yet, based on prior research, this study assumes that grievances can contribute to violent mobilization, whilst not being the only explanatory factor of why some people turn towards violence, while others remain peaceful.

Causal story

Having defined the theoretical elements necessary to explain the study’s research question of how the level of education influences the occurrence of electoral violence, the paper presents the theoretical argument, centred around the relation between education and the occurrence of electoral violence.

As established by the literature, an increase in the average years in education for a certain population, will result in an increase of the population’s cognitive performances. Likewise, a decrease in the average years of education for a certain population, consequently results into a decrease of the population’s cognitive performances. Given the assumed linear relationship between years in education and cognitive performance, one of the sub-elements of one’s cognitive performance, one’s ability to reason, can also be expected to decrease or increase based on the amount of time spend in education (Opdebeeck, Martyr, and Clare 2016). This ability has been strongly associated with one’s acceptance of disinformation; the more one engages with pieces of disinformation by reasoning, the less likely is one to accept this disinformation as valid (Pennycook and Rand 2019). On a population level, this entails that disinformation is more likely to be accepted when average cognitive reasoning abilities are low, compared to when these abilities are on average

(18)

18

high. Hence, given the influence of education on cognitive reasoning, one could expect that disinformation is more likely to be accepted when the number of average years of education are lower, compared to situations when this is higher.

As disinformation campaigns have been linked to the spread of grievances against the electoral process (Smidt 2020; Human Rights Watch 1995), an increased acceptance of disinformation during the electoral cycle will consequently increase the anti-election grievances. Grievances, albeit not coming close to being the sole factor in explaining violent mobilization, do provide a narrative of how citizens turn into perpetrators of violence. As such, an increased number of grievances in a certain population contributes to parts of the population mobilizing violently, in this case against the electoral process and those involved. In conclusion, this paper asserts that one of the factors explaining electoral violence mobilization is the lower average years of education in a certain population, enabling an increased acceptance of disinformation, resulting into anti-election grievances and mobilization. Consequently, the following hypothesis can be posited:

H1: The lower the average number of years in education, the higher the occurrence of electoral violence.

It should be noted that the causal argument moves from the group level (average years in education) to the individual level (acceptance of disinformation) and then returns to a group level outcome again (electoral violence).

Additionally, concentrating on the causal chain, one can also expect to witness an increase in mobilization against the elections, hence the following hypothesis:

H2: The lower the average number of years in education, the higher the number of people mobilized for electoral violence.

Causal diagram

Followingly, the theory’s scope conditions have to be introduced. As provided for in the study’s defintion of electoral violence, this theory applies to cases of electoral violence conducted by non-

Average years in education Occurence of electoral violence

Potentially lower average cognitive performance

Increased acceptance of disinformation on the election

Increased grievances against the election

Increased mobilization against the election

(19)

19

state actors before and during election day, as violence by government forces is not dependent on the mobilization process laid out in the causal story. Post-election violence has been associated with contesting fraudulent electoral behaviour (Young 2020), and hence does not have to follow from disinformation campaigns. The theory aims to account for sub-national variation in the occurrence of electoral violence, examining the association between the variables at a sub-national level.

Different factors could serve as confounders to the theorized relation between education and the occurrence of electoral violence. The socio-economic development of a region could both impact the years in education (Bils and Klenow 2000; Córdoba and Ripoll 2013) and the occurrence of electoral violence. Lower socio-economic development could result in more violent mobilization, including mobilization against elections. Studying sub-national variation, differences between rural and urban areas have to be taken into account, as rural regions are likely to have lesser developed educational institutions compared to urban areas (Graetz et al. 2018), yet might be more likely to experience grievances against the centralized government (Cederman et al. 2011).

It is important to note that the causal story this study presents, is only one causal story among many, explaining how violent mobilization during the electoral cycle occurs. This research aims to examine whether this causal story is credible when tested for in the empirical world. It does not posit that in general lower educated human beings are more prone to be violent, but rather examines the influence of disinformation on the association between education and violent mobilization. A wide array of distinct pathways that lead to violent mobilization exists.

(20)

20

Research design

In the previous section, the hypothesized relation between education and the occurrence of electoral violence has been outlined. Consequently, a research design which examines the empirical significance of this proposed association is introduced. This study applies a negative binomial regression model on the sub-national cross-country variation of average years in education and electoral violence occurrence. The proposed causal mechanism, the role of disinformation on the mobilization process for electoral violence, is not assessed, although by testing the number of people mobilized for electoral violence the study aims to examine the last part of the causal chain.

By constructing a new hypothesis on the mobilization process for electoral violence and testing this hypothesis empirically, this thesis hopes to combine both components of theory-building and theory-testing, aiming to further research on the mobilization process for electoral violence. In the following sections, the research design necessary for this study is presented.

Universe of cases and unit of analysis

The study’s universe of cases are national elections, held in multi-party regimes, where voters are able to cast their votes for more than one party. Elections in one-party states, such as Eritrea or China are not taken into account for the theorized relation between education and the occurrence of electoral violence. The unit of analysis of this thesis is the sub-national unit per national election cycle.

The time period covered in the thesis ranges from 2004 until 2012. The study covers 433 national election rounds in 157 countries, measuring 4667 sub-national units in total.

The range of the time period, from 2004 until 2012, is due to data limitations and does not impact the hypothesis of the study. It will be assumed that potential findings can be generalized to extent to the entirety of the post-Cold War Era, including the current day (2021). Due to the time period covered and the type of cases this study is concerned with, being national election rounds, the study’s universe of cases contains differing amounts of elections per country. While most countries have multiple national election rounds between 2004 and 2012, some countries have one election round in the time period covered. As visible in figure 2, the largest number of countries (66) has two national election rounds in the time period. Countries with more than four election rounds are scarce: twenty countries qualify for this criterium. Low outliers are also minimal as only 16 countries have one national election in the time period measured. Hence, although the range of the time period does create a variation in the amount of national elections in the data, the amount of national elections is relatively evenly distributed and can provide generalizable results.

(21)

21

Figure 2: Distribution of the amount of national election rounds per country

In order to control for the different frequencies in national elections, which could influence the variation of electoral violence based on the particular characteristics of the given countries, country fixed effects are added to the model. This will be expanded on in a later segment. Similarly, control variables will be included in order to control for differences between regions. Due to data limitations, only national election cycles can be examined. However, as argued for in the theory section, it is assumed that the causal mechanism and association studied in this thesis can be generalized unto local elections as well, even though the data allows for an examination of national election cycles only.

The ‘sub-national unit’ entails the largest sub-national administration level per country; for some countries these levels are called provinces while others might refer to them as states. Different countries have different amounts of the largest sub-national unit, populated by different numbers of inhabitants. In order to account for possible variation in the dependent variable due to the differing population sizes, the study will control for population, as elaborated on in a subsequent section. Given the significant variation in within-country education rates (Graetz et al. 2018), examining sub-national education rates is critical if one aims to study the association between education and electoral violence. The study applies observational data.

(22)

22

Choice of method

As this paper hypothesizes on a linear relationship between the education level and the occurrence of electoral violence and aims to study the correlation between these two variables, an Ordinary Least Squares (OLS) regression might seem like an appropriate choice. OLS regression provides an estimated linear correlation between the independent variable and the dependent variable by minimizing the difference between the observed values and the predicted values. OLS models assume normally distributed residuals, and as such, can potentially include negative values, providing a challenge for data that conceptually does not allow for negative predicted values. The data on the dependent variable for this study classifies as count data, containing the counted number of occurrences of electoral violence. The variable’s distribution is discrete. Theoretically, it is impossible for values of the dependent variable to go below 0, clashing with the aforementioned OLS distribution. Hence, potentially providing skewed predictive values in directions theoretically considered impossible, OLS is not suitable as the main regression model of the study. Figure 3 displays the distribution of the study’s dependent variable, the number of occurrences of pre-election and during election day(s) violence, highlighting how the DV’s distribution does not contain any negative values due to the data structure.

Figure 3: Distribution of the frequency of events of pre-election during election day(s) violence

Based on the DV’s discrete distribution, the Poisson model provides an alternative. Poisson models assume the absence of overdispersion in the distribution; the variance and the mean are assumed

(23)

23

to be equal (Fisher, Hartwell, and Deng 2017). However, due to the structure of count data, a check on the presence of overdispersion needs to be conducted. The results of this test (visible in the appendix) indicate that the DV’s distribution is overdispersed: the variance is significantly larger than the mean. Therefore, instead of the Poisson model, the negative binomial distribution should be considered. Given its abilities to account for discrete variables where the mean and the variance are not equally distributed (Fisher, Hartwell, and Deng 2017), a negative binomial regression is be employed as the study’s main regression model.

Nevertheless, as visible in figure 3, the DV’s distribution contains a large number of regions where no electoral violence took place during the election cycle. These zeros account for a majority of the cases. Next to overdispersion, count data is frequently challenged by zero-inflation, indicating a variable with an excess number of zeros (Fisher, Hartwell, and Deng 2017). Zero-inflated (ZI) regression models aim to capture these ‘structural zeros’ by assuming they are generated by a different process than the process that explains positive variation and ‘sampling zeros’ (Fisher, Hartwell, and Deng 2017). For instance, the number of fish caught by campers can be zero because of two distinct processes: the campers decided not to go fishing (structural zeros) or the campers’

skills proved to be inadequate to catch anything (sampling zeros) (Fisher, Hartwell, and Deng 2017). Different data generating processes give rise to the zero. As such, ZI models provide two distinct regressions: one predicting the structural zeros and one predicting the positive outcomes and the sampling zeros. ZI models have been effectively applied in conflict research (Holmes, Gutiérrez De Piñeres, and Curtin 2007), yet have also faced scrutiny. Allison (2012) has argued that structural zeros can be included into negative binomial distributions, whose model fit is not significantly different from ZI models. Most importantly however, Fisher et al. (2017) emphasize that the employment of a ZI model should fit the data structure; the structural zeros need to be accounted for by a separate process in order for a ZI model to be adequate. Although the distribution of the dependent variable of this study is skewed towards zero, no theoretical underpinning on why electoral violence cannot hypothetically occur in the alleged ‘structural zeros’

exists. Supported by Allison’s (2012) findings on the comparative model fit of negative binomial regression models, this study employs the latter.

The negative binomial regression necessitates several model assumptions to hold in order to provide valid estimates. Linearity between the model’s independent variable and dependent variable needs to be assumed, and given the thesis’ hypothesis, can expected to be present. The individual observations are assumed to be statistically independent. In order to guarantee this, control variables are employed on sub-national level and on country level, including the application of fixed effects. These steps limit possible bias in the model’s estimations and ensure that there is no

(24)

24

correlation between the error term and the independent variable. Tests for multicollinearity are employed. Constant variance (homoscedasticity) is not assumed by the negative binomial distribution (Fisher, Hartwell, and Deng 2017) and hence, does not need to be controlled for.

Instead, the assumption of overdispersion is confirmed through the previously referenced test (see appendix).

Justification of the method

The choice for an observational quantitative research design is guided by several advantages, yet also faces limitations. Most importantly, the choice is guided by the research paradigm of positivism, assuming that a single tangible reality exists and can be measured (Reiter 2015). From these measurements, generalizations about the world can be drawn (Reiter 2015). As such, a large- N study allows for generalizability of the results of this thesis, even though the time-series included in this thesis covers cases between 2004 and 2012. As no previous research has examined the potential covariation between education levels and the occurrence of violence during the electoral cycle, the possibility to map out this potential covariation is a critical advantage of a large-N study.

However, the concentration on the covariation between the two variables limits the study’s ability to probe into the hypothesized causal chain and for instance, qualitatively compare two election rounds in order to gain an understanding on the role of disinformation campaigns on violent mobilizations during the electoral cycle. The choice for a quantitative large-N research design ensures that this thesis has to assume the presence of disinformation campaigns during electoral cycles.

Operationalizations and data

Having established the scope of cases of the study, a closer look into the variables at play is provided. Operationalizations of the independent variable and the dependent variable are presented in tandem with the data employed by the research. Thereafter, the control variables are introduced through the same procedure.

Firstly, the independent variable of education level is operationalized as the average years of education on a population level. Given the scope of the study, this entails the average years of education per region per election year. The variable includes both primary education, secondary education and other forms of education such as university degrees and does not make a distinction between the different educational stages. Additionally, the quality of the education is not included in the operationalization, potentially providing the same value for different forms of average years in education per region. This could challenge the study’s hypothesized causal step from more

(25)

25

education to stronger cognitive performance. However, given the wide-ranging data on the average years of education on a sub-national level, this study adopts an operationalization of education level solely looking at the average years of attended education. While this might not capture the entirety of the educational picture, it does allow for cross-country sub-national comparison. Moreover, although educational quality is not equal to educational attendance rates, it is assumed that there exists an overlap between both variables; stronger educational systems are likely to be attended for longer periods of time by a populace. Therefore, higher average years in education contribute to stronger cognitive performance. Hence, this thesis employs the average years of education per region as its independent variable.

As this thesis applies a large-N study to investigate the covariation between sub-national data on the average years in education and the sub-national occurrence of electoral violence, data on both the independent and dependent variable has to be coded sub-nationally. The programme ArcGIS has been employed to structure the data accordingly.

The independent variable of the study, the average years of education per region per year, is obtained from the Sub-National Human Development Index (Smits and Permanyer 2019). The index aggregates household surveys from the Demographic and Health Surveys, conducted by ICF International (Smits and Permanyer 2019). These surveys cover around 3000 to 30 0000 households per region per year and allow for aggregation to a population level, as done by the index (Smits and Permanyer 2019). For years of which survey data on education years is lacking, the index has established a method of extrapolating from existing data, and as such, is able to produce an estimation of the value. However, as this method is solely based on existing data, data scarcity can challenge the reliability of the estimation model. Due to missing observations for certain data points, the study requires the use of the estimation model. In order to obtain a balance between an adequate number of cases and reliable data on the independent variable, this thesis limits the amount of extrapolation to five years, meaning that the index is able to estimate observations on the average years of education up to five years from the last observed data point. Doing so, allows the study to use the data on average years of education per region from the year 2004 to the year 2012.

Secondly, the occurrence of electoral violence, the main dependent variable of the study, is operationalized as the sub-national number of occurrences of pre-election and during election day(s) violence per national election year. No specific forms of electoral violence are excluded.

Following the causal story, electoral violence committed by government forces and international actors is excluded from the study’s operationalization of electoral violence. Additionally, the study

(26)

26

assumes the spread of disinformation targeting elections during the election cycle, based on the hypothesized causal mechanism. In order to account for the possibility of violent protests sparked by the incumbent’s fraudulent behaviour during the elections, electoral violence committed before and during the election day(s) is computed as the dependent variable, excluding post-election violence. Although limiting the amount of electoral violence variation potentially explained by other mobilization processes which are not part of the causal story (for instance protesting electoral fraud), this methodological decision does not completely limit the possibility to explain the outcome variable by other means. A secondary dependent variable, encompassing pre-election, during election day(s) and post-election violence is employed as a robustness check.

Aggregating particular occurrences of electoral violence into the number of electoral violence instances per sub-national unit allows for large-N comparison on electoral violence frequencies globally. However, this aggregation limits the amount of information stored on particular instances of electoral violence; the type of violence, the perpetrator, the target and the duration of the violence are not taken into account in this model. Assuming that there exists variation in types of electoral violence, this study incorporates the number of people mobilized per particular instance of violence and aggregates these amounts into one number on the number of people mobilized for electoral violence per sub-national unit per election year. Doing so, touches upon the last element of the thesis’ hypothesized causal mechanism, and allows for a test on the second hypothesis of the study: The lower the average number of years in education, the higher the number of people mobilized for electoral violence.

Various datasets on the occurrence of electoral violence exist. Due to the thesis’ scope condition on specific actors, a dataset which differentiates the actors involved in the electoral violence is necessary. Additionally, the dataset has to code the occurrence of electoral violence geographically, in order to allow for conversion into data points per region through ArcGIS. Both of these prerequisites are met in the Electoral Contention and Violence (ECAV) dataset (Daxecker, Amicarelli, and Jung 2019). ECAV covers all the instances of electoral violence from 1990 till 2012 in national elections rounds of unconsolidated multi-party regimes (Daxecker, Amicarelli, and Jung 2019). Unconsolidated regimes are defined as every non-OECD state in 1990 (Daxecker, Amicarelli, and Jung 2019). As the scope of the dataset surpasses that of the study, several transformations of the dataset are employed: (1) the instances of electoral violence are limited to the years 2004-2012, (2) electoral violence committed by government actors and international forces, such as NATO forces, is omitted from the data due to the thesis’ hypothesized causal mechanism and (3) instances of electoral violence for which regional geographic coordinates are not available are omitted from the data. The particular instances of electoral violence per sub-

(27)

27

national unit per election year are aggregated to provide frequencies on the number of electoral violence occurrences. Given ECAV’s concentration on election cycles where electoral violence occurred, the study employs the NELDA election dataset to add the missing observations (Hyde and Marinov 2012). These are election cycles were no electoral violence occurred during national election cycles. As such, the data on the DV is compatible with the data on the IV.

The variable concerning the number of people mobilized for all events of electoral violence per sub-national unit per national election cycle is estimated through data in the ECAV dataset. As ECAV provides ordinal data on the number of electoral violence participants, the median of the corresponding amount of participants per ordinal value is used. For instance, an event coded with the ordinal value of 1 corresponds to a number of participants ranging from 1 to 10. Followingly, the median (in this case 5) is used for this study and aggregated together with all the other estimated numbers of mobilized people per event of electoral violence. Due to the ordinal data structure of ECAV on the mobilization per event, this approach can only provide an estimate and one has to assume that the mobilization numbers, hiding behind ECAV’s ordinal values, follow a normal distribution with 50% of the values being less than the assumed median and 50% being greater than the assumed median. Otherwise, the calculated estimation cannot serve as a reliable aggregation of the number of people mobilized per election cycle per sub-national unit.

ECAV data on mobilization numbers per event of electoral violence is lacking for a majority of the cases of electoral violence. Therefore, the computed values within the study’s data on mobilization for electoral violence per sub-national unit per national election cycle provide a limited overview on the amount of electoral violence mobilization. These data limitations will be taken into account during the analysis. Nevertheless, the available data on variation in electoral violence mobilization offers an insight into the hypothesized causal mechanism, probing into the association between education and the amount of mobilization for electoral violence.

The ECAV dataset has been coded manually by human coders based on news reports of electoral violence. Although a wide array of different search terms has been applied in the process, the usage of news coverage to provide an empirically valid description of the occurrence of electoral violence worldwide faces challenges. As emphasized by von Borzyskowski and Wahman (2019), measurement error in the study of electoral violence is a common phenomenon, yet with potentially critical outcomes, by for instance establishing covariation where no actual covariation exists.

Skewed results due to media-bias, resulting in over- and underreporting, have been examined in the wider field of peace and conflict studies (Eck 2012). Measurement validity is endangered due to differences in the reported numbers of electoral violence and the actual occurrence of electoral

(28)

28

violence, if the latter is mistaken for the former. Additionally, and relevant for the focus of this thesis, Borzyskowski and Wahman (2019) establish that especially electoral violence incidents in rural areas are prone to be underreported. The solution proposed by the two authors entails a move towards expert judgment when examining electoral violence through small-N studies (Borzyskowski and Wahman 2019). However, for large-N enquiries into electoral violence, such an approach is not feasible and given the prominent role of media-based datasets on electoral violence, the only way forward is defined by awareness and carefulness when describing and interpreting results.

Control variables

Given the previous research on the determinants of electoral violence, and the potential for variables which could explain variation in both the independent variable and the dependent variable and variables which could variation in the dependent variable, the study includes several control variables.

On a sub-national level, the study includes four variables which could serve as possible alternative explanations to the occurrence of electoral violence or as confounding variables for variation in both the independent variable and the dependent variable. The latter could be the case for the gross income per capita per region. Regional differences in economic prosperity covary with regional differences in the strength of educational institutions (Acemoglu et al. 2008) . Therefore, it can be assumed that in regions with lower gross income per capita levels, financial support from the state towards educational institutions will be lower, decreasing the average years spend in school.

Additionally, lower regional gross income increases the possibility for the occurrence of violence, including electoral violence, by decreasing people’s cost to engage in violent action (Collier and Hoeffler 2004). Hence, the regional gross income per capita could serve as a confounding variable, influencing both the occurrence of electoral violence and the education level in a region.

The second sub-national variable included in the model is the regional urbanization rate, defined as the percentage of the population living in urban areas per region. As previous research has pointed out how urban areas are more likely to experience electoral violence (Söderberg Kovacs and Bjarnesen 2018 pp.87-113), it is critical to control for the moderating effect the urbanization rate might have on the association between education level and the occurrence of electoral violence.

This moderation effect is based on the assumption that more urban areas are more likely to have better educated residents than more rural areas, and that hence, following the hypothesis of the study, they are simultaneously more likely and less likely to experience electoral violence. By controlling for the effect of urbanization on the occurrence of electoral violence, the thesis can

(29)

29

move away from this contradiction and focus on the possible association between education and electoral violence. Nevertheless, it should be noted that the regional urbanization rate does not display within-region differences between urban centres and rural areas per region, providing a drawback to urbanization as a control variable. Due to the unit of analysis of the study, controlling for within-region differences is not possible. Distinct national measurements on the urbanization rate provide an additional hurdle to its inclusion within the study’s model. Countries define ‘urban’

differently; while the Democratic Republic of Congo considers settlements with a population of more than 2500 to be urban, Mali puts the limit for urban settlements at more than 25 000 inhabitants (Ritchie and Roser 2018). Such distinct operationalizations threaten the global comparative strength of the urbanization rate variable. Accepting the discrepancy between national operationalizions of the urbanization rate, this thesis cautiously employs the variable in order to control for its effect on the occurrence of electoral violence.

The third sub-national variable that is added into the regression model aims to capture the challenge posited by the operationalization of the dependent variable; electoral violence as measured through the occurrence of electoral violence. Sub-national differences in population numbers can strongly influence the electoral violence instance rate. Therefore, by controlling for the population per region, the study is able to examine the possible association between education and electoral violence without potentially incorporating the influence of the amount of region inhabitants on the occurrence of electoral violence.

The last sub-national variable to be included into the regression model is the presence of conflict at the sub-national unit during the year of national elections. The presence of conflict serves as a estimator for the general stability in the region, and could function as a predictor for the outbreak of electoral violence. Additionally, conflict presence could have a negative effect on the average years of attended education by for instance threatening and damaging the strength of educational institutions. Therefore, it might operate as a confounding variable, negatively effecting the independent variable and positively effecting the dependent variable and should be included within the model. Following the UCPD, the study operationalizes the presence of conflict as state-based armed conflict which results in at least 25 battle deaths per year, where the government is one of the warring parties (Sundberg and Melander 2013). Data on this variable is obtained from the UCDP GED dataset (Sundberg and Melander 2013).

In order to account for variation in the occurrence of electoral violence due to country-specific factors, including previously established determinants of electoral violence such as institutional strength and electoral design, the study’s regression model includes country fixed effects. Doing

Education, Disinformation And Electoral Violence