Abstract
Documenting election violence is central to safeguarding electoral integrity, but collecting such data is difficult. While citizen crowdsourcing is often seen as a cost-effective alternative to traditional monitoring, we argue that monitors offer unique advantages through their training and insulation from local pressures. Our field experiment during Côte d’Ivoire’s 2020 presidential election assessed whether monitors enhance election violence documentation when used alongside citizen reporting. We found that the presence of a monitor increased the likelihood of violence being reported by 10.7 percentage points without affecting citizen behavior. Monitors with more geographic experience were more likely to report incidents, regardless of their proximity to their home communities. These findings highlight the importance of monitors in revealing latent violence and in enhancing the scope and depth of reports. This demonstrates the critical value of monitors, even as crowdsourced data become more prevalent.
Introduction
The global spread of democracy has enabled millions to participate in elections, but violence can emerge where partisan rivalries, fragile institutions, and fraught histories intersect (Birch 2020; Gillies 2011; Hafner-Burton et al. 2014; Straus and Taylor 2012). The international community has sought to mitigate election violence through interventions like electoral management reform and dispute resolution (Birch and Muchlinski 2018; UNDP 2009). However, efforts to assess and improve these interventions have been limited by a lack of reliable information on where and when violence occurs (Daxecker et al. 2019; Fjelde and Höglund 2022; Von Borzyskowski and Wahman 2021). Accurately documenting election violence remains a challenge because it requires data that can detect incidents across an entire country (breadth) and fine-grained enough to distinguish varied forms of coercion that differ in political significance (depth). 1
The act of reporting is at the heart of violence measurement. All measures of election violence, including event data, are byproducts of reporting: the process by which an individual who witnesses or learns about a violent act transmits information to an authority or organization that makes it public. Effective policy interventions depend on reliable information on the occurrence of violence. Therefore, advancing our understanding of violence reporting is a necessary step toward safeguarding election integrity.
Election violence reporting has traditionally relied on monitoring. For decades, the international community has deployed monitors from civil society and multilateral organizations to document violent incidents (Garber 2020). However, some studies suggest that monitors can inadvertently displace violence (Asunka et al. 2019; Daxecker 2014; Ichino and Schundeln 2012), while others argue that political actors adapt their tactics in response to observation (Hyde 2011; Kelley 2012). Whatever their impact, monitors remain consequential because international recognition of election outcomes often depends on their assessments.
Concerns about limitations in violence reporting have prompted experimentation with alternative mechanisms. Policymakers and practitioners have increasingly adopted a hybrid approach that combines monitoring missions with crowdsourced citizen reporting. Interest in crowdsourcing reflects its potential to harness large-scale citizen participation where institutional capacity is limited (Bott et al. 2014). Advances in communication technology have made it cheaper and faster to recruit citizens to submit reports through online or mobile platforms (Bailard and Livingston 2014; Tuccinardi and Balme 2013; Van der Windt and Humphreys 2016). Such crowdsourcing reduces financial costs and political negotiations associated with monitor deployments. Platforms like Ushahidi, developed after Kenya’s violent 2007–08 elections, demonstrated the potential for citizen-generated data to map violence in near real time (Okolloh 2009; Rotich 2017).
Practitioner guidance and donor investments encourage crowdsourced citizen reporting. The National Democratic Institute (NDI) (2023) claims that citizen-generated reporting improves timeliness and strengthens coordination. Such guidance is increasingly followed, as seen in the eMonitor+ platform used during Jordan’s 2024 parliamentary elections to document electoral violence (UNDP 2024). This growing adoption of crowdsourcing has made it nearly as common as conventional monitoring. Ushahidi supported election-related reporting in 27 countries over the past 2 years, 2 while the European Union deployed election observation missions to 24 countries during the same period. 3
Despite the growth of crowdsourcing violence reports during elections, we know relatively little about the comparative effectiveness of citizens versus monitors. Are citizens better positioned to observe violence within their own communities, or are monitors better equipped to recognize incidents citizens might overlook? Could monitor presence influence whether citizens report violence? Understanding these dynamics is critical not only for improving measurement but also for informing the design of electoral accountability systems. If crowdsourced reporting and professional monitoring generate systematically different data about when and where violence occurs, then the choice has far-reaching implications for electoral integrity.
We argue that traditional election monitoring continues to add measurement value over citizen reporting alone. Monitors possess distinct advantages. Monitors are trained to identify coercion and observe elections neutrally and professionally. Over multiple election cycles, they gain experience in detecting and interpreting violent acts in politically diverse communities. By contrast, citizen reporting often depends on limited awareness of what constitutes electoral violence and subjective interpretations shaped by partisan loyalties or fear of reprisal. Accordingly, we expect that monitoring will increase the likelihood of detecting and reporting election violence compared to relying solely on citizen reporting.
We test these expectations through a field experiment that randomized the modality for collecting reports of election violence across communities in Côte d’Ivoire, a country where elections have been routinely marred by violence. We collected violence reports surrounding the 2020 presidential elections from 238 locations across the eight most populous departments and the two most populous communes of the de facto capital of Abidjan. These locations were randomly assigned to citizen crowdsourcing or trained monitoring alongside citizen reporting. The citizen-reporting component of our study used a crowdseeding approach in which a preselected panel of residents was contacted throughout the election period. This design allows us to compare violence reporting from locations that had only citizen reports to those that had both citizen and monitor reports across a range of political, demographic, and geographic conditions.
Our analyses show that assigning a monitor to observe a community, alongside citizen reporting, significantly increases the likelihood that election-related violence is reported. Assigning a single monitor increased the likelihood of reporting violence by 10.7 percentage points. These findings suggest that monitors do not induce citizens to report more. Instead, monitors themselves report violence where citizens do not. We do not find that monitor reports change the reported violence profile. Our data show no difference in the type of violence reported by monitors versus citizens. Suggestive evidence indicates that monitors aggregate reports from citizens; they are more likely to report violence when citizens also report first-hand experiences of violence, although the inverse relationship does not hold. We also find evidence that monitors’ experience matters: those with more prior observation experience were more likely to generate reports.
This study contributes to research on election violence by providing new experimental evidence on the comparative effectiveness of professional monitoring and citizen crowdsourcing. Our findings underscore that monitors continue to play an indispensable role in addressing the measurement dilemma faced by scholars and policymakers in producing data that are broad in coverage and precise in detail. Investing in monitoring capacity not only improves violence data quality but also strengthens the broader reporting infrastructure on which election credibility depends. Although deploying monitors in our experiment cost approximately $50,000 compared to $41,000 for citizen-based reporting, this investment produced broader reporting coverage of election violence that otherwise would have gone undetected.
In what follows, we first lay out a framework for understanding why monitors may systematically outperform citizens in detecting and reporting election violence. We then describe our research design, present the experimental findings, and conclude with broader implications for violence reporting and electoral integrity.
Capability and Insulation in Measuring Election Violence
Tracking election violence is challenging because its occurrence varies widely across space, time, and severity. We define election violence as lethal or nonlethal coercive acts against people (citizens, candidates, officials) or infrastructure intended to influence electoral processes and outcomes (Birch et al. 2020). These acts include threats, assaults, killings, property destruction, and disruption of election-related activities. This definition aligns with scholarly usage, emphasizing political intent and temporal proximity to elections while accommodating diverse manifestations (Höglund 2009; Von Borzyskowski and Wahman 2021). Policymakers and practitioners likewise include intimidation, physical attacks, and other coercive acts used to deter participation, influence results, or protest perceived unfairness (NDI 2023; United Nations Department of Political Affairs 2016). 4
Measuring election violence requires achieving breadth by spatially capturing events and depth by detecting less visible coercion alongside overt attacks. This depends on how information moves through a chain of individual decisions. For an act to become reportable, it must first be detected by someone present and attentive enough to recognize wrongdoing. However, detection alone is insufficient. Reporting also depends on observer willingness to speak up and on communication channels that convey credible information. At each stage—detection, willingness, and transmission—some incidents are captured while others are lost, shaping data available to researchers and practitioners.
Building on these challenges, we propose a framework for understanding how capability and insulation shape how violence is detected, reported, and transmitted. Capability reflects an observer’s ability to accurately recognize and classify violent acts, while insulation reflects protection from pressures that might deter reporting. These mechanisms govern the reach and precision of election-violence data. By highlighting capability and insulation, we explain gaps in existing reporting systems and clarify why professional monitors enhance reporting beyond citizen crowdsourcing alone.
How Citizens and Monitors Report Violence
To understand how capability and insulation shape violence reporting, we distinguish two approaches: professional monitoring and citizen-based reporting. Both can turn latent violent acts into observable data, but those involved differ in training, social position, and institutional support. These differences map directly onto reporting mechanisms.
Election monitors, typically affiliated with civil society organizations (CSOs) or international missions, are trained observers tasked with documenting electoral irregularities, including violence (Kelley 2008). Domestic and international monitors operate under formal mandates and rely on structured observation protocols, government accreditation, and established communication networks to facilitate data collection and verification. International monitors also benefit from external affiliations and media attention, making their claims harder for governments to dismiss and insulating them from local pressures (Daxecker 2012; Hyde and Marinov 2014; Kelley 2012).
We focus on domestic monitors because they combine professional training with local familiarity, enabling them to detect incidents international missions might overlook (Bjornlund 2004; Nevitte and Canton 1997; Park 2025). They are better equipped to interpret ambiguous situations accurately (Bush and Prather 2022; Grömping 2017; Leeffers and Vicente 2019), making them especially effective in reporting violence in rural areas and beyond predictable urban hotspots, where intimidation is often fiercest (Von Borzyskowski and Wahman 2021). Domestic monitors thus represent a relatively high-capability, high-insulation model that combines training with organizational backing and formalized procedures, enhancing breadth and depth of reporting.
Crowdsourcing represents a bottom-up approach in which citizens directly collect and share information (Kahl et al., 2012; Shayo, 2017). It democratizes data collection by enabling real-time reports from communities that may be inaccessible to monitors (Dowd et al. 2020). Citizen-based initiatives take different forms (Kahl et al. 2012): some use open platforms inviting broad participation, like Ushahidi’s SMS-based system in Kenya (Rotich 2017), while others rely on crowdseeding to recruit pre-selected participants to submit reports (Van der Windt and Humphreys 2016).
We focus on crowdseeding because it offers a citizen-based structure comparable to monitoring. By controlling who reports within each community, we can systematically compare between citizen and monitor reporting, which is less feasible with platforms that rely on anonymous submissions. For our purposes, crowdseeding represents a low-capability, low-insulation model that relies on citizens’ proximity to events and personal initiative rather than professional preparation or institutional protection.
Capability Differences Between Citizens and Monitors
Detecting and classifying election violence depends on the capability of those reporting it. Capability reflects knowledge, skills, and tools that raise the odds an observer will correctly detect and accurately classify an incident. Observers with greater capabilities are more likely to spot subtle forms of coercion, distinguish signal from noise, and document events with precision. In this respect, professional monitors have significant advantages over citizen reporters.
Citizens, by virtue of their proximity and personal stakes, often have unparalleled knowledge of violence in their communities because they are direct witnesses or even victims. During Nigeria’s 2011 elections, for instance, CSOs partnered with the Independent National Electoral Commission to crowdsource violence reports through social media, facilitating rapid crisis responses (Meier 2012). Such proximity makes citizen reporting especially valuable for identifying incidents in hard-to-access areas where monitors are less likely to be deployed.
However, citizen reporting also has limitations. Citizens must rely on their everyday experiences rather than formal training to detect and classify different types of violence. While many will recognize outright assault, they may overlook seemingly minor acts of harassment or intimidation. Citizens typically report only what they personally experience or witness, but they rarely gather corroborating details (Paulussen and D’heer 2013). Without training in structured protocols, citizen reports tend to vary in quality and are harder to verify. Thus, citizen reporting may lack the depth and precision of trained monitors, despite advantages in the geographic breadth of election-violence detection.
Monitors have greater observational capabilities due to professional norms and cumulative expertise. They receive formal training that equips them to recognize incident typologies, legal violations, and ambiguous coercive acts (Donno 2010; Norris 2013). They follow standardized reporting protocols that reduce vagueness and enhance verifiability relative to citizen-generated submissions. Their training also emphasizes continual improvement, including more robust detection and documentation practices over time (Yukawa and Sakamoto 2024). Monitors are often selected for their knowledge of law and human rights, experience in particular localities, or their standing as civic leaders (Anglin 1998). Repeated deployments further improve their ability to liaise with communities and detect impending violence (Bush and Prather 2022; Grömping 2017; Leeffers and Vicente 2019).
Capability differences between citizens and monitors have clear implications for both breadth and depth of election-violence reporting. Regarding breadth, monitors are more likely to document incidents in peripheral constituencies or remote areas where citizen reporting is limited. On depth, professional training enables monitors to identify low-salience acts, like verbal harassment, that casual observers may miss. Consequently, where monitors are deployed, we should expect a detectable increase in violence reporting relative to relying solely on citizens.
Insulation Differences Between Citizens and Monitors
While capability governs what observers can detect, insulation determines whether what they detect becomes public. Reporting requires willingness to speak despite risks. Insulation is an observer’s freedom to report sensitive information without fear of retaliation, social sanction, or personal harm. Higher insulation increases the likelihood that violent acts are transmitted into data.
Citizens report from their own communities. Their proximity to events gives them intimate knowledge of local violence, but it also exposes them to risks. Perpetrators of violence may be neighbors, friends, or local leaders. Identifying such individuals as perpetrators can invite personal retaliation. Fear of retribution discourages many from speaking up, particularly in politically charged or repressive environments. Citizens may then choose to self-censor and leave incidents unreported, despite detecting them.
Citizens’ local political ties can generate partisan biases that affect reporting patterns. Individuals may be more willing to report violence by political rivals while downplaying offenses by co-partisans (Davenport 2009; Davenport and Ball 2002). Therefore, reporting may vary across political contexts. In incumbent strongholds, citizens may hesitate to report state violence for fear of sanctions or social pressure. In opposition strongholds, where anti-incumbent sentiment is common, citizens may be less likely to report opposition violence. Reporting incentives should be strongest in politically competitive areas where violence can be used to deter participation or shape outcomes (Daxecker et al. 2025). But these same contexts often expose citizens to greater coercion and internal conflict over reporting.
Even when motivated to report, citizens often lack structured channels to do so safely. Crowdsourcing platforms designed to gather citizen reports can flood data systems with valuable firsthand accounts but also with background noise, like duplicate messages and unverifiable claims (Bader 2013; Croicu and Kreutz 2017; Roberts and Marchais 2018). In this sense, although citizen engagement broadens the potential information pool, insulation-related vulnerabilities limit reliability of reports generated through crowd-based systems (Sambuli et al. 2013).
Monitors, conversely, typically operate with greater insulation from local pressures. They are often assigned outside their home communities to maintain social distance. Sleeping in a different village or leaving soon after election day reduces their exposure to reprisals. Organizational affiliation provides additional protection. Accredited monitors benefit from legal status, official identity badges, and escalation channels that allow incidents of harassment to be flagged to authorities or publicized in media. Common knowledge that retaliation against a monitor could trigger broader scrutiny can deter suppression of their reporting.
Monitors are also accountable to professional standards, whereas citizens are not. Structured reporting forms require monitors to sign their names to incident reports, creating reputational consequences for inaccuracies or exaggerations. Validation committees often review submissions for credibility, further reinforcing expectations that monitors only report verifiable events. This formalized system of professional obligations, reinforced by risks of harming organizational credibility, encourages monitors to report comprehensively but responsibly (Bădescu et al. 2004; Nevitte and Canton 1997).
Insulation primarily enhances the breadth of reporting by empowering monitors in areas where intimidation is high and local citizens might self-censor. In politically sensitive or hard-to-reach locales, adding insulated monitors can illuminate otherwise hidden violence, expanding the geographic footprint of verified incidents. Insulation also affects depth, although its effect should be most evident for high-salience but high-risk events—e.g., where known perpetrators can be named and documented—that citizens might otherwise suppress. In these ways, insulation does not necessarily sharpen perception, but it reduces self-censorship, enabling more reliable documentation of violence. Thus, while monitors’ greater capability improves diagnostic precision of election-violence data, their greater insulation ensures that sensitive information reaches public records.
When combined with capability, the insulation mechanism implies that monitors supplement rather than crowd out information generated by citizen crowdsourcing. Monitor presence is unlikely to induce or discourage citizen reporting during the short data collection windows of election campaigns. Because reporting depends on individuals’ detection, willingness, and transmission decisions, monitors primarily improve reporting through their own behavior rather than by altering citizens’ behavior. Most citizens do not strategically weigh whether to report based on monitor presence. Their behavior reflects immediate personal circumstances: fear of reprisals, partisan ties, or access to communication channels. Introducing a monitor should therefore increase total report volume through their own efforts without affecting citizen reporting.
Testable Hypotheses
The capability and insulation framework yields several testable hypotheses. 1. Geographic Breadth: Localities assigned a monitor, alongside citizens, should have a higher likelihood of generating any violence report compared to those relying solely on citizen crowdsourcing. 2. Identification Depth: Localities assigned a monitor, alongside citizens, should be more likely to report less conspicuous forms of violence (e.g., verbal harassment). 3. Limited Substitution: Introducing a monitor should neither increase nor decrease the likelihood of citizen-generated reports. 4. Experience Differential: Monitors with greater experience should have a higher likelihood of submitting reports. 5. Insulation Differential: Monitors should be as likely to report violence regardless of geographic proximity to their home communities.
Scope Conditions and Theoretical Boundaries
This framework applies best to multiparty electoral environments where violence is a credible threat. It assumes civil society and election monitors are permitted to operate. Under such conditions, trained monitors should generate more reports of violence than citizens. However, in stable democracies, or highly repressive autocracies, monitors lose their relative advantage: citizens can report freely in democracies, while both reporter types struggle in autocracies due to state restrictions.
The framework remains agnostic about whether monitors deter violence, which could reduce latent incidents available for detection. Prior research shows that monitors can reduce some forms of manipulation (Callen and Long 2015; Hyde 2007, 2011) and prompt strategic adaptation by political actors (Asunka et al. 2019; Daxecker 2014; Simpser and Donno 2012). Although some studies find that monitors may provoke violence by exposing misconduct (Brancati and Penn 2023), most evidence indicates that monitored areas experience less election violence (Beaulieu 2014; Daxecker 2012; Hyde and Marinov 2014; Luo and Rozenas 2018). Because deterrence operates on underlying violence rather than on reporting behavior, we treat it as conceptually distinct from mechanisms through which monitors transmit information. Still, if monitors reduce latent incidents, this may complicate efforts to separate reporting dynamics from behavioral responses. Fully disentangling these dynamics remains a challenge for future research. Thus, our focus is on reporting behavior conditional on whatever violence occurs.
By centering capability and insulation, our approach reframes how to improve election-violence reporting: the key issue is not whether monitors substitute for citizens, but when and where their skills and protections expand the informational net to capture otherwise hidden violence. Even with technology-enabled crowdsourcing, trained monitors sharpen detection and widen coverage, providing enduring measurement value wherever competitive elections risk violence.
Election Violence Reporting in Côte d’Ivoire
To evaluate how our framework operates, we examine election violence reporting in Côte d’Ivoire. The Ivorian case provides variation in both violence exposure and institutional infrastructure for observation. These features enable us to examine how capability and insulation shape reports of election-related abuses.
Despite the reintroduction of multiparty competition in the early 1990s, Côte d’Ivoire continues to contend with a civil war legacy (2002–2007) that divided the country along ethnic, regional, and partisan lines. The 2010 presidential election was meant to mark a democratic restoration, but instead resulted in thousands of deaths. Current President Alassane Ouattara was installed in office only after international intervention. The 2015 election was less violent but still marked by intimidation (HRW 2015).
The 2020 election underscored the fragility of electoral peace. Ouattara’s announcement that he would seek a third term, despite constitutional term limits, sparked protests and a coordinated opposition boycott. Nationwide violence left at least 85 people dead. On election day, observers recorded violent clashes between partisans (Banégas and Popineau 2021; CNDH 2021; van Baalen and Gbala 2023). While the most extreme episodes drew international headlines, many incidents took quieter forms, such as gangs preventing voters from reaching polling centers.
The environment in Côte d’Ivoire poses significant measurement challenges. Election violence is geographically uneven and qualitatively diverse, requiring observers who can operate in high-risk areas where coercion might be more likely while identifying overt assaults along with easily obscured coercive acts. These challenges are exacerbated by a politicized media environment and risks faced by ordinary citizens. The detection and reporting of election violence thus depend heavily on the characteristics of those observing.
Citizen Reporting in Côte d’Ivoire
Government, civil society, and international initiatives have produced a decentralized ecosystem for crowdsourcing citizen reports. During the 2020 elections, citizens could report incidents to emergency hotline numbers operated by the security forces or CSOs like the West Africa Network for Peacebuilding (WANEP-CI). They could submit reports to CSOs via online platforms, Facebook, or WhatsApp. Citizens could also report incidents in person to offices of the Independent Electoral Commission (CEI), WANEP-CI, and other CSOs.
These initiatives sought to widen the reporting net but faced challenges in producing a comprehensive and reliable picture of election-related violence. Pervasive fear was a major obstacle: Afrobarometer surveys reveal that approximately half of Ivorians fear becoming victims of election-related violence (Afrobarometer 2023). This widespread apprehension may discourage individuals from reporting, particularly in areas with heightened political tensions. Moreover, the lack of standardized training for citizen reporters could lead to inconsistencies in data quality (Bader 2013).
Monitor Reporting in Côte d’Ivoire
Monitors in Côte d’Ivoire operate within structured programs that provide both the breadth of geographic coverage and the depth of diagnostic precision. Domestic CSOs have coordinated observation efforts since the early 2000s, building monitoring capabilities through training and securing insulation through institutional protections. The most prominent observation organization, WANEP-CI, requires monitors to have a minimum educational background to ensure they can analyze complex events, and it formally secures government permission through the CEI prior to deploying monitors across the country.
In preparation for the 2020 presidential election, two major organizations, Initiative de Dialogue et Recherche Action pour la Paix and WANEP-CI, deployed over 1,000 monitors. Monitors participated in multi-day training workshops that covered incident typologies, the legal framework for elections, and the use of geotagging apps to document and transmit reports. Trainees engaged in scenario-based exercises designed to improve judgment in ambiguous situations, such as whether the presence of armed youth near a polling site constituted intimidation.
WANEP-CI’s records indicate that monitors reported a range of acts: violent demonstrations involving confrontations between partisans; attacks on CEI staff and destruction of polling materials; roadblocks limiting access to polling sites; and presence of non-state armed groups intimidating voters (WANEP-Côte d’Ivoire 2020). These reports came from across the country, reflecting geographic breadth. Their detailed classification of both overt and subtle forms of violence demonstrated diagnostic depth.
The 2020 election monitoring experience further shows how insulation facilitates violence reporting. Monitors operated under the CEI’s “Charter for the Observation of Elections in Côte d’Ivoire,” which granted accredited observers legal recognition, official badges, and the right to request cooperation from local authorities (Commission Électorale Indépendante 2020). By codifying impartiality, mandating nonintervention, and encouraging objective reporting, the Charter created an institutional buffer that empowered monitors to document politically sensitive events without the threats ordinary citizens might face. Additionally, monitors received training on personal safety and psychological resilience and were often posted outside their home communities.
Côte d’Ivoire’s experience suggests that establishing multiple channels for crowdsourcing citizen reports may not be sufficient to overcome structural constraints, like fear of retribution and a lack of standardized training. By contrast, CSO-coordinated monitors benefit from both capability and insulation advantages. These mechanisms allow monitors to detect subtle forms of coercion and operate in regions where citizens may self-censor.
Research Design
The case of Côte d’Ivoire demonstrates how differences between citizens and monitors shape the breadth and depth of election violence reporting. We designed a field experiment to assess whether the presence of trained monitors affects the volume and composition of reported violence as well as whether monitor reports differ from citizen reports. We hypothesize that capability enables monitors to detect more incidents, while insulation empowers them to report potentially politically sensitive events—both of which should increase the observed incidence of violence in areas where monitors are deployed.
Sampling and Randomization
To assess variation in electoral violence reporting during the 2020 election, we employed a multi-pronged sampling strategy using a cross-section of locations that reflect political and sociological diversity.
5
We sampled from among the country’s most populous departments: Abengourou, Bouaké, Daloa, Gagnoa, Korhogo, Man, San Pedro, and Yamoussoukro (the official political capital); see Figure 1. Additionally, we sampled from two of Abidjan’s most populous communes with distinct partisan profiles: Abobo (ruling-party stronghold) and Yopougon (opposition stronghold). The selected departments have considerable variation in partisan affiliation, economic development, and past exposure to violence. Selected departments and locations in Côte d’Ivoire.
From these departments and communes, we randomly selected a sample of communities. The unit of analysis is analogous to a neighborhood or village. 6 Our sampling strategy attempted to minimize spillover in reporting among locations. In rural areas, we created a buffer of 8 km between villages and sampled villages that were at least this distance apart. In urban areas, we divided each neighborhood into 40,000-square-meter hexagons and sampled hexagons that were at least 800 m apart. 7 Random selection helps ensure representativeness at the local level.
Summary statistics - locations
Citizen and Monitor Samples
Citizen respondents were selected through a baseline in-person survey. Enumerator teams were sent GPS coordinates to initiate the survey and used a standard random walk to recruit respondents. 8 Eligible participants were literate adults of voting age who had access to a mobile phone, which ensured they could participate in follow-up SMS surveys. Respondents who met these criteria were invited to provide mobile phone numbers for recontact. This design follows a crowdseeding approach in which pre-identified respondents are recruited to facilitate more consistent reporting over time.
Monitors were adults of voting age affiliated with CSOs accredited by the CEI. They were invited to participate in our study through their respective CSOs as government-approved monitors. We assigned one monitor to each randomly selected treatment locale for the citizen-plus-monitor condition. 9 To align monitors’ observation areas with the citizen survey, monitors were given the same GPS starting points as baseline survey enumerators. Monitors then traveled to their assigned locations to observe. These visits were undertaken for our project and were not part of a broader monitoring initiative.
There are notable demographic differences between the citizen and monitor samples, summarized in Tables A1 and A2. Monitors are more educated than citizens: 75 percent of monitors completed high school compared to 40 percent of citizens. Monitors are relatively wealthier: they own more assets than citizens and are less likely to have gone without water, cash, food, or medicines. Monitors are more likely to live in urban areas and follow the news more regularly than citizens. These distinctions reflect the self-selection that shapes civil society engagement. Individuals who volunteer to observe elections are more resourced, informed, and civically engaged—traits that may partly explain why monitors report incidents that citizens either fail to detect or feel constrained from disclosing.
Data Collection for Unit of Analysis
The unit of analysis for this study is the location-time period. We collected weekly panel data in treatment locations during four distinct periods: October 30–November 1 (election day was October 31), November 8–10, November 14–16, and November 21–23. Figure 2 visually summarizes the data collection process. Data collection process.
Citizen reports were collected via Telerivet, an online platform for sending and receiving SMS. Each participant received a weekly five-question survey via SMS and was asked to respond using simple yes/no responses. To incentivize participation, all participants received 500 FCFA (US$0.82) in mobile phone credit weekly, regardless of whether they reported violence. Credit was transferred electronically and exceeded response costs (10 FCFA per message).
If a citizen reported witnessing or hearing about violence, a research assistant called them to administer a structured Qualtrics survey to elicit details: incident type, frequency, and context. In a typical location, two citizens submitted violence reports during the study period, though up to five did so in some areas. Individuals could report multiple types of violence. The maximum number of citizen reports in any location was 8 on election day and 17 during the post-election period. The median number of citizen reports per location-time period was zero.
Monitors submitted their reports in parallel during each reporting window. Monitors were prompted via WhatsApp to travel to their randomly assigned location and complete a Qualtrics survey. These surveys documented events they witnessed or heard about. Monitors received a modest weekly stipend to defray transportation and other routine costs that their organizations typically covered. Like citizens, monitors submitted relatively few reports. The maximum number of violence reports generated by monitors in any location was four on election day and three in the post-election period. The median number of monitor reports per location-time period was zero.
Balance table - citizen demographics aggregated
Note. Citizens responses to baseline questions aggregated to the location level.
Two-tailed t-test for differences in means between treatments.
Outcomes of Interest
The primary outcome is the number of violence reports generated in each sampled location during two key time periods: election day and post-election weeks. 10 It should be noted that we analyze reports of violence rather than unique violent events. These reports represent what citizens or monitors claim to have witnessed or heard, which may vary in terms of accuracy, salience, and duplication.
We categorize reports into four types: physical violence, verbal harassment, election-related violence, and protest. Physical violence involves direct attacks, including killings, disappearances, and property damage. Verbal harassment involves threats, insults, or intimidation without physical force. Election-related violence involves acts that disrupt the electoral process like blocking access to polling places and destroying ballots and voter IDs. Protest includes demonstrations, marches, and riots.
We use an expansive definition of election violence that is consistent with scholarship and practice. We include a category for verbal harassment and intimidation because they are widely recognized as manifestations of election violence (e.g., Birch 2020; Höglund 2009). Given their potential to influence voter participation (Burchard 2015; Gonzalez-Ocantos et al. 2020; Ley 2018), international observation organizations instruct monitors to treat threats as indicators of electoral conflict. NDI trains monitors to report threats against election officials and voters (NDI 2023). Similarly, the International Foundation for Electoral Systems includes threats as “election-related violence.” 11
We also collect reports of protests. While protests are not included in our technical definition of election violence, such events during electoral periods in Côte d’Ivoire and elsewhere often involve violent confrontations with security forces or between rival groups. Protests are considered a form of social unrest that can be an early warning sign of future violence (Demarest and Langer 2018; Raleigh et al. 2023). NDI urges monitors to closely observe protests and rallies during uncertain electoral outcomes or perceived unfairness, as these can spark election violence (NDI 2023). In Côte d’Ivoire, 18 percent of protests between 1997 and 2020 resulted in at least one fatality (Raleigh et al. 2023). Most importantly, during the 2020 election cycle, protests were declared illegal under COVID-19 regulations, facilitating use of force by security forces and resulting in violence against civilians. Further, opposition demonstrations also turned violent due to clashes between partisan groups (Banégas and Popineau 2021; Van Baalen 2024). Since we did not know, a priori, whether protests in Côte d’Ivoire would turn violent during the 2020 elections, we asked our monitors and citizens to report them, along with other forms of violence.
Our main analyses use an indicator coded 1 if any form of violence is reported in a location-time period and zero otherwise. For ease of interpretation, all violence categories are binary outcomes. The appendix presents additional specifications, including a count of reports per location and a variable measuring the proportion of total reporters in each location who reported violence (Tables A8, A9, A10).
We also distinguish reports based on whether they are firsthand or secondhand. Firsthand reports come from individuals who directly observed or experienced the event. Secondhand reports are based on information passed through others, like neighbors or local leaders. We consider firsthand reports to be more reliable indicators of violence in a location, whereas secondhand reports depend on the reporter’s quality, potentially reflecting broader dynamics but also subject to rumor. Because monitors visit their assigned locations only once per reporting period, they are more likely than citizens to rely on secondhand information.
Our focus on reports rather than events differentiates our approach from event-based datasets like ACLED (Raleigh et al. 2023), which rely on media coverage to document episodes of violence (Daxecker et al. 2019; Van Baalen 2024). We adopt this approach for three key reasons. First, our framework captures forms of coercion, including verbal harassment, that often go unreported in traditional datasets. Second, event-based data tend to over-represent large, urban, or visible incidents while overlooking more diffuse or rural occurrences (Clarke 2023; Demarest and Langer 2018; Weidmann 2016). Third, our design allows for a direct comparison of how ordinary citizens and trained monitors report violence across diverse settings.
Reporting Trends
Reporting incidence by violence type and time period
Note. Cells represent percentage of locations with at least one report of violence. Physical violence includes property damage and violence between individuals, including killings or disappearances. Verbal violence includes harassment or threats of violence. Protest includes demonstrations, marches, or riots. Election-related violence includes prevention of voting and destruction of ballots or voter IDs.
Because survey-based studies can raise concerns about demand effects, particularly when respondents are compensated, we assess the plausibility of systematic over-reporting. If compensation pressured citizens or monitors to report violence even when none occurred, most localities would register at least one report. Instead, reporting was far from universal. In the citizen-only condition, 46 percent of locations produced no citizen reports. In the citizen-plus-monitor condition, 54 percent of locations generated no citizen reports, and 66 percent produced no monitor reports. These patterns suggest that respondents did not feel obligated to report violence to justify compensation. Although monitors might report at somewhat higher rates because they were primed to identify violence, their training and experience should help guard against systematic over-reporting.
The presence of a monitor might dampen citizen reporting if citizens defer to the monitor’s authority. If monitors discouraged citizen reporting, we would expect citizen reports to decline in the combined treatment arm. We find no evidence of this dynamic: citizen reporting rates do not differ significantly between locations assigned to the citizen-only condition and the citizen-plus-monitor condition. Alternatively, it is possible that some respondents chose not to report violence to receive compensation with minimal effort. But the willingness of both citizens and monitors to complete follow-up reporting — despite the increased burden — suggests genuine engagement. Generally, we find no credible reason to believe that our data systematically overstate or understate the incidence of violence across study locations.
Challenges to Measuring Violence Reports in a Pure Control Group
Our experimental design included a pure control group: in locations not assigned to citizens or monitors outcomes were measured through media reports. Research assistants coded articles from major Ivorian news outlets for any mention of violence in both treated and control locations. This strategy produced no reports of violence in any rural sites. To assess whether this reflected a lack of violence or limited media coverage, we examined ACLED data as an external benchmark. ACLED likewise recorded violence only in urbanized locations (See Figure A2 in the Appendix). Although our study generated reports of violence in rural locations, ACLED contained none. The absence of media or ACLED-reported violence in pure control sites is therefore implausible and more likely reflects underreporting than actual lack of violence.
We have no reason to believe that pure control locations were less violent than treated villages, since treatment status was randomly assigned, and without citizen or monitor presence, we cannot confirm their exposure to violence. This limitation underscores the vulnerability of relying exclusively on media-based reports to measure election violence, particularly in rural areas without media coverage.
We also note that we did not include a pure monitoring treatment arm for both theoretical and practical reasons. Implementing a monitor-only arm would have required withholding citizen reporting, which is inconsistent with the prevailing practice of hybrid observation systems. Because citizen reporting is increasingly incorporated as an integral component of election observation, our goal was to estimate the added value of monitors above citizen-based platforms. Field logistics also made a monitor-only arm infeasible: cost constraints, tight timelines, and deployment delays with partners left insufficient resources to field a third treatment. Our design, therefore, focuses on the policy-relevant contrast that mirrors current practice: citizen reporting supplemented by domestic monitors. This allows us to establish a baseline for future research that incorporates all three arms, thereby further disentangling the contributions of trained monitors and untrained citizens.
Estimation
We estimate treatment effects using ordinary least squares (OLS) models with robust standard errors clustered at the location level, the unit of treatment assignment. All models include fixed effects for the time period (election day or post-election) to account for potential differences in reporting behavior across time. These models estimate whether assigning a monitor to a location increases the probability that any form of violence is reported. Alternative specifications and robustness checks can be found in Tables A6, A7, A8, A9, A10. The results presented in the main text remain substantively consistent across estimation strategies.
Monitors Add Breadth and Depth to Election Violence Reporting
The empirical analysis tests whether deploying monitors alongside citizen crowdsourcing increases the number of election violence reports. We expect monitors to improve reporting through two mechanisms: greater capability to detect and classify violence and greater insulation from local pressures that inhibit disclosure. These expectations generate three key hypotheses: H1 (Geographic Breadth), H2 (Identification Depth), and H3 (Limited Substitution).
Effect of citizen + monitor treatment on reports of any violence
Note. *p < 0.05; **p < 0.01; ***p < 0.001. Outcomes are binary variable that indicates any violence reported by all reporters (1), citizens only (2), monitors only (3). Each column represents a separate OLS regression. Robust standard errors in parentheses, clustered by location with fixed effects for time period.
Models 2 and 3 disaggregate reporting by source to identify who drives the effect. Model 2 shows that monitor assignment has no statistically significant effect on citizen reporting. Citizens are neither more nor less likely to report violence when a monitor is assigned to their location. This result is consistent with H3, suggesting that monitor effects operate independently rather than altering citizen behavior. By contrast, Model 3 shows that locations assigned to citizen-plus-monitor treatment are 22.3 percentage points more likely to report violence than citizen-only locations (p < 0.001). This effect is mechanical: assigning a monitor increases the likelihood of monitor reporting. This clarifies that Model 1 is driven by monitor reports, not changes in citizen behavior.
These experimental findings are reflected in simple descriptive statistics. Across the 121 locations assigned both citizens and monitors, monitors documented violence in 33 percent of locations compared to 24 percent for citizens in the same locations. This gap highlights the value added of monitors: they uncover incidents that citizens either do not report or do not observe.
The concern remains that monitors might inadvertently discourage citizens from reporting. The results in Table 4 show no such effect. To probe for the possibility of crowding out, we examine whether citizens were less likely to respond to our SMS surveys in the citizen-plus-monitor locations. If monitors deterred citizen reporting, we would expect citizens in these treatment areas to submit fewer reports of no violence. This would signal disengagement prompted by monitor presence. Table A5 in the Appendix estimates the effect of the treatment on the count of “no violence” reports. We find no statistically significant relationship between monitor assignment and citizen reporting that no violence occurred. This finding is consistent with H3, reinforcing our theoretical claim that monitors operate through their own detection rather than by altering citizens’ behavior.
Effect of citizen + monitor treatment on firsthand reports of violence
Note. *p < 0.05; **p < 0.01; ***p < 0.001.
Outcomes are binary variables indicating whether any violence of the type specified was directly witnessed or experienced by either monitors or citizens. Each column represents a separate OLS regression. Robust standard errors in parentheses, clustered by location with fixed effects for time period.
Monitors may outperform citizens in reporting violence because they can aggregate information across social networks. Unlike citizens who may primarily report violence targeted at them or their close contacts, monitors often have broader networks through their work in civil society. As described in Côte d’Ivoire, organizations like WANEP-CI purposely train their monitors to seek out and verify information through multiple sources. This outreach may facilitate the collection of reliable secondhand reports. Consistent with this expectation, we find that 78 percent of monitors who did not personally witness an incident, but still reported it, learned about it secondhand from their contacts.
Figure 3 provides additional evidence for monitors’ aggregation function. We observe positive, statistically significant correlations between citizen firsthand reports and monitor reports for the same violence types within a location. There is a positive, statistically significant correlation between citizen firsthand protest reports and monitor protest reports (Table A4 in the Appendix reports correlation coefficients and Figure A1 reports correlations between monitor and citizen reporting, including all citizen reporting). These correlations indicate that monitors’ reports are not generated in isolation but are plausibly informed by citizen observations. Monitors thus appear to transmit local information that might otherwise go unreported. Correlation matrix: Firsthand reporting by citizens and all reporting by monitors.
Appendix Tables A11-A14 present additional analyses of treatment effects disaggregated by violence type (physical, verbal, election-related, protest). These results show that assigning a monitor, in addition to citizen reporting, does not differentially affect the types of violence reported in a location. The primary effect of monitor assignment is to increase the probability of reporting in that location, not to shift the type of violence reported. The monitor effect enhances breadth and depth in reporting without distorting the underlying distribution of violence types. Again, this is consistent with the expectation that monitors add information without altering citizen behavior.
Capability
We argue that monitors are likely to outperform citizens because they possess greater capability. Monitors develop observational skills and situational awareness that helps them accurately document violence. We have emphasized that these capabilities are acquired through training and experience. This leads to a clear empirical expectation: monitors with greater experience should be more likely to identify violence.
We first examine prior experience in our monitor sample. Sixty-five percent of monitors reported prior election observation experience, and some monitors previously worked in up to 11 departments across Côte d’Ivoire. This variation provides an opportunity to test whether accumulated experience sharpens monitors’ ability to convert latent violence into observable reports. Appendix Table A3 presents descriptive statistics on monitor experience.
Monitors’ geographic experience increases reporting likelihood
Note. Table reports coefficients from two separate regressions. In both regressions, the outcome is a binary variable whether the location had a monitor report of violence. Both regressions include time fixed effects and standard errors are clustered at the location-level.
Insulation
We argue that insulation shapes monitors’ effectiveness as election violence reporters. Insulation depends on availability of social and institutional protections that reduce reporting costs. Unlike citizens, monitors operate under organizational affiliations, possess formal accreditation, and often work outside their immediate home communities. These protections provide a buffer against local pressures that might discourage reporting. Monitor reporting should therefore be unaffected by geographic proximity to home.
We test this insulation implication by examining whether distance from a monitor’s home community predicts reporting behavior. Fifty-two percent of monitors in our sample worked within 10 km of their homes. In principle, these monitors could face social and political pressures like those experienced by citizens. If fear of reprisal limits reporting, we should see lower reporting rates from monitors observing in or near their home locations. By contrast, monitors assigned to locations to observe farther from home should enjoy greater insulation and therefore be more likely to report violence.
Monitors’ geographic proximity does not affect reporting likelihood
Note. Table reports coefficients of interest from two separate regressions. Distance traveled categorical variable: (0–5 km = 1, 6–10 km = 2, 11–15 km = 3, 16–20 km = 4, more than 20 km = 5). In both regressions, the outcome is a binary variable whether the location had a report of violence. Both regressions include time fixed effects and the standard errors are clustered at the location-level.
Conclusion
This study provides experimental evidence that trained monitors substantially improve election violence reporting beyond what citizen crowdseeding alone can achieve. Supplementing citizen reporting with monitors significantly increases the likelihood that violence will be reported in a location. This increase is not attributable to changes in citizen behavior. Monitors neither encourage nor displace citizen reports. Rather, their added reporting reflects advantages in capability and insulation. Monitors were more likely to submit firsthand reports, and their reporting did not vary with proximity to home communities. Monitors with broader geographic experience were also more likely to be sole reporters of violence, underscoring how cumulative experience enhances reporting. Monitors thus improve both breadth and depth of reporting by capturing incidents across wider areas and with greater diagnostic precision.
The Côte d’Ivoire case highlights the stakes of measurement in contexts where electoral violence is spatially dispersed and politically sensitive. In such environments, citizen crowdsourcing may be limited by fear, bias, or misinformation. The fact that monitors operated effectively across these settings, including rural areas, speaks to the practical value of investing in structured monitoring systems even in the age of decentralized reporting technologies.
Future research should explore how hybrid systems can maximize the strengths of both approaches. For example, citizen alerts could trigger follow-up investigations by trained monitors, expanding geographic coverage while ensuring more accurate documentation. Additional work is needed to understand how these mechanisms function in areas with different partisan compositions. As civil society and electoral commissions continue to adopt new technologies and reporting infrastructures, careful attention to detection and disclosure mechanisms remains critical. Strengthening the quality of violence data is not merely technical; it is foundational to protecting electoral integrity.
Supplemental Material
Supplemental Material - Seeing what Citizens Miss: How Monitors Improve Election Violence Reporting
Supplemental Material for Seeing what Citizens Miss: How Monitors Improve Election Violence Reporting by Leonardo R. Arriola, Arsene Brice Bado, Justine M. Davis, Allison N. Grossman, Aila M. Matanock in Journal of Conflict Resolution.
Supplemental Material
Supplemental Material - Seeing what Citizens Miss: How Monitors Improve Election Violence Reporting
Supplemental Material for Seeing what Citizens Miss: How Monitors Improve Election Violence Reporting by Leonardo R. Arriola, Arsene Brice Bado, Justine M. Davis, Allison N. Grossman, Aila M. Matanock in Journal of Conflict Resolution.
Footnotes
Acknowledgments
The authors thank our research assistants at CERAP in Abidjan and IPSOS Côte d’Ivoire for the survey work. We also extend gratitude to David Dow for assistance with the urban sampling procedures. We thank attendees at APSA, SPSA, and SoWEPS for their helpful comments and suggestions. We also thank staff at the Center for African Studies at University of California, Berkeley for support.
Ethical Considerations
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: We acknowledge financial support from the Carnegie Corporation of New York.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
All data and code needed to reproduce these results is available in the Harvard Dataverse (https://doi.org/10.7910/DVN/ARIFVH), 
. Preregistration details are available in the online appendix.
Supplemental Material
Supplemental material for this article is available online.