When Do Voters See Fraud? Evaluating the Effects of Poll Supervision on Perceptions of Integrity

Abstract

What shapes voter perceptions of election outcomes? Recent disputes in Malawi and Kenya highlight the vulnerability of local vote counts to accusations of malfeasance, which often generate negative public perceptions of vote reliability. Election monitoring in these countries is thought to crucially affect both the quality of the election and voters’ perceptions of the same. To date, most research on this topic has focused on the effect of non-partisan electoral observers. However, in many countries, two other interest groups also monitor the vote-counting process: political party agents and government election officials. Does the presence of these actors also affect voter perceptions of election integrity? To answer this question, I conducted a conjoint experiment in Malawi and Kenya in which voters evaluate the reliability of vote counts from hypothetical polling stations where the presence of party agents, non-partisan observers, and election officials is varied. I find that the presence of each of these groups does indeed shape voter perceptions: voters are more likely to view vote counts as reliable when they are co-signed by a party agent, election official, or non-partisan observer. Further, these preferences persist regardless of the voters’ own party affiliation or trust in electoral institutions.

Introduction

When electoral losers make allegations about biased vote counts, voters often have limited information with which to evaluate misconduct claims. Voting on election day constitutes the extent of any possible witnessing of the vote count for most voters (Norris et al., 2015). Besides election officials, the only other groups with constitutional access to vote-counting procedures both within and across polling stations are political party agents (Ascencio and Rueda, 2019) and non-partisan election observers (Asunka et al., 2019), representing candidate and citizen interests, respectively. As the sole source of electoral accountability on election day, the presence or absence of these poll actors at a polling station has significant potential to influence voter perceptions of election integrity (Kerr, 2018). Whether (in practice) voters use information about poll monitoring to amend their evaluations of election integrity is less clear.

In this article, I test these ideas using a visual conjoint experiment to compare voter preferences for election results from polling stations where an interest group is present versus one where they are absent. The visual conjoint simulates vote tally sheets from Malawi and Kenya, which are single, one-page documents that compactly communicate a station’s location, vote count and the signatures of the election officials, party agents, and observers present. These monitoring details make these sheets a useful template for conjoint “candidate profiles” for any given polling station, and the use of images allows me to visually manipulate many polling details without presenting participants with long tables of polling information. Here Malawi and Kenya’s recent electoral histories of judicial disputes over vote tally irregularities make them useful case studies on the importance of station-level actors for election integrity (Akinkugbe and Gathii, 2020). Furthermore, I test whether voter characteristics like partisanship and political knowledge underpin voters’ valuation of poll actors as causal mechanisms (Kerr, 2013; Robertson, 2017).

My results contribute to our understanding of voter perceptions of electoral integrity in two ways. First, I find that the presence of every polling actor (election officials, party agents, and non-partisan observers) matters for evaluations of vote reliability. Work on perceptions of integrity predominately explores the impact of non-partisan observer presence on voters’ trust in elections (Bush and Prather, 2018; Kerr, 2018). The results additionally show that election officials and party agent presence is salient to voters. Second, I find that the descriptive voter characteristics of partisan affiliation, trust in Election Management Boards (EMBs), and knowledge of observers do not condition respondents’ preferences. This casts doubt on the extent to which “subjective,” voter-level factors can be used to explain voter preferences in polling supervision (Hernández-Huerta and Cantú, 2022).

Perceptions of electoral integrity

Existing research on election officials, party agents, and non-partisan observers informs the hypotheses on voters’ prospective responses to the presence of each interest group.

Election officials are the administrative personnel sent by EMBs to organize voting at every polling station (ACE, 2012). In Malawi and Kenya, these officials are called “Presiding Officers” (PO) and the majority of people (98% and 97.3%, respectively) know the name and role of their respective EMBs (Afrobarometer, 2021). ¹ The importance of the role suggests that their absence would be a point of concern if interpreted as reduced oversight from a government body. However, not all voters may think this way. Analysis of LatinoBarometer responses to the question “would you say that elections in your country are clean and fair” by Rosas (2010) indicates that partiality and lack of EMB independence negatively affect perceptions of electoral integrity. Alternatively, PO’s absence may be interpreted as a logistic failure on the part of the EMB, which is associated with lower perceptions of election quality among AfroBarometer respondents (Kerr, 2013). So, while election officials are important, past EMB performance and practices may lead voters to ignore their presence during evaluations of integrity (Erlich and Kerr, 2016; Elklit, 2019).

Hypothesis 1

1. a. Respondents who have higher trust in their EMB will have higher trust in the reliability of vote tabulation if a PO is present.

2. b. Respondents with lower trust in their EMB will be indifferent to the presence of a PO in their evaluations of vote reliability.

Political party agents protect the interests of candidates by monitoring polling places for faults made during vote counting (ACE, 2012). Ascencio and Rueda (2019) show that parties strategically allocate their agents to stations where opposition agents will be present, suggesting that party agents play a largely protective role. The shared political identity and interests between voters and their copartisan agents may boost confidence in the validity of results at stations where their party agents are present (Robertson, 2017). However, shared political identity alone may be insufficient in predicting voters’ response to the presence of their party agents. Cantú and García-Ponce (2015) used an exit poll of Mexican 2012 elections to evaluate the presence of party agents against a question asking voters about their confidence that their vote will be “respected and counted.” They find a weak, insignificant association, suggesting that voters can be indifferent to party agent presence. So, while there is strong evidence of the importance of partisanship in framing voter evaluations, there is also evidence that partisanship may not be the core basis for how voters respond to partisan monitors.

I adopt the Cantú and García-Ponce (2015) intuition (that shared political interest between voters and their copartisan agents matters), but contend that voters may not register the physical presence of agents when voting (Kerr, 2018). I propose using a metric that captures voters’ relative preferences when presented with the option of monitored vs. unmonitored stations, which may yield stronger findings than head counts. Given that partisanship strongly shapes how voters process election information (Edelson et al., 2017; Anderson et al., 2005), I argue that partisan identification will lead voters to prefer results from stations with copartisan agents.

Hypothesis 2 Respondents will have greater trust in the reliability of vote tabulation if copartisan agents are present.

Non-partisan observers are the neutral group between party agents and election officials, who record and report any misconduct in the vote-counting process (Asunka et al., 2019). However, we should not assume that the average voter knows about the deterrent role played by observers, especially because non-partisan observers are a recent development relative to representatives of parties and EMBs (Lehoucq, 2003). Dawn (2014) uses a field experiment to show that when provided information about the role of observers, voters tended to rate the election as being more “fair and free.” Work on Tunisia by Bush and Prather (2018) comparing domestic, regional, and international observers also shows that receiving knowledge about election observers impacts voter perceptions of observers. While Malawi and Kenya host multiple international observer teams, each country has one or two core domestic non-partisan observer organizations providing the majority of observers on election day. ²

I posit that the relative novelty of non-partisan observation implies that voter awareness of observer groups is limited (Lehoucq, 2003). This intuition is substantiated by preliminary survey findings by Macdonald et al. (2022) in Zambia, Kenya, and the Gambia. They found that respondents saw election observation as important, but most respondents needed to be given information about observers first. Notably, few people had knowledge of what observers do or knew the names of any local, regional, or international observer group. The general lack of awareness leads me to expect that people’s knowledge of observer groups strongly conditions their reactions to information about observers.

Hypothesis 3 Respondents who have prior knowledge about non-partisan observers will have greater trust in the reliability of vote tabulation if non-partisan observers are present.

To study how these personnel shape perceptions of electoral integrity, I use a visual conjoint experiment. Much existing work measures voters’ perceptions of electoral integrity through survey questions often worded: “was X election fair and free.” In asking voters to evaluate an entire election, the response is a respondent’s aggregate evaluation of the election that does not capture: (1) what part(s) of the election cycle are influencing these evaluations and (2) how voters are attributing electoral rights and wrongs for every part of the electoral process. Using the visual conjoint setup allows me to both constrain respondents’ information to station-level activities and compare perceptions across polling actors.

Research design

I take advantage of the format of election forms to narrow the scope of election information being used to make evaluations of integrity. Since election forms contain the names and signatures of all the monitor groups, they provide a micro-level synopsis of the scenario at a given station and usefully accomplish this on an A4-page layout. A standard, fully filled-out vote tally sheet from Malawi can be seen in Figure 1, where: the number of votes for each candidate is highlighted in yellow, the details of political party agents and non-partisan observers are highlighted in blue, and the PO (the election official from the EMB) is circled in red. Similar documents also exist for Kenya. There are a lot of details on the page, so to hone a respondent’s focus on the personnel, a simpler version of the election form was created (Figure 2). It only contains the election information I am interested in manipulating while simulating the format of the original form. ³

Figure 1. Example of an actual election form (from Malawi’s 2019 election). Details relevant to the experiment have been highlighted for clarity.

Figure 2. Example of a simplified form (Malawian version) shown during the experiment.

In the context of the conjoint experiment, the presence or absence of a signature on a form (from an agent/election official /observer) is used to imply the presence or absence of their associated interest group. The primary assumption is that voters will interpret the presence/absence of a signature on a form, as the presence/absence of a person. This is a challenge if, for example, a respondent sees a form with a missing signature, but does not interpret it as missing person: e.g. they may assume that their party agent was present, but just forgot to sign the form. I try to overcome this challenge using framing, by telling respondents that a signature is a useful signal that this agent/official/observer was involved during the vote count.

Data and methods

The goal of the design is to capture evaluations of vote reliability for polling scenarios where the interest groups present are varied. I designed two online surveys in Qualtrics that were distributed to respondents in Kenya and Malawi from late March to mid-April 2022. ⁴ Respondents were asked a battery of demographic questions and then shown real examples of the election forms similar to Figure 1 (with similar highlighting). This was done to draw attention to specific parts of the form, with wording that was careful not to induce any positive or negative biases toward the information. The core of the survey is a forced-choice experiment, where respondents were shown two (randomly selected) election form images. Here, respondents had to choose which form was the more reliable document. Specifically, they were asked: “Between these two vote tally sheets, click on the one that you believe has more reliable information about the vote outcomes at its polling station.”

Unit of analysis

The unit of analysis is the survey respondent. Each respondent reviews three forced-choice questions, meaning that they see and evaluate six election forms. Survey participants were recruited from Qualtrics’ in-house pools of survey respondents in Malawi and Kenya. ⁵ The forced-choice format, together with 140 Malawian respondents and 250 Kenyan respondents evaluating 3 pairs of documents, resulted in a sample of approximately 2300 observations, clustered at the respondent level. ⁶

Independent variable

The presence or absence of a PO, party agent, or non-partisan observer signatures on the hypothetical election form is the main independent variable. I add to these attributes two additional ones: the presence or absence of vote count errors on the tallies of the two main candidates ⁷ . This enables an additional evaluation of how much more or less relevant the people at a station are when a voting irregularity (the crossing out of vote counts) occurs. ⁸ In terms of attributes of the conjoint, there are 6 total attributes (to be randomized, see Table 1 for summary). This is because the presence of party agents and vote errors is deconstructed into incumbent party (DPP and Jubilee) vs. opposition party (MCP and ODM) agents and errors.

These represent the treatments that a respondent would be exposed to as they review the hypothetical election form. The total number of combinations that arise from the above-mentioned treatments is a set of 2⁶ = 64 hypothetical form images for Kenya and Malawi. ⁹ . The images were created using a Python Image library and uploaded to Qualtrics. At the beginning of the survey, the Qualtrics software randomly selects six form images to be displayed in forced-choice pairs to each respondent.

Dependent variable

The outcome variable from this survey is the choice that the respondent makes during the forced-choice experiment. Since respondents have to click on what they believe is the more reliable form, their selection is turned into a binary variable called “chosen.”

After the forced choice, there are also four follow-up questions asking respondents about the reasoning behind their selection (see Table 2).

Table 1. A summary of the conjoint attributes and treatment levels. The Average Marginal Component Effect baseline attribute level is marked with (b). Appendix A2 provides an example of the altered form

Table 2. To what extent

The questions are asked for both forms that were presented in the forced choice. This information is used to understand why respondents may find the presence of an agent/observer/PO salient, providing a secondary source of evidence for the causal pathways between the experimental treatments and forced-choice outcomes.

Validity concerns

One implication of conducting an online survey (especially since both Malawi and Kenya have low internet penetration) is that the sample skews more educated. This impacts the design in two ways: firstly, observer awareness is likely tied to education and a skewed sample reduces the number low awareness respondents relative to what a more representative sample would have provided. Secondly, this also implies that the results hold strongest for more peri-urban/urban, educated, and/or politically engaged Kenyan and Malawian voters. Nonetheless, I would not discount that less educated or rural voters may also have concerns surrounding poll monitoring.

Results

The results from the forced-choice experiment illustrate which treatments, across all the forms, were the most salient for respondents (holding all other attributes constant). This is the Average Marginal Component Effect (AMCE) (Hainmueller et al., 2014). The AMCEs plotted are of the combined data of both Malawian and Kenyan respondents. ¹⁰ Data are combined since the hypotheses are not country-specific and the larger dataset provides more observations for each group in the subgroup analyses. Appendix A1 lists the number of observations per subgroup.

Full results

Figure 3 provides the AMCE for all conjoint attributes, without subsetting by the hypothesis-related respondent characteristics like trust, partisan identity, and observer awareness. In general, the signature of each poll monitor increased perceptions of tally sheet reliability. Substantively, this implies that voters respond positively to evidence of election monitoring from government, partisan, and non-partisan groups at a polling station. The subsequent subgroup analyses explore whether the hypothesized respondent characteristics underpin these results (Appendix A1 discusses relevant changes to statistical power).

Figure 3. Average Marginal Component Effect results for all conjoint attributes. All monitor groups are relevant for respondent perceptions of tally reliability.

Presiding officers

Figure 4 shows that the presence of POs is a salient feature across all trust levels. Among respondents with the lowest levels of trust in their EMB, the presence of a PO is still relevant to evaluations of integrity, as forms with PO signatures were preferred to those without PO signatures. The same outcome is seen among those with moderate and high trust. This conflicts with the hypothesis that voters’ response to the presence of election officials at a station is conditional on their trust in electoral institutions. The outcome additionally suggests (contrary to expectation) that the penalties for past EMB misconduct are not being extended to election officials. Alternatively, it could be that respondents believe that PO presence is important in general, and that a station having PO is a basic procedural expectation of election management for voters across all levels of institutional trust (Estévez et al., 2008).

Figure 4. Forms with presiding officer signatures were preferred, regardless of prior trust in Election Management Boards.

Political party agents

Figure 5 offers evidence that partisan monitoring is a salient feature, with respondents preferring forms where copartisan agents are present. However, respondents also preferred forms with other-party signatures, challenging the theoretical assumption that shared party identity is the main determinant of whether respondents considered the presence of partisan monitors. Additionally, the large estimate (AMCE = 0.34) for a combination of both copartisan and non-copartisan agent signatures suggests that having more eyes on the poll from different parties has the strongest (positive) impact on respondent perceptions of tally reliability. This preference for forms with non-copartisan agents suggests that factors beyond shared party identity may shape also voters’ attitudes toward party agents.

Figure 5. ‘Agents Present’ provides estimates for forms with one copartisan vs. one non-copartisan vs. both (copartisan and non-copartisan) agent signatures relative to a baseline of no agent signatures. Respondents preferred forms with the signatures of a copartisan and forms with other-party agent signatures.

Non-partisan observers

Lastly, Figure 6 shows that both respondents with and without prior awareness of observer groups considered the presence of observers salient, contrary to the hypothesis. Observer awareness was assessed by asking respondents to identify a domestic non-partisan observer group from a list of 5 organizations. A measurement consideration for this outcome is that while some struggled to recall specific names, the term “observer” itself, along with the survey’s description of them as “witnesses of vote counting,” may have provided enough information for unaware respondents to vaguely understand and consider observers’ role.

Figure 6. Respondents preferred forms with signatures of observer groups, regardless of their awareness of domestic non-partisan observer organizations.

Why do monitors matter?

Evidently, the presence of each polling station interest group is salient to voters, but not for the reasons suggested by the hypotheses. Respondents with different levels of institutional trust, partisan identities, and knowledge of observer groups display similar preferences in election monitoring. Accounting for the cited measurement caveats, the results suggest that conditioning perceptions of tally reliability on individuals’ characteristics is too deterministic an approach to understanding the logic of voter perceptions of electoral integrity. The effect of poll actor signatures is consistently positive, with range of 10–20 percentage-point increases in respondents’ perception of tally reliability, with effect sizes similar to those in experimental work on perceptions of partisan monitoring in the US by Sheagley and Cohen (2023).

Interestingly, the strong preference for two partisan agent signatures in Figure 5 highlights a vote preference for more eyes on the poll. This raises the broader question of whether respondents may just be counting signatures, basing their forced-choice selections on the absolute number of people at the station, and not necessarily distinguishing between the individual identities of the poll actors involved. To speak to both the potential complexity of respondent decision-making and evaluate the idea of inattentive respondent behavior, I look at all the conjoint attributes more broadly. I compare all poll actor combinations using the responses from the follow-up questions that asked about people’s reasons for preferring one collection of monitors over another. This is summarized in Figure 7.

Figure 7. Means and standard errors of concern level. The three-digit numbers in the x-axis, reading left to right, signify presiding officer (PO) presence, agent presence, and observer presence. Here, 0 = absent, 1 = present. For party agents specifically, 1 = one agent present and 2 = two different agents present. For example, “021” is: no presiding officer, two agents from opposing parties, and one observer.

Figure 7 highlights that more eyes on the poll do matter, but this is related to the protections conferred by additional poll actors (especially partisan ones), rather than the absolute count of poll actors. Some examples where the counting speculation is challenged are when forms with the same number of monitors yield different integrity outcomes, and also where forms with fewer monitors receive more or equally positive assessments. Notably, forms with only two agents (020) received a similar average rating as a form with a PO, an agent, and an observer (111) – despite the latter having a greater number and diversity of individuals. Additionally forms with only a PO and observer (101), which similarly have only two individuals, were rated almost the same as 020 except for the higher rating for “unprotected party interests,” illustrating a particular respondentʼs sensitivity to the presence and lack of party agents. This aligns with experimental findings by Sheagley and Cohen (2023), where scenarios in which there is only one, opposing-party agent present are viewed less favorably by respondents than instances where agents from both parties are present. Additionally, these observations both cast doubt on the idea of indifferent respondent behavior and show a sophisticated logic of respondents’ beliefs about the specific protections offered by each interest group. While the gold standard is understandable that all groups are present (121), the variability in the responses to each misconduct concern within and across each combination of polling actors suggests that respondents distinguished between poll monitor identities.

Conclusion

The finding that each of these groups matters for voter perceptions of integrity has implications for studying and creating policy around election monitoring.

First, it suggests that polling supervision should be accounted for when modeling surveys that elicit voter responses to information on polling conditions. While voters may not notice or respond to the presence of monitoring groups on voting day, my experimental findings in Kenya and Malawi, similar US results from Sheagley and Cohen (2023), and adjacent field data in Zambia, Kenya, and the Gambia (Macdonald et al., 2022) show that the presence of election officials, party agents, or observers remains salient when people are provided information about conditions at a polling place and have to make judgments about election quality. This also means that it is important to control for monitor presence in experimental studies of perceptions of election integrity where respondents can observe variation in poll supervision.

Further, the relevance of election officials, agents, and observers challenges existing literature that has found agent and observer presence irrelevant to perceptions of integrity. This perhaps speaks to the utility of deconstructing how voters’ perceptions of integrity are measured, and of exploring some of the more basic elements of electoral integrity (like the information on election forms). Future electoral research may benefit from extracting voter response data that is specific to the part of the election cycle being studied. However, the null results on voter characteristics suggest that the mechanisms behind voter perceptions of integrity extend beyond individual identities. Figure 7 highlights the added utility of exploring: voters’ understandings of the protections conferred by each actor; where they see opportunities for malpractice; and how these factors subsequently inform their evaluations of vote reliability.

Finally, there is the policy consideration that poll supervision and the identities of those involved can affect how the integrity of elections is perceived by voters. As election disputes draw public attention to polling station conditions, one way EMBs can maximize public confidence in elections is by facilitating a diversity of poll monitors. The experimental outcomes suggest that diversity of party agents is particularly useful for ameliorating misconduct concerns, which points to potential merits of encouraging election observation from both partisan and non-partisan groups.