Ballot measures, complexity, and cognitive biases

We aim to expand the ballot proposition literature’s engagement with complexity. Specifically, we explore how individuals evaluate multiple provisions within a single ballot proposition, and how those component evaluations translate into the overall support for the proposition. Prior works have considered complexity in terms of the number of ballot measures a voter considers in an election (Bowler and Donovan Reference Bowler and Donovan1998; Selb Reference Selb2008; Stadelmann and Torgler Reference Stadelmann and Torgler2013), the counts of words or lines within a measure (Nicholson Reference Nicholson2003), as well as the complexity of the language within the ballot (Reilly and Richey Reference Reilly and Richey2011). We shift the question of complexity to the dimensions of policy content in a single proposition and offer an alternative mechanism in this context.

Our perspective stems from a broad and deep literature on the structure of electoral processes (e.g., Dahl Reference Dahl1961; Riker Reference Riker1986; Schattschneider Reference Schattschneider1942), where substantial attention has been directed to the consequences of variation in ballot features (e.g., Engstrom and Roberts Reference Engstrom and Roberts2020; Hamilton and Ladd Reference Hamilton and Ladd1996; Miller and Krosnick Reference Miller and Krosnick1998; Reynolds and Steenbergen Reference Reynolds and Steenbergen2006; Taebel Reference Taebel1975; Walker Reference Walker1966). In an important study in this context Glaser (Reference Glaser2002) finds less support among Mississippians in 1999 for an omnibus proposal on school district improvements than for some of the features in it. He posits that the mechanism is the additional precision in expressing preferences on interlinked features (Lacy Reference Lacy2001) and the corresponding feelings of greater control, which allow individuals – particularly whites on this historically racially charged issue – to change their minds. In this vein, the format in which individuals make decisions affects the decisions themselves.

Previous works examining ballot measures have found that when voters evaluate ballot measures, they tend to engage in some form of cognitive bias against their passage. However, the underlying mechanism of this bias is still debated. Most notably, voters have been shown to be susceptible to a status quo bias in opposition to the measure – i.e., a natural opposition to the measure absent additional information (Bowler and Donovan Reference Bowler and Donovan2002). This bias is especially prevalent when multiple measures are on the ballot (Augenblick and Nicholson Reference Augenblick and Nicholson2016), measures are complex (Hessami and Resnjanskij Reference Hessami and Resnjanskij2019), or voters are largely uninformed regarding the contents of the measure (Barber et al. Reference Barber, Gordon, Hill and Price2017). Dyck and Pearson-Merkowitz (Reference Dyck and Pearson-Merkowitz2019) further argue that status quo bias in this context stems from negativity bias in response to media framing. By assigning voters to brief summary arguments of ballot propositions, their survey experiment finds that voters are swayed more by opposing than supporting frames. Of course, not all ballot issues can be boiled down to a single issue, nor do all voters enter the polls with clear signals (from the media or otherwise) on the pros and cons of every ballot proposition, let alone each provision within each ballot proposition.

Building on the theoretical insights and empirical findings in the literature, we posit that negativity bias may be endemic in evaluations of ballot propositions – i.e., not solely a product of their coverage or frames for or against a ballot proposal, but a function of the multiple provisions frequently detailed in its text. Thus, we move from considerations of the information environment around ballots to their exact text, from simplifications of ballots’ main features to their full set of provisions, from complexity as an overall measure of cognitive burden to complexity as aggregating multiple evaluations. In short, our theoretical perspective and research design consider the possibility for negativity bias to arise from a different source than previously offered. In line with the aforementioned literature, we expect a negativity bias, but posit a different mechanism for it: the aggregation of evaluations on multiple provisions within a single ballot proposal.

Negativity bias is the well-developed phenomenon that negative information has a greater influence on a host of evaluations than positive information (Kahneman and Tversky Reference Kahneman, Tversky, MacLean and Ziemba2013; Kanouse and Hanson Reference Kanouse, Hanson, Jones, Kanouse, Kelley, Nisbett, Valins and Weiner1987). Particularly relevant for our work, studies of impression formation have noted that the negative traits of evaluated individuals are weighted more heavily than their positive ones (Birnbaum Reference Birnbaum1972; Feldman Reference Feldman and Feldman1966; Fiske Reference Fiske1980). For example, Anderson (Reference Anderson1965) finds that positive trait evaluations are virtually averaged in the overall evaluation of the individual, while the overall evaluation of an individual with negative traits is more negative than their average (see also Ito et al. Reference Ito, Larsen, Smith and Cacioppo1998).

Because most modern ballot propositions contain multiple provisions, there is more than one evaluation to conduct for each proposition – not unlike traits within an individual. We expect voters to weigh the provisions with negative evaluations more heavily than those with positive ones in the construction of their overall support for a ballot measure. Thus, when looking at the overall support for a ballot measure with multiple provisions, we expect individuals to be negatively biased simply by virtue of the multiple components. Provided, of course, they view one or more components negatively, individuals should weight their overall score of a proposition downward in accordance with their least favored provision within it. In addition, we expect the provision-driven negativity bias to be strongest on the most complex and least polarized ballot propositions. In the following section, we describe the research design we created to test these hypotheses.

Data and design

To test our hypotheses, we conducted a survey of likely voters within the state of Missouri.Footnote ³ Overall, our survey included a sample of 1,107 respondents, largely representative of Missouri’s general population.Footnote ⁴ We randomly assigned each respondent to either a treatment or control condition with different levels of provisional focus. We first presented the entire ballot text for three potential ballot measures in both conditions. The prompts for each measure included the components of the multi-provisional ballot laid out in a bullet point format, as well as the projected cost of the measure to the state. Figure 1 presents an example prompt for one of the measures used in our research design. We used this standardized format to mimic how the measures were expected to appear on the November ballot and avoid additional ballot wording complexity. While most proposed ballot measures in Missouri follow this bulleted format, we slightly altered the text to improve clarity and better group components.

Figure 1.

Sample ballot measure prompt.

Figure 1. Long description

The document begins with the header: A potential ballot measure asks.

The main question is: Do you want to amend the Missouri Constitution to:

* require votes to be counted by hand not by machines

* require all ballots to be paper ballots

* require a state-issued I D or driver license to vote

* permit early voting no more than fourteen days prior to election

* allow observers inside and outside polling stations to take pictures

* allow the state General Assembly to adjust or void any vote totals it determines appropriate in presidential elections and

* create the crime of treason against persons for voter fraud, failure to report results timely and certain poll worker actions?

The text concludes with a fiscal note: State and local governments estimate ongoing costs of at least $24 million per election, but the total cost is unknown. Local governments estimate savings of at least $80,000 annually and at least $42,000 for each general election in a presidential election year. State governmental entities estimate no savings.

Immediately following each ballot measure text, we asked respondents randomly assigned to our control group how much they supported the measure on a four-point Likert scale. We recorded their response as an Overall score. For respondents randomly assigned to our treatment group, we instead asked them to rate each component of the measure on the same four-point Likert scale. To generate a comparable score between our treatment and control, we created a Composite score for these individuals by taking the average of the components. By calculating our composite score in this manner, we can test whether individuals’ overall support for a ballot measure is a simple aggregation across the multiple provisions or whether they use alternative methods to formulate their overall support.Footnote ⁵ Additionally, we can examine the correlation of support for each provision with the others. To avoid question ordering effects, we randomized the order in which the components were presented in the question. The assignment of treatment was kept consistent across all three measures – i.e., individuals were not asked for their general knowledge for one issue and the component for another or any such combination. As part of the larger survey in which our design was embedded, we also asked respondents a series of pretreatment demographic questions, which we use later in our multivariate analysis.

To enhance the external validity of our study, we leveraged the fair ballot text from three Missouri ballot measures in the process of gathering signatures in the spring of 2022. The first measure, Election Security, proposed a series of changes to the election process, including requiring identification to vote, poll watchers, and, notably, allowing the legislature to adjust vote totals. The second measure, Marijuana, proposed the legalization and recreational use and sale of marijuana. The final measure, Election Reform, proposed reforms to the state’s primary and general elections through open primaries and ranked-choice voting. Of the three measures, the marijuana legalization and election reform measures made it to the signature-gathering stage, but only the former made it onto the November ballot.Footnote ⁶

Beyond increasing our study’s external validity, our choice of ballot measures allows us to leverage the inherent policy differences between each measure. Specifically, our three measures vary regarding their policy complexity, saliency in news coverage, and partisan association. These features may make it easier or harder for individuals to conceptualize the specific policies and influence how they evaluate the ballots (Dyck and Pearson-Merkowitz Reference Dyck and Pearson-Merkowitz2019; Lupia Reference Lupia1994; Magleby Reference Magleby1984). In terms of complexity, both the marijuana and election security measure prescribe traditionally straightforward policy changes (e.g., the legalization of a substance, the implementation of additional requirements). However, policy changes such as ranked-choice voting or open primaries – two major provisions of the election reform measure – are likely far more abstract for voters. Similar differences exist in terms of each measure’s news coverage and interest, with both the election security and marijuana measures receiving more consistent policy coverage as compared to the election reform measure.Footnote ⁷ Finally, all three measures vary in terms of their partisan associations, with both the marijuana and election security measures having established – yet opposite – partisan signals and our election reform measure lacking a similar cue (Kimball et al. Reference Kimball, Anthony, Scott, Santucci, Drutman and Strano2022).

Composite versus overall measures of support

To test our hypothesis on provisional complexity, we compare the generated Composite scores to the Overall scores. If the scores are indistinguishable, it suggests that individuals’ overall score is a particular aggregation of their support for each component of a multi-provisional ballot – specifically, likely voters evaluate ballots by averaging across the multiple provisions within them. If, however, the scores are different, it suggests that they engage in an alternative weighting of components when calculating their overall support for a ballot proposal.

Figure 2 presents the average overall, composite, and component scores for each ballot measure, along with their 95% Bonferroni-corrected confidence intervals.Footnote ⁸ Importantly, across all issues individuals appear to weigh each component differently, with some components rated significantly more/less favorably than others, which suggests that voters may be sufficiently competent to evaluate even complex multi-provisional ballots in a manner consistent with their policy beliefs. However, across the issues we also find differences.

Figure 2.

Differences in average support by ballot measure and treatment group. Results include 95% Bonferroni-corrected confidence intervals. Composite support scores are the average of a measure’s component support scores.

Figure 2. Long description

The dot plot displays support scores on the X axis ranging from 2.0 to 3.2. The Y axis lists propositions grouped into three sections separated by solid horizontal lines. Each data point includes a 95 percent Bonferroni-corrected confidence interval represented by horizontal error bars. Points are color-coded by type: black for Overall, red for Composite, and blue for Component.

* Election Security section: The Overall score is approximately 2.7. The Composite score is slightly higher at 2.75. Component measures vary widely: Adjust Vote Totals is the lowest at 2.05, Prohibit Voting Machines is at 2.5, and Require Voter I D is the highest at 3.1.

* Marijuana section: The Overall score is 2.8. The Composite score is 2.9. Component measures include License Lottery at 2.6, while Legalization, Cultivation Registration, Expunge Records, License Cong. Dist. Equally, and Sales Tax all cluster between 2.9 and 3.0.

* Election Reform section: The Overall score is 2.75. The Composite score is 2.95. Component measures show Ranked Choice Elections at 2.7, Open Primaries at 3.0, and Paper Record at 3.1.

Legend at the bottom indicates Type: Overall (black circle), Composite (red circle), and Component (blue diamond).

For both our election security and marijuana measures, the composites are more supportive than the overall scores, though the difference is not statistically significant ( $ p=0.25 $ and $ p=0.21 $ , respectively). For these measures, individuals’ overall assessments resemble the averages of their support across the components.Footnote ⁹ However, for the election reform measure, the overall and composite scores are statistically different and substantively large. Here the overall score is 0.21 points lower than the average of the provisions. Respondents’ overall scores for election reform is not a simple average of their support for each component. Instead, the overall score is best correlated with the most negatively rated component, ranked-choice voting. Indeed, a respondent’s overall support appears closest to effectively weighing their least-preferred provision three times as much as all other components.Footnote ¹⁰

Overall, we find that down-weighting exists across all three ballot measures, though it is only significant and substantively large on the election reform issue. When faced with a multi-provisional ballot, individuals may not always weigh their opinion equally between the components. When they do vary in this context, respondents appear to weigh negatively viewed components more heavily than those they view positively.Footnote ¹¹

Issue-specific heterogeneity

The striking result from the analysis above is the difference between the overall and composite support scores. Respondents’ overall scores are more correlated with their most negatively rated component than the composite score. Notably, this negativity bias, while present for all propositions, only manifests in a significant difference in support for the election reform measure. What makes the election reform measure different? While our design precludes an explicit test for the various factors that lead to different composite and overall evaluations (for the same individual), we can exploit the correlates of support to gain some leverage on the different natures of the policies.

The election reform measure sits in a potentially unique situation compared to the other measures in our study. First, despite their prevalence within political science, ranked choice voting and open primary elections – two core components of the measure – received far less national or local news coverage than the other measures, and voters may be less informed regarding their meaning and details (Chapp and Aehl Reference Chapp and Aehl2021; Fahey, Weissert, and Uttermark Reference Fahey, Weissert and Uttermark2018).Footnote ¹² Second, unlike either the election security or marijuana measures, the election reform does not benefit from pre-existing partisan signals regarding the policy changes (Kimball et al. Reference Kimball, Anthony, Scott, Santucci, Drutman and Strano2022). Without these signals, voters cannot lean on their partisan identities when forming an opinion like the other two measures. Thus, being uninformed regarding a complex topic and unable to rely on their political identity, the literature suggests that respondents’ evaluations may reflect a cognitive bias.Footnote ¹³ In the multi-provisional context, we expect the public to disproportionately weight down their overall evaluations based on their least favorable component(s), or exhibit negativity bias, particularly when the issue is complex and lacks salience and partisan cues.Footnote ¹⁴

Motivated by these expectations, we construct an ordered logistic regression model for each measure’s overall support score. We test for correlation with an individual’s politics through party identification, ideology, and self-reported importance of partisan identity. To examine issue complexity and saliency, we also test for correlation with media consumption, political sophistication, and education. While these sets of correlations pertain only to respondents’ support for each measure, they aid us in contextualizing the previous results on the different degrees of negativity bias across the issues.

The results presented in Figure 3 largely corroborate the issue-specific expectations. Foremost, an individual’s partisan identity appears to influence evaluations of only the election security measure, but not evaluations for the more complex and less politicized election reform measure. Similarly, party identification importance does not correlate with support for election reform.Footnote ¹⁵ While the sophistication relationship is consistently negative across the measures, education has a uniquely positive relationship with support for election reform, suggesting a higher level of complexity on this issue. Media consumption is related to support for both marijuana and election security, but not election reform. The lack of an effect is likely due to the relatively sparse media coverage surrounding the election reform measure and its policies.Footnote ¹⁶ Overall, these correlations provide preliminary support for our expectations that individuals are more likely to engage in provision-based negativity bias – disproportionately weighting down their overall evaluations based on their least favorable components – when the ballot measures are complex, non-salient, and without clear partisan cues.

Figure 3.

Individual-level support for the overall ballot. The plot presents the change in logged odds from individual ordered logistic regressions with 95% Bonferroni-corrected confidence intervals. Full regression results and complementary composite-score regressions are in the Supplementary Material.

Figure 3. Long description

A coefficient plot with the y-axis labeled Covariate and the x-axis labeled Logged Odds Ratio ranging from minus 0.5 to 0.5. A vertical dashed line marks the zero point. Six covariates are listed from top to bottom: Education, Political Sophistication, Media Consumption, Republican, Ideology, and P I D Importance. Each covariate has three horizontal error bars representing different Ballot Measures: Election Reform (green), Election Security (orange), and Marijuana (blue).

* Education: All three measures cluster near zero with narrow intervals.

* Political Sophistication: All three measures are slightly negative, with Marijuana being the least negative.

* Media Consumption: Election Security and Marijuana show positive logged odds, while Election Reform is centered near zero.

* Republican: This category shows the highest variance. Marijuana is negative, Election Reform is slightly positive, and Election Security shows a strong positive logged odds ratio with a wide confidence interval extending toward 0.8.

* Ideology: Marijuana and Election Reform are negative, while Election Security is positive.

* P I D Importance: Marijuana is slightly negative, while Election Security and Election Reform are slightly positive.

A legend at the bottom identifies the markers: a green circle for Election Reform, an orange circle for Election Security, and a blue diamond for Marijuana.