Introduction
A well-functioning democracy requires that citizens are able to make an informed decision when they vote. Voting for a party, and even more so for a ballot measure in a direct democratic vote, places high demands on voters. While direct democracy has attracted increasing public and scholarly interest, important questions remain about whether initiative and referendum campaigns enable voters to make coherent choices. Previous studies in the United States have highlighted the ‘educative effects’ of direct democracy (see, for example, Smith and Tolbert Reference Smith and Tolbert2004), but few have examined how and through what channels voters form their opinions during campaigns. This paper addresses this gap by examining the processes through which citizens learn and align their votes with policy arguments and party recommendations over the course of direct democratic campaigns.
Drawing on dual-process theories from social psychology (Eagly and Chaiken Reference Eagly and Chaiken1993; Petty and Cacioppo Reference Petty and Cacioppo1986), recent scholarship suggests two main pathways for opinion formation in direct democratic contexts (Barbieri et al. Reference Barbieri, Petitpas and Sciarini2025,Boudreau and MacKenzie Reference Boudreau and MacKenzie2014; Bullock Reference Bullock2011; Colombo and Kriesi Reference Colombo and Kriesi2017; Kriesi Reference Kriesi2005; Nai Reference Nai2014): heuristic and systematic processing. Heuristic processing relies on shortcuts, such as following the voting recommendation of one’s preferred party, while systematic processing engages voters in a more sophisticated treatment of political information provided during referendum or initiative campaigns. This article aligns with the view that campaigns foster voter learning and coherence in decision making (see, for example, Arceneaux Reference Arceneaux2006; Hobolt Reference Hobolt2009; Kriesi Reference Kriesi2005; Reference Kriesi and Kriesi2012a; Lupia Reference Lupia1994; Selb et al. Reference Selb, Kriesi, Hänggli and Marr2009). Consequently, we build a comprehensive model that combines the two pathways of information processing. We focus on these two dimensions because they capture distinct but equally important components of informed decision making in direct democratic settings: party cues help voters to translate broader political orientations into concrete choices, while policy arguments speak to the ability to reason about issue content. Both represent meaningful forms of opinion formation in the context of complex ballot measures.
Our first set of hypotheses proposes that, as the campaign unfolds and voting day approaches, voters become more knowledgeable about party cues and more engaged with policy arguments both in favor of and against a given ballot measure. We similarly expect that the likelihood of voters aligning their vote with their preferred party’s recommendation – what we refer to as in-line voting – increases over the course of the campaign. A similar pattern is hypothesized for consistent voting, defined as casting a vote that aligns with one’s stance on the key pro- and contra-policy arguments related to the measure – that is, voting Yes when one supports the pro-arguments and rejects the contra-arguments, and voting No when one rejects the pro-arguments and supports the contra-arguments. Our second set of hypotheses links these learning processes to voting behavior. Specifically, we posit that increased knowledge of party cues leads to a rise in in-line voting, while a greater ability to take a position on policy arguments fosters more consistent voting.
We test these two sets of hypotheses on two citizen-sponsored initiatives in California (Propositions 26 and 27 on sports betting) and two referendums in Switzerland (introduction of the Organisation for Economic Co-operation and Development (OECD) minimum tax rate and revision of the Climate law). California and Switzerland share an extensive experience in direct democracy, but differ with respect to elite involvement in direct democratic processes (Kriesi Reference Kriesi, Nahrath and Varone2009). This, together with the variety of the four ballot measures under consideration, allows us to test our hypotheses in a range of situations.
A common weakness in the literature on both sides of the Atlantic is the lack of attention to the dynamics of opinion formation. To capture the day-to-day dynamics of how voters engage with information and form opinions during direct democratic campaigns, we rely on rolling cross-section (RCS) data. The RCS design has been used in some national election studies (see, for example, Johnston and Brady Reference Johnston and Brady2002; Johnston et al. Reference Johnston, Partheymüller, Schmitt-Beck, Wessels, Rattinger, Rossteutscher and Schmitt-Beck2014; Partheymüller and Johnston Reference Partheymüller, Johnston, Schmitt-Beck, Rossteuscher, Schoen, Wessels and Wolf2022; Schmitt-Beck and Staudt Reference Schmitt-Beck, Staudt, Schmitt-Beck, Rossteuscher, Schoen, Wessels and Wolf2022), but very rarely in direct democratic campaigns (for exceptions, see Faas Reference Faas2015; Schoen et al. Reference Schoen, Glantz, Teusch, Feld, Huber, Jung, Welzel and Wittreck2011), and it has never been applied to a comparative study of within-campaign dynamics. To assess these dynamics, we conduct both an aggregate-level and an individual-level analysis of daily data. At the individual level, we fit structural equation models that test the mediating effects of knowledge gains on in-line and consistent voting.
Literature Review
Extensive research on electoral campaigns has shown that they are critical moments for political learning (Arceneaux Reference Arceneaux2006; Brady et al. Reference Brady, Johnston, Sides, Brady and Johnston2006; Hansen and Pedersen Reference Hansen and Pedersen2014; Holbrook Reference Holbrook1996). While much of this literature focuses on party or candidate elections, its insights are also relevant to direct democratic contexts. Like electoral campaigns, direct democratic campaigns offer structured opportunities for citizens to encounter political information and form an opinion. However, they also differ in key ways – for example, their single-issue character, the binary architecture of the choice (Yes v. No), and the general unfamiliarity of ballot measures (de Vreese Reference De Vreese2007). This latter point is particularly important, as it may heighten opinion volatility during referendums due to less entrenched attitudes (Farrell and Schmitt-Beck Reference Farrell and Schmitt-Beck2002; LeDuc Reference LeDuc, Farrell and Schmitt-Beck2002).
In the United States, an important stream of literature highlights the ‘educative effects’ of the availability and use of direct democratic institutions (Smith and Tolbert Reference Smith and Tolbert2004), which have a positive impact on voters’ political knowledge, interest in politics, efficacy, and turnout (Bowler and Donovan Reference Bowler and Donovan1998; Smith and Tolbert Reference Smith and Tolbert2004; Smith Reference Smith2001; Reference Smith2002); for a diverging view, see Barth et al. (Reference Barth, Burnet and Parry2020), Dyck and Lasher (Reference Dyck and Lascher2009), and Schlozman and Yohai (Reference Schlozman and Yohai2008). However, only a few studies have examined the dynamics of opinion formation during initiative campaigns. Moreover, these studies focus mainly on the effects of ballot characteristics and campaign activities on mobilization and participation – in presidential, mid-term, local, or initiative elections (Donovan et al. Reference Donovan, Tolbert and Smith2009; Dyck and Seabrook Reference Dyck and Seabrook2010; Hillygus Reference Hillygus2005; Niven Reference Niven2004; Parry et al. Reference Parry, Barth, Kropf and Jones2008; Tolbert et al. Reference Tolbert, Bowen and Donovan2009); they do not examine whether and to what extent campaigns help voters to learn about ballot measures and contribute to the quality of their voting decision.
As the country with the most experience in direct legislation at the national level, Switzerland is often referred to as a ‘laboratory’ (Kriesi Reference Kriesi2005; Sciarini and Tresch Reference Sciarini, Tresch, Emmenegger, Fossati, Häusermann, Papadopoulos, Sciarini and Vatter2024). Accordingly, there is a rich literature on campaign effects and vote choice in direct democratic votes. A first set of studies has applied Zaller’s (Reference Zaller1992) model of opinion formation to post-referendum survey data, supporting the view that reception and acceptance of elite communication depends on the interaction between voters’ level of political awareness and their partisan predispositions (Bützer and Marquis Reference Bützer, Marquis, Farrell and Schmitt-Beck2002; Kriesi Reference Kriesi2005; Marquis and Sciarini Reference Marquis and Sciarini1999; Sciarini and Tresch Reference Sciarini and Tresch2011). A second aspect of direct democracy that has been extensively researched is whether voters make decisions based on a systematic processing of information or on heuristics. Kriesi (Reference Kriesi2005) finds that the partisan heuristic, that is, voting in line with the recommendation of the political party to which the voter feels closest, is a well-developed strategy (see also Milic Reference Milic2020). However, systematic decision making based on a more sophisticated processing of the information about the pros and cons of a ballot measure is also commonly used (Milic Reference Milic2020).
Consistent with the view that US citizens base most of their decisions on simple and easily accessible information (Lupia and Matsusaka Reference Lupia and Matsusaka2004; Magleby Reference Magleby1984), Lupia (Reference Lupia1994) shows that the use of information shortcuts (knowledge of the voting recommendations of major insurance companies) enables low-competence voters to imitate the voting behavior of high-competence voters (see also Bowler and Donovan Reference Bowler and Donovan1998). Several studies across Europe confirm that partisan heuristics and systematic processing play a role in vote choice, either separately (Elkink and Sinnott Reference Elkink and Sinnott2015; Suiter and Reidy Reference Suiter and Reidy2013) or jointly (Bergman Reference Bergman2020; Hobolt Reference Hobolt2005; Reference Hobolt2007). These studies show that cue-taking and issue-based considerations can co-exist and jointly inform voting behavior.
Apart from dual process theories, another framework that is often used to assess the relative importance of party cues and policy information in voters’ decisions in direct democratic votes is the theory of partisan-motivated reasoning. According to this theory, voters process information with a partisan bias, that is, they tend to blindly follow the cue of their preferred party and ignore policy information. However, the empirical evidence is mixed. For example, an experimental study by Boudreau and MacKenzie (Reference Boudreau and MacKenzie2014) finds that when voters receive substantial policy information that contradicts their party’s position on a given citizen-sponsored initiative, they tend to shift their opinions away from the party’s position and show opinions that are no different from the control group. In contrast, Barbieri et al. (Reference Barbieri, Petitpas and Sciarini2025) show that during a California initiative campaign, party cues override policy information even when contradicting initial vote intentions. Their pre–post experiment demonstrates that partisan-motivated reasoning shapes opinion change, especially among voters with weak prior attitudes and strong party ties.
The studies mentioned so far are based on post-election surveys or on experimental surveys, and are static in nature – for an exception, see Slothuus & Bisgaard (Reference Slothuus and Bisgaard2021). On the one hand, studies based on post-election surveys cannot, by definition, capture campaign-specific shifts in public opinion. On the other hand, while experimental studies help to solve the problem of causal attribution, they cannot create real campaign stimuli, and thus identify potential rather than actual effects (Brady et al. Reference Brady, Johnston, Sides, Brady and Johnston2006).
To assess the dynamic nature of opinion formation, some studies have turned to panel survey data (Bernhard Reference Bernhard2018; Colombo and Kriesi Reference Colombo and Kriesi2017; Kriesi Reference Kriesi and Kriesi2012b; Selb et al. Reference Selb, Kriesi, Hänggli and Marr2009). They highlight the high degree of instability of individuals’ voting intentions. Bernhard (Reference Bernhard2018) finds that post-campaign levels of knowledge about specific ballot measures are higher than pre-campaign levels. Relatedly, Selb et al. (Reference Selb, Kriesi, Hänggli and Marr2009) show that referendum campaigns help citizens to adjust their vote intentions to their partisan orientation (see also Kriesi Reference Kriesi and Kriesi2012b). This partisan alignment, interpreted as a sign of cue taking, is helpful for all voters, and especially among voters who are ambivalent at the outset of the campaign. Testing the claim of motivated reasoning theory in the Swiss context, Colombo and Kriesi (Reference Colombo and Kriesi2017) find that both policy arguments and party cues influence vote intentions, but that voters tend to align their position on arguments with the position of their preferred party during the campaign. Therefore, party cues influence vote choice directly and indirectly, by shaping how arguments are processed.
Data from a panel survey conducted during the 2016 Italian constitutional referendum in Italy address the dynamics of heuristic versus systematic processing during referendum campaigns, but they do not focus on party cues. Instead, they test the role of the government trust heuristic (de Angelis et al. Reference De Angelis, Colombo and Morisi2020) and the status quo heuristic (Morisi et al. Reference Morisi, Colombo and De angelis2021). The former study highlights the overall predominance of government cues while also underscoring the complementary role of systematic processing across different voter types. The latter also provides a nuanced perspective, showing that policy information attenuates the influence of cues, particularly among right-wing voters.
In sum, while existing studies have shown that both party cues and policy arguments influence voting choice in direct democratic contexts, the relationship between the two remains theoretically and empirically ambiguous. Some evidence points to their co-existence, but the underlying mechanisms are not yet fully understood. In this study, we bring party-based and argument-based voting together in a single analytical framework that also integrates voters’ political knowledge, allowing us to assess how these different pathways contribute to voters’ decision making. Most importantly, we move beyond static designs by leveraging RCS data to examine the dynamics of opinion formation throughout the campaign. This enables us to trace how voters come to align with cues and arguments over time, offering new insights into the foundations of voting.
Hypotheses
Democratic campaigns are information-rich events that provide voters with the opportunity to learn about the ballot measures at stake and thus have an enlightenment effect (Bernhard Reference Bernhard2018; Bowler and Donovan Reference Bowler and Donovan2002; Kriesi Reference Kriesi and Kriesi2012a; Selb et al. Reference Selb, Kriesi, Hänggli and Marr2009). This tends to mitigate voters’ initial lack of knowledge regarding the often unfamiliar and complex ballot measures (Bergman Reference Bergman2020; Bowler Reference Bowler2015; Hobolt Reference Hobolt2007), thereby reducing attitudinal uncertainty and enabling voters to make decisions aligned with their predispositions (Bowler et al. Reference Bowler, Dobbs and Nicholson2020; Dermont and Stadelmann-Steffen Reference Dermont and Stadelmann-Steffen2019; Hobolt Reference Hobolt2005; Selb et al. Reference Selb, Kriesi, Hänggli and Marr2009).
Specifically, according to dual-process theories, initiative and referendum campaigns can have two types of informational effects. First, following the heuristic path of information processing, campaigns can help voters identify which political parties support or oppose the ballot measure, and thus learn their preferred party’s voting recommendation (Bowler and Donovan Reference Bowler and Donovan2002, Denver Reference Denver2002, Selb et al. Reference Selb, Kriesi, Hänggli and Marr2009). Second, following the systematic path, campaigns provide voters with detailed information about the strengths and weaknesses of ballot measures (Bowler Reference Bowler2015; Colombo and Kriesi Reference Colombo and Kriesi2017). Over time, an increasing number of voters are likely to recognize their preferred party’s cue and gain awareness of the arguments for and against ballot measures. These insights inform the first two hypotheses:
HYPOTHESIS 1a Voters’ knowledge of their preferred party’s voting recommendations increases over the course of the campaign.
HYPOTHESIS 1b Voters’ ability to take side on the arguments for and against a ballot measure increases over the course of the campaign.
Turning to the dynamics of vote intentions, party endorsements can have a significant impact on voters’ choices (Kriesi Reference Kriesi2005; Selb et al. Reference Selb, Kriesi, Hänggli and Marr2009; Walder and Strijbis Reference Walder and Strijbis2022). As Kriesi (Reference Kriesi2005) bluntly puts it in the Swiss context, party cue is ‘the quintessential shortcut in direct democratic votes’. In a context of limited information, party cues can be used as heuristics to sort through the information demands and help voters make reasonable decisions (Hobolt Reference Hobolt2007, Milic Reference Milic2020, Suiter and Reidy Reference Suiter and Reidy2015). To the same extent that campaigns help voters to learn about their preferred party’s vote recommendation (hypothesis 1a), they should also help them to cast a vote that is in line with their preferred party’s cue.
HYPOTHESIS 1c Voters’ ability to vote in line with the vote recommendation of their preferred party (‘in-line voting’) increases over the course of the campaign.
Similarly, campaigns may help voters to cast a consistent vote, that is, a vote that is consistent with their position on the arguments for and against a given ballot measure (Lanz and Nai Reference Lanz and Nai2015; Lauener Reference Lauener2020; Milic Reference Milic2012). Indeed, as a campaign unfolds, its intensity grows, and policy arguments become increasingly visible to the public. This makes it easier to assess the pros and cons of the ballot measure (Bowler Reference Bowler2015; Kam Reference Kam2006). Moreover, the increase in the amount of accessible information motivates voters to think about the issue and thus to make a decision with a higher degree of elaboration (Hobolt Reference Hobolt2005). As a result, by the end of the campaign, voters should be better able to cast a vote consistent with their underlying policy preferences.Footnote 1
HYPOTHESIS 1d Voters’ ability to cast a vote that is consistent with their position on the pro and con arguments (‘consistent voting’) increases over the course of the campaign.
Finally, it should be clear from the development above that the knowledge variables and in-line or consistent voting are closely related. As election day approaches, greater knowledge of party cues increases the likelihood of in-line voting. Similarly, as voters’ knowledge of policy arguments increases over time, they are more likely to vote consistently with their argument position. In other words, knowledge gains act as a mediator in the relationship between time as a measure of campaign progress, on the one hand, and in-line and consistent voting, on the other (see Figure 1).
Mediation model.
HYPOTHESIS 2a Increased knowledge of the party cue over the course of the campaign contributes to an increase in voting in line with the cue of one’s preferred party.
HYPOTHESIS 2b The increased ability to take sides on the pro and con policy arguments over the course of the campaign contributes to an increase in voting in accordance with one’s position on policy arguments.
In this paper, we do not see the heuristic and systematic paths as mutually exclusive. According to our more realistic view, voters may draw on both party cues and substantive arguments simultaneously (Kriesi Reference Kriesi2005, Milic Reference Milic2020).
Methodological Framework
Cases
Our analysis is based on four cases drawn from two distinct institutional contexts, each featuring two direct democratic votes on different topics. The central question we pose is whether, despite these contextual and topical differences, similar patterns of opinion formation emerge – specifically, increases in issue-specific knowledge, and consequently, in in-line and consistent voting.
Starting with the institutional context, according to Kriesi’s (Reference Kriesi, Nahrath and Varone2009) threefold typology, the citizen-sponsored initiatives at work in California represent the unmediated (or ‘populist’) variant of direct democracy. In this variant, popular votes follow a bottom-up logic. Ballot measures emanate from demands raised by interest groups or initiative committees, which helps interest groups and social movements bypass state authorities (governors and legislatures). Moreover, legal restrictions on involvement, support, and spending mean that state governments do not actively engage in initiative campaigns, in which interest groups take the lead and political parties play a secondary role.
In contrast, the Swiss direct democracy belongs to the ‘mediated’ variant, where referendum and initiative votes are more tightly controlled by the elite – government, legislature, and governing parties (Kriesi Reference Kriesi, Nahrath and Varone2009).Footnote 2 In particular, unlike in US states, the government and governing parties actively campaign for or against ballot measures in order to influence voters’ decisions.Footnote 3 The role of interest groups is less visible and operates mainly through the financial support of political parties’ campaign activities (for example by funding ads in newspapers or on the internet).
Furthermore, while direct democracy is cognitively demanding in both California and Switzerland, the sources of complexity differ. In California, direct democratic votes are held concurrently with presidential, midterm, and/or state-level elections. The simultaneity of multiple ballot measures and multiple electoral contests requires voters to make numerous decisions at once (Bowler Reference Bowler2015). In such a context, the challenge lies not only in cognitive complexity but also in time constraints, as voters must be motivated to engage with multiple elections within a short timeframe. Consistent with this, the prevailing view in the United States is that citizens tend to have limited knowledge about the ballot measures submitted to them (Barth et al. Reference Barth, Burnet and Parry2020; Boudreau and MacKenzie Reference Boudreau and MacKenzie2021; Bowler and Donovan Reference Bowler and Donovan1998; Lupia Reference Lupia1994). In Switzerland, by contrast, direct democratic votes occur outside the context of national elections, but in some cantons, they may coincide with state-level or local elections. Moreover, federal ballots are held more frequently than in US states: citizens are called to the polls three to four times per year to vote on one or several federal ballot measures – and often on additional local measures as well. This requires long-term, sustained motivation.
Turning to the characterization of ballot measures, in California, we study two ‘combined initiated constitutional amendments and state statutes’ to legalize sports betting that were submitted to voters in parallel to the November 2022 midterm elections.Footnote 4 The two ballot measures competed against each other. Proposition 26 aimed to legalize sports betting exclusively in Native American casinos, while Proposition 27 aimed to legalize online sports betting – and to create in parallel a homelessness prevention fund. The campaign became a competition among interest groups aiming to gain exclusive control over a future market of tens of millions of consumers. It resulted in the most expensive initiative campaign ever in California ($400 million).Footnote 5 While the two ballot measures addressed the same policy issue and had some similarities in content, they differed in terms of partisan support. The Republicans opposed both ballot measures, whereas the Democrats opposed Proposition 27 but did not take a position on Proposition 26. Ultimately, voters overwhelmingly rejected both ballot measures (67 per cent against Proposition 26 and 82.3 per cent against Proposition 27).
In Switzerland, our study covers two referendums held in June 2023. The first was a constitutional amendment and was therefore subject to a mandatory referendum. The amendment, promoted by the government and parliament, aimed to implement a global tax reform initiated by the OECD and the G20, by introducing a minimum tax of 15 per cent on large multinational corporations. All parties except the Socialist party supported the reform, while the Greens let the freedom of vote. The broad partisan consensus in favor of a rather technical and unfamiliar ballot measure, together with the mainly platonic socialist opposition, resulted in a low-intensity and highly one-sided referendum campaign.Footnote 6
The second vote resulted from a referendum launched by the Swiss People’s Party against the revision of the Climate law. The revision created a comprehensive framework to achieve climate neutrality by 2050 and increase energy security through various means and incentives, such as reducing fossil fuel consumption by subsidizing building insulation and promoting renewable energies. Of the six main parties, only the Swiss People’s Party advocated a No vote, but it invested significant resources in the referendum campaign, which reached a fairly high level of intensity by Swiss standards.Footnote 7 In the end, both proposals were approved by the people, with an overwhelming support to the OECD reform (78.5 per cent Yes), and a large margin for the Climate law (59.1 per cent Yes).
Overall, then, in addition to the differences between the two direct democratic contexts described above, the differences between ballot measures and related contexts (institutional type, campaigns, and configuration of party cues) provide an opportunity to examine the dynamics of citizens’ attitudes in a variety of situations.
RCS Survey Data
This paper is based on the second wave of a three-wave panel survey conducted in California and Switzerland, with the first wave taking place before the campaign, the second during the campaign, and the third after the vote.Footnote 8 The second wave took the form of a RCS. Each day during the forty days before the vote, we drew a random sample of wave 1 respondents (see Appendix A for sample representativeness). The average number of respondents per day was forty (SD = 13) in California and fifty-nine (SD = 13) in Switzerland, for a total of 1,624 and 2,408 interviews, respectively.
The RCS design is an appropriate tool for tracking changes and trends in public opinion over the course of a campaign, and offers some advantages over panel studies (Brady et al. Reference Brady, Johnston, Sides, Brady and Johnston2006). In particular, as a result of the coarse granularity of the panel, observed changes between two waves may be due to events occurring in-between; the greater the gap between waves, the harder it is to identify campaign effects as competing explanations accumulate. RCS’s daily interviews help overcome the problem, providing detailed information on the dynamics of knowledge acquisition and vote intentions.Footnote 9
Measures
A key factor in this study is the time that elapses as the campaign progresses, specifically the date of each respondent’s interview during the RCS. This is expressed as a continuous variable ranging from −40 to 0, with 0 being the day closest to the election day. We use time as a variable in the remainder of the paper, bearing in mind that time is our measure of campaign progress.
The first mediator is knowledge of party cue. It is based on two questions. In the first wave of the panel, respondents were asked which party they felt closest to (see Appendix A for survey questions and descriptive statistics). In wave 2 (RCS), respondents were then asked if they knew the cue of their preferred party with the following question: ‘To your knowledge, what is the [preferred party’s] vote recommendation for the following ballot measures?’ The variable is coded 1 if the respondent gave a correct answer and 0 if they gave an incorrect answer or did not know.Footnote 10 Note that we focus only on respondents who indicated a preferred party; by definition, respondents who did not feel close to any party could not be asked about the voting recommendation of their ‘preferred party’. The proportion of partisans is 75 per cent in Switzerland (including the six major parties) and 78 per cent in California.
The second mediator is the voter’s ability to take a position on policy arguments. For each ballot measure, we formulated six arguments – three in favor and three against – reflecting the positions promoted by the Yes and No camps during the referendum or initiative campaigns (see Appendix A for the full list of arguments).Footnote 11 To mitigate the risk of ex-post rationalization – that is, the tendency of respondents to align their evaluation of policy arguments with their vote intention or vote choice – we carefully formulated the arguments to avoid any direct cues with a particular voting position, ensuring that respondents could not readily infer whether an argument supported or opposed the measure. Moreover, we deliberately withheld information about the source of each argument – such as the sponsoring political actors behind the initiative in California or the referendum in Switzerland – to prevent respondents from relying on source cues when evaluating the arguments. Finally, we presented the argument-related questions prior to the vote intention or choice question, thereby avoiding the risk that voters would adjust their evaluations of arguments to conform with their stated vote.
All six argument items followed the same standardized question format: ‘To what extent do you agree or disagree with the following arguments regarding [name of ballot measure]?’, with responses ranging from ’strongly agree’ to ’strongly disagree’, as well as a ‘don’t know’ option. We sum the number of ‘don’t know’ responses provided by respondents for each ballot measure and use this as a measure of voters’ ability to take side on policy arguments.
The first dependent variable measures whether the respondent’s vote intention or decisionFootnote 12 is in line with their preferred party’s recommendation. This variable is coded as 1 if the respondent’s vote is in line with the recommendation and 0 if it is not.
The second dependent variable measures whether the respondent’s vote is consistent with their opinion on policy arguments. This measure is based on the respondent’s position on an argument scale ranging from -12 (extremely against) to 12 (extremely in favor) that summarizes their opinion on the six arguments per ballot measure mentioned above. For example, a respondent with a score of 12 ’strongly agrees’ with the three arguments in favor of the ballot measure and ’strongly disagrees’ with the three arguments against the ballot measure. To derive our measure of voting according to position on the arguments, we match this scale to the respondent’s vote intention/choice. The binary variable is coded 1 if the voting decision is consistent with the arguments and 0 if it is not. For example, if a respondent votes ‘Yes’ and is in favor of the ballot measure according to their position on the argument scale (for example position 7 on the scale), the variable is coded as 1. In contrast, the variable is coded as 0 if the respondent’s vote is inconsistent with their opinion on arguments, that is, if they vote ‘Yes’ (‘No’) but are against (for) the ballot measure according to their position on the argument scale. More specifically, on the argument scale, we consider respondents to be in favor of the ballot measure if they are between + 2 and + 12 and against if they are between -2 and -12. The vote of respondents with an ambiguous position on the arguments scale (position between -2 and + 2) is coded as inconsistent. The same applies to those (few) respondents whose vote intention is undecided.
As control variables, we first include a measure of respondents’ issue-specific knowledge based on factual knowledge questions about each ballot proposal – two questions in California and three in Switzerland. This issue-specific knowledge scale represents the number of correct answers given by the respondent. Second, we measure the subjective importance that respondents assign to each proposition on a scale of 0–10, where 0 is ‘not at all important’ and 10 is ‘very important’. We also include age, gender, and education level as socio-demographic controls.
Empirical Strategy
Our empirical strategy consists of two steps, each based on between-individual analyses of daily data. First, to test the hypotheses H1a to H1d, we conduct an analysis at the aggregate level. That is, we aggregate the data for each day of the RCS to obtain the percentages or means for each variable of interest. We then plot these values over time and derive the best-fit lines using linear models.Footnote 13
Second, to test hypotheses H2a and H2b at the individual level, we specify a structural equation model (SEM) for each ballot measure. To build this model, we start with four equations to predict the two mediators and the two outcomes:
$$\matrix{ & {\rm{KnowledgePartyCue}} = {a_1} \cdot {\rm{time}} + X \cdot {\beta _1} + {\varepsilon _1} \cr & {\rm{Inline}} = {c_1} \cdot {\rm{time}} + {b_1} \cdot {\rm{KnowledgePartyCue}} + X \cdot {\beta _2} + {\varepsilon _2} \cr & {\rm{Arguments}} = {a_2} \cdot {\rm{time}} + X \cdot {\beta _3} + {\varepsilon _3} \cr & {\rm Consistent}=c_{2}\cdot {\rm time}+b_{2}\cdot {\rm Arguments}+X\cdot \beta _{4}+\varepsilon _{4}} $$
where KnowledgePartyCue is the knowledge of the party cue, Arguments is voters’ ability to take side on policy arguments (number of ‘don’t knows’), In-line is voting in line with the party cue, Consistent is voting consistent with one’s positions on arguments, a, b, and c letters denote direct paths, β indicates parameters, X is a vector representing the control variables, and
$\varepsilon$
the error terms.
To account for the non-independence between the mediators and the outcomes, we also introduce the following covariance terms into the model:
$$\matrix{ & {\rm{Cov}}\left( {{\rm{KnowledgePartyCue}},\;{\rm{Arguments}}} \right) \ne 0 \cr & {\rm{Cov}}\left( {{\rm{Inline}},\;{\rm{Consistent}}} \right) \ne 0 \cr} $$
We then estimated the following indirect and total effects corresponding to H2a and H2b:
$$\matrix{ & {\rm{Indirect}}\;{\rm{path}}{:}\;{\rm{time}}\; \to {\rm{KnowledgePartyCue}} \to \;{\rm{Inline}} = {a_1} \cdot {b_1} \cr & {\rm{Indirect}}\;{\rm{path}}{:}\;{\rm{time}}\; \to {\rm{Arguments}} \to \;{\rm{Consistent}} = {a_2} \cdot {b_2} \cr & {\rm{Total}}\;{\rm{effect}}\;{\rm{Inline}} = {c_1} + \left( {{a_1} \cdot {b_1}} \right) \cr & {\rm{Total}}\;{\rm{effect}}\;{\rm{Consistent}} = {c_2} + \left( {{a_2} \cdot {b_2}} \right) \cr} $$
This SEM allows us to test for direct and indirect effects, considering that the learning mechanisms, as well as the likelihood of in-line and consistent voting, may be correlated.
Findings
Aggregate-Level Analysis
Before testing our hypotheses, we present a simple cross-tabulation of the two dependent variables (Table 1). For the Climate law and Proposition 27, in-line and consistent voting go largely hand in hand: nearly half of voters cast votes that were both in line with their party cue and consistent with their opinions on arguments. The corresponding proportion is lower for the OECD Act and Proposition 26, yet it still accounts for approximately one-third of cases. At the other end of the typology, the share of voters who cast a vote that was neither in-line nor consistent is also substantial – around one-fifth, with some variation across votes. Finally, for all four votes, a significant proportion of voters – ranging from 5 per cent on the Climate Law to 45 per cent on Proposition 26 – voted either in-line but not consistently, or consistently but not in-line.
Proportion of in-line and consistent voting in the four ballot measures
A caveat is in order regarding Proposition 26. Because the Democratic Party took no official position on this measure, any vote choice can technically be interpreted as in-line with the party’s stance. This explains the very high share of in-line voting reported in Table 1 (78 per cent in total). However, this should not be understood as a genuine ‘in-line’ vote – at least not in the same sense as for the other three ballot measures. One could even argue that in the absence of a party cue, Democratic voters could not have cast an in-line vote at all.
To account for this ambiguity, Table 1 includes supplementary figures (shown in square brackets) based on an alternative, more stringent definition of in-line voting, under which all Democratic votes on Proposition 26 are considered not in-line. Of course, this stricter classification alters the results considerably: a majority of voters cast ballots that were neither in-line nor consistent, and only a small minority (8 per cent) – only Republican voters – cast votes that were both in-line and consistent.Footnote 14
Figure 2(a) shows the evolution (by day) of the percentage of voters who know the cue of the party they feel close to for each ballot measure. Figures 2(b), 2(c), and 2(d) do the same for the ability to take side on policy arguments, the ability to vote in line with the party cue, and the ability to vote consistently, respectively.
Evolution of voters’ attitudes by day. (A) Proportion of knowledge of party cue. (B) Number of DKs on the argument questions. (C) Proportion of votes in-line with party cue. (D) Proportion of votes consistent with the argument position.
Starting with the knowledge of party cues (Figure 2(a)), the four figures show an increase in the share of voters correctly identifying their preferred party’s vote recommendation. The increase is, however, small for Proposition 26, but still statistically significant (see Appendix B1 for the regression tables). The lack of vote recommendation for Proposition 26 by the Democrats, which has remained unknown to most party supporters, explains this result. Despite the learning process, a majority of responses about the party cues for the two Californian ballot measures were still incorrect the day before the election. In Switzerland, by contrast, the proportion of respondents who knew the vote recommendation was just over 50 per cent at the end of the campaign for the OECD-led amendment and peaked on the Climate law (80 per cent). In the latter case, a majority of correct answers was already present at the start of the study period.
The negative trend in the number of ‘don’t knows’ on the six argument questions also delivers a positive message (Figure 2(b)): as time passes, voters have been increasingly able to take side on arguments. For this variable, the evolution over time is significant for all four ballots. By the end of the campaign, the average number of ‘don’t knows’ was far less than one.
In Switzerland, the increase of in-line voting (Figure 2(c)) is significant for both the Climate law and the OECD-led reform, but with differences in levels between the two ballots: for the Climate law, in-line voting already outweighed non-in-line voting forty days before the election day, whereas for the OECD act this became true during the campaign.
In California, the results for Proposition 27 show a slight increase in in-line voting over time, but the increase fails to reach statistical significance (p = 0.15). The high proportion of in-line voting, despite low levels of party cue knowledge (Figure 2(a)), suggests that the two phenomena do not necessarily go hand in hand. On the one hand, voters who are unaware of their preferred party’s cue may still vote in accordance with it, owing to shared ideological orientations or foundational values. On the other hand, in some cases, the correspondence between a voter’s choice and the party’s recommendation may arise from unrelated factors, such as a status quo bias or mere coincidence.
In the case of Proposition 26, caution is again warranted when interpreting the results due to the Democrats’ (largely unnoticed) free vote recommendation. While the proportion of Democratic supporters voting in accordance with their party’s recommendation is very high – and increases over the course of the campaign – this pattern is largely an artifact of our permissive coding of in-line voting for this particular ballot measure.Footnote 15
Finally, the figures for consistent voting (Figure 2(d)) show a positive trend. For all four ballot measures, the share of respondents who cast a vote consistent with their position on the argument scale significantly increased as the election day drew near. The increase is highest for the two Swiss ballots and lowest for Proposition 26. Further, we again observe differences across ballots with respect to the level of consistent voting. At the end of the campaign, an overwhelming majority of voters voted consistently on the Climate law and on Proposition 27. For the OECD provisions, however, the proportion of consistent and inconsistent votes was about the same, whereas the proportion of inconsistent votes still exceeded the proportion of consistent votes on Proposition 26.
Overall, the aggregate-level analysis supports hypotheses 1a to 1d. All four indicators (party cue knowledge, ability to take side on policy arguments, in-line voting, and consistent voting) increased over the course of the campaign, though notable differences in levels emerged across ballot measures. In Switzerland, knowledge and in-line or consistent voting reached high levels by the end of the campaign for the Climate Act, whereas for the OECD Act this pattern applies mainly to ability to take side on arguments and in-line voting, but less so the party cue knowledge and consistent voting.
For Proposition 27, both the ability to take a position on arguments and the rate of consistent voting increased significantly over the course of the campaign. However, by the end of the campaign, a majority of voters still lacked knowledge of their preferred party’s vote recommendation. The situation is even more complex for Proposition 26. Democratic voters largely remained unaware that their party had taken no official position, and the high share of in-line voting is, in fact, an optical illusion. Since Democratic voters had no party cue to rely on, one would expect them to base their vote on policy considerations. Yet, according to our data, a majority cast a vote that contradicted their argument position, that is, a vote misaligned with their underlying policy preferences.
Individual-Level Analysis
While the aggregate-level analysis conducted so far provides information on the evolution over time of knowledge measures and in-line and consistent voting, it does not say anything about the possible relationships between these variables at the individual level. Therefore, we turn to a structural equation model as described in the Methodological Framework section.
Starting with the direct effect of time – our measure of campaign progress – on in-line or consistent voting, there is no theoretical reason to expect that time per se would influence citizens’ vote choice. Consistent with this expectation, Figure 3(a) shows no direct effect of time, except in one ambiguous case: in-line voting on Proposition 26. Figure 3(b), by contrast, supports hypotheses 2a and 2b that the effects of time on voters’ ability to vote in-line or consistently operates indirectly, through the increase of knowledge. All effects but one are statistically significant, and all are in the expected direction: knowledge increases over time, which in turn increases in-line and consistent voting.
Mediation analyses of in-line voting and consistent voting, direct and indirect effects. (A) Direct effects. (B) Indirect effects.
For in-line voting, the indirect effect is not significant for Proposition 26, which is again unsurprising given the lack of Democratic Party endorsement. For consistent voting, the indirect effect reaches statistical significance for all four ballots, and it has a similar magnitude in each case. Thus, both pathways of opinion formation (cue taking and policy information) are at work in our data. From a comparative perspective, it is noteworthy that the specific type of direct democracy at stake appears to have minimal impact. With the exception of the peculiar case of Proposition 26 (no cue from the Democrats), knowledge acquisition has the expected (virtuous) effects on the likelihood of in-line and consistent voting in both California and Switzerland. In that sense, our analysis suggests that initiative and referendum campaigns have the expected enlightening effect, helping voters to learn and cast votes aligned with their partisan orientation and policy preferences.
This positive conclusion is, however, tempered by the results of the aggregate-level analysis, which revealed some important differences across contexts and ballot measures regarding the degree of knowledge acquisition, and resulting degree of in-line and consistent voting.
Robustness Tests
We conducted three robustness tests. First, the analyses presented thus far include all respondents, regardless of their intention to participate. To test the sensitivity of our findings, we restricted the sample to respondents who had either already voted or indicated they would ‘certainly’ vote. This excludes respondents who reported they would ‘likely vote’, ‘likely not vote’, or ‘certainly not vote’, as well as those who responded ‘don’t know’. The results remain unchanged under this restriction (see Figures C1.2a to C1.2d and C1.3 in Appendix C1).
Second, we take a closer look at the ambiguous case of in-line voting on Proposition 26 to address the absence of a vote recommendation from the Democratic Party, by testing the implications of an alternative coding, in which all Democratic voters are classified as casting a not in-line vote. As shown in Figure C2a.2c, this alternative approach unsurprisingly produces a reversed pattern compared to Figure 2(c), resulting in a very high proportion of not in-line voting – and no trend over time. However, and more importantly, the results of the mediation analysis remain unchanged: the indirect effect of time on in-line voting for Proposition 26 operating through knowledge of party cues remains insignificant.
Additionally, we conducted a further robustness test by excluding Democratic voters entirely, thereby focusing solely on Republican respondents. Figure C2b.2c shows that the share of in-line voters is roughly equal to the share of not in-line voters, with only a modest increase in in-line voting observed over the course of the campaign. Interestingly, despite the small number of Republican voters in California, which results in wide confidence intervals, the indirect effect operating through knowledge of party cues becomes significantly positive. This finding further underscores the importance of knowing (or at least being able to infer) the cue of one’s preferred party in order to cast an in-line vote, assuming such a party cue is available in the first place.
Third, our measure of consistent voting considers respondents with a position between –2 and 2 on the argument scale to have ambiguous preferences and automatically assigns them to the group of inconsistent voters. To assess the extent to which the results are sensitive to these thresholds, we conducted two tests. The first adopts a more permissive coding strategy for consistent voting, classifying as consistent those voters who favor the ballot measure (those who have a value above 0 on the argument scale) and vote Yes, as well as those who oppose it (value below 0) and vote No. Inconsistent voters are those who do not meet this criterion, along with respondents positioned at 0 on the argument scale, regardless of their vote. As expected, this more lenient approach leads to higher levels of consistent voting (see Figure C3a.2d in the Appendix). The second test, by contrast, employs a more restrictive coding strategy, limiting the category of consistent voters to respondents positioned between –6 and –12 on the argument scale who vote No, and those between +6 and +12 who vote Yes – thus, respondents located between –5 and +5 are classified ex ante as inconsistent voters. Again unsurprisingly, this stricter approach yields lower levels of consistent voting (see Figure C3b.2d). Crucially, however, neither specification alters the results of the mediation analysis (see Figures C3a.3 and C3b.3).
Conclusion
This paper examines how voters’ knowledge and attitudes evolve during initiative and referendum campaigns in two different direct democratic contexts. Through a comparative analysis of two Californian initiatives on sports betting and two Swiss referendums on tax and climate policy, we offer insights into the dynamic nature of voter opinion formation in two different democratic settings. Our empirical approach using RCS survey data to examine how knowledge of party cues, voters’ ability to take side on arguments, in-line voting, and consistent voting evolves in the run-up to the ballot represents a novel way of understanding voter decision making in direct democratic votes.
Our results provide both good and bad news with respect to voters’ knowledge acquisition during campaigns and ability to vote in-line or consistently. On the one hand, the findings highlight several positive aspects of the dynamics of voter attitudes during campaigns. First, there is a clear pattern of increasing voter knowledge as campaigns progress. This trend is evidenced by the growing familiarity with party cues and policy arguments related to the ballot measures. Second, mediation models suggest an indirect effect of time on voters’ ability to connect their underlying attitudes to their vote. As voters become more informed about the ballot measures during the campaign, their vote becomes more aligned with their knowledge of these measures. The dynamics of citizens’ attitudes underscore the enlightening role of direct democratic campaigns, through both cue taking and processing of substantial policy information.
Despite the demanding nature of decision making in direct democracy and the complex environment in which it takes place, many voters show an ability to assimilate and use information over the course of a campaign. The result that voters are increasingly able to cast a vote consistent with their argument position, which holds in all four votes, is especially worth considering. Similarly, the expected increase in in-line voting resulting from voters’ increased knowledge of their preferred party’s cue holds for the three ballot measures where parties provided such a cue – the null result for Proposition 26 being due to a lack of party cue among Democrats. The two pathways of opinion formation thus seem to work similarly in both the unmediated (Californian) and mediated (Swiss) variants of direct democracy.
On the other hand, some findings paint a more pessimistic picture of voters’ skills. First, significant differences in knowledge levels were observed across both ballot measures and contexts. Presumably due to the stronger involvement of party elites, both party cue awareness and the degree of in-line voting were higher in Switzerland than in California by the end of the campaign. Moreover, while consistent voting reached similar levels on Proposition 27 and the two Swiss ballot measures, it was notably lower on Proposition 26. For Proposition 26, a majority of voters cast inconsistent votes, that is, their vote contradicted their argument position. This discrepancy may be attributable to the particularly complex electoral environment in California, where initiative elections are held concurrently with national and state candidate elections – contests that typically attract greater public attention – and where political parties play only a limited role in initiative campaigns. In contrast, our findings support the view that the Swiss context provides voters with a steady flow of arguments and voting cues, thereby ‘allowing them, in principle to make enlightened choices – that is, choices which are in line with their preferences’ (Kriesi Reference Kriesi and Kriesi2012b: 239).
Second, the high proportion of voters aligning with party cues on Proposition 27 appears to be largely an optical illusion, as it coincides with low awareness of the cues themselves. Third, even in the more favorable Swiss context, our data reveal differences in the level of knowledge acquisition, as well as in the degree of in-line and consistent voting, between the controversial and intensively discussed Climate Law and the more technical, less mediatized OECD-related act.
Further, while we interpret the growing alignment with party cues as a sign of voter learning and coherent decision making, it could also reflect a disciplining effect where voters follow party lines in a more dogmatic or uncritical manner, potentially reinforcing polarization and inhibiting deliberation. Similarly, reliance on policy arguments is normatively desirable only if those arguments are well-founded and fact-based. In our case, the use of grounded, high-quality arguments provided to respondents supports the interpretation of these effects as indicative of substantive learning rather than misinformed reasoning.
Finally, despite – or perhaps because of – its pioneering character, a limitation of our study is that it covers only four ballot measures. A replication on other Californian and Swiss initiatives and referendums would help to gather finer-grained and more systematic insights into the respective role of ballot- and context-related factors. Further research could also explore additional contextual variables, such as the impact of digital media on voter knowledge. Such studies would offer deeper insights into the dynamics of opinion formation and voter behavior in direct democratic votes.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0007123425101105.
Data availability statement
Replication data for this article can be found in Harvard Dataverse at: https://doi.org/10.7910/DVN/LHPMSU
Acknowledgments
This paper was presented at the University of Geneva ’staff seminar’ in April 2024. Thanks to the participants and to Anke Tresch for discussing the paper. We also presented it at the ‘Direct Democracy Workshop’ in Zürich in June 2024, organized by Iva Srbinovska, Arndt Leininger, and Lucas Leemann. Special thanks to them and to Alice El-Wakil for discussing the paper. Thanks to the other participants for a highly stimulating workshop.
Financial support
Swiss National Science Foundation grant 100017_201119.
Competing interests
None.
Ethical Approval
This study was approved by the CUREG (research ethics committee at the University of Geneva) on 2 June 2021.
AI statement
LLM was used solely to improve the clarity and grammar of the manuscript text; no scientific content was generated or altered.
Open access
