How does information affect vote choice in open-list PR systems? Evidence from a survey experiment mimicking real-world elections in Switzerland ✩

List proportional representation with candidate voting can facilitate policy representation in multiple dimensions. However, candidates with deviating positions may not benefit if cues such as shared socio-demographics drive candidate choice instead. Does this use of cues reflect a lack of policy-related information or a preference for descriptive representation? We study this question in a real-world context, using a survey-embedded experiment that emulates actual vote choice shortly after the 2019 Swiss elections. We vary the level of information on candidates’ policy positions in zero, one or two dimensions (left–right, environment). Our results show that spatial proximity voting increases with better information on the secondary (but not the first) dimension, indicating that information can improve the alignment of (environmental) policy views between voters and candidates. In turn, same-gender and same-age voting slightly decreases when more information is available. The preference for local candidates remains strong. Our results inform debates regarding citizens’ preferences for different types of representation and how electoral systems moderate their expression.

Unlike single-seat-district or closed-list electoral systems, the widely used preferential-list proportional representation (PLPR) systems give voters the opportunity to support individual politicians independently from choosing a party (Gallagher and Mitchell, 2005;Renwick and Pilet, 2016;Rudolph and Däubler, 2016): 1 In principle, this feature of the electoral system facilitates policy representation when there are multiple dimensions, since voters can support an individual candidate sharing their view on issues that cut across the main axis of party competition.Results of survey experiments (Blumenau et al., 2017;Bräuninger et al., 2022) suggest that citizens do take into account the positions of individual candidates and may even switch parties when lists are open rather than closed.In addition, there is real-world empirical evidence that vote-seeking parties diversify their lists by nominating ✩ Authors' note: The order of authors reflects the principle of rotation over multiple joint projects; contributions were equal.We thank Thomas Bräuninger, Oliver Pamp, participants of the European Political Science Association 2021 Annual Conference, and the German Political Science Association Working Group 'Handlungs-und Entscheidungstheorie' Annual 2021 Meeting for helpful comments.We thank Michael Erne and Daniel Schwarz (Smartvote) for providing data on candidates and their issue positions and Janek Bruker and Colin Walder for their excellent research assistance.Replication data and code to reproduce the analyses are available at the Harvard Dataverse via doi:10.7910/DVN/DD3QQN.The study was approved by the Ethics Commission of ETH Zurich (decision EK 2019-N-43).
E-mail addresses: franziska.quoss@gesis.org(F.Quoß), lukas.rudolph@uni-konstanz.de(L.Rudolph), thomas.daubler@ucd.ie(T.Däubler). 1 In this paper, we do not consider whether variation in list flexibility within the PLPR family may affect candidate voting behavior.We speak of open lists in the sense of any non-closed lists but acknowledge that the voter calculus may be different if candidate voting affects intra-party seat allocation less strongly.
2 See also Riera and Cantú (2022), who report that left-right proximity voting is weaker under candidate-centered electoral systems.
We examine these questions in the context of environmental politics.Not least due to climate change, questions concerning the adequate scope of environmental protection measures are gaining in political importance (Kenny, 2021).At the same time, this issue dimension creates challenges for policy representation at the party level since it does not necessarily coincide with the dominant axis of political competition (Green-Pedersen, 2007;Spoon et al., 2014;Hooghe et al., 2002;Carter et al., 2018), also from the perspective of voters (Kenny and Langsaether, 2023).In addition, policy representation on the environmental dimension has two interesting links with descriptive representation.First, the politics of the 'green transition' can create conflict between urban and rural areas, with the latter incurring a larger share of the associated costs (Stokes, 2016;Arndt et al., 2022).Voting for local candidates may also reflect a motivation to elect politicians who represent place-based interests in this conflict.Second, some socio-demographic groups -especially women and younger citizens -tend to have more pro-environmental views (Gifford and Nilsson, 2014) but are disadvantaged in terms of relative numbers in most legislative bodies (Inter-Parliamentary Union, 2021b,a).Together, these patterns make it particularly interesting to examine how rural, female, and younger voters value policy representation on the environmental dimension compared to descriptive representation overall. 4 On the theoretical side, we focus on the role of policy-related information in candidate voting.Assuming that the knowledge about the policy positions of individual candidates is low, we expect such policy-related information to increase the degree of spatial proximity voting.This effect should be larger for the environmental dimension than for the left-right axis due to less ex-ante knowledge and higher levels of intra-party heterogeneity on the secondary dimension.Also, we expect that the information on policy positions reduces the attractiveness of 'matching candidates,' i.e., candidates who share sociodemographic characteristics.The extent to which this happens will reflect how strongly voters value policy representation in comparison with descriptive representation.
Our empirical evidence comes from an experiment embedded in a population-representative survey in which we vary information about policy positions while emulating an actual PLPR choice.To maximize ecological validity, we conducted it shortly after a real-world parliamentary election -in Switzerland, in 2019 -using mock ballots with party labels and actual politicians.In the studied election, environmental issues played a very strong role (Bernhard, 2020).At the same time, record numbers of women (42.0%compared to 32.0% four years earlier) (BfS, 2021) and candidates aged 40 years or younger (25.4% compared to 20.1%) (Bundesversammlung, 2021) were elected to the National Council, the lower chamber of the Swiss parliament.
3 From a supply-side perspective, candidates are more likely to have local ties when personal vote-seeking incentives are stronger (Shugart et al., 2005). 4A link between the under-representation of younger generations and slow responses to climate change has also been postulated in public discourse, for example as part of the ''OK boomer'' meme (Sundström and Stockemer, 2021, 195).''OK Boomer'' was ranked second behind ''Klimajugend'' (climate youth) in the list of the (German-language) 'words of the year' in Switzerland in 2019, the context of our experiment (see https://www.zhaw.ch/de/linguistik/wortdes-jahres-schweiz/#c111775).
Randomly exposing some survey participants to information about candidates' actual positions on the general left-right and/or the environmental policy dimension (information on the positions comes from candidate replies for a voting advice application) allows us to infer to which extent spatial proximity voting increases and preferences for alike candidates (in terms of gender, age, and place of residence) decrease, compared with a control group.
We find that providing information about candidate positions considerably increases spatial proximity voting for the environmental dimension (but only marginally for the left-right dimension).Even in a favorable context like Switzerland in 2019, where the secondary dimension of interest was highly salient, citizens seem not to know much about the policy views of individual candidates.We also observe that preferences for matching candidates decrease with policy information, although the patterns here are more complex.While citizens become less inclined to vote for candidates who share the same gender or a similar age if information about environmental (but not left-right) positions is increased, voters' preferences for local candidates remain strong in the presence of any type of policy information.
Our findings, which are based on an experimental design that refers to a real-world election context, show that the availability of information is crucial for reaching better policy representation in PLPR systems.The salient secondary dimension of environmental politics becomes more influential for vote choice if citizens know where to locate parties and candidates.Our study also ties in with other literature suggesting that more attention should be paid to which aspects of representation citizens themselves are interested in (Cowley, 2013;Wolkenstein and Wratil, 2021).From a practical perspective, while efforts of improving descriptive representation are of course important, the question of how to facilitate policy representation should not be neglected.

Information on policy positions and candidate voting
In a PLPR system, citizens can choose among candidates within parties.The opportunity to choose from several candidates of the same party should make it easy for voters to take into account candidate features, including socio-demographic characteristics but also candidate-specific policy positions (Blumenau et al., 2017;Bräuninger et al., 2022).However, citizens' knowledge about the individual candidates may, in practice, not go beyond the information provided on the ballot papers (Brockington, 2003).Therefore, it is not surprising that we frequently observe a preference for candidates who share the voters' characteristics (for the PLPR context, e.g., Holli and Wass, 2010;Jankowski, 2016;Kukołowicz, 2013;van Erkel, 2019).What we do not know is whether this pattern reflects simply a lack of other, including policy-related information, or a genuine preference for descriptive representation.Put differently, the question is what would happen if citizens knew more about the policy positions of candidates.
Spatial proximity is a core component in explaining voting decisions (Downs, 1957;Adams et al., 2005).At the same time, making a vote choice based on distance in the policy space is cognitively demanding.This is reflected in findings that spatial proximity voting is more common among more sophisticated (educated, knowledgeable, or informed) voters (Tomz and Houweling, 2008;Jessee, 2010;Lau et al., 2014;Singh and Roy, 2014;Tiemann, 2022).Hence, assuming that there are indeed knowledge gaps ex-ante and that citizens care about voting for a candidate whose views are close to their own, spatial proximity should become more important for vote choice when more information about policy positions is available.A priori, some people do not know anything about the positions and, therefore, cannot take them into account.Others may have some knowledge but feel uncertain about the exact locations in the policy space; better information enables this group to consider policy congruence with greater confidence.This expectation can be summarized as: H1: Providing information on candidates' policy positions increases the extent of spatial proximity voting.
The magnitude of such information effects should depend on the degree of prior knowledge (e.g., Fowler and Margolis, 2014;Presberger et al., 2023).They will be smaller if citizens already have a good understanding of actors' locations on a policy dimension.Hence, we would expect that information provision leads to bigger effects on a secondary dimension (like environmental protection) than for the primary dimension of political competition.To begin with, voters should have a good understanding of where parties are located on the leftright axis (e.g., Van der Brug and Van der Eijk, 1999;Vegetti et al., 2017).While knowledge of the positions of individual candidates is likely limited, this is less relevant on a primary dimension since withinparty heterogeneity of positions is also not as pronounced (Tromborg, 2019).In contrast, on a secondary dimension like the one related to environmental protection, citizens may already struggle to locate the parties (Thomassen, 2012;Dahlberg and Harteveld, 2016;Banducci et al., 2017;Dejaeghere and Van Erkel, 2017). 5The positions of parties that attach less importance to the policy area may be less clearly developed and are less often communicated.In addition, there will be more within-party variation in policy views.Weaker sorting of politicians into parties (or pruning by selectors) on secondary dimensions leads to more heterogeneous views within the candidate field (Carroll and Kubo, 2019;Tromborg, 2019) -as we illustrate below in Fig. 1, we also do find this to be the case for candidates in the Swiss national election.This yields the second expectation: H2: The effect posited in H1 -that information on candidates' policy positions increases the extent of spatial proximity voting -is stronger for secondary dimensions (like environmental policy) than for the left-right dimension.
Another question is what will happen to the support of alike candidates once citizens have more information about policy positions.As a starting point, it is useful to reflect upon different possible mechanisms behind such a choice.The motivation may be psychological (and noninstrumental), in the sense that a voter may hold positive feelings toward an in-group candidate (Plutzer and Zipp, 1996, 33).In contrast, a matching candidate may also be seen as more likely to possess valuable character traits, such as honesty, dedication (Campbell et al., 2019), or trustworthiness (Gay, 2002).Such candidate valence factors may also subsume that a politician from the voter's geographic area will be better equipped to represent that area's material interests (e.g., by allocating particularistic goods (Eulau and Karps, 1977;Golden and Min, 2013)).Finally, voters may have a preference for balanced descriptive representation at the aggregate level (Cowley, 2013;Campbell and Heath, 2017;Wäckerle, 2023).
Given this wide range of possible motivations, it is, of course, possible that voters continue to support candidates who are ''like them'', even if additional information on their policy positions becomes available.However, we expect that voting for matching candidates will become less relevant in any of the following three circumstances.First, candidate similarity (in addition to list rank) may be the only cue that citizens can use in practice if they lack any further information about candidates.Once policy-related information is provided, it readily replaces similarity as a decision-making criterion.Second, common characteristics may serve as proxies for shared policy views (Arnesen et al., 2019;Jones, 2016;Däubler et al., 2021) in the first place.In the sense of a similarity heuristic, voters choose a candidate who is like them since they hope that this candidate is more likely to hold policy views close to their own.6With information on policy positions becoming available, the need for a proxy disappears.Third, citizens may value both descriptive representation and policy congruence with the candidate.However, if the set of candidates to choose from (within the preferred party) is small, it may not be possible to find a candidate who delivers both.If voters more often favor spatial proximity to descriptive representation when facing such a trade-off, voting for similar candidates will also become less prevalent if the information on candidates' positions is directly available.Hence, as long as any of these three mechanisms prevail over a preference for descriptive representation as such, we expect: H3: Providing information on candidates' policy positions reduces the attractiveness of candidates who match the voter in socio-demographic characteristics.
Answering the counterfactual question of whether voting behavior would change if citizens were better informed is notoriously difficult (Rapeli, 2018).We experimentally vary the available information about candidate policy positions.This will allow for a direct test of the three hypotheses concerning the choice of candidates in a PLPR context.

Context: Environmental politics and the swiss ''climate'' election of 2019
In elections to its National Council, the lower chamber of the parliament, Switzerland generally uses a PLPR electoral system (Vatter, 2016, 73-75).The number of votes corresponds to district magnitude and, in 2019, varies from one (six cantons and half-cantons with small population sizes) to thirty-five (in Zurich). 7The electoral rules allow the expression of preferences in a very flexible way.Voters can cast a list vote, they can vote for candidates from more than one list (panachage), and individual candidates can be assigned one or two votes.While parties present ranked lists, post-electoral seat allocation within parties is based on preference votes alone. 8The flexibility provided by the electoral rules is also used by many voters in practice.In the official panachage statistics by the National Statistical Office, 44.97% of valid ballots consist of a party list as given, 43.47% of an altered party list, and 8.55% of self-created lists that do not have the name of one party on them (BfS, 2020).
While questions of environmental protection have been part of the Swiss political agenda for a number of decades, they have recently gained in public salience.Following a party split in the mid-2000s, Switzerland now has two major environmental parties, the Green Party of Switzerland (GPS) and the Green-Liberal Party (GLP).In terms of policy positions, a major difference is that the former is economically leftleaning while the latter promotes centrist economic positions (Bakker et al., 2020).In the national election of 20 October 2019, environmental and climate topics played a crucial role (Bernhard, 2020).The Swiss Radio and Television (SRF) afterward referred to it as the ''Climate Election'' (SRF, 2019).The two environmental parties polled 21% of the vote, leading one of the most-read Swiss quality newspapers, Neue Zürcher Zeitung, to declare a ''Green Triumph'' (Hehli and Thelitz, 2019).The continuing salience of the topic was also evident in three controversial referendums (on CO 2 -emissions, on the use of pesticides, and on drinking water quality) in June 2021.
From the perspective of political representation, environmental protection is a particularly interesting dimension because it is not that well aligned with the main (left-right) axis of competition.In Fig. 1, the density contours display the distribution of citizens in a twodimensional space with a left-right and an environmental dimension. 9 The points represent the mean party positions and the crosshairs the middle 80% of candidate positions (for the set of candidates included in the experiment, see below).The graph points out that the positions of citizens are not that strongly correlated (−0.55).While the mean party positions are more strongly associated, we can see that especially the parties in the center and right of the spectrum include candidates with quite dissimilar views on environmental protection.This bears two implications for our analysis.First, it is hard to pin down what these parties stand for in the area of environmental politics.Second, it is worthwhile to know the positions of individual candidates since it will make it easier for some voters to find spatially proximate candidates.

Research design
The research design for our study relies on a population-representative survey experiment.This allows us to maximize internal and external validity (Mutz, 2011): the former by experimentally manipulating whether ballots inform about candidate policy positions on one 9 The subsequent section describes the operationalization in detail.
or two dimensions (or not at all); the latter by working with an addressbased random draw of the population, which is shown electoral ballots mimicking those from the national elections that occurred just a few weeks earlier.
Closely related survey experimental work has often run conjoint experiments asking participants to choose between just two candidate profiles.With full randomization of party labels and policy positions, this can lead to combinations that are very rare in real life, reducing the external validity of the design (De la Cuesta et al., 2022).Some studies have thus chosen not to include party labels (for example Arnesen et al., 2019), but this raises the risk of violating information equivalence (Dafoe et al., 2018).
Our approach instead takes a given party list, including candidates' socio-demographic background, and varies the amount of information respondents have on those candidates' policy positions at the ballot level.Compared to previous work that presented party lists with fictitious candidates (e.g.Bräuninger et al., 2022;Blumenau et al., 2017), we show real candidates from a real election with their real policy positions and therefore have higher confidence that our results allow direct conclusions about real-world voting behavior.

Data
The experiment was embedded in the fourth wave of the Swiss Environmental Panel (Rudolph et al., 2020a), whose fieldwork was conducted between November 18th, 2019 and February 2020 (the vast majority of respondents completed the survey in December 2019). 10he survey was fielded in dual mode; respondents were first asked

Experimental design
In our study, we experimentally vary the information regarding candidates' policy positions.We place a strong emphasis on ecological validity, i.e., we designed ballots in a way that the decision-making task of participants closely resembles a real-world vote choice under PLPR, with regard to three key aspects.This concerns, first of all, timing: We conducted the experiment shortly after the Swiss 2019 elections, which took place on October 20th, i.e., in a still politicized environment, where the experience of the electoral campaigns and the exposure to candidates was recent.Secondly, we showed respondents a ballot with real candidates matched to the electoral district (the canton) in which they reside.We include all information that voters in Switzerland usually see on their ballots (name, i.e. also gender, party affiliation, residence, age, list position).Voting behavior can hence rely on the standard set of cues available to voters in the voting booth (Brockington, 2003;Shugart et al., 2005;Lutz, 2010). 14hirdly, in line with the complex Swiss electoral system that provides for panachage and cumulative voting, we allow respondents to split their votes (three votes to up to three candidates) or give up to two votes per candidate (just as in the real elections).Appendix Section A.4 gives a descriptive overview of the voting behavior of our respondents by party, including panachage and cumulative voting, which shows that respondents make ample use of these options, just as they do in real-world elections.Hence, compared to other survey-experimental studies with fictitious ballots (e.g.Blumenau et al., 2017;Bräuninger et al., 2022), we estimate effects from a survey environment with very high ecological validity, which should closely mirror real-world voting behavior.
In order to keep the level of information presented to respondents and the survey time manageable, we simplified the ballots somewhat: respondents are shown a reduced set of parties -only the six largest parties of their canton -and within those, candidates from the main list.We also limit each list to three candidates and do not introduce apparentments (list connections) between different lists/parties to respondents.Nevertheless, we display 18 candidates, clustered in six parties, to each respondent, and hence present them with a complex choice set.Importantly, the displayed candidates are an idiosyncratic random draw from all main party list candidates.We thus base our results on a broad set of real-world candidates and observe candidate choice over the competitive range of the 2019 Swiss national parliamentary election.The procedure for the random draw of candidates is described in detail in Appendix Section A.5.
Our main experimental treatment concerns variation in the ballot presented to respondents, with four experimental conditions (see Table 1).In the control arm (group 1), respondents see only standard information as printed on the real-world ballot.Under treatment, we reveal the environmental policy position of all candidates to group 2 via a visual scale and printed text.In group 3, we show, in a similar manner, the left-right position of all candidates.In group 4, we simultaneously display both the environmental policy and the left-right position of the candidates.Fig. 2 presents an exemplary ballot from the fourth experimental group.Smartvote, a non-profit organization that runs Switzerland's main vote-advice application (see also Fivaz and Nadig, 2010;Ladner et al., 2012;Pianzola et al., 2019;Benesch et al., 2023), provides information on candidates' policy positions. 1585% of all political candidates for the National Council filled out their survey on policy positions, among which are the vast majority of main party list candidates.

Operationalization of candidate and respondent policy positions
The environmental policy position of candidates is taken from the Smartvote voting advice application and based on policy questions linked to the environmental dimension.We list the underlying items in Appendix Table A.1.For respondents, we directly replicate the Smartvote approach based on the inclusion of a subset of the Smartvote items in our questionnaire. 16 Concerning the left-right position of candidates, we similarly draw on a scale provided by Smartvote.It is mainly but not exclusively based on policy questions about economic distribution, protection of private property, the welfare state, and taxation of income and wealth. 17Participants in the experiment report their left-right position on a standard 11-point scale (wording in Appendix Section A.2).
We recode both the environmental and the left-right dimension to a 0-10 range for respondents and candidates; candidate positions are 15 See https://www.smartvote.ch. 16Smartvote's candidate scores rely on the long version of the policy questionnaire; Smartvote also provides a ''short version'' for rapid assessment (which we use in the population survey).Scores based on the full and the shortened instrument correlate very highly (0.94) for candidates. 17It represents the first dimension of a correspondence analysis of all policy questions included in Smartvote, see https://sv19.cdn.prismic.io/sv19%2Ffcff287e-aa5e-4f5c-ae46-f3d53d6e0b30_methodology_smartmap_de.pdf.also visually and textually presented to respondents in this manner (see Fig. 2).The absolute distance between both positions is used as the core independent variable to explain respondents' vote choice.As shown in Appendix Figures A.4 and A.5, the experimental design ensured that almost all respondents select from a choice set including at least one candidate who is spatially proximate on the respective dimension.Appendix Section A.2 reports the operationalization or question wording of all the explanatory variables we use.

Estimation strategy
For statistical modeling, we start from a random utility maximization framework.The utility of voter  derived from (candidate) option  consists of a valence component   , weighted absolute policy distances on the two dimensions, and an error term:18 The valence part   consists of several sub-components.We are especially interested in the contributions of indicator variables that equal one if voter  matches candidate  in terms of age, gender, and place of residence.In addition, valence will also include party identification (specific to combinations of  and candidates  from the same party), candidate features like list rank, and party dummies (to capture overall differences in party valence).The candidates presented on the ballot are the choice alternatives, and we explain vote choice via attributes of these alternatives and variables capturing the relationship between respondents and alternatives.
In our setting, respondents can either give two votes to one candidate and assign the remaining vote to another candidate, or they can allocate one vote to up to three candidates.In essence, respondents provide us with a candidate ranking, but we observe ties if there are several candidates who received one vote by the same respondent (and also for any candidates who were not chosen, but this is less consequential).
We use a rank-ordered logit model (Allison and Christakis, 1994) to estimate the underlying utility function from the ranked choice data.This follows from assuming that the error terms   in Eq. ( 1) follow a Gumbel distribution, like in the conditional logit model (McFadden et al., 1973;Thurner, 2000), and that the most preferred alternative is ranked first, following by the best among the remaining ones, and so on.We use the exact marginal likelihood approach (Allison and Christakis, 1994, 206-208) for ties, which considers all possible orderings of candidates who received exactly one vote.Standard errors are bootstrapped (500 replications) for most models and clustered by respondents. 19 The attributes we use are the presented ballot information on candidates' party affiliation, gender, age, place of residence, display position (first, second, or third on the survey ballot), and rank on the realworld party list as well as the experimentally revealed candidate and party policy positions.We also include indicators on whether candidates' party affiliation and socio-demographic characteristics match the respondents' own characteristics (e.g., we code a binary gender match if respondent and candidate gender align; a similar approach is taken for municipality (zip codes align), age (within range of up to 5 years; within range of 6-10 years) and party identification).We additionally let party dummies enter the model to capture baseline differences in valence.
To measure proximity in policy space, we use the absolute difference between the respondent's and the candidates' policy positions on each dimension (left-right and environmental).This proximity measure is included for all experimental conditions, including the control group.While the candidate positions are not displayed to the control group, its spatial proximity term reflects the baseline level of policy voting based on ex-ante knowledge.For the control group, we can think of the ''true'' spatial proximity term that we include in the model as a noisy measure 19 As bootstrapped standard errors are computationally very intensive for a large sample like ours, some supplementary analyses use standard errors derived from asymptotic theory (indicated in table notes).For our main models, we ascertained that both approaches lead to similar standard errors.
F. Quoß et al. of citizens' individual (and erroneous) perceptions of policy distance.After the treatment brings perceptions in line with ''true'' distances, vote choices more clearly reflect these ''true'' patterns of proximity (cp.Calvo et al., 2014, 98).In addition, there is a ''psychological'' effect.With the information provided, citizens can more confidently rely on spatial proximity in their choice, and its weight in the utility function should increase.To capture the effect of the experimentally revealed policy positions, we employ split sample regressions that estimate the above-mentioned model for each of the four experimental groups.We report formal tests for the statistical significance of differences between model 1 (control group) and treatment groups (Models 2-4) by interacting (for all respondents) treatment group status (i.e., information provision) with all other variables.

Candidate-level policy distance explains vote choice, and more strongly so with information on policy positions provided
Fig. 3 presents results from the test of H1 and H2, focusing on spatial proximity voting: how candidates' left-right (left panel) and environmental (right panel) policy scores affect vote choice, given these are displayed to voters (models 2, 3, 4) or not (model 1, control group).In the upper two panels, the graph displays the change in the predicted probability of preferring a candidate when the candidate's distance is increased by one point on the 0-10 scale for the left-right dimension (left side) and the environmental dimension (right side). 20The bottom two panels display the difference between the control group and the three treatment group coefficients.
As can be seen from the upper panels, a one-point increase in the distance leads to a decrease in choice probabilities irrespective of dimension and treatment condition, controlling for a wide array of additional factors (see Appendix Table A.3, candidate, and respondent party ID, ballot order, and gender, age, and residency of the candidate in relation to the respondent).In particular, this implies that already in the control condition (left-most coefficient within each panel), when candidates' policy positions are not communicated to voters, spatial proximity matters.Also respondents in the control group have some understanding of their distance to the choice alternatives, and at least implicitly choose on this basis.Predictive spatial proximity effects are substantial for both dimensions but stronger for distances in the left-right (about −0.8 percentage points per unit increase in distance) compared to the environmental dimension (about −0.4 percentage points per unit increase in distance). 21ext, we turn to the question of how the weight of spatial proximity in vote choice changes with information provision: if the policy standpoints of candidates are revealed to respondents, does this cause an increase in the penalty implied by the policy distance term?In the bottom two panels, we see that citizens indeed react to the treatment.When the exact policy position of candidates is displayed, and citizens can gauge the distance of candidates to them personally more precisely, they react more strongly to distance.However, for the leftright dimension, the differences to the control group are small when respective information is provided (see lower left panel, the two rightmost coefficients), and the confidence intervals cover zero.For the environmental dimension, the relative increase in spatial proximity voting compared to the control group is substantial, and the confidence intervals exclude zero.This is the case both when showing respondents only the environmental position (lower right panel, leftmost coefficient; then, a one-point increase in distance leads to about a 0.2 percentage point additional decrease in choice probabilities, i.e., citizens react about 50% more strongly to distance), and when showing it in combination with the left-right position (see lower right panel, right-most coefficient; at similar substantive size).Note also that we observe some crowding out for the left-right dimension: when only displaying information on the environment, citizens react less to distance on the first dimension (lower left panel, left-most coefficient; borderline significant result); for the environment, no corresponding picture emerges. 22aken together, this provides evidence for H1 -that providing information on candidates' policy positions increases the extent of spatial proximity voting -though we have to note that the influence of spatial proximity voting is already strong at baseline and does not increase much for the core dimension of politics in Switzerland, left-right.We see clear evidence for H2: the effect posited in H1 is stronger for the secondary dimension of environmental policy than for the left-right dimension.

Policy position information decreases the use of matching based on socio-demographic cues
We now turn to the question of how the provision of policy position information changes the support for candidates who resemble the voter in terms of socio-demographic variables, i.e., whether we observe a trade-off between substantive and descriptive representation (H3).Similar to the previous graph, Fig. 4 illustrates how the predicted preference for ''matching'' candidates changes when information on policy positions becomes available. 23urning to the upper panel and the control group condition (leftmost coefficients), we observe that descriptive cues are important for vote choice: the probability of being preferred increases by about 1 percentage point for a candidate of the same gender, by a similar amount for a candidate in the same age group (≤ 5 yrs.), and most strongly for a candidate with the same place of residency (3 percentage points).
The informational treatments tend to reduce the weight of a match in the utility function.Looking at the changes in the vote bonus for matching candidates caused by the treatment (bottom row of the graph), we can see that seven out of the nine differences (and six out of six for gender and age) have a negative mean.The strongest effect can be found for the effect of environmental policy information on the preference for candidates of a similar age.Respondents prefer such candidates in the control group, but when only the environmental policy score is presented, the vote bonus becomes very small and statistically insignificant.The confidence interval for this difference (left entry in the middle panel of the bottom row) just about covers zero.It is interesting to note that this effect is less pronounced for the treatment showing positions in both dimensions when candidates of a similar age retain their bonus.
In summary, we find some evidence in support of Hypothesis 3, which states that the provision of policy position information reduces the support of candidates who share the same socio-demographic background.We can interpret this as an indication that voters with better information about candidate positions switch demand from descriptive to policy representation.However, we cannot statistically distinguish the coefficients for age and gender matches between the control and treatment groups.This is potentially due to insufficient statistical power.Note that we present respondents with real-world lists from an election that took place just one month earlier.Hence, respondents may have additional information from the actual campaign at their disposal, and compared to survey experiments with fictitious candidates, power demands for treatment effects are likely higher in our setup.Last, we find that the preference for hometown candidates remains strong regardless of the policy information provided.
Detailed results are presented in Appendix Table A.3, estimated by treatment condition.These also allow us to briefly discuss the evidence for correlates of vote choice (i.e., the standard information displayed on real-world ballots), which, as expected, matters greatly for candidate choice: we find statistically significant associations for all cue indicators.Respondents in the control group are more likely to choose candidates who come from the same municipality, have the same gender, or have a similar age.The predictive effects of sharing the same gender (odds ratio of 1.18; information on odds ratios calculated from the logit coefficients are presented in Appendix Table A.4) or age (1.20) are of a size comparable to (prior knowledge of) a one-point distance on the left-right dimension.The preference for a candidate with the same place of residence is even stronger (odds ratio of 1.69).
As expected, a match in party identification sees a very strong association with choice, increasing the odds of voting for a candidate by a factor of 4.03.List order also has a relevant influence: both a lower place on the survey ballot and the display of a lower actual list position imply a penalty for these candidates, with the odds of choosing lowdisplayed or low-ranked candidates decreasing by up to 50%.Finally, there are unobserved differences with regard to party valence.Relative to the SVP (baseline category), choosing candidates from the SP, FDP, CVP, GP, glp, and CSP is more likely, and for those from BDP and PdA less likely.

Policy information leads to a stronger reduction in matching in specific subgroups
Finally, we explore whether particular subgroups of the population drive the decrease in preference for descriptive representation due to information provision.Appendix Table A.5 shows that the shift away from candidates with shared characteristics does not occur for all voters equally.Rather, when distinguishing subgroups of respondents by gender, age, and rurality, we see that it is female voters who are significantly more likely to rely on gender cues in the control group and consequently reduce their reliance on cues more strongly (although not statistically significant) with the informational treatment, predominantly with the environmental (and the combined) treatment. 24 similar pattern appears for age matches: it is young voters who use age cues significantly more strongly in the control group but less so (to the point of non-significance, although the interaction effects are Fig. 4. Changes in the prediction of the candidate chosen by candidate attribute (candidate gender matching respondent (left panel); candidate age ≤ 5 yrs.matching respondent (mid panel); candidate residency matching respondent (right panel)) and treatment group (x-axis, control group; environmental policy score displayed; left-right policy score displayed; both displayed).95% confidence intervals displayed.Based on coefficients from a rank-ordered logistic choice model of outcome (candidate choice) on candidate attributes within the treatment groups (see Appendix Table A.3). not significant) with the informational treatment.Regarding locality matches, reliance on cues stays strong in the treatment groups, but it also slightly decreases for all three groups (urban, periurban, rural) with the environmental treatment.For the left-right and joint treatment, we find mixed effects, with some groups showing weaker but others stronger preferences for candidates from the same area.None of these differences are statistically significant.Hence, the provision of information seems to reduce preferences for alike candidates, especially among traditionally under-represented groups, namely female and young voters.

Conclusion
We have examined how the survey-experimental provision of realworld information on actual candidates' positions on the environmental and left-right dimensions affects vote choice in a PLPR context.Drawing on a large population-representative sample just weeks after the 2019 Swiss federal elections, our experiment emulated real-world vote choice, contributing causal evidence with high ecological validity to a literature mostly based on survey experiments with fictitious candidates or observational data.Our study produced two main findings.First, in line with our expectations, information on policy positions increases spatial proximity voting on the environmental policy dimension.Information revelation induces little change in spatial proximity voting on the left-right dimension; there, policy distances already matter in the control group (which did not receive any information on policy positions).Second, voters are more likely to support candidates who share their gender, age, and especially their place of residence.This pattern is stronger for women, young people, and citizens from rural areas.The preference for same-gender or same-age candidates becomes somewhat weaker with more information on policy positions, while that for hometown candidates persists.This suggests that female and young voters, two groups traditionally associated with support for descriptive representation (Sevi, 2021;Rudolph et al., 2022), might in part follow descriptive cues as a stand-in for policy proximity (Däubler et al., 2021).
The findings show that even in a context where a second policy dimension is highly salient, like the environmental dimension in the Swiss elections of 2019, additional information can produce a better alignment of policy views between voters and candidates.Even more, this indicates that lack of information is a barrier to the selection of politicians in line with population preferences on the environmental policy dimension.More generally, less-than-perfect information indeed seems to constitute an obstacle to the full realization of choice opportunities under PLPR.If people were better informed about candidate and party positions on the cross-cutting dimension, here, environmental policy, this information would become more influential in their vote choice.This result also complements some earlier findings regarding the causal effects of Voting Advice Applications (VAA).Also, for our case, Switzerland, Pianzola et al. (2019) find that VAA users report a larger consideration set of parties than non-users, and Benesch et al. (2023) infer that the availability of VAAs promotes ticket-splitting.Together with our findings, these results suggest that information on policy positions, as provided by VAAs, can improve policy congruence.Even in a multi-dimensional policy space, citizens can process and make good use of such information.
Our findings provide a counterpoint to several studies based on aggregate-level candidate vote data, which suggest the absence of an electoral bonus (van Erkel, 2023) or even the presence of a malus (Folke and Rickne, 2020;Isotalo et al., 2020;von Schoultz and Papageorgiou, 2021;Isotalo et al., 2022) for candidates with deviating positions.There are several potential explanations for this discrepancy, which include cross-country differences or unobserved confounding in the aggregate studies.Our approach has the advantage that it is based on experimental evidence at the micro-level and that it can take into account voter-candidate distance for each dyad.
We also found that Swiss citizens like to support candidates who share their socio-demographic characteristics.The fact that these preferences are only somewhat (gender, age) or not at all (place of residence) reduced as further policy information becomes available suggests that voters indeed attach utility to similarity -hence, similarity is more than just a shortcut for shared policy views.Comparing the relative weight given to candidate similarity across features is worthwhile since it tells us something about citizens' own preferences regarding representation (Cowley, 2013;Wolkenstein and Wratil, 2021).Samegender and same-age voting is due to young and female voters, two groups that are traditionally underrepresented.Hence, these groups likely make use of the candidate vote to achieve better descriptive representation, be it with or without knowledge on (some) policy positions of candidates -while we also note that young and female voters swap part of their voting for candidates of matching age or gender with voting for candidates of closer spatial proximity.It is striking that the strong preference for local candidates, which is observable in both urban and rural settings (but stronger in the latter), is hardly affected by policy information.We know comparatively little about the reasons behind this preference (Campbell et al., 2019).Its robustness to the provision of additional information makes the call for further research into it even more relevant, also because addressing climate change may create political tensions between urban and rural areas (Stokes, 2016;Arndt et al., 2022).The representation of local interests -beyond overall policy differences regarding the adequate scope of environmental protection -possibly exacerbates such political divides.

Fig.
Fig. Exemplary presentation of a ballot with revealed left-right and environmental policy position of candidates (experimental group 4) in the paper questionnaire subsample (header translated).Respondents indicate their vote by placing 1-3 crosses in the tick boxes on the right.

Fig. 3 .
Fig. 3. Changes in the prediction of the candidate chosen by candidate attribute level (policy distance candidate-voter on left-right dimension [left panel] and environmental dimension [right panel]) and treatment group (x-axis, control group; environmental policy score displayed; left-right policy score displayed; both displayed).95% confidence intervals displayed.Based on coefficients from a rank-ordered logistic choice model of outcome (candidate choice) on candidate attributes within the treatment groups (see Appendix Table A.3).

Table 1
Rudolph et al. (2020b)to candidates in the survey experiment.Party, name, age, residence, list position + left-right + environmental position to take part in the survey online and later sent a printed questionnaire together with a second invitation letter. 11ix Section A.1 describes the ethics of our study on human subjects.The sample of invitees is a random draw of the Swiss resident population, and hence population-representative. As noted inRudolph et al. (2020b)overall survey take-up 12 is not strongly related to socio-demographics.As we subset the sample to those eligible to vote, living in cantons that use OLPR, we also show in Appendix Section A.3 that the respondents of the experiment do not differ from this subset of the population.We are therefore confident that we report results that speak for the general Swiss population.Similarly, vote choice by survey respondents is by and large comparable to aggregate electoral results (see Appendix TableA.1).Note that while the sampling included non-citizen Swiss residents, we base our results on Swiss citizens who are eligible to vote.As this article is concerned with vote choice in open lists, we did not field the experiment to voters in cantons that use single-member districts or lists with just two candidates.13Overall,we work with 6268 respondents and hence have a well-powered study.