Existing Theoretical and Empirical Analyses

Facing mounting disaffection among voters, party elites have increasingly turned to intra party mechanisms of democracy, especially after electoral letdowns (Hazan and Rahat Reference Hazan and Rahat2010; Katz Reference Katz, William and Richard2013; Sandri and Seddone Reference Sandri and Seddone2021). In particular, extensive scholarly work illustrates the increasing use of inclusive candidate selection methods in Western Europe for both legislative office and party leadership (Cross and Blais Reference Cross and Blais2012; Pilet and Cross Reference Pilet and Cross2014; Sandri, Seddone and Venturino Reference Sandri, Seddone and Venturino2015). Similar trends have been documented in Latin America (Carey and Polga-Hecimovich Reference Carey and Polga-Hecimovich2006; Kemahlioglu, Weitz-Shapiro and Hirano Reference Kemahlioglu, Weitz-Shapiro and Hirano2009; Siavelis Reference Siavelis2012) and Africa (Ichino and Nathan Reference Ichino and Nathan2022; Casey, Kamara and Meriggi Reference Casey, Kamara and Meriggi2021).

More inclusive candidate selection procedures can appeal to voters for several reasons. First, broader findings show that most people support the general notion of democracy (Inglehart Reference Inglehart2001; Voeten Reference Voeten2016; Adserà, Arenas and Boix Reference Adserà, Arenas and Boix2023). intra party democratization, in particular, can improve democratic values such as accountability, procedural legitimacy, and responsiveness (Carey and Polga-Hecimovich Reference Carey and Polga-Hecimovich2006; Cozza and Somer-Topcu Reference Cozza and Somer-Topcu2021; Shomer, Put and Gedalya-Lavy Reference Shomer, Put and Gedalya-Lavy2016; Wauters and Kern Reference Wauters and Kern2021; Smith and Tsutsumi Reference Smith and Tsutsumi2016). Second, studies show that primaries raise the likelihood that parties nominate more electable candidates (Adams and Merrill III Reference Adams and Samuel Merrill2008; Serra Reference Serra2011; Snyder and Ting Reference Snyder and Ting2011; Ascencio Reference Ascencio2023). Indeed, while party elites prioritize loyalty and core constituencies, more inclusive selection increases candidates’ perceived legitimacy and representation of broader preferences (Casey, Kamara and Meriggi Reference Casey, Kamara and Meriggi2021; Schwenk Reference Schwenk2023). Finally, bottom-up selection procedures typically gain media visibility that draws public attention to their democratic nature and candidate positions (Pedersen and Schumacher Reference Pedersen, Schumacher, William and Pilet2015; Mahéo Reference Mahéo2017; Put and Shomer Reference Put and Shomer2023b).

Other works, however, argue that more inclusive candidate selection can have potential downsides. First, a growing debate links primaries to greater polarization. Primary candidates often adopt more ideologically extreme positions than the general electorate when they compete for the support of the party’s core voters and activists (Lehrer Reference Lehrer2012; Rosenbluth and Shapiro Reference Rosenbluth and Shapiro2018; Broockman et al. Reference Broockman, Carnes, Crowder-Meyer and Skovron2021; Amitai Reference Amitai2024). Second, general voters might perceive primaries as corrupt ‘window dressing’ mechanisms, with the added peril of potential vote buying (Charron and Schwenk Reference Charron and Schwenk2023). Finally, there is mixed evidence about the linkage between primaries and the selection of historically disadvantaged groups (Hinojosa Reference Hinojosa2012; Astudillo and Lago Reference Astudillo and Lago2021; Ichino and Nathan Reference Ichino and Nathan2022).

Given the existing competing explanations, are inclusive candidate selection methods electorally effective in practice? The evidence is limited and mixed. On the one hand, several studies indicate that such rewards do exist in certain contexts. In Latin America, for example, parties that select presidential candidates by open primaries gain an estimated 4–6 per cent electoral bonus (Carey and Polga-Hecimovich Reference Carey and Polga-Hecimovich2006). In Canada, Germany, and Spain, parties with closed nomination procedures for legislative candidates face an electoral penalty if rival parties use more inclusive methods (Astudillo and Lago Reference Astudillo and Lago2021). Timing also matters: multiple findings show a short-term boost in polling scores for parties embracing open selection procedures for either leadership or parliamentary candidates, deeming them particularly beneficial closer to the general elections (Cozza and Somer-Topcu Reference Cozza and Somer-Topcu2021; Pedersen and Schumacher Reference Pedersen, Schumacher, William and Pilet2015; Ramiro Reference Ramiro2016).

On the other hand, several works dispute the empirical link between inclusive candidate selection and electoral outcomes. In France and Italy, parties that select parliamentary and presidential candidates by primaries exhibit inconsistent electoral patterns (De Luca and Venturino Reference De Luca and Venturino2017). Similarly, experimental data from Belgium suggest that open leadership selection does not yield higher voter support (Wauters and Kern Reference Wauters and Kern2021). Democratic candidate selection can also have negative electoral effects. For instance, primaries might hurt parties’ general performance when they put the limelight on internal disputes and legislative dissent that are often weaponized during internal campaigns (Hazan and Rahat Reference Hazan and Rahat2010; Shomer Reference Shomer2017).

Legislative Candidate and Leadership Selection

The literature discussed above gives equal attention to two arenas: party leadership contests and the choice of legislative candidates. While they select candidates for different offices, the two processes share motivational and behavioral assumptions about how and why party elites decide to democratize primaries and the benefits they hope to reap. As Kenig, Rahat and Hazan (Reference Kenig, Rahat and Hazan2016: 39) argue, ‘both leadership […] and candidate selection share many commonalities in terms of the relevant dimensions for their delineation, and both are going through a clear process of democratization exhibiting somewhat similar consequences.’ Indeed, for our purposes, the signaling of responsiveness, openness to the electorate, and procedural legitimacy remains fundamentally similar in both candidate and leadership contests.

Given this similarity, we henceforth focus on legislative candidate selection rather than party leader nomination. First, legislative candidates are typically selected relatively close to election day, increasing the procedure’s potential effect on voter behavior. Party leaders, by contrast, are often chosen long before the next elections while their nomination method has only short-lived influences (Cozza and Somer-Topcu Reference Cozza and Somer-Topcu2021; Pedersen and Schumacher Reference Pedersen, Schumacher, William and Pilet2015). Second, while party leaders may enjoy greater national visibility and personalization (Wauters et al. Reference Wauters, Thijssen, Van Aelst and Pilet2016), legislative candidates represent voters more directly, are personally accountable to specific communities and districts, and maintain more immediate contact with their constituents. Finally, from a practical standpoint, the limited number of leadership nomination reforms constrains robust empirical conclusions about their impact. Nevertheless, as we discuss in the article’s conclusion, we believe that our findings could generalize to leadership selection as well.

In sum, whether democratized candidate selection sways voters in the polls remains an open question. Although existing studies examine different levels of candidate selection – party leaders, legislators, and local candidates – two issues stand out in particular. First, empirically, the current literature offers mixed evidence based on limited case selection, contexts, and data. Second, theoretically, it focuses primarily on the presumed appeal (or lack thereof) of inclusive methods to voters. This perspective neglects the subsequent link from isolated preferences to actual voting decisions. Next, we discuss the latter issue by suggesting a broader theoretical perspective. Subsequently, addressing the empirical puzzle, we test our argument using comparative observational and experimental data covering multiple countries and contexts.

Data and Variables

Our observational analysis leverages party reforms in candidate selection procedures to causally identify whether shifts toward more inclusive methods yield electoral returns. To do so, we compiled data on party organizations and election results in thirty-eight democracies between 1970 and 2019. To ensure a comparable democratic baseline, the countries in our sample were selected by two criteria. First, we only include Organisation for Economic Co-operation and Development member states. Second, we require a score of 0.75 or higher in the Varieties of Democracy (V-Dem) Additive Polyarchy Index (0–1) in a given election year (Coppedge et al. Reference Coppedge, John Gerring, Lindberg, Jan Teorell, Fabio Angiolillo, Borella, Cornell, Steven Fish, Linnea Fox, Haakon Gjerløw, Ana Good God, Allen Hicken, Krusell, Marquardt, Kelly McMann, Juraj Medzihorsky, Anja Neundorf, Daniel Pemstein, Rydén, von Römer, Brigitte Seim, Svend-Erik Skaaning, Sundström, Tzelgov, Wang, Wig, Wilson and Ziblatt2024). We include all parties that gained at least 5 per cent of electoral votes in each relevant election. Our overall sample includes 576 parties with 2,421 party-election observations. Supplementary Appendix (SA) A.1 lists the sample’s countries and election years.

Our main outcome of interest is each party’s electoral performance. Our dependent variable is the vote share that party $i$ gets in election $j$ , drawing on ParlGov data Döring and Düpont (Reference Döring and Düpont2020).Footnote ⁴ Our primary explanatory variable is the candidate selection inclusiveness level of party $i$ in election $j$ . We measure inclusiveness using a shortened version of Hazan and Rahat’s (Reference Hazan and Rahat2010) five-point scale. Specifically, we collapsed the latter into an ordinal three-point index: (1) candidates are selected top-down by the party leader or an executive committee of party leadership; (2) candidates are selected by delegates of local/regional party organizations; or (3) candidates are selected by party members or all registered voters via primaries/caucuses.Footnote ⁵ This classification is measured using candidate nomination data from the Varieties of Party Identity and Organization (V-Party) V2 dataset (Lindberg et al. Reference Lindberg, Nils Düpont, Berker Kavasoglu, Marquardt, Michael Bernhard, Allen Hicken, Juraj Medzihorsky, Ora John Reuter, Weghorst, Nina Wiesehomeier, Nazifa Alizada, Lisa Gastaldi, Garry Hindle, Johannes von Römer, Pemstein and Seim2022).

Method

Isolating the impact of candidate selection methods on electoral results poses two inferential challenges. The first challenge is that most parties maintain stable candidate selection procedures over time, making their influence difficult to disentangle from other fixed party attributes. To deal with this issue, we leverage party reforms that switch between different candidate nomination categories. This empirical strategy provides two advantages. First, it examines variations in candidate selection methods while holding other party and country attributes constant. Second, candidate selection methods are likely more visible and politically salient right after such reforms, heightening their potential electoral impact. Since greater inclusiveness is ordinal and relative, we use two alternative measures of change in our scale: moderate inclusiveness from category 1 to 2/3 (leadership $ \to $ delegates, members, or voters) and strict inclusiveness from categories 1/2 to 3 (leadership or delegates $ \to $ members or voters).

The second inferential challenge is that party elites choose to reform candidate selection neither randomly nor independently of other factors affecting party vote share, including the party’s prior electoral performance, expected electoral prospects, and institutional incentives (Hazan and Rahat Reference Hazan and Rahat2010; Katz Reference Katz, William and Richard2013; Shomer Reference Shomer2017; Sandri and Seddone Reference Sandri and Seddone2021). Thus, a simple cross-sectional estimation could be confounded by other factors that explain both the reform and the party’s electoral outcomes.

To overcome this problem, we employ a difference-in-differences estimation of matched time-series cross-sectional data (Imai, Kim and Wang Reference Imai, Song Kim and Wang2023). This method compares the electoral performance of ‘treated’ parties, which reformed their candidate selection, against a matched set of ‘control’ parties that share parallel electoral trends and similar observable attributes but did not change their selection procedure.

The method involves several steps, summarized in Figure 1. First, we identify parties that democratized their candidate selection based on our scale. Second, we match them with parties that did not make such a change but share a parallel prior trend in both the treatment (pre-change candidate selection method) and the dependent variable (pre-change electoral performance). Third, we refine the matched sets’ balance on the lagged treatment status, lagged outcome, and additional party covariates that could otherwise affect electoral performance. Finally, with sufficient balance on these factors, we use a difference-in-differences estimator to infer the Average Treatment effect on the Treated (ATT) of such reforms on vote shares in the subsequent election. We calculate the standard errors using block bootstrapping (1,000 iterations).

Figure 1. Steps for difference-in-differences estimation using matching with time-series cross-sectional data, based on Imai, Kim and Wang (Reference Imai, Song Kim and Wang2023).

The design addresses several additional concerns. First, while most matching techniques assume strictly cross-sectional data, ours are both cross-sectional and time-series with multiple elections per party. Hence, matching on both dimensions fits this data structure better. Second, candidate selection reforms occur sporadically and occasionally more than once, ruling out standard difference-in-differences specifications that assume a single concurrent change in all treated units. By contrast, our approach can accommodate staggered or repeated treatments as part of the matching parameters. Third, since different countries run elections in different years, our observations do not share identical timelines. Moreover, not all parties compete in all election years due to splits, mergers, and alliances. Our selected method bypasses this problem by matching parties on their relative trajectories over time – that is, their ordinal continuum of elections in our data (election 1, 2, 3, etc.) regardless of exact years – before a reform took place.Footnote ⁶ Finally, matching on electoral outcomes in the cycles preceding the treatment reinforces the parallel trend assumption.

Since election cycles are several years apart, we match parties by their selection method and electoral performance in the two elections prior to the reform ( $L = 2$ ). We further refine the matching sets by several relevant covariates that could generate differences in party vote shares. First, we use V-Party data to balance treated and control parties by their positions on three key issues – economic left-right views, cultural liberalism (gender equality), and immigration – shaping electoral competition across the West in recent decades (Kriesi et al. Reference Kriesi, Edgar Grande, Dolezal, Bornschier and Frey2008; Hooghe and Marks Reference Hooghe and Marks2018). Second, we balance parties by their incumbency status prior to the election, which often adds an electoral advantage (Ansolabehere and Snyder, Jr. Reference Ansolabehere and Snyder2002; Cox and Katz Reference Cox and Katz1996), using a dummy variable indicating government membership after the prior election based on V-Party data. Third, we balance parties on whether they compete as part of an electoral alliance or not, which can artificially inflate electoral gains by pooling voter bases from multiple parties. Fourth, we balance on country-level electoral institutions, which influence strategic incentives for both elite and voter behavior (Cox Reference Cox1997; Kedar Reference Kedar2009), using V-Dem’s classification of lower-chamber electoral systems as majoritarian, proportional, mixed, or other. Fifth, to account for the electoral influence of national economic conditions (Duch and Stevenson Reference Duch and Stevenson2008), we balance parties on country-level GDP per capita in the treated election year. Finally, since candidate selection reforms may be a reaction to other parties’ selection methods (Astudillo and Lago Reference Astudillo and Lago2021), we include a dummy variable indicating whether any other party in the country reformed its selection method in the prior election cycle.Footnote ⁷ SA A.2 summarizes the descriptive statistics of all dependent, treatment, and matching variables. We refine the matching using propensity score weighting.Footnote ⁸

The above procedure yields several matched treatment–control sets. The group of treated parties that implemented moderate inclusiveness reforms – that is, their switched from leadership-based selection to delegates, members, or voters – includes 20 cases with matched sets ranging from 0 to 84 control parties per treatment (the median control set has 59 parties). For strict inclusiveness reforms, which shifted from leadership/delegates to selection by members/voters, there are 7 treated parties. This lower number trades off positively with larger matched control sets, which range from 29 to 169 per treatment (the median control set has 57 parties). Overall, the moderate measure’s sample includes $n = 1,037$$ parties and the stricter measure’s sample includes $n = 632$ parties. SA A.4 elaborates on the treated parties and matched control sets.

Findings

Our analysis calculated the difference-in-differences estimations of the ATT of increasing candidate selection inclusiveness on vote share (Imai, Kim and Wang Reference Imai, Song Kim and Wang2023). We examined the subsequent election after the reform, assuming that their direct influence on voters should be immediate and considering the long periods between election cycles. Figure 2 plots the point estimates with 95% and 90% confidence intervals (thin and thick horizontal lines, respectively) by our two treatment measures.

Figure 2. The estimated average treatment effects of candidate selection inclusiveness on treated parties’ vote share. The thin and thick horizontal lines indicate 95 per cent and 90 per cent confidence intervals, respectively.

The findings corroborated Hypothesis 1. Candidate selection democratization did not have a statistically discernible electoral benefit. Parties that underwent moderate democratization of candidate selection had a near-zero ATT of −0.52 points in vote share compared to matched control parties. The estimated ATT of stricter reforms was slightly higher, reaching 2.83 per centage points more than control parties, but had noisy error margins that failed to reach conventional levels of statistical significance at the 95% and 90% levels. Moreover, substantively, a 2-point gain is relatively small except in very close races and contexts (for example, winner-takes-all rules or parties close to the electoral threshold). Thus, although stricter reforms did showcase a weak potential signal, we did not find sufficient evidence that either a moderate or a strict shift to more inclusive candidate selection had a clear electoral benefit.

These null findings are robust to several alternative specifications, detailed in SA A.5. First, despite the lengthy period between election cycles, we extended our estimation to include the second election after the reform. A longer and/or lagged effect would be consistent with alternative explanations underscoring indirect and gradual mechanisms, such as better candidate recruitment and internal recalibration of party operations. The results, however, remained null, suggesting no delayed or longer-term effects on vote shares. Second, the null findings held when we matched on the parties’ vote share in three prior cycles instead of two. Third, to verify that the importance of candidate selection methods did not change over time, we re-estimated our analysis while also balancing parties by the particular decade in which their elections took place. The results remained substantively unchanged. Finally, given the noisy but positive signal for strict reforms, we re-estimated the confidence intervals with 10,000 bootstrap iterations instead of 1,000 but found similar outcomes.

Data and Case Selection

What explains this insignificant effect? Our hypotheses suggest two main reasons. First, voters may have weak isolated preferences for greater candidate inclusiveness. Second, voters may give a low relative priority to candidate selection inclusiveness in their voting calculi. We examined both mechanisms using a preregistered conjoint experimentFootnote ⁹ conducted in three developed democracies with different electoral systems: France, Israel, and Sweden. All three countries held elections in 2022, offering a similar temporal distance from the most recent electoral cycle.

Existing research suggests that institutional and cultural differences across countries – such as electoral rules, electoral traditions, democratic and social norms, and so on – may influence citizen preferences for candidate selection methods (Cross Reference Cross2008). Thus, fielding the experiment in three different institutional and cultural contexts adds external validity and assuages concerns about idiosyncratic political vagaries due to these differences.

Institutionally, France represents a majoritarian electoral system. Although France’s two-round majority-plurality rules are not a pure first-past-the-post system, its multiparty system offers greater variation in candidate selection procedures and lower public association of this issue with particular parties. In the United Kingdom, for instance, candidate selection reform has become publicly identified with the Conservative Party (Alexandre-Collier Reference Alexandre-Collier2016) and could thus introduce partisan bias. In Canada, another potential case study, all dominant parties (Conservative, New Democrats, and Liberals) use party members as the selectorate with little variation. Israel, by contrast, is a closed-list proportional system with many small to medium parties, where voters choose a single-party ballot with a preset candidate list determined by the parties. Sweden has a preferential-list proportional system, which has become more popular over the past decades (Renwick and Pilet Reference Renwick and Pilet2016). The country’s flexible-list procedure allows voters to choose either a generic party ballot with preset candidates or a specific candidate from one of the lists.The three countries also differ in the type and internal variation of candidate selection methods used by their parties, as detailed in SA B.2. Finally, the cases also differ in various cultural aspects, including gender norms, ethnic heterogeneity and multiculturalism, religion and secularism, political trust, and others.

The survey samples include approximately 1,300 respondents per country (1,301 in France, 1,312 in Israel, and 1,337 in Sweden).Footnote ¹⁰ Surveys in France and Sweden were both fielded by Dynata on 9–15 October 2023, 2023. Due to the 7 October terror attack and subsequent War in Gaza, the Israeli survey was fielded belatedly on 29–30 January 2024, using the Midgam Project Panel. Both Dynata and Midgam sampled respondents from regularly maintained online panels using nationally representative quotas by age, gender, and region. Due to practical sampling limitations of Arab-Israeli citizens, the Israeli sample includes only Jewish respondents. For quality assurance, we screened respondents with two attention checks at the beginning and middle of the surveys. SA B.1 presents the samples’ demographic distributions and compares them with the general population in each country.

Experimental Design

Our experimental study uses a conjoint design, a popular instrument for studying multidimensional preferences (Hainmueller, Hopkins and Yamamoto Reference Hainmueller, Hopkins and Yamamoto2014; Bansak et al. Reference Bansak, Hainmueller, Hopkins, Yamamoto, James and Donald2021). Conjoint experiments ask respondents to select their preferred option from a pair of hypothetical choices with randomly assigned attributes. In so doing, they estimate whether the inclusion of particular attributes increases the odds of choosing a given alternative over others. This design is particularly fitting to simulate voting decisions, which consider multiple party characteristics as part of a single choice at the ballot box.

The experiment asked respondents to choose between two profiles of fictional parties, Party A and Party B, assuming an election was held today. We did not use real-world party labels to ensure cross-country comparability and to prevent bias due to partisan attachments and idiosyncratic influences (Brutger et al. Reference Brutger, Kertzer, Renshon, Tingley and Weiss2023).Footnote ¹¹ The profiles randomly varied by five party attributes, summarized in Table 2. Whereas parties can differ by many more dimensions, we limited the experiment to five attributes to ease respondents’ cognitive burden. We chose a mix of attributes that include ideology, party institutions, and valence, which can apply across cases.

Table 2. Conjoint experiment: randomized attributes and levels

^aIn the analysis, ideological leaning is recoded to capture their distance from respondents by subtracting the latter’s self-identification on the same scale.

First, as our main attribute of interest, we presented respondents with each fictionalized party’s candidate selection procedure (leadership selection, party delegates, party members, or all registered voters). This attribute was listed third rather than first to avoid revealing our goals.Footnote ¹² Second, to capture the influence of ideology and partisanship, the parties randomly differed by their ideological leaning (left, center-left, center, center-right, right). After the survey, we recoded this variable to reflect each fictional party’s relative ideological distance from respondents’ own self-identified position on the same 5-point scale, gauged earlier in the survey.Footnote ¹³

Finally, our experiment included three attributes that capture additional valence considerations: the party leader’s professional background and experience (politics, business, or the media), the party’s political status (incumbent, opposition, or new party), and the party’s female representation (above or below the national average). These attributes convey key aspects of party valence – descriptive representation, experience in government, seasoned leadership, and potential outsider appeal – using real-world cues rather than crude statements about their corruption or competence levels.Footnote ¹⁴ SA B.4 presents the full questionnaire.

Each respondent completed three consecutive selection tasks.Footnote ¹⁵ The dataset’s observations consist of all randomized party profiles presented to every respondent. The outcome variable indicates whether a profile was chosen or rejected and the explanatory variables are its randomly assigned attribute levels. Given the number of profile pairs shown per respondent, the samples include 6,924 observations in the French survey, 7,842 in the Israeli survey, and 6,996 in the Swedish survey.

Voter Preferences

We begin by examining voter preferences for inclusive candidate selection in isolation: ceteris paribus, do voters favor parties with more inclusive candidate selection methods over parties with more closed procedures? To answer this question, we regressed the outcome (whether a hypothetical party is chosen over its paired alternative) on a series of dummies indicating all attribute levels listed above. We then calculated each level’s Average Marginal Component Effect (AMCE) compared to their attribute’s baseline condition (Hainmueller, Hopkins and Yamamoto Reference Hainmueller, Hopkins and Yamamoto2014; Bansak et al. Reference Bansak, Hainmueller, Hopkins, Yamamoto, James and Donald2021). Our main focus was on whether more inclusive selection methods increase the likelihood of party choice compared to leadership-based selection, the baseline level, while controlling for other attributes. Recall that, unlike the observational analysis above, in the experimental setting we break down primaries into more categories to better simulate hypothetical choices.Footnote ¹⁶

Figure 3 plots the model’s estimates. Our findings supported Hypothesis 2.a and rejected Hypothesis 2.b. All else equal, more inclusive candidate selection procedures increased the odds of voting for a party compared to leadership-based selection methods. This pattern was particularly pronounced for parties with open primaries for all registered voters, whose choice probability was higher by 5.4 to 15.5 per centage points (depending on the country) than parties with leadership-based methods. Primaries for party members also raised the odds by 3.2 to 7.1 per centage points. Despite some differences in effect magnitude, the clear replication across all three countries attests to this finding’s generalizability and robustness.

Figure 3. AMCEs of different party attribute levels. The dots and horizontal lines indicate point estimates with 95 per cent confidence intervals. Standard errors clustered at the respondent level.

Our analysis also revealed preferences on the other attributes. As expected, ideological distance had a significant and strong effect. The gap in selection likelihood between parties with a perfect ideological fit to respondents and parties with the highest distance was 33.3 to 52.7 per centage points. Higher female representation also mattered, raising a party’s selection odds by 5.7–10.7 per centage points compared to below-average representation. Leader experience showed a less consistent pattern. While Swedish and Israeli respondents prefered parties whose leaders have meaningful political or business experience over a background in the media, French respondents favored business experience over both. Finally, party status showed the weakest direct influence. Respondents in France and Israel did not favor a particular incumbency status, whereas Swedish respondents were more likely to vote for an incumbent party over new parties by 7.1 per centage points.

As a robustness test, our surveys also included a second conjoint experiment. To further control for the strong influence of ideological considerations, the second experiment asked respondents to choose between pairs of institutional reform packages for a party with which they already agree ideologically. The reform’s randomized attributes included more/less inclusive candidate selection (our main attribute of interest), more national/local focus, greater/lower parliamentary discipline, donor/member-based funds, and more/fewer female candidate quotas. The results corroborated our main findings. In all three countries, respondents prefered institutional reforms that increase candidate selection inclusiveness. SA B.8 elaborates on this test.

Relative Priority

How does the positive preference for more inclusive candidate selection align with the insignificant electoral outcomes we found earlier? Hypothesis 3 suggests that voters put low relative priority on democratized candidate selection compared to other considerations. To test this possibility, we wanted to see how respondents prioritized different attributes when forced to choose between them rather than hold them constant.

To do so, we leveraged the conjoint experiment’s paired forced-choice design and created four subsetted datasets. Each subsample retains only observations with less appealing levels on one other attribute, in turn, than their paired alternative. Thus, each dataset represents profiles that necessarily trade off another dimension if chosen: farther ideological distance (first subsample), lower leader experience in politics (second subsample), no incumbency status (third subsample), or less female representation (fourth subsample).Footnote ¹⁷ The subsetted datasets ranged from 1,764 to 2,747 observations. SA B.6 demonstrates the subsetting procedure using a hypothetical example.

In each subsample, we regress the outcome – whether the suggested party, which necessarily sacrifices another attribute, was preferred – on all other attribute levels and estimate their marginal means (MMs). Contrary to AMCEs, which measure relative differences compared to an attribute’s baseline level, MMs in paired-choice conjoints can be interpreted as predicted probabilities (Leeper, Hobolt and Tilley Reference Leeper, Hobolt and Tilley2020). An attribute level with an MM value above 0.5 indicates a favorable probability (over 50%) of selecting a party featuring it over an alternative. Conversely, an attribute level with MM values below 0.5 signifies a greater likelihood of losing to competing parties, whereas 0.5 represents a null influence (that is, akin to a coin toss). In this case, since each subset forces a loss in one other attribute, candidate selection methods with MMs above 0.5 indicate greater priority than the former. By contrast, MMs below 0.5 suggest unwillingness to sacrifice the other attribute even at the cost of more exclusive selection methods.

Figure 4 showcases the results in each of the four subsetted datasets. For brevity, we only display the coefficients for candidate selection levels.Footnote ¹⁸ Each panel exhibits the estimated MMs of different candidate selection methods given a forced trade-off with another dimension. As a baseline reference, the top-left panel shows the MMs of candidate selection methods in the full and unconstrained sample.

Figure 4. MMs of different candidate selection methods under forced trade-offs with different attributes. The dots and horizontal lines indicate point estimates with 95 per cent confidence intervals. Standard errors clustered at the respondent level.

The findings reveal the dominance of ideological distance. As the top-center panel shows, parties that face a competitor with closer ideological proximity to respondents always lose – that is, have an estimated MM below 0.5 – regardless of their candidate selection method. Although the most inclusive selection procedure has slightly higher MMs than more exclusive ones, its values remain below 0.5 and reflect a 30 to 36.6 per cent probability of being selected. Meanwhile, the most closed method, selection by leadership, exhibits the lowest MM estimates (14.5 to 29.5 per cent).

Female representation poses another notable counter-force, albeit not as strong as ideology. Any selection method except by all voters is more likely to lose if it trades off with lower female representation. Facing this sacrifice, the selection probabilities of these procedures ranged from 38.6 to 46 per cent depending on the exact method and country. At the same time, opening candidate selection to all voters, the most inclusive procedure, balances the loss of female representation. In such as case, neither alternative – a party with exclusive selection and high female representation versus a party with open primaries for all voters but lower female representation – consistently wins over the other (MM values statistically indistinguishable from 50 per cent at a 95 per cent significance level).

Leadership experience and incumbency status showed the weakest priority compared to candidate selection. Although the MM values of selection methods decreased slightly under a leader with lower political experience, the differences from the full baseline model (top-left corner) were relatively small. Nevertheless, they did neutralize the positive edge of open primaries for all voters. The loss of incumbency status, meanwhile, did not have a noteworthy effect on MM values compared to the baseline model, indicating that it was dominated by candidate selection methods within respondents’ priorities.

Taken together, these findings supported Hypothesis 3. When selecting a party, respondents prioritize its ideological position above all other considerations. Although parties with open primaries fare slightly better than more closed ones, they are still likely lose to ideologically closer parties. Additionally, female representation dominates most selection procedures except for all-voter primaries. This result is meaningful given this method’s rarity in the real world: only three parties in our observational data opened their primaries to all registered voters.Footnote ¹⁹ At the same time, candidate selection is not the weakest consideration, especially compared to leadership experience and party status, indicating some non-zero priority. These results, too, replicated across all three countries despite their different institutional and political contexts.

Real-World Knowledge

Alongside its advantages, our conjoint design is limited by its explicitness, which presents voters with clearly stated party attributes. In reality, however, voters do not pay equal attention to all party characteristics, a sign of their internal priorities in itself. Thus, some respondents may overstate dormant preferences when asked about them directly but overlook them in real life. Real-world attention, therefore, provides an additional indication of an attribute’s relative priority and cognitive availability to voters when thinking about electoral politics.

To address this limitation and reaffirm our experimental findings, we concluded our survey with a simple multiple-choice knowledge question. We first requested that respondents state the party for which they voted in the most recent elections, which were held a year prior (2022) in all three countries, and estimate their attachment to it on a 0–10 scale. Next, we asked whether they recall this party’s candidate selection method in the last elections, as follows: ‘Think of the party that you voted for in the last election. To the best of your knowledge, how did this party select its candidates for parliament? If you cannot remember or are unsure, please mark the appropriate answer (option 4) without guessing or checking the correct answer.’ The answer scale included five options: (1) ‘Candidates were selected by the party’s leaders’; (2) ‘Candidates were selected by an internal body of party representatives, like a committee or a convention’; (3) ‘Candidates were selected by primary elections’; (4) ‘I do not know or currently cannot remember’; (5) ‘I did not vote in the last election.’ We then collected real-world data on party procedures in each country and recoded the answers into the following categories: a correct answer, a wrong answer, or a stated lack of knowledge/memory.Footnote ²⁰

Figure 5 presents the distribution of answers. To reduce bias, we employed survey weights representing each country’s adult population by gender, age, and education. The results aligned with our earlier findings. In all three countries, only a minority of voters knew/remembered the candidate selection method used by the party for which they voted only one year earlier. This pattern was most pronounced in France and Sweden, where less than a quarter of respondents (24.4 and 17.1 per cent, respectively) answered correctly. Israeli respondents fared better, with 43.9 per cent answering successfully, but the majority still answered wrongly or could not remember.Footnote ²¹ The findings, therefore, supported our expectation that voters devote low attention and cognitive resources for candidate selection methods when supporting a party.

Figure 5. Respondents’ recall of their elected party’s candidate selection method in the most recent election.