Measuring attitudes and assessments regarding democratic legitimacy is a central aspect of the literature devoted to measuring public opinion in relation to democracy (Inglehart and Welzel, 2005). Survey research has exhaustively documented the bases for (dis)contentment with democracy, revealing public opinion about the workings of democracy. For instance, open-ended survey questions have shown that people mostly define the meaning of democracy in terms of political freedoms, civil liberties, and rights rather than institutional and procedural or socioeconomic features (Dalton et al., 2007; Shin, 2017). Furthermore, open-ended survey questions conducted in the United Kingdom and in Australia inform us that citizens mostly define democracy in terms of freedom and human rights, elections and procedures, as well as having a voice in government (Wike and Fetterolf, 2021).

Closed-ended survey questions further show that opinions on how well democracy is working relate to whether people believe their most fundamental rights are being respected (Wike et al., 2021). Furthermore, the extent to which citizens consider their political elites and institutions corrupt has been consistently found to exert a strong effect on both their trust in institutions and, thereby, their evaluation of the workings of democracy (Maciel and de Sousa, 2018; Mauk, 2021).

Despite these findings, it is often difficult to clearly assess the causes of variations in satisfaction with democracy (Ariely, 2014), as well as what representations of democracy are salient in public opinion (Dahlberg et al., 2020). The literature has also revealed significant cross-cultural variations in the understanding and expectations towards democracy (O'Donnell, 2007; Ariely and Davidov, 2011). Overall, democracy remains an “essentially contested concept” (Ariely, 2014, p. 624) whose meaning can vary greatly in diverse cultural settings.

In addition, the use of a single measure of support for democratic performance holds important weaknesses, specifically because citizens’ views of democracy impact the way this question is answered (see some critics of using a single measure: Norris, 1999; Canache et al., 2001; Schaffer, 2014; Ferrin, 2016). For instance, the SWD item (i.e., “How far are you satisfied with the way democracy is working in your country?”) is one of the most frequently used indicators in public opinion research. This single-item measure is used in the European Social Survey (ESS) and the World Values Survey (WVS) with slightly different phrasing (“On the whole, how satisfied are you with the way democracy works in the [country]” and “How satisfied are you with the way democracy is developing in our country,” respectively) and answer options (10-point scale and 4-point scale, respectively). Nevertheless, Quaranta (2017) showed that the comparison of an index based on several items related to democracy evaluation and the single item SWD provides a reliable convergence across the measures at the country level and presents similar findings when used at the individual level.

To improve cross-country comparisons, sets of indicators have been developed to measure which aspects of democracy are supported and contested by citizens. A summary of the prominent dimensions for measuring citizens’ understanding and evaluation of the workings of democracy can be found in the ESS documentations for rounds 6 (Kriesi et al., 2013) and 10 of the survey (Ferrin, 2018). Overall, six dimensions aim to capture different components of democracy: electoral, liberal, social, direct democracy, inclusiveness, and type of representation dimensions. The development of these sets of indicators enables researchers to better understand what citizens want democracy to be and what adaptations are needed.

The use of other data sources (e.g., social media content, comments from online news portals) might help to obtain more precise indicators of citizens’ democracy concerns. For instance, Dahlberg et al. (2020) showed that investigating democracy dimensions can be usefully applied for extracting relevant understandings from news and social online sources and, thereby, for analysing how the concept of democracy is used in its “natural habitat.” Reveilhac and Morselli (2022) have also shown the potential of relying on complementary text classification methods to classify democracy frames along the survey dimensions from the ESS.

Political discussion constitutes a core component of democratic life that is consequential for public assessments of the workings of democracy, as political discourse is heavily contextually and culturally loaded. Previous research has demonstrated that political elites’ views correlate with citizens’ broader attitudes. For instance, early expectations regarding the influence of political discourse on citizen opinions have been extensively discussed in the context of foreign policy, especially by Zaller (1992), who explains that elite discourse is the key to explaining support for war. We follow a similar top-down approach suggesting that political discourse is closely related to and, under some circumstances, might affect citizens’ attitudes towards policy issues in general and towards the working of democracy.

Zaller’s (1992) classic models of opinion formation underline the importance of opinion leaders on public opinion (Katz and Lazarsfeld, 1955), describing the role of opinion leaders in the filtering of political information on traditional media. Zaller has argued that, in the context of a democratic system, opinion leaders play a key role in informing citizens about policy issues and democratic questions. By emphasizing the importance and specific aspects of democratic life, opinion leaders can inspire individuals to change their views based on the messages they receive.

This top-down approach to public opinion is thus well suited to examine the relationship between political discourse and public understandings of democracy. In general, political discourses are an important part of a country’s political climate (Flores, 2018), which might be related to people’s understanding of reality (Careja, 2016). Evaluations of the role of political elites in shaping attitudes towards societal issues have been conducted on different topics. For instance, Czymara (2020), relying on opinion survey data and party manifesto data, has investigated how political elites can play a key role in fostering or impeding immigrant integration by shaping public opinion. The impact of elite discourse on citizens’ attitudes has also been investigated in other policy domains, such as trade agreements (Dür and Schlipphak, 2021), gender equality (Jones and Brewer, 2020) and European policy (Kortenska et al., 2020).

The perspective of political elites’ discussions and views can act as drivers of the public evaluation of democracy (Dalton et al., 2007; Norris, 2017) also makes sense when extended to the realm of social media political discourses. Indeed, users who are highly active and vocal on social media tend to form an elite network, particularly in the case of Twitter (Blank and Lutz, 2017), which can influence opinions circulating online but also affect their dissemination in public opinion offline. For instance, political users can gain access to exposure on traditional media by voicing their positions and interests on social media (Dubois and Gaffney, 2014), thereby indirectly affecting public knowledge about policy issues. Social media content is indeed increasingly reported in news media by journalists (McGregor, 2019), facilitating the diffusion of the political content discussed on these platforms even to citizens who are not active on social media. In addition, Newman et al. (2017) demonstrated that up to 42% of the European population follows at least one politician on social media.

Social media provides new opportunities for political communication by allowing a closer interaction between the public and political elite (Ceron and Memoli, 2016; Newman et al., 2017). In particular, voices that contest prevalent democratic principles can reach a wide audience via social media platforms (Hameleers, 2018). Most notably, the challenge to moral principles and institutional arrangements in stable democracies has been associated with the rise of populist ideas (Rooduijn et al., 2014; Norris and Inglehart, 2016; Norris, 2017).

Therefore, in addition to the top-down approach, our study draws from the observation that citizens can endorse selective attention to political content. For instance, Wüest (2018) demonstrated that citizens’ attention to given policy issues and their related subtopics is partly correlated with respondents’ media consumption preferences. This suggests that citizens might give different amounts of attention based on the accessibility (or resonance) of political information based on shortcuts and affinities.

With respect to political discourse, it can be expected that when faced with complex policy issues, citizens are more receptive to messages from their preferred politicians (or politicians with whom they share a similar ideology) to reach informed preferences (McDermott, 2005). Citizens might thus be more likely to incorporate political information that is congruent with their existing political orientations into their own views (Steppat et al., 2022). This is also in line with the extensive literature on the role of social media in promoting polarized discourse (see review by Kubin and von Sikorski, 2021). The extent to which citizens are receptive to political messages might also depend on their level of interest in politics (Otjes and Rekker, 2021).

In this study, we merge social media and opinion survey data sources at the individual level based on citizens’ political orientation to better understand the relationship between political discourse and public SWD. We adopt a cross-country perspective, investigating a selection of diverse democratic systems. To minimize cultural confounding, although preserving some heterogeneity, we limited our selection of countries to the Western European context. This focus is on countries within the Schengen area to ensure a consistent context of freedom of movement and common policies that might influence political discourse and public opinion. The inclusion of Switzerland and the UK, although non-EU countries, is justified by their significant political and economic interactions with the EU, making their political dynamics relevant for comparative analysis. More specifically, we choose countries belonging to the Schengen Area that took part in round 10 of the ESS (European Social Survey) and that did not undergo a parliamentary election during the ESS fieldwork.

A final selection of eleven countries met these criteria: Austria, Finland, France, Germany, Greece, Italy, the Netherlands, Spain, Sweden, Switzerland, and the United Kingdom. Table 1 reports the Economist Intelligence Unit (EIU)’s democracy index (2021)¹, showing that the selected countries rank high on the list of the 167 investigated countries.

This selection allows us to cover a range of different democratic systems. For instance, in Switzerland, voters are called to the polls up to four times annually, while combined ballots are exceptional measures that take place only every few years in other countries (e.g., France and Sweden). However, the average voter turnout in Switzerland is relatively low as a reflection of the frequency and complexity of the objects. Furthermore, participation in (action) groups (such as trade unions or civic organizations) is comparatively low in Switzerland. These differences can affect rankings of democracy quality, such as the EIU’s democracy index, as these processes might not be completely reflected in expert indexes.

To determine citizens’ SWD in their country, we used data from round 10 of the ESS. In particular, we retrieved the SWD item “How satisfied with the way democracy works in country.” The response scale ranges from 1 “Not at all satisfied” to 10 “Very satisfied.” This item constitutes the dependent variable.

The specific module of round 10 of the ESS emphasizes different dimensions of democracy while also differentiating between the meaning (e.g., support for the idea that it is important to live in a country governed democratically) and the satisfaction/evaluation (e.g., how far people think democracy lives up to this ideal in practice) of these dimensions (see Table 2). Politicians’ discourse on democracy was analysed to assess the extent to which the emphasis (or salience) of specific dimensions relates to citizens’ SWD. We use emphasis as a good approximation of the extent to which a democracy dimension is central to party communication and, thereby, the extent to which it is accessible to the public. To do so, we classified the tweets according to the democracy dimensions described by Kriesi et al. (2013) and Ferrin (2018). More specifically, we focused on the following democracy dimensions (see third column of Table 2): competition (or fairness of electoral procedures), representation, accountability, voice, institutional participation, social and political equality, efficiency, social and political fairness, institutional participation, voice, freedom and rule of law, responsiveness to citizens, and responsiveness to other stakeholders. We also included sovereignty (e.g., referring to defending a nation’s political autonomy, preserving cultural identity, and shielding the domestic economy), as well as political efficiency (e.g., referring to political authorities’ corruption and mismanagement of resources but also to the guaranty of parliament and government independence) as additional dimensions.

The level of mean public satisfaction with respect to each of these democracy dimensions is given in Table 3. We see that “responsiveness to citizens,” “institutional participation” and “voice” are the dimensions on which satisfaction is the lowest on average across countries.

To capture the main framings associated with national and partisan democracy opinions, we extract tweets by elected politicians in each country. Figure 1 summarizes the different steps involved in the collection, filtering, and classification of politicians’ tweets. The list of politicians’ accounts was obtained from the Politicians on Social Media project (Haman and Školník, 2021) and updated manually for countries in which a more recent parliamentary election has been held. We focused only on politicians with an active political mandate. Twitter data were collected for a period of 6 months before the starting date of the ESS fieldwork in the respective countries.

Regarding data collection, we used Twitter’s Academics API to access politicians’ full tweet history (original tweets, retweets and replies). We relied on the academictwitteR wrapper (Barrie and Ho, 2021) from the R programming language. For each politician, we also specified his political affiliation.

As our focus is on tweets related to democracy ideas and perceptions, we filtered tweets containing specific keywords. The keywords were first generated based on Wordnet and the RelatedWords² webpages to retrieve synonyms and words associated with the term “democracy.” A list was elaborated based on the identified candidate terms in English. Then, the authors independently coded the candidate terms for their inclusion or exclusion in the final list (Cohen’s kappa 0.76). The divergences were discussed, and a final list (see Supplementary Table S1 for the English terms) was elaborated and translated into the different country languages by native speakers. Our final corpus of selected tweets related to democracy contains 16,505 tweets from 167 Swiss politicians and 474,751 tweets from 3,075 other politicians.

After selection based on keywords related to democracy, preprocessing steps were conducted, which included the removal of URLs as well as the division of concatenated hashtags (e.g., “#RuleOfLaw” becomes “rule of law”). On this basis, the tweets were translated into English using Google Translate. Once translated, the tweets are lemmatized using the udpipe library from the R programming language (Wijffels et al., 2018). We also considered only words with more than two characters. Thus, after removing any unconventional features, we converted the remaining text to lowercase. We also removed any stop words from a document (e.g., “the,” “and” “are”), which have little bearing on the overall meaning of the document. Additionally, we removed words that are specific to Twitter terminology (e.g., “&amp,” “rt”).

Then, to assess the prevalence of the specific democracy dimensions (see Table 2), we draw on a custom dictionary containing 1,072 words validated in a previous study (Reveilhac and Morselli, 2022). The dictionary was applied to the collected tweets using the liwcalike function from the quanteda.dictionarie package for the R programming language. The function tags each tweet with the proportion of words that fall into each dimension. We randomly sampled 1,000 tweets from each democracy dimension to train the classification model. To label each tweet, we selected the dimension corresponding to the higher proportion. We did not assign a label to tweets having the same proportion on several dimensions. The tweets that could be labelled by the dictionary can be considered as “emblematic” of each democracy dimension and were used as a training set to train a classification model based on the BERT (Devlin et al., 2018) algorithm for Python. The classification accuracy for each democracy dimension is given in Table 4. The resulting classification model was applied to unseen tweets.

The chosen classification model involves unsupervised pretraining based on word embeddings (Mikolov et al., 2013; Joulin et al., 2017). More specifically, our custom BERT model was pretrained on our training dataset to generate context-specific embeddings. We then conducted multinomial regression on the top of the BERT representation for predicting democracy dimensions. We assessed the classification accuracy by applying the model to a held-out test dataset (which represents 20% of the labelled tweets).

We relied on the ESS items asking respondents to mention their proximity to major parties in each country. For each respondent, we assigned the preferred party preference. If no party was mentioned, we assigned the political orientation measured on a 10-point scale ranging from 1 “Extreme left” to 10 “Extreme right,” which we recoded into three categories (1–4 as “left,” 5–6 as “centre,” and 7–10 as “right”). We also assigned a similar left/centre/right categorization to each party mentioned.

For each country, we merged survey and social media data using political orientation as an aggregating variable. Individual-level responses to the ESS were averaged by the respondent’s left/centre/right classification. Similarly, Twitter data were classified as left/centre/right depending on the party orientation of the politician. We aggregated only tweets emitted 6 months prior to respondents’ date of survey completion. We add one column per democracy dimension, and each column represents the share (or salience) of tweets dedicated to the specific dimensions. When merging the data, we used party preference when available or the political left-centre-right orientation if otherwise.

Our analytical strategy aims for a detailed exploration of the role that political information plays in a real-world assessment of democracy. To do so, we rely on a multistep strategy:

First, we rely on Spearman rank correlation between survey answers (satisfaction) and salience in politicians’ tweets of the democracy dimensions.

Second, we display the distance between surveyed attitudes towards democracy dimensions (items about the “importance” of each dimension) and the prevalence of political discourse on Twitter on the same dimensions.

Third, we rely on density plots to highlight the (theoretical) exposition of respondents to political messages (tells us how likely democracy dimensions are accessible to citizens given their political orientation).

Fourth, we use random forest-based variable importance analysis to determine how politicians’ emphasis on democracy dimensions is related to citizens’ SWD.

The reliance on variable importance analysis is inspired by previous studies that have used social media data to understand public opinions and attitudes (e.g., climate change beliefs as in Kirilenko and Stepchenkova, 2014 or Cody et al., 2015) and to elucidate the constituent factors that contribute to individuals’ opinions (Bennett et al., 2021). Variable importance analysis is based on the idea that the more a model relies on a variable to make predictions, the more important this variable is for the model. The relative importance for a single independent variable can be determined by deconstructing the model weights. This enables us to reveal the relative relevance (or degree of association) of a certain explanatory variable for the response variable. Random forest offers several advantages. First, it can account for nonlinear relationships and interactions in the data without the need to specify them (Molina and Garip, 2019). Second, it can handle cross-level data, as we are studying individual-level SWD and contextual emphasis on democracy. Third, random forest allows us to quantify the importance of a variable.

In our study, we used the average Gini impurity to measure how well the data were split, that is, how well it correctly predicts respondents being satisfied or dissatisfied with democracy (i.e., Gini = 0 if a covariate perfectly splits the data, Gini>0 for an imperfect split). Specifically, we use a random forest model (as implemented in the R package randomForest from Liaw and Wiener, 2002) that estimates the importance of variables in a model: if the trees in the forest split the sample by dimension A more than by dimension B, then dimension A is more important to the model. We use the results from the variable importance analysis to assess which democracy dimension most affects SWD with mean Gini impurity as the threshold when other control variables (e.g., gender, age, education and political interest) are included in the model. This allows us to make comparisons across the effect of the emphasis of the democracy dimensions on citizens’ SWD.

The analysis of the correlation between citizens’ satisfaction with each democracy dimension and the salience of these dimensions in politicians’ tweets is calculated based on Spearman rank correlation. We only included democracy dimensions that were also surveyed (thus, efficiency and sovereignty are excluded from Table 5). There is variability across the countries, as the correlation is high for Germany, Austria, and Finland (ρ >0.3). It is moderated for the United Kingdom and Greece and low for Italy, Switzerland, the Netherlands, France and Sweden. It is even negative for Spain. This might suggest that the dimensions that are considered the least satisfactory from citizens’ perspective are not necessarily the ones emphasized in politicians’ tweets. Low and negative correlations might even indicate a disconnect between citizens’ preoccupations and politicians’ emphasis.

Figure 2 presents a measure of distance between the attitude towards each democratic dimension (item asking survey respondents to what extent the dimension “is important”) and the salience in politicians’ tweets by political orientation. Both measures are standardized between 0 and 1. The smaller the distance between surveyed attitudes and tweet salience, the higher the resonance between public opinion and politicians’ discourse, while a large distance implies a gap between citizens’ opinions and politicians’ discourse. We only included democracy dimensions that were also surveyed (thus, efficiency and sovereignty are excluded from Figure 2). Furthermore, the item on “responsiveness to other stakeholders” is reversed.

Figure 2 shows a smaller distance among right-wing oriented citizens and politicians. Wide gaps were found among the left, especially in relation to rules, fairness, responsiveness to other stakeholders, competition, and representation, with politicians massively underdiscussing the dimensions compared to the importance given to them by their electorate. It is also worth noting that survey responses on these dimensions indicated a higher importance among left-wing respondents than centre and right-wing respondents. Accountability was the only dimension to be overstressed by (right and centre) politicians when compared to their electorate.

To further assess which political messages citizens are more likely to be exposed to, Figure 3 displays the density plots of the saliency of each democracy dimension in the tweets of politicians, according to their political orientation. All democracy dimensions are considered, and Figure 3 provides the average scores across countries. One-way ANOVA test with country as the grouping variable shows significant differences across countries for accountability (with Finland displaying the highest salience and the United Kingdom displaying the lowest salience, p < 0.05), competition (with Germany displaying the highest salience and Spain displaying the lowest salience, p < 0.05), efficiency (Netherlands highest and France lowest, p < 0.05), fairness (Spain highest and Greece lowest, p < 0.05), institutional participation (Switzerland highest and Finland lowest, p < 0.05), representation (Sweden highest and Germany lowest, p < 0.05), responsiveness to citizens (Switzerland highest and Germany lowest, p < 0.05), responsiveness to other stakeholders (Switzerland highest and Sweden lowest, p < 0.05), responsiveness to other stakeholders (Switzerland highest and Sweden lowest, p < 0.05), rules (Spain highest and Sweden lowest, p < 0.05), and voice (Netherland highest and Austria lowest, p < 0.05).

Figure 3 and one-way ANOVA with political orientation as the grouping variable show that citizens with a right-wing orientation are more likely to be exposed to discourses about accountability, representation, institutional participation, rules and sovereignty. Citizens with a left-wing orientation are essentially exposed to the dimensions related to fairness, responsiveness to other stakeholders, and voice. Exposition specific to centrist citizens is less clear-cut, but electoral competition, efficiency and responsiveness to citizens appear as salient dimensions. Some dimensions are salient across ideological preferences, such as competition and representation.

Figure 4 displays the results of variable importance to predict respondents’ SWD (controlled by gender, age, years of education, political interest, and country). The grey vertical line corresponds to the mean variable importance for the variables related to the democracy dimensions, while the red line also accounts for the control variables. The democracy dimensions that have values above these thresholds are the most important for predicting SWD.

Figure 4 shows that rules and responsiveness to other stakeholders and to citizens are the most important dimensions for predicting citizens’ SWD across countries. These results suggest that discourses from the right generally have more influence on predicting citizens’ SWD. It is also noticeable that the voice, sovereignty and accountability dimensions are impactful if we take the mean threshold for all democracy dimensions. The least important dimensions to predict overall satisfaction with democracy are voice, representation and efficiency. The other dimensions remain below the threshold values but can nevertheless be more impactful in some countries. For instance, institutional participation is moderately important, but analyses for individual countries (not shown here) demonstrate that this dimension can be salient in some contexts (e.g., in France, political discussions asking for more direct democracy procedures were prevalent in relation to the Yellow Jackets claims).