The preservation of a sphere of activity that is outside of politics is important if one is to have the balanced participation of the civic culture. – Almond and Verba (Reference Almond and Verba1963)
Introduction
Affective and media polarization are key drivers of political discontent in Western democracies. While some studies question the polarising effects of partisan media (for example, Guess et al. Reference Guess, Barberá, Munzert and Yang2021; Wojcieszak et al. Reference Wojcieszak, de Leeuw, Menchen-Trevino, Lee, Huang-Isherwood and Weeks2023), strong evidence suggests that partisan media exposure fuels affective polarization by promoting positive feelings toward members of political ingroups and negative feelings toward outgroups (Kubin and Sikorski Reference Kubin and von Sikorski2021; Lelkes, Sood and Iyengar Reference Lelkes, Sood and Iyengar2017; Levendusky Reference Levendusky2013). In turn, affective polarization erodes interpersonal trust (Carlin and Love Reference Carlin and Love2018) and diminishes satisfaction with democracy (Dassonneville and McAllister Reference Dassonneville and McAllister2020). This dynamic is often depicted as a vicious circle in which escalating media polarization deepens public opinion divides, while partisan identities fuel the demand for increasingly polarized media content (Wilson, Parker and Feinberg Reference Wilson, Parker and Feinberg2020), leading to growing discontent with the nature and quality of democratic life. While considerable scholarly attention has been devoted to understanding the factors behind rising discontent with democratic institutions (Beek 2019), limited knowledge exists regarding methods to slow down or reverse this trend or the role that the media can play in reducing civic polarization and disaffection. For example, emerging research suggests that decreasing exposure to national news, while increasing engagement with local news, could reduce polarization in the USA (Darr, Hitt and Dunaway Reference Darr, Hitt and Dunaway2018, Reference Darr, Hitt and Dunaway2021; Martin and McCrain Reference Martin and McCrain2019; Moskowitz Reference Moskowitz2021). This observation leads to a broader implication: if the media can have such negative effects on public opinion, it might also play a role in the reduction of polarization, and the restoration of satisfaction with democracy.
This study explores such a mechanism in the context of falling media polarization during the global coronavirus pandemic in the UK. We find that the onset of the stay-at-home (lockdown) order brought a sharp depoliticization of media content, as newspapers shifted from coverage of divisive domestic political subjects towards coverage of neutral, non-political areas such as entertainment, home improvement, sports, or leisure. Concurrent with this shift in the media landscape, we find both a sharp reduction in surveyed attention to politics and an increase in overall satisfaction with the democratic political system. These changes can be linked at the individual level to changes in news exposure and were especially pronounced among readers of media aligned with the political opposition, vis-à-vis readers of newspapers normally siding with the governing party. In contrast to conventional theories that focus on the role of collective mobilisation in securing satisfaction with the democratic process, we argue that the key to reducing partisan divides was a form of parochial de-mobilisation, as politically aligned citizens became less exposed to divisive political topics and less interested in politics on the whole.
We rely for this study on original cross-sectional survey data from a daily tracker of satisfaction with democracy in the UK (including over 200,000 respondents) and on data from X (formerly known as Twitter) for the contents of the newspapers people surveyed read. We use a Latent Dirchlet Allocation algorithm (LDA) to identify the topics, and the AFINN method to identify the tone of the contents each respondent was exposed to. After matching each respondent to the content and sentiment of the newspaper they report reading, we then use that information to explain variation in satisfaction with democracy. Our findings suggest that media content changed during the pandemic, particularly among newspapers ordinarily critical of the governing party, and that, at the same time, their readers’ satisfaction with democracy increased. We also identify the specific topics that display weaker or stronger associations with levels of expressed satisfaction with democracy. While we cannot establish definitive causal links, our analysis fills existing gaps in the literature on the possible influence of the media on satisfaction with democracy by identifying significant correlations between exposure to specific news and changes in the perceived functioning of democratic institutions.
Accordingly, the rest of this article is structured as follows. We begin by reviewing the research on the influence of mass media on democratic attitudes. Subsequently, we discuss the existing literature on the shifts in democratic attitudes observed during the pandemic. Following this, we present our data and the methodology employed in this study. After presenting our findings, we engage in a thorough discussion of our results and their limitations, accompanied by suggestions for future research. Finally, we conclude with a reflection on the significance of our findings in relation to democratic theory.
Mass Media, Selective Exposure, and Democratic Attitudes
The literature on the impact of mass media on democracy broadly offers two dominant opposing perspectives. A substantial body of research presents a ‘media malaise’ hypothesis, highlighting the potential role of the media in fostering civic cynicism and a decline in civic engagement (Putnam Reference Putnam2000; Patterson Reference Patterson2002; Mutz and Reeves Reference Mutz and Reeves2005). By contrast, other studies argue that media exposure can enhance citizens’ political interest and participation (Norris Reference Norris1999) and even contribute to the reduction or prevention of political polarization (Melki and Pickering Reference Melki and Pickering2014; Darr, Hitt and Dunaway Reference Darr, Hitt and Dunaway2018).
One means of reconciling these divergent perspectives would be to posit cross-cutting effects, resulting in a non-monotonic relationship between media consumption and democratic functioning. At early stages of economic and social development, increases in news consumption improve political knowledge and civic agency, resulting in a positive information-satisfaction relationship at the global cross-country level (Wegscheider and Stark Reference Wegscheider and Stark2020). However, in developed democracies, these benefits are largely exhausted as news saturation leads to diminished marginal returns, and further increases in media consumption carry risks as well as gains. Firstly, in line with the psychological literature on ‘information overload’ (Schmitt, Debbelt and Schneider Reference Schmitt, Debbelt and Schneider2018), higher levels of news exposure may exceed the cognitive threshold at which citizens can process additional facts and form new judgements. Secondly (and as a means of reducing cognitive burden) greater media intake may covary with confirmation bias, further reinforcing civic misperceptions. Many scholars apply this latter thesis to today’s media landscape, which not only offers citizens a wide range of choices in their media consumption but also algorithms that reinforce the role of partisanship in source selection. As individuals gravitate towards media sources that reinforce their pre-existing political leanings, this reinforces selective exposure or the behaviour where individuals actively seek out messages that align with their political beliefs (Iyengar and Hahn Reference Iyengar and Hahn2009; Mutz and Martin Reference Mutz and Martin2001; Steppat, Castro Herrero and Esser Reference Steppat, Herrero and Esser2022). Meanwhile, the choices of editors and newsroom staff in selecting and presenting news also augment this effect due to the association between content partisanship and viewership appeal. Media providers not only deliver information on specific issues but also play a significant role in shaping the perceived importance of those issues through the amount of coverage and positioning they receive. This phenomenon, commonly referred to as ‘agenda-setting’ (McCombs and Shaw Reference McCombs and Shaw1972), underscores the influence of television and the press in determining which issues receive attention and are deemed important by the public. As the relative prominence of news issues significantly impacts the public attention they are given (Dearing and Rogers Reference Dearing and Rogers1996), it follows that the media agenda has the potential to shape citizens’ considerations when evaluating the performance of democracy, in both positive and negative regards.
To date, much of the literature on partisanship and satisfaction with democracy focuses not on media effects per se but, rather, the broader ‘winner-loser gap’ in satisfaction with democracy separating voters for the winning candidate or party and supporters of the losing side (Anderson and Guillory Reference Anderson and Guillory1997; Blais and Gélineau Reference Blais and Gélineau2007; Singh, Lago and Blais Reference Singh, Lago and André Blais2011). However, this difference can be attributed in part to information exposure biases, as citizens actively seek out information that aligns with their pre-existing views, thereby reinforcing partisan differences in perceptions of institutional performance (Nadeau, Daoust and Dassonneville Reference Nadeau, Daoust and Dassonneville2021). Furthermore, given the varying priorities and emphases of different media outlets, we can anticipate parallel agenda-setting dynamics, wherein audiences form distinct conclusions about the political landscape based on the information they selectively consume. Thus, the satisfaction levels of election losers are influenced not only by the outcome of the election but also by the issues highlighted by the news sources of their choice. By catering to specific ideological affiliations, media outlets contribute to the formation of distinct political narratives and interpretations of current events. This, in turn, fosters partisan gaps in political satisfaction, as individuals align themselves with media sources that reinforce their pre-existing beliefs. Consequently, this partisan-driven media consumption behaviour serves to reinforce and amplify existing political beliefs, potentially exacerbating political polarization and shaping citizens’ perceptions of the political landscape they inhabit.
In addition to ‘issue salience’, or the relative importance of a topic in the media, it is equally important to understand how these issues are discussed (Balmas and Sheafer Reference Balmas and Sheafer2010). The concept of ‘negative news’ and its influence on individuals’ perceptions and evaluations is a fertile area of research in communication and political science (Lengauer, Esser and Berganza Reference Lengauer, Esser and Berganza2012). For instance, exposure to negative news about the European Union has shown a corresponding depressive effect on citizens’ evaluations of the institution (Desmet, Spanje and Vreese Reference Desmet, van Spanje and de Vreese2015; Brosius, Elsas and Vreese Reference Brosius, van Elsas and de Vreese2019; Foos and Bischof Reference Foos and Bischof2022). This suggests that negative media coverage can shape public perceptions and attitudes towards political entities, including national democratic institutions, particularly in cases where partisan news outlets contribute to fostering negative perceptions of democratic functioning.
If it is true that the degree and framing of media exposure in contemporary Western democracies have produced adverse consequences, then the ‘suspension’ in partisan news coverage during a major national crisis, such as the COVID-19 pandemic, offers a unique opportunity to observe how changes in media coverage, including a potential reduction in partisan bias, could possibly influence public perceptions of democratic performance. Examining such possible influence of media coverage during this critical period will provide us with valuable insights into how the media can shape citizens’ evaluations of democratic institutions.
Political Satisfaction During the Pandemic
A range of studies suggest that the global coronavirus pandemic led to a rise in support for existing leaders and enhanced trust in political institutions (for an early literature review, see Devine et al. Reference Devine, Gaskell, Jennings and Stoker2021; Adam et al. Reference Adam, Urman, Arlt, Makhortykh, Gil-Lopez and Maier2023; León et al. Reference León, Makhortykh, Gil-Lopez, Urman and Silke2023). The implementation of lockdown measures not only bolstered voting intentions for the governing party but also increased levels of confidence in the government and overall satisfaction with democracy in a number of European countries, including the UK (Bol et al. Reference Bol, Giani, Blais and Loewen2021; Davies et al. Reference Davies, Lalot, Peitz, Heering, Ozkececi, Babaian, Davies Hayon, Broadwood and Abrams2021). Such increases have been led by segments of the population that typically display lower levels of trust, indicating a catch-up effect that reduced enduring perception divides (Hegewald and Schraff Reference Hegewald and Schraff2022). Moreover, this increased trust spilled over to institutions that were not directly involved in crisis management, indicating a broader impact on public perceptions of the government (Baekgaard et al. Reference Baekgaard, Christensen, Madsen and Mikkelsen2020).
Despite the large number of studies documenting this improvement in both trust and political satisfaction during the COVID-19 pandemic, there remains a surprising lack of research examining its underlying mechanisms. Some researchers have characterized the rise in confidence as a ‘rally-round-the-flag’ effect (Baekgaard et al. Reference Baekgaard, Christensen, Madsen and Mikkelsen2020; Hegewald and Schraff Reference Hegewald and Schraff2022), which describes the tendency of public opinion to become more favourable toward political leaders in times of crisis (Mueller Reference Mueller1970). Yet this leaves unexplained the means by which such an effect has been produced.
We suggest that changes in mass media coverage can help to explain why such a rally effect took place. The proliferation of partisan news outlets and the prevalence of a high-choice media environment has often been attributed to exacerbating divisions and creating a polarized atmosphere (Lelkes, Sood and Iyengar Reference Lelkes, Sood and Iyengar2017), while Chang (Reference Chang2018) provides evidence of a negative association between reading newspapers and satisfaction with democracy. Yet, if the media has contributed to this increase in political dissatisfaction, it might also have contributed to the restoration of trust and satisfaction with democracy, once its content shifted in partisan balance and tone.
Throughout the COVID-19 pandemic, the portrayal of the rapidly evolving situation by the media significantly influenced public perceptions of risk and the broader discursive environment. In particular, national newspapers play a crucial role in shaping how other media outlets cover the same subject across different platforms (Ball and Maxmen Reference Ball and Maxmen2020; Mach et al. Reference Mach, Reyes, Pentz, Taylor, Costa, Cruz and Kerronia2021). The influence of newspaper coverage of the pandemic on public opinion has been attested by Mach et al.’s (Reference Mach, Reyes, Pentz, Taylor, Costa, Cruz and Kerronia2021) comparative study between the USA and the UK. In the USA, newspapers aligned with opposition parties adopted a more sensationalist approach to reporting on the pandemic, contributing to its politicization. Conversely, the situation in the UK exhibited a distinct pattern: sensationalism was low in British media, and it did not significantly vary across news outlets in the UK, regardless of their ideological alignment (Mach et al. Reference Mach, Reyes, Pentz, Taylor, Costa, Cruz and Kerronia2021). In fact, studies have shown that affective polarization was lower in the UK, compared to other countries such as the USA (Flores et al. Reference Flores, Cole, Dickert, Eom, Jiga-Boy, Kogut, Loria, Mayorga, Pedersen Eric, Pereira, Rubaltelli, Sherman, Slovic, Västfjäll and Van Boven2022), most likely because health policy experts were treated in a non-partisan manner, thus reinforcing their credibility in the eyes of the electorate. This leads us to question whether the lack of sensationalism and polarization in media coverage, particularly regarding partisan divides, could have played a role in reducing political dissatisfaction.
Argument and Hypotheses
Building upon the existing body of literature, this study aims to investigate the role of mass media in the observed increase in satisfaction with democracy during the COVID-19 pandemic, focusing on the UK. The UK has been characterized by a notable surge in political dissatisfaction that reached historical highs in late 2019 due to escalating political polarization along partisan lines and the Brexit divide (Foa et al. Reference Foa, Klassen, Slade, Rand and Collins2020, Hobolt et al. Reference Hobolt, Popa, Van der Brug and Schmitt2022). The sudden reversal of this downward trend in satisfaction with democracy presents a unique opportunity to investigate the underlying mechanisms responsible for this restoration.
While we do not make any strict causal claims, our study aims to address a significant research gap by specifically examining the relationship between media exposure and changes in satisfaction with democracy. While several studies have investigated the influence of media coverage on satisfaction with democratic institutions at the European Union level (Desmet, Spanje and Vreese Reference Desmet, van Spanje and de Vreese2015; Foos and Bischof Reference Foos and Bischof2022), there is a notable scarcity of research exploring the impact of media coverage on satisfaction with the functioning of national democratic institutions (Maurer Reference Maurer2023; Ceron and Memoli Reference Ceron and Memoli2016; Chang Reference Chang2018). Furthermore, to the best of our knowledge, while a few studies have analysed the ‘rally-round-the-flag’ effect on governmental support during the COVID-19 pandemic (Hegewald and Schraff Reference Hegewald and Schraff2022), no study has so far explained how traditional media sources act as the main mechanisms through which such effect is diffused. Similarly, to our understanding, there is no research on the relationship between the tone, or sentiment, associated with specific news topics, and satisfaction with democracy. Therefore, our research seeks to fill this gap and shed light on how differences in media coverage might relate to the observed restoration of democratic satisfaction during the pandemic. While alternative explanations for the media effect exist (Kritzinger et al. Reference Kritzinger, Foucault, Lachat, Partheymüller, Plescia and Brouard2021; Erhardt et al. Reference Erhardt, Freitag, Filsinger and Wamsler2021), the scope of our study is limited to the role of traditional media outlets.
Understanding the relationship between reporting by mainstream media outlets and citizens’ democratic attitudes requires an examination of the polarized nature of the British media landscape. Unlike in the USA, the UK print media market continues to be characterized by the dominance of a few national newspapers that serve as significant sources of political information (Brynin and Newton Reference Brynin and Newton2003). The British newspaper industry has played a notable role as a political actor, marked by its controversial and often overtly partisan nature (Wring and Deacon Reference Wring and Deacon2010). Consequently, we should expect substantial differences in content between newspapers supporting the Conservative government and those aligned with the opposition Labour Party.
Against this backdrop of polarization, the emergence of the COVID-19 pandemic had the potential to bring partisan groups closer as a common threat to all citizens regardless of their political views. For example, in the first weeks of the pandemic, even opposition parties expressed a more positive tone in parliamentary debates (Louwerse et al. Reference Louwerse, Sieberer, Tuttnauer and Andeweg2021). This convergence also might have taken place in the media landscape: as discussed earlier, British media showed no significant variation across news outlets in the first months of the pandemic, irrespective of their ideological alignment (Mach et al. Reference Mach, Reyes, Pentz, Taylor, Costa, Cruz and Kerronia2021). Building upon this evidence from the UK context, we propose the following hypothesis:
H1: Media content polarization decreased in response to the pandemic.
This hypothesis suggests that the COVID-19 pandemic may have prompted media outlets to reduce polarization in their content, thereby potentially contributing to the restoration of democratic satisfaction. It is important to consider that the dynamics of media polarization during the pandemic may have evolved over time, with potential variations across different phases of the crisis. Therefore, we will test this hypothesis by tracking content polarization throughout the pandemic.
The editorial stance of different news outlets shapes the extent and emphasis of particular issues, which in turn influence how citizens evaluate the effectiveness of democratic institutions (Desmet, Spanje and Vreese Reference Desmet, van Spanje and de Vreese2015; Foos and Bischof Reference Foos and Bischof2022). Moreover, the relative prominence of different news topics can significantly impact individuals’ perceptions of the political landscape (Hegewald and Schraff Reference Hegewald and Schraff2022). Given these considerations, we anticipate that alterations in the media content to which citizens are exposed may be linked to the observed increase in political satisfaction during the pandemic:
H2: Shifts in media content during the pandemic towards more politically neutral topics are associated with increases at the individual level in satisfaction with the democratic political system.
This hypothesis aligns with the argument that partisan dissatisfaction is often influenced by media bubbles and echo chambers, where individuals selectively consume information that aligns with their political beliefs, leading to polarization (Stroud Reference Stroud2010). However, if media contents converge and become less polarized during the pandemic, it is reasonable to expect that the satisfaction gap between media bubbles will decrease.
Finally, in order to further refine our understanding of the mechanism by which media correlates with satisfaction with democracy, it is important to consider the salience of different topics in media coverage. During the pandemic, the global health crisis dominated news agendas and overshadowed the polarized political climate in the UK, following protracted Brexit negotiations and a snap general election in December 2019. This raises the question of how exposure to specific negative or neutral news itself may have affected individuals’ satisfaction with democracy. Based on this consideration, we propose the following hypothesis:
H3: Exposure to specific negative news during the pandemic is associated with corresponding decreases in satisfaction with democracy.
This hypothesis suggests that the nature and tone of news coverage during the pandemic are associated with changes in individuals’ perceptions of democratic functioning. By examining the relationship between exposure to different topics in the news and satisfaction with democracy, we aim to shed light on the role of specific news topics in possibly shaping citizens’ attitudes and satisfaction with the democratic political system. Overall, our three hypotheses provide a comprehensive framework for investigating the correlation between media consumption and democratic satisfaction during the COVID-19 pandemic. By analysing changes in media content polarization (H1), the possible effects of content exposure (H2), and the role of specific news topics (H3), we aim to contribute to a deeper understanding of the complex interplay between media, democracy, and citizen satisfaction.
Data and Methods
Survey Data and Linkage Analysis
In this study, we scrutinize the evolution of citizen satisfaction with democracy between November 2019 and December 2021 in the UK.Footnote 1 We rely on cross-sectional data from the YouGov tracker of satisfaction with democracy, with 201,144 respondents through that period of time.Footnote 2 Founded over two decades ago with just 1,000 participants, the YouGov web panel is now one of the largest in the world with over two million British adults, of which approximately 0.5m are active respondents. Its core instrument is the Omnibus survey, in which a demographically representative sample of members are asked daily on a wide range of topics. The items included in this project were fielded three times per week, with respondents asked about their level of satisfaction with democracy and their newspaper consumption.Footnote 3 Satisfaction with democracy was measured using the classic formulation: ‘On the whole, are you very satisfied, fairly satisfied, not very satisfied or not at all satisfied with the way democracy works in Britain?’ For ease of coefficient interpretation and effect estimation, we re-factor this ordinal dependent variable on satisfaction with democracy to a dichotomous format (‘satisfied’ or ‘dissatisfied’), in line with existing studies in the literature such as Anderson and Guillory (Reference Anderson and Guillory1997), Singh (Reference Singh2014), or Mayne and Hakhverdian (Reference Mayne and Hakhverdian2017). We then coded a variable for the partisanship of newspaper exposure, classifying as left-leaning The Guardian, The Mirror, and The Independent, and as right-leaning The Telegraph, The Times, The Sun, and the Daily Mail.Footnote 4 Additionally, the survey also included a battery of questions concerning respondents’ personal characteristics such as household income, education, region, and how they voted in the last general election or in the Brexit referendum (see appendix for a summary of the data).
We use self-reported newspaper readership to build a dataset of exposure to media content by matching readers with the news content of the newspaper read in the days leading to the interview, as further explained below. The idea of constructing individual measures of exposure to relevant media messages was first explicitly described and implemented by Miller, Goldenberg and Erbring (Reference Miller, Goldenberg and Erbring1979). Such an approach allows us to take into account the actual media content individuals consume instead of assuming the kind of media messages the respondents were exposed to. This type of study, commonly defined as ‘linkage analysis’, combines content analysis (media message variables are aggregated for each media outlet and, possibly, time period) and survey data (containing self-reported media consumption and the outcome/s of interest). One of the main challenges in studying media effects is that errors in media use self-reports attenuate media effect estimates (Ansolabehere and Iyengar Reference Ansolabehere and Iyengar1995; Bartels Reference Bartels1993; Zaller Reference Zaller2002; Scharkow and Bachl Reference Scharkow and Marko Bachl2017). Even more importantly, there are limitations related to the coding and classification of contents respondents are exposed to. Scharkow and Bachl (Reference Scharkow and Marko Bachl2017) argue that automatic content analysis is particularly beneficial for linkage analysis, as it enables the coding and classification of huge amounts of text and allows for the estimation of misclassification error probabilities. Furthermore, one can easily employ an ensemble of multiple classifiers for the same message coding and increase the coding quality (Hillard, Purpura and Wilkerson Reference Hillard, Purpura and Wilkerson2008). Following this line of thought, we rely on automated text analysis techniques to analyze the contents our respondents were exposed to.
Media Content: Topic Modelling and Sentiment Analysis
Twitter (now known as X) has become an important platform for sharing news and political information, making it a suitable proxy for measuring the salience of news in the public sphere. With the declining reach of print and online newspapers in recent years, especially among younger age groups, social media has emerged as a dominant source of news for many individuals. The statistics from Ofcom in 2022 suggest that 46 per cent of UK adults use social media for news consumption, whereas the reach of newspapers decreased from 47 per cent in 2020 to 38 per cent in 2022. This indicates that social media platforms, like Twitter, play a crucial role in disseminating news from mainstream media. Therefore, we rely on news disseminated on Twitter by national newspapers to assess the type of news that dominates the public agenda. We scraped 1,564,350 tweets between October 2019 and December 2021 from all major British newspapers (Daily Mail, Daily Mirror, Financial Times, The Guardian, The Independent, The Sun, The Telegraph, and The Times), using the Twitter (X) API. Building on previous studies (Barberá et al. Reference Barberá, Casas, Nagler, Egan, Bonneau, Jost and Tucker2019), we use newspaper tweets as a proxy to study the salience of topics that are discussed in the public sphere. In our case, we use them as indicators of what each newspaper focused on in the days leading up to the YouGov respondents being interviewed.
Once we compiled and cleaned our corpus of circa 1.5 million tweets, we applied a topic modelling approach to identify the issues covered by these media outlets. As Grimmer and Stewart (Reference Grimmer and Stewart2013) succinctly explain, topic models are a type of unsupervised learning method useful when approaching a problem without a predetermined categorisation scheme. To put it in other words, they are useful methods to apply in an exploratory context without a predetermined list of topics (Quinn et al. Reference Quinn, Monroe, Colaresi, Crespin and Radev2010). Out of different topic modelling approaches, such as ‘Dynamic’, ‘Correlated’, and ‘Structural Topic Models’, the ‘Latent Dirichlet Allocation’ (LDA) developed by Blei, Ng and Jordan in Reference Blei, Ng and Jordan2003 is the most widely used approach. Following previous studies, we apply LDA in our research (Barberá et al. Reference Barberá, Casas, Nagler, Egan, Bonneau, Jost and Tucker2019; Blei, Ng and Jordan Reference Blei, Ng and Jordan2003; Blei and Lafferty Reference Blei and Lafferty2007; Grimmer and Stewart Reference Grimmer and Stewart2013).Footnote 5
The LDA algorithm, being an unsupervised method, only requires the specification of the number of clusters (
$K$
) as input. As Grimmer and Stewart (Reference Grimmer and Stewart2013) note, ‘determining the number of clusters is one of the most difficult questions in unsupervised learning’ (19). Scholars have suggested several metrics to select the optimal
$K$
number in LDA, based on various measures of statistical fit, including held-out likelihood (Arun et al. Reference Arun, Suresh, Veni Madhavan, Narasimha Murthy, Zaki, Xu Yu, Ravindran and Pudi2010; Cao et al. Reference Cao, Xia, Li, Zhang and Tang2009; Deveaud, Sanjuan and Bellot Reference Deveaud, Sanjuan and Bellot2014; Griffiths and Steyvers Reference Griffiths and Steyvers2004; Mimno et al. Reference Mimno, Wallach, Talley, Leenders, McCallum, Barzilay and Johnson2011; Taddy Reference Taddy2012). However, Quinn et al. (Reference Quinn, Monroe, Colaresi, Crespin and Radev2010) advise that rather than focusing on statistical fit, the primary criteria for
$K$
selection should be the ‘substantive and conceptual’ fit (216), that is, the identification of homogeneous and well-defined clusters (news topics, in this case) (Chang et al. Reference Chang, Gerrish, Wang, Boyd-Graber, Blei, Bengio, Schuurmans, Lafferty, Williams and Culotta2009). The final number of clusters can be reached through an iterative process, based first on statistical fit, and then, upon the aggregation of clusters closest to one another (by Euclidean distance) to form fewer more coherent groups. We take a comprehensive approach in our selection of
$K$
clusters, by following a four-step approach. First, we use four different measures of statistical fit to identify the ideal
$K$
number of topics (see appendix for more detail). The four measures we use to do so seem to converge, for our corpus, around the ideal number of nineteen topics. In a second step, we explore the automatically generated clusters by scrutinizing the most common words in each cluster of tweets and thus identify the substantive issues in each cluster. Third, we aggregate the clusters with shared topics (for example, we group three clusters of tweets on COVID-19) for further analysis, resulting in twelve topics (see appendix for more details).
Footnote 6,Footnote 7
In Figure1, we plot the percentage of tweets on each topic between November 2019 and December 2021. We find that tweets relating to COVID-19 increased substantially in March 2020, and remain at a substantial level (above 10 per cent of total tweets) until the end of our period of analysis. Similarly, tweets related to political violence spiked in November 2019, June 2020, and August 2021, following the London Bridge attack in London, the murder of George Floyd in the USA, and the takeover of Kabul by the Taliban, respectively. However, tweets about American politics increase around the time of the 2020 US presidential election and the January 6 assault on the Capitol. Finally, increases in sports-related news are clearly visible in July and August 2021, during the UEFA European Football Championship and the Tokyo Summer Olympics. Overall, our modelling seems quite reliable in terms of its ability to describe changes in coverage of news from across the world as reported in UK newspapers.Footnote 8
Average Issue Salience Across Newspaper Tweets. Source: X (Twitter) API.
As discussed previously, we are not only interested in knowing what topics each newspaper is focusing on but also how they talk about different topics. By quantifying the positivity or negativity by which topics are covered in the media, we are able to assess our first hypothesis (H1) regarding whether differences in sentiment between left- and right-wing newspapers may have narrowed during the pandemic. Thus, in addition to the unsupervised modelling, we also conducted sentiment analysis to understand how newspapers covered each topic. Sentiment analysis is an approach to text analysis used to identify and evaluate opinions of issues through automated methods (Zimbra et al. Reference Zimbra, Abbasi, Zeng and Chen2018). As an established method in the analysis of tweets, we apply the AFINN technique to understand how different newspapers in the UK reported over time on the topics identified using LDA (Nielsen Reference Nielsen2011). The integration of topic modelling and sentiment analysis is a well-established approach to Twitter data analysis (Xiang and Zhou Reference Xiang and Zhou2014; Saif, He and Alani Reference Saif, He and Alani2012; Huang et al. Reference Huang, Peng, Li and Lee2013; Tan et al. Reference Tan, Li, Sun, Guan, Yan, Bu, Chen and He2014).
The adoption of a dictionary-based AFINN approach reflects a preference for transparency and ease of replication, as though some alternative automated sentiment analysis algorithms are better performing, they are also opaque and harder to reproduce (Zimbra et al. Reference Zimbra, Abbasi, Zeng and Chen2018). Another clear advantage of AFINN is that it was developed specifically for the analysis of Tweets (Nielsen Reference Nielsen2011). AFINN scores individual words from
$ - 5$
to
$5$
, with the former indicating very negative emotions, and the latter having a more positive affect. The closer word scores approximate to zero, the more neutral they are in tone. We further validate our application of the AFINN method by comparing how different topics are associated with specific sentiments, and how distinctively (or not) newspapers from across the spectrum report on certain issues (see the Appendix).
As reported in Figure 2, some topics are systemically associated with more negative sentiment, whereas others are more positively framed. Validating our prior intuition, the topics with the lowest affective baseline are crime and political violence, followed by reporting on climate change, coronavirus, and both domestic and US politics. However, softer topics such as sports or entertainment oscillate around a neutral mid-point, at times reaching positive emotive scores. Although sentiment within topics seems stable to a large extent, short shifts in sentiment are described by the data. For instance, we find a sharp decrease in positive coverage of the British royal family in November 2019 following an interview with Prince Andrew attempting to address allegations of sexual misconduct, and then again in early 2020 following Prince Harry and Meghan Markle’s announcement to withdraw from official duties. Thus, also our sentiment analysis seems to be reliable as an instrument to understand how different topics were covered in the news in the UK for the period under study.
Average Sentiment Across Topics. Source: X (Twitter) API.
Modelling Strategy
Having clarified the origin of our Twitter data and the methods to measure it, we turn to our two-step methodological approach. With the first step, we aim to explain patterns in satisfaction with democracy in the UK, investigating its relationship with party affiliation and newspaper readership, thus testing H2. In a second step, we seek to disentangle how exposure to specific topics might correlate - and possibly influence - satisfaction with democracy, and therefore address the expectations set out in H3.
Multilevel Models
We use the YouGov tracker data to estimate multilevel models with random slopes and intercepts by week of observation for the key variables of consideration, such as partisan affiliation (past vote) or newspaper readership (left- or right-leaning), as well as fixed effects for age, gender, education, region, and income. A multi-level model is a statistical model applied to data collected at more than one level, with slopes varying by group (Luke Reference Luke2020; Gelman and Hill Reference Gelman and Hill2007). By grouping coefficient estimates by week of observation, such a modelling approach allows us to disentangle the relationship over time between satisfaction with democracy and other relevant factors during this period, such as political party support, media exposure, or Brexit position, while still controlling for relevant respondent attributes. As such, multilevel models offer an effective means of analysing trends in data, as they allow for the estimation of separate slopes by period between individual attributes and outcomes of interest while taking into account demographic fixed effects and short-term sample biases (Skrondal and Rabe-Hesketh Reference Skrondal and Rabe-Hesketh2004; Singer and Willett Reference Singer and Willett2003; Steele Reference Steele2008; Wright and London Reference Wright and London2009). As noted, fixed effects for age, gender, education, region, and income are estimated at the first level, while at the second level, we include random slopes for party support and media preference. This allows us to parse out changes in satisfaction with democracy by group within Britain from November 2019 to early 2022 in order to identify changes over time in the independent association of partisan identity or news readership with perceptions of democratic performance.
Multilevel models were estimated according to the standard specification:
$$SW{D_{ij}}=({\beta _{0j}}+{X_{0j}})+{\beta _1}\,{A_{ij}}$$
Where
$SW{D_{ij}}$
represents the score of subject
$i$
on satisfaction with democracy in period
$j$
,
${X_{0j}}$
denotes the random effects design matrix consisting of ones in the first column (corresponding to the estimation of random slope intercepts) and second-level variables in the other columns,
${\beta _{0j}}$
to the set of random slope coefficients for each time period
$j$
,
${A_{ij}}$
to a matrix of first-level independent variables including a constant term, for which time-invariant coefficients are provided by the vector
${\beta _1}$
. As random slopes estimated by week, we include the constant term together with three variables that allow us to measure the changing association between partisanship and satisfaction with democracy during the period under consideration: political party support (based on respondent vote preference in the previous general election); whether the respondent had supported remaining in the European Union (based on respondent vote choice in the 2016 referendum); and whether the respondent is a reader of a left- or right-leaning news source. Fixed-effect results of the multilevel models are shown in Appendix Table (A.2), and weekly random slopes by key variable are charted in the results section below.
Logit Models
In addition to the multilevel model described above, as a second step, we seek to explain how exposure to specific topics relates to satisfaction with democracy. To do that, we match each respondent to the retweets by topic on the days previous to the one in which they were interviewed for the newspaper that they usually report reading.Footnote 9 We then develop a robust research design to explain the relationship between satisfaction with democracy and exposure to certain topics by taking the average number of retweets in each newspaper for selected theoretically relevant topics 2, 10, 15, 20, and 25 days prior to the interview (that is, the lags of retweet by topic/newspaper) as a proxy for topic salience in each newspaper over a certain period. The following model should thus help answer the question about what topics seem temporally to predate a shift in satisfaction with democracy.
We therefore estimate the following general logit model given the binary nature of our dependent variable:
$$SW{D_{i,t}} = {\beta} Retweet{s_{i,t - n}} + \gamma PastVote{s_i} + \delta Persona{l_i} + \zeta Newspape{r_i} + \iota Pol. Attitud{e_i} + {\xi _t} + {\rm \varepsilon _{i,t}}$$
where
$SW{D_{i,t}}$
is the binary outcome variable for satisfaction with democracy for individual
$_i$
at time
$_t$
,
$\beta Retweet{s_{i,t}}$
is a vector of mean retweets by topics each day in the matched newspaper in the t-days before the survey took place,
$\gamma PastVote{s_i}$
a vector of political variables (party voted in last general election and Brexit referendum),
$\delta Persona{l_i}$
a vector of personal covariates, such as sex, income, education, age and region.
$\zeta Newspape{r_i}$
is type of newspaper readership (left or right-wing),
$\iota Pol. Attitud{e_i}$
is political attitude (that is, the self-reported level of attention to politics),
${\xi _t}$
the week fixed effects, and
${\varepsilon _{i,t}}$
is the standard error, with subscripts
$i\,=\,1,....,n$
;
$t\,=\,22/11/2019, ...,06/12/2021$
.
Results
We begin our empirical analysis by exploring the evolution of media content and citizen attitudes during the two years under study. As shown in Figure 1, the salience of different topics has varied dramatically over time. By contrast, the sentiment of each topic remains quite stable, with some topics being consistently negative while others remain neutral or moderately positive (see Figure 2). Combining both dimensions of our content analyses, we estimate the combined salience of negative and neutral topics. We consider ‘negative’ the topics with negative average sentiment scores across all newspapers (climate change, COVID-19, crime, daily news, UK politics, US politics, political violence, and weather). In contrast to these, news in the categories of entertainment, animals and pets, royal family and sports have sentiment values above-0.5 and reach positive average sentiment scores for some newspapers. We classify these as ‘neutral’ topics.Footnote 10
Figure 3 shows the trend lines during the pandemic across three key variables: the share of media stories that focus on ‘polarizing’, rather than neutral, topics (for example, domestic politics rather than lifestyle or entertainment); surveyed levels of attention to politics (self-reported on a 1-10 scale); and levels of surveyed dissatisfaction with the functioning of democracy. As the pandemic struck in March 2020, we found the confluence of three trends: media content became less polarizing, interest in politics declined, and satisfaction with democracy increased.
Media Content, Attention to Politics and Democratic Satisfaction: Key Trends. Sources: YouGov, X (Twitter).
Concurrent with this reduction in media attention to divisive topics, we also discover a marked narrowing of the partisan political divide. Figure 4 plots the differences between partisan groups across three variables: the share of left vs. right-leaning media stories that focus on ‘polarizing’ rather than neutral topics (for example, UK domestic politics rather than lifestyle or entertainment); the gap in surveyed levels of attention to politics (self-reported on a 1–10 scale) between left- and right-leaning newspaper readers; and the gap in levels of surveyed dissatisfaction with the functioning of democracy between readers of left- and right-wing media. In line with H1, we find that media content polarization decreased during the pandemic: left-wing media made the largest movement away from coverage of politically divisive subjects, such as domestic politics, and towards more neutral content such as lifestyle, travel, or personal interest stories. Simultaneously, the descriptive trends suggest that levels of attention to politics have equalized, and so has the winner-loser gap in satisfaction with democracy.
Depolarization During the Pandemic: the Closing Left-Right Media Gap. Trendlines during the pandemic Sources: YouGov, Twitter (X).
Figure 5 zooms in on the narrowing gap in the salience of negative topics between left- and right-wing newspapers, as a result of left-wing newspapers decreasing their coverage of negative news, thus supporting our predictions as set out in H1.
Gap in Salience of Negative Topics between Types of Newspaper.
These descriptive trends in Figures 4 and 5 indicate a parallel decrease in attitudes across partisan lines as well as in newspaper content differences. In our second hypothesis, we posit that the increase in political satisfaction during the pandemic is associated with changes in individual exposure to media content. In order to thoroughly investigate the association between news exposure and readers’ satisfaction with democracy at the individual level, as well as track the evolution of this relationship during the pandemic, a more sophisticated approach is warranted. We apply a linear mixed-effects model to our survey data, with random slopes for the last party voted for and the type of newspaper readership by week. Figure 6 plots the weekly association between reading left- or right-leaning newspapers and feeling satisfied with the functioning of democratic institutions in the UK. This allows us to assess the extent to which the narrowing partisan divide in perceptions of democratic performance at the macro level is independently associated at the individual level with either partisan support (past vote) or partisan media exposure, respectively.
Satisfaction with Democracy by Type of Newspaper Readership and Party Affiliation (Past Vote).
Changes in these estimated coefficients by week suggest a difference of trend in the association of each variable with democratic satisfaction over time. While at a descriptive level, left-wing respondents became more satisfied with democracy over the course of the pandemic, as shown by the steady reduction of the left-right satisfaction gap in Figure 4, after controlling for newspaper readership, the independent association between political orientation and satisfaction with democracy was found to remain stable. This is shown by the series line for past Labour vote in Figure 6 which plots period-specific coefficients that fluctuate around a-10 percentage point midpoint. Alternatively stated, after controlling for media exposure, Labour voters remained around ten percentage points less satisfied with democracy over the course of the period.
By contrast, coefficient magnitudes for newspaper readership follow an upward trendline, with the implication that (ceteris paribus) consumers of left-wing media became steadily less negative in their evaluation of democratic performance over time. At the pandemic’s outset, such readers were estimated to have a 10 percentage point shortfall in their rate of satisfaction with democracy, over and above the estimated electoral partisanship effect. However, the newspaper readership coefficient moderated by several percentage points over the course of the first COVID-19 lockdown from April to July 2020 before declining further the following year. As a result, the model coefficients imply that a substantial proportion of the overall reduction in the winner-loser gap during the pandemic seen in Figure 4 – that is, the left-right gap in satisfaction with democracy – can be associated with changing attitudes by newspaper partisanship rather than electoral party of support.Footnote 11
For inferential purposes, these findings have two implications. Firstly, in line with hypothesis H2, increases in satisfaction with democracy during the pandemic exhibit an independent association with media exposure. Given the moderation of partisan media negativity bias revealed by Figure 5, this provides a basis to explore further the connection between individual content exposure and respondent attitudes. Secondly, the stability of the negative partisanship coefficient is inconsistent with the alternative hypothesis that a narrowing winner-loser gap reflected reversion to baseline following the heightened political mobilisation of the December 2019 election. Were this the case, we would expect to observe a diminution of the partisanship coefficient, especially given its operationalisation using respondent votes in that very electoral contest. However, this coefficient remained stable after controlling for the choice of respondent media.
The Role of Negative News: Survey-Weighted Generalized Linear Model
The analyses presented so far suggest that media content changed during the pandemic and that there were divergent trends between progressive and conservative outlets. The multilevel models, in turn, demonstrate the existence of a statistically significant relationship between these content changes and the changing relationship between newspaper readership and satisfaction with democracy. To further substantiate this association and test our H3 regarding the specific effect of negative news on satisfaction with democracy, in this section, we turn to the effect of exposure to specific news in different newspapers.
Our results, presented in Table 1 using a survey-weighted generalized linear model (logit), show the relationship between exposure to different news topics and individuals’ attitudes. Controlling for other factors, we find that individuals who were exposed to media content covering issues relating to partisan political matters, such as UK politics, political violence, and climate change, all negatively correlate with satisfaction with democracy. Note that these effects control for partisanship and are relative to identical individuals who differed only in not being exposed to such media content in the days prior to the interview. Therefore, our analysis provides robust support for Hypothesis 3, which posits that exposure to negative news during the pandemic is associated with a decrease in satisfaction with democracy.
Survey-Weighted Generalized Linear Model (Logit)
Notes: *p < 0.1; **p < 0.05; ***p < 0.01.
In contrast to the negative effects of some topics, we find that exposure to media sources with greater neutral content – notably ‘personal interest’ stories (the daily news category), entertainment, and crime show no consistent effect. Feelgood stories about animals and pets seem to correlate positively, though, only in the short term. Meanwhile, exposure to content regarding non-partisan ‘unifying’ topics such as responses to the COVID-19 pandemic or news about the British royal family is associated with higher levels of democratic satisfaction relative to those whose media sources did not cover such issues. The models include individuals who were surveyed and who reported not reading any news. For these respondents, the exposure to topics is estimated at zero. To clarify, an increase in the salience of a topic here is measured as the average of retweets by topic for a given newspaper in the 2 to 25 days prior to the survey.
We can easily transform the log odds in Table 1 to changes in the predicted probabilities to express satisfaction with democracy as a function of the salience of each topic. For example, a one standard deviation increase in retweets of British politics or climate change news over a period of 20 days (211 and 279 RTs, respectively) is associated with a 3 percentage point decrease in the probability of being satisfied with democracy. By contrast, the salience of COVID-19 or the royal family predicts substantially higher levels of satisfaction with democracy. The probability of being satisfied increases by one percentage point for each one-standard-deviation increase in retweets of royal family news (178 RTs). The estimated effect is even larger for coronavirus-related news, with a one standard deviation rise in salience (of 635 RTs) predicting a three percentage point increase in the probability of feeling satisfied with democracy. These estimates show that the usual changes in the salience of topics are associated with non-negligible changes in democratic evaluation, often altering the probability of being satisfied above or below 50 per cent, depending on the sign of the effect. Thus, news exposure seems to be a significant predictor of citizens’ satisfaction with democracy.
The time lags in the results suggest that, consistent with theory, the greater the delay between exposure to specific topics in the news and the interview, the less of an association that exposure will have with a person’s perceptions of democratic performance. In other words, there might be an effect of exposure to certain news on satisfaction with democracy but for some topics, at least, it is only a temporary one. For instance, news stories about animals and pets appear to lose significance very quickly, while COVID-19, climate change, and UK politics retain a significant negative association almost one month subsequent to exposure. This suggests that, during the period of study, news exposure on these topics consistently correlated with lower satisfaction with democracy, to the extent that outlets focused on those issues.Footnote 12
All the individual-level logit models are consistent with the trends observed in the multi-level models regarding changing media exposure during the pandemic and the reduction in the partisan affect divide. As media content became more balanced across sources between divisive and non-divisive topics, individuals who previously were ‘primed’ to be dissatisfied with democratic performance became less so, and we can confirm that this association exists at the individual level after being matched with personal news source content. Moreover, this association is robust to the inclusion of political and demographic controls and shows the estimated effect of individual exposure relative to other individuals with similar beliefs and attitudes who were not similarly exposed in the days prior to the interview. Our study suggests that a relationship between media exposure and satisfaction with democracy exists, and these results hold important consequences for our understanding of democratic functioning in an era of heightened polarization.
Conclusion
In recent years, political scientists have dedicated greater attention to the study of rising media polarization, animosity across partisan divides, and the associated deterioration in civic satisfaction (Boxell, Gentzkow and Shapiro Reference Boxell, Gentzkow and Shapiro2022). However, less attention has been given to broader policy questions that arise from this analysis. For instance, is rising media polarization a secular process or are there periods in which tensions can subside? If so, how do such periods arise and for how long can they endure? And, finally, could reforms to the informational environment (for example, requiring news filter neutrality, extension of legal liability for third-party content, or tagging of contested claims) aid in maintaining civility and balance in the public sphere, or is civic polarization ultimately a result of socioeconomic frustrations that cause citizens to seek polarizing content? These important questions belong to a broader research agenda. Yet to the extent that our study contributes, it is by observing that large shifts in media polarization are possible, can endure from one year to the next, and may be associated with substantial narrowing of partisan divides over democratic evaluation.
By focusing on partisan variation in news exposure, we were able to link such attitudes to the nature of content to which respondents were exposed, and while we refrained from any direct assertion of causality, our results invalidate the expectations of key alternative hypotheses. Meanwhile, our analyses also identify a significant shift in the nature of content exposure, namely, the ‘depoliticization’ of topic selection. During the pandemic, newspapers moved away from covering divisive political issues and instead focused on neutral and non-political story content. This shift in the media landscape led to an increase in overall satisfaction with the democratic political system. At the individual level, these changes can be attributed to variations in content exposure. Specifically, readers of newspapers aligned with the political opposition experienced more pronounced shifts compared to readers of newspapers that typically support the governing party. These findings suggest that the content exposure experienced by individuals might have played a role in shaping their perceptions and satisfaction with the democratic political system.
Our results also touch on a broader finding, which is that, as the UK was shifting to home-working and social distancing in 2020 and 2021, media content, in general, appears to have shifted its ‘topic balance’ from more divisive issues, such as UK domestic politics, towards more neutral areas, such as entertainment stories, personal interest, or sport. As this occurred, the shift in exposure was associated with higher levels of satisfaction with democracy – not only at the individual level but also nationally, as average satisfaction levels increased. This effect has persisted beyond the pandemic, with a majority of UK respondents continuing to evaluate the country’s democratic functioning positively from March to April 2022, according to data collected by the Pew Research Center. Yet the implications of this macro-level shift for democratic life merit debate. If democratic performance depends upon ‘critical citizenship’, then such a change may be interpreted negatively (Norris Reference Norris2004). However, another interpretation would be that of Gabriel Almond and Sidney Verba, who, in their 1963 classic text, The Civic Culture, argued that while the functioning of representative democracy depends upon the presence of a sufficiently large number of individuals ready to participate in partisan contests through election campaigns, pressure groups, and public debate over salient national issues, it also depends upon the presence of two further groups: firstly, a bulwark whose loyalty remains subject to the political system as a whole, rather than any party or faction within it; and, second, a residual number whose focus lies upon improving conditions in their ‘parochial’ circle of agency, such as their household, neighbourhood, or local community (Almond and Verba Reference Almond and Verba1963).
In a similar vein, our findings suggest that while news consumption may play an important role in raising access to political information and overall citizen engagement with politics, in a partisan media landscape, continuous exposure could possibly undermine satisfaction with democracy. During the pandemic, as media content shifted to more ‘parochial’ topics – such as entertainment or ‘human interest’ stories – individual respondents exposed to such content reported higher levels of institutional satisfaction. This, in turn, opens another possible mechanism by which crisis events, such as a global pandemic, can restore citizen confidence in the political system: not simply through the reduction of partisanship in news coverage but also via a reduced focus on politics per se, or a change in the tone of such coverage towards less polarization – and a rebalancing of interest towards other domains of life.
Supplementary Material
The supplementary material for this article can be found at https://doi.org/10.1017/S0007123424000395.
Data Availability Statement
Replication files for this study are available on the Harvard Dataverse at: https://doi.org/10.7910/DVN/GS6N40. The original files from YouGov’s Tracker on Satisfaction with Democracy, part of its Omnibus Survey, and Twitter(X) cannot be shared directly due to copyright. However, access to the former can be requested from YouGov. The relevant contact at YouGov is Joel Rogers de Waal at joel.rogers@yougov.com. Twitter(X) data can be accessed via the Twitter(X) API: https://developer.x.com/en/docs/x-api. Replication code for this article can be found in the Harvard Dataverse at: https://doi.org/10.7910/DVN/HL0NKK.
Acknowledgements
We thank the editor and the anonymous reviewers for their valuable comments that considerably improved this article. We are also grateful to seminar participants at the University of Cambridge Bennett Institute for Public Policy seminar series, the PSPE Work in Progress Seminar at The London School of Economics, the World Association for Public Opinion Research, and Workshop participants at Lancaster University. We would also like to thank Sarah Jewett for her pivotal role in the early stages of this research, and Paul Sullivan and the LSE School of Public Policy for their support. Finally, we would like to thank Joel Rogers de Waal for his patience, dedication and commitment to this project, and Stephan Shakespeare for providing support to the YouGov-Cambridge Centre for Public Opinion Research. The views expressed in this article are solely those of the authors and do not necessarily reflect the opinions of YouGov Plc.
Financial Support
We gratefully acknowledge the support of the UK Cabinet Office to the Cambridge Centre for the Future of Democracy team that led this research, in the context of the Global Democracy Insights For Cabinet Office project. For generous financial support, Xavier Romero-Vidal gratefully acknowledges the support of the Spanish Ministry of Science, Innovation and Universities and the call for grants relating to the Requalification of the Spanish University System in the Universidad Carlos III de Madrid (Royal Decree 289/2021) for a Margarita Salas postdoctoral grant at the Carlos III University Madrid.
Competing Interests
None.
Open access
