American media, Scandinavian audiences: Contextual fragmentation and polarisation among Swedes and Norwegians engaging with American politics

The rise of digital media has sparked debates about media fragmentation and its implications for society. Media fragmentation is driven by the proliferation and diversification of media channels, platforms, and content in the digital age. The high-choice media environment enables audiences to select content that aligns with their preferences and interests, which may lead to increasingly fragmented and tailored media consumption patterns (Prior, 2005). From this perspective, fragmentation results from the growing abundance of media options and the ability of users to actively choose and curate their media experiences (Napoli, 2011; Webster, 2005). Fragmentation is a concern because it can lead to audience segregation, echo chambers, and an erosion of shared cultural experiences, potentially undermining the democratic exchange of ideas and the formation of a well-informed citizenry (Pariser, 2011; Sunstein, 2007).

Prior (2005) investigated the emergence of a high-choice media environment due to technological advancements, namely cable and satellite television and the Internet, focusing on news and entertainment content. Using data from the American National Election Studies as well as Nielsen Media Research, Prior found that high-choice media environments led to a widening gap in political knowledge and participation between politically interested and disinterested individuals. Prior argued that the proliferation of media and the ability to make active choices allow individuals with a strong interest in politics to consume more news content, while those with a low interest in politics are more likely to choose entertainment, which, he found, led to lower voter turnout.

Other scholars have argued that while audiences might be able to make active choices, they still converge around many of the same media channels, outlets, and events (Hindman, 2009; Webster & Ksiazek, 2012). Mutz and Young (2011: 1028) summarised the phenomenon this way:

In an extremely high-choice environment, with hundreds of program options as is now common, it is improbable to suggest that a viewer pick up the remote control and channel-surf among all possible programs in order to decide which program to watch at a given sitting. There are simply too many options to make this approach feasible.

Cultural trends, recommendations from friends, and social environments still prevail, particularly when confronted with overwhelming choice. In other words, media proliferation does not result in proportionally fragmented media consumption (Webster, 2005).

Webster and colleagues have proposed that media use should be understood through the concept of audience duplication (Webster, 2005; Webster et al., 2006; Webster & Ksiazek, 2012). Duplication occurs when different audience segments consume and engage with the same media content or sources. This distinction has been visualised in Figure 1, showing that the number of media outlets available does not necessarily equal fragmentation. The concept of duplication suggests that common media experiences can still exist within high-choice environments (Webster & Ksiazek, 2012). Thus, Prior’s entertainment-preferring viewers are not entirely cut off from news. In this view, digital media platforms may facilitate the emergence of new shared media experiences and cross-cutting exposure to diverse content, fostering a degree of overlap in audience consumption patterns.

FIGURE 1

Drawing on Webster and Ksiazek (2012), Fletcher and Nielsen (2017) further investigated audience fragmentation and duplication in seven countries, using data from the Reuters Institute Digital News Report. Their findings challenge the universality of Prior’s earlier finding that news audiences are becoming increasingly fragmented due to media proliferation. Instead, Fletcher and Nielsen found that media fragmentation varies significantly across countries and platforms, and that audience duplication is more common than previously thought.

Among the findings relevant to this article are Fletcher and Nielsen’s observation of higher duplication in the US than in Denmark. This is unexpected given the common assumption that the American media environment is more fragmented due to the immense proliferation of outlets. Fletcher and Nielsen found that despite the abundance of available media, Americans still consume news from a core set of popular sources. In contrast, Denmark has fewer options and less duplication. This finding is not conclusive, though. Steppat and colleagues (2022), for example, found that Denmark is much less fragmented than the US and other European countries.

Other comparative studies have found that different countries have experienced the proliferation of media in different ways (Gaol et al., 2020). In their study of news consumption gaps in Europe, Aalberg and colleagues (2013) found significant variation across European countries – even in the Nordic region, where Norway’s news consumption has declined and Denmark’s is steady. These studies highlight that fragmentation is influenced by various factors such as national media systems, the strength of public service media, and political and cultural differences. However, Fletcher and Nielsen (2017: 478) have pointed out that the underlying factors are not entirely clear: “our findings [… underline] the need for further comparative research to develop our understanding of the interplay between structural differences in media systems and audience difference in media use in different countries [emphasis added]”.

A concept closely aligned with media fragmentation is the concept of political polarisation. This issue has received particular attention in the US, where some scholars have tied partisan media outlets, particularly on the right, to political division in the country (Benkler et al., 2018; Jurkowitz et al., 2022; Mutz, 2006; Mitchell et al., 2014; Prior, 2013). Polarisation has also become an increasing concern in Brazil, India, and many European democracies, where far-right and nativist–populist parties have been gaining ground among voters, leading to concerns that audiences are moving into different media spheres.

In their study “Red Media, Blue Media”, Iyengar and Hahn (2009: 20) drew a connection between media proliferation and audience polarisation in the digital age:

Given this dramatic increase in the number of available news outlets, it is not surprising that media choices increasingly reflect partisan considerations. People who feel strongly about the correctness of their cause or policy preferences seek out information they believe is consistent.

The scholars found clear distinctions in the preferred news brands of Democrats (CNN, NPR) and Republicans (Fox News) – even on coverage of apolitical topics – which they attributed to ideological differences (“cause or policy preferences” in the quote above). In effect, this is a form of selective exposure that creates preferences for not just individual news items but entire media brands (Arendt et al., 2019).

The concern over fragmentation of audiences along political lines has been further driven by the affordances of online networks. Particularly influential in these discussions are the concepts of “echo chambers” (Sunstein, 2007) and “filter bubbles” (Pariser, 2011), which theorise that digital platforms – and in the case of filter bubbles, the algorithms that platforms use – allow or encourage people to interact with similarly minded people and information that aligns with their existing views. These concepts suggest that political divisions may be exasperated by digital media.

Yet empirical investigations into the phenomena of filter bubbles and echo chambers often reveal that people’s digital practices are messier than these theories imply (Bruns, 2019; Dahlgren, 2021). A literature review by Ross Arguedas and colleagues (2022: 4) found that even in the US, “people have relatively diverse media diets”, and those who get all their news from partisan sources were a small minority. Studies have found that online news consumption exposes individuals to diverse viewpoints through incidental exposure (Bakshy et al., 2015; Flaxman et al., 2016), that consumers may have more varied diets than their declared brand preferences imply (Weeks et al., 2016), and that online environments may be less ideologically segregated than people’s offline environments (Gentzgow & Shapiro, 2011). In a study on Twitter users in Norway, Enjolras and Salway (2022) found that while users tend to retweet content from those who share their political orientation, political networks overall did not exhibit high levels of ideological homophily.

Moreover, preferences may not be cultivated through careful evaluation of a media brand’s political position. Mutz and Young (2011), reflecting on Iyengar and Hahn’s “Red Media, Blue Media” study, interpreted the results as people not so much actively looking for media that aligns with their policy preferences, but as people sticking to habits learned in their social environments, much closer to the social identity frameworks that Iyengar and colleagues (2012) argued are at work in affective polarisation (see also Arendt et al., 2019). Perhaps one of the most interesting findings for this study is that selective exposure may be highly context-specific. In a cross-national experiment, Steppat and colleagues (2022) found that participants from countries with low media fragmentation and polarisation exhibited low tendencies toward selective exposure in their own environment, but they were more likely than others to seek out ideologically congruent sources when placed in given the option of many partisan sources.

Yet the existing literature on audience polarisation presents certain limitations and challenges. First, much of the research has been conducted in an American context, which is a two-party system and may more easily lend itself to a divide between two poles. This context is different from most of the parliamentary systems in Europe, including the Scandinavian countries.

Second, the literature presents certain ontological challenges; as noted by Lelkes (2016), the concept of polarisation is not deployed consistently. It has been used to describe a tendency toward homophily (Enjolras & Salway, 2022; Webster, 2005), as a synonym for partisanship (Prior, 2013), and to mean the process by which groups become more extreme in their beliefs (Gentzkow, 2016; Isenberg, 1986). At times, it is not clear whether it is a dependent or independent variable (e.g., are echo chambers a synonym for polarisation, or a cause of polarisation?), which further complicates interpretation of the body of research.

Due to the ongoing confusion around polarisation as a concept, I would like to distinguish in this article between three different forms: ideological, affective, and audience polarisation. I understand ideological polarisation

Ideological polarisation as defined in this article is similar to what Lelkes (2016) calls “ideological divergence”. In his conception, ideological polarisation can also take the form of “ideological consistency”, in which people’s various ideological standpoints clearly align with one side or the other.

Finally, despite the concern about the role of digital media in contributing to fragmentation and polarisation, relatively little research has considered one of the most common media practices in digital media: sharing links. I discuss link sharing as a source of data for measuring fragmentation and polarisation further in the next section.

With the growth of social media platforms, the concept of media “engagement” has emerged as a framework for understanding audience behaviour. Previously, audience studies primarily focused on media consumption – generally measured in terms of minutes of exposure – an approach that has faced criticism for neglecting the more active, participatory aspects of media use (Livingstone & Markham, 2008). In contrast, according to Ha and colleagues (2018), news engagement goes beyond consumption and embraces the wide range of participatory activities, including posting content, commenting, liking, and retweeting (see also Hermida et al., 2012; van Dijck & Poell, 2013), activities that reflect the blurred boundaries between media production and consumption (Bruns, 2018).

News sharing, a key aspect of media engagement, is a prime way that news media is distributed in networks (Bakshy et al., 2015; Bruns & Stieglitz, 2012; Grinberg et al., 2019; Hedman et al., 2018; Rauchfleisch et al., 2020). Trilling and colleagues (2017) found that the logics of “shareworthiness” are largely similar to the journalistic logics that dictate “newsworthiness”. While studies find that mainstream media are often the sources being shared, Bruns (2018: 140–141) argued that the act of sharing – and often, commenting – has lowered the threshold to political discussions, as link sharing is “not the domain of news and political junkies only, but a fundamental everyday activity on social media”.

On Twitter, links were shared in various ways, including by pasting a link into a tweet, retweeting a link shared by another user, or by tweeting directly “via” a news website. These functions may have different motivations. Studies in the US and Norway have found that users often retweeted content that agrees with their views, while tweets with original commentary may be more critical of the link (Conover et al., 2011; Enjolras & Salway, 2022), or what Bruns (2018) calls gatewatching. Yet all of these forms extend the reach of the media outlet and help “curate” the personalised news flows of other users (Thorson & Wells, 2016). “Exposure to any given message”, write Thorson and Wells (2016: 313), “depends on a person’s position within the multiplicity of intertwined message flows”. In this way, users help each other sort through the “abundance” of high-choice environments, and curate a “collection of manageable size” (Thorson & Wells, 2016: 313).

However, we must consider that news sharing is fundamentally different from consumption, and in some ways may involve less exposure – namely, that users often share news articles without reading them (Ward et al., 2022). In a similar vein, researchers have found that that link sharing is shaped by both a desire to share information and influence others, as well as a desire to conform to the expectations of others (Bhagat & Kim, 2022), making it a highly social and socialised act (Wall, 2015). More generous theories of news sharing have framed it as an affective act and a way of showing solidarity (Papacharissi, 2015), as well as a performative act, or a way of signalling an identity to other users in the network (Bruns, 2018).

Thus, because link sharing occurs in socio-technological spaces, it is important to keep in mind that link sharing may have different underlying barriers and motivations compared with previous measures of news consumption. Despite these limitations, link-sharing data can be a powerful tool for tracking media flows on digital platforms – in part because this data is able to capture the universe of options now available to media consumers.

As previously discussed, the emergence of digital media has presented new questions about how people find and consume information, opening up possibilities for global information sharing that crosses the traditional boundaries of national media systems (Castells, 2015; Rauchfleisch et al., 2020). Online platforms facilitate not only the exchange of information, but potentially also political identities and affect (Appadurai, 1996; Papacharissi, 2015). The degree to which digital networks potentially exacerbate issues of fragmentation and polarisation is unclear, but it is evident that the impact of political communication crossing borders should not be overlooked in understanding the complexities of media audiences.

Traditional cross-national comparative research has been one way of exploring differences in media systems and their effects on public opinion and behaviour (Urman, 2020). However, this approach may be insufficient to capture the fluidity of the global media environment. To address this, I argue for a new kind of comparative research that focuses on the same population of people in different media environments – what I refer to as intranational comparative research. This approach has been used effectively in other contexts, such as research on diaspora populations (Tripodi & Potocky-Tripodi, 2006), but has yet to be fully explored in media research. By comparing audiences in different media environments, researchers can better understand how different media systems and platforms shape attitudes, and to what extent these differences are shaped by what Fletcher and Nielsen (2017) called structural or audience differences. This approach also allows researchers to overcome some of the limitations of traditional cross-national comparisons, such as differences in methodology, sampling, and context.

In line with this approach, in this article I use a global event, the 2020 American presidential race, as a window into how the same population of people in different media environments perceive and engage with political discourse. American elections are notable for their global impact and extensive coverage worldwide (Boyon, 2020; Vliegenthart et al., 2010), making the election an ideal subject for intranational comparative research. This approach builds on a previous study of mine where I, together with Gunn Enli (2022), mapped the “deterritorialisation” of elections in the global information space of Twitter, highlighting the need for closure research on audiences and users themselves. By examining the same population of people in different media environments, it is possible to explore how the global flow of information and media shapes attitudes towards political candidates, issues, and events.

Swedish and Norwegian users were identified through a combination of language (Swedish and Norwegian) and geoparsing. Geoparsing involves natural language matching of place names and has been used in previous studies on the geography of Twitter (Bruns et al., 2014; Bruns & Enli, 2018; Hänska & Bauchowitz, 2019; Leetaru et al., 2013; Rauchfleisch et al., 2020; Robinson, 2022; Sloan & Morgan, 2015). In this case, the self-reported location information in user profiles was matched against a list of 4,059 Swedish and Norwegian place names, including country names, regions, historic regions, provinces, cities, and the neighbourhoods and suburbs of major cities, compiled from tables on Wikipedia, Airbnb.com, and national census bureaus (SSB and SCB). This is drawn from Bruns and colleagues’ (2014) approach for determining “Australianness” in mapping the Australian Twittersphere. Where applicable, both the English- and Scandinavian-language versions were included, for example, Göteborg and Gothenburg.

For Norway, national, regional, and county-level names in nynorsk [New Norwegian] were also included.

Users identified through geoparsing were then selected if they tweeted at least once in Norwegian or Swedish, or if they had a Norwegian or Swedish vowel their account profiles. After selecting the tweets that included a URL, this resulted in four datasets grouped by country and language: Swedish users in Swedish (N = 3,093); Swedish users in English (N = 2,682); Norwegian users in Norwegian (N = 1,216); and Norwegian users in English (N = 1,397). These sets are later denoted as Swe-Swe, Swe-Eng, Nor-Nor, and Nor-Eng.

About a third of users had tweets in both their national language and in English (30.7% of users in Norway and 31.1% in Sweden), meaning that these users appear in both datasets for their country. One option would have been to study only these users. However, it is likely that tweets in multiple languages appear from these users because they tweeted more; dual-language users are responsible for 69.5 per cent of the tweets by Norwegian users and 70.1 per cent of the tweets from Swedish users. Thus, selecting only dual-language users may inadvertently select for Twitter power-users.

Similar disproportionate usage rates have been documented in other Twitter analyses (Wojcik & Hughes, 2019).

Media outlets were identified through the URLs embedded in the collected tweets. To identify the base domains of the media outlets (e.g., cnn.com or nrk. no), Regular Expression (Regex) commands were applied. Similar URLs (e.g., video.cnn.com with cnn.com) were identified computationally through text matching; these were then consolidated, ensuring that different subdomains of the same outlet were treated as a single entity. Before the subsequent analysis, URL shorteners (e.g., bitly) and platforms such as Twitter, Instagram, and YouTube were removed, as these serve as conduits to other content of unknown origin.

Because this data includes any potential URL that users linked to, the potential links are as limitless as the Internet. This results in a very long tail of URLs, with many shared by only one user. Webster and Ksiazek (2012: 48), who used browser-tracking data, addressed this issue by applying a threshold: Websites had to be visited by 3 per cent or more of users. Here, because sharing creates a slightly higher barrier to clear, I used a 1 per cent threshold for each country–language group in order to maintain the potential diversity of sources.

Fletcher and Nielsen (2017), who used survey data, addressed this issue by using only the top-14 media outlets in all countries they studied. However, this could artificially abridge the American media environment especially, since it is characterised by many more media outlets, and allegedly, more fragmentation – and in fact Fletcher and Nielsen (2017: 485) pointed this out as a potential limitation of their study.

Media fragmentation was measured by examining the diversity and concentration of information sources shared by Scandinavian users. Users’ URL shares were converted into media-to-media networks based on overlapping sharing. For example, let’s say User 1 shares a link to Outlet A in one tweet and a link to Outlet B in another tweet; in the media-to-media network, this creates a tie between Outlet A and Outlet B. The data was unweighted on a user level, meaning that subsequent tweets by User 1 that linked to Outlet A or B would not increase the strength of the tie between the two outlets. This decision avoids the above-mentioned issue of power-users dominating the data. Moreover, it ensures methodological consistency with previous studies, in which calculations (e.g., “reach”, see below) are based on distribution of use across audiences rather than how individuals distribute their media use (Webster & Ksiazek, 2012; Fletcher & Nielsen, 2017). Even so, this choice potentially impacts the results, as discussed further on.

One of the challenges of measuring audience fragmentation is that any single user can establish a tie between two outlets, effectively creating a low barrier to duplication. To address this, only ties that are statistically significant are included. This was calculated using the method described by Webster and Ksiazek (2012), in which the observed ties between each pair of outlets is compared with the number of ties that would be expected by chance based on the reach of each outlet. In this case, “reach” refers to the number of users who shared a link to the outlet out of all users who shared links. In the previous example, if Outlet A has a reach of 2 per cent and Outlet B has a reach of 5 per cent, then we could expect that 0.1 per cent of users (0.02 × 0.05) would have shared links to both by chance. This is the expected value. Ties between outlets whose observed ties were below the expected values were removed.

Drawing on Fletcher and Nielsen (2017), several network statistics were calculated on the media-to-media networks. Network density, which compares the number of potential ties to existing ties in the network (Himelboim, 2017), was used to measure the level of fragmentation or duplication. Lower density scores indicate higher levels of media fragmentation, while higher scores suggest higher duplication. This is complemented by centralisation, a measure of the degree to which the network is organised around certain hubs. Generally, we expect networks that have high density (duplication) to have low centralisation. I also calculated diameter and transitivity. Diameter is a measure of the shortest path between the two most distant (in terms of connections) outlets in the network. Transitivity (here measured by the clustering coefficient) is the degree to which outlets in the network are clustered together, meaning that neighbours will share connections, as measured by the number of closed triangles in the network (Fletcher & Nielsen, 2017). Network centralisation was calculated in igraph, an R library for network analysis; the other statistics were calculated in the network visualisation program Gephi.

To measure audience polarisation, defined here as alignment between media use and political orientation, community clusters were identified in the media-to-media networks using Gephi’s modularity class algorithm, which identifies divisions in a network (Blondel et al., 2008). The algorithm iteratively sorts nodes – in this case, media outlets – into communities where connections are stronger than would be expected by chance. The output is a set of subnetworks – “modularity classes” – effectively grouping media outlets based on the level of duplication they share with other outlets, and thus categorising them according to their shared audience (Guerra et al., 2013).

Second, a random sample of 1,200 Scandinavian users was classified by political orientation. Researchers have used various methods to determine social media users’ political affiliation, such as follower–followee networks (Himelboim et al., 2013; Narayanan et al., 2018); users’ stated political ideologies in their bios (Himelboim et al., 2016); hashtag usage (Bovet et al., 2018; Conover et al., 2011); linguistic patterns found through machine learning (Clark et al., 2018; Colleoni et al., 2014); and retweet networks (Wong et al., 2016). Many of these methods involve a certain level of national similarity (e.g., that American users will largely use the same hashtags or share linguistic patterns), or even ideological homophily (e.g., that people will follow people they agree with) or political passion (e.g., that they will indicate their politics in their bio).

Because the data used here involves different national milieux, these methods were deemed inapplicable. Instead, I employed manual analysis; manual annotation of political orientation is a standard part of machine-learning methods, and though non-scalable, is considered reliable. Following Colleoni and colleagues (2014), I classified users based on tweet content into the following categories: pro-Biden, anti-Trump, pro-Trump, anti-Biden, neither, or unknown/informational. A reliability assessment with a second coder showed an acceptable kappa value of .864 (p < .005). The pro-Biden and anti-Trump users are later grouped as Biden-leaning; pro-Trump and anti-Biden are grouped as Trump-leaning.

Chi-square tests were then used to test the correlation between the identified modularity classes and the political orientation of users who shared content from outlets within these communities. A significant correlation suggests that users in the network tended to share the same content as other users with the same political views, and thus audience polarisation. Cramer’s V is used as a measure of the degree of audience polarisation.

Table 1 shows the results of measurements taken on the media-to-media networks. In summary, media duplication is relatively high in all four networks, with densities ranging from 0.67 to 0.88 (cf. Fletcher & Nielsen, 2017: 488). Transitivity is also high across all networks, indicating a high level of interconnectedness. The diameter is consistently low at 2 for each network, suggesting short paths between outlets, while centralisation varies across the networks, with the Norwegian-language network (Nor-Nor) having the highest, and Sweden’s English-language network (Swe-Eng) having the lowest centralisation. Modularity, a measure of community structure, is extremely low in the Scandinavian languages, and higher in English, a point I return to in the next section on polarisation. Overall, these findings suggest that there is a considerable amount of overlap in the media outlets used in these networks, supporting the media duplication argument.

Comparing the Scandinavian-language tweets (Nor-Nor and Swe-Swe) with the English-language tweets (Nor-Eng and Swe-Eng), we can observe that the national languages have slightly lower duplication (0.67 and 0.86) compared with their respective English-language tweets (0.81 and 0.88), an indication that there is somewhat higher fragmentation in the Scandinavian-language networks. Although, interestingly, the Norwegian-language network has the least media duplication of the four networks (67%) and is also more centralised. It is also larger than Sweden’s in terms of outlets, despite being a smaller country. This suggests a few dominant outlets, followed by a longer tail in Norway, reflecting a few major players in an otherwise diffuse media market – possibly a reflection of Norway’s support of local newspapers that serve the country’s dispersed population (Syvertsen et al., 2014: 53–56).

While transitivity is high for all networks, the Scandinavian-language networks have slightly lower transitivity values (0.81 and 0.90) compared with the English-language networks (0.89 and 0.92). This implies that the Scandinavian-language networks have somewhat less interconnectedness among their media outlets. The Scandinavian-language networks also exhibit higher centralisation (0.31 and 0.16) than the English-language networks (0.19 and 0.12), which suggests that the Scandinavian-language networks have a more concentrated distribution of influence or power within their networks compared with the English-language networks.

This is further reflected in the network visualisations in Figures 2a–d. Following previous studies (Fletcher & Nielsen, 2017; Narayanan et al., 2018; Webster & Ksiazek, 2012), these graphs have been created using the Fruchterman Reingold algorithm, which places the outlets roughly equidistant to each other. The outlets have been sized according to the number of connections they have with other outlets (degree). In the Norwegian- and Swedish-language networks, we see highly read tabloid newspapers at the centre (VG and Expressen), and relatively centrally located public broadcasters (NRK and SVT) and daily broadsheets (Aftenposten and Aftonbladet), while the English-language networks are tightly clustered, but have no clear, centralised source.

FIGURE 2A

FIGURE 2B

FIGURE 2C

FIGURE 2D

These visualisations illustrate the overall findings: The Scandinavian-language networks (Nor-Nor and Swe-Swe) are characterised by smaller network size, lower duplication densities, lower modularity, slightly lower transitivity, and higher centralisation compared with the English-language networks (Nor-Eng and Swe-Eng). These differences indicate that the English-language networks have a greater variety of media outlets and a more distinct community structure, while the Scandinavian-language networks have a more concentrated distribution of influence or power within their networks. More to the point, there is more media fragmentation in the Scandinavian-language networks than in English.

Polarisation is operationalised as the strength of association between clusters of media outlets and the political orientation of the users who shared links to those outlets. As we saw in Table 1, the modularity values for the Scandinavian-language networks are lower (0.10 and 0.05) than those for the English-language networks (0.24 and 0.23), indicating that the Scandinavian-language networks have a less distinct community structure compared with the English-language networks.

After running the modularity class algorithm, a chi-square test of independence was performed to determine the relationship between the different modularity classes and user political orientation. The results of these tests for each country–language network is shown in Figures 4a–d, along with graphical representations of the bivariate relationship. The associations are statistically significant (p < 0.05) in each case, though the Norwegian-language results are only slightly below this threshold (p = 0.033).

The tests suggest a certain amount of audience polarisation in each network, but with greater polarisation in the English-language networks. There are also interesting differences between the two countries. The Norwegian-language polarisation is the lowest (.378), but in English, Norwegians have the highest level of polarisation of the four networks (.7). The Swedish users in their national language are already relatively polarised (.506), and more so in English (.672). Examination of standardised residuals indicate that it is the Trump-leaners in the national languages that mainly have distinct media tastes, while in the English-language networks, both the Biden- and Trump-leaners stick to certain clusters of media. Interestingly, we also see in Figure 4b that the Swedish-language tweets are dominated by Trump-leaners, though this is not the case in English. It may be a sign of language preference among Trump-supporting Swedes, or an indication that Trump supporters shared more links to Swedish-language content than Biden’s Swedish supporters.

To further understand the polarisation in the networks, I revisit the media-to-media networks, but this time use a network visualisation algorithm, ForceAtlas 2, that spaces nodes according to shared relationships between nodes (Jacomy et al., 2014). These visualisations are shown in Figures 3a–d. This makes it easier to see the social distances between outlets, as constructed by their audiences. In these network images, we can clearly see two groupings in the English-language networks. While there are still connections bridging the two sides, it appears that Scandinavian users critical of Trump and/or favourable to Biden tend to share links to outlets like the Washington Post and The Hill, while users critical of Biden and/or favourable to Trump are more likely to share links to Fox News and Russia Today. This is similar to findings about Americans’ own preferences for media, as found by Iyengar and Hahn (2009), the Pew Research Center (Jurkowitz et al., 2022; Mitchell et al., 2014), and the Reuters Institute (Newman et al., 2017). Meanwhile, in the Swedish-language network, right-wing alternative outlets like Samtiden and Fria Tider are in the same cluster with the public broadcaster and other mainstream outlets. In Norway, we do see that the populist Resett, Document.no, and Lykten are in their own cluster, but the people sharing links to these media are no more likely to be Trump-leaning. I discuss these findings further in the next section.

FIGURE 3A

FIGURE 3B

FIGURE 3C

FIGURE 3D

FIGURE 4A

FIGURE 4B

FIGURE 4C

FIGURE 4D