Ideological biases in social sharing of online information about climate change

share. The aim is to understand how people engage with, and share online media content about climate change on social media (specifically Twitter). We have found that amongst the communities of shared URLs, right-wing and climate sceptic views are strongly correlated, as are left-wing and environmentalist views. This correla- tion has been observed in individuals before using survey-based methods (e.g. [3, 51, 52]). Our study shows that the association extends to media outlets, specifically to the content produced by a large set of online news providers (Fig 4), and is a major feature of how online content is shared. Recent work by Hornsey et al. [52] suggests that the scepticism-conservatism link may be strongest in the US. This finding is supported here, where we observe two large communi- ties based mainly on US sources; one more left-wing and environmentalist, the other more right-wing and sceptic. Although we did not examine where the users in our dataset were located geographically, the dominance of US sources in two of the communities, and of UK sources in one of the communities, suggests that these communities are largely formed around users from those countries. Our overall focus on English-language tweets and media content, plus large numbers of Twitter users in these countries (nearly 15% of online US adults used Twitter in 2017 [12]), are consistent with this conjecture. In this study we find two levels of community structure. At the highest level, there is polari- sation and segregation with a large left-wing/environmentalist grouping and a smaller right-

wing/sceptic grouping clearly visible in the sharing network after the application of the force- directed layout algorithm. Within the left-wing/environmentalist group, algorithmic commu- nity detection finds (typically) four smaller groupings, characterised by further analysis as above. It is interesting to note that the right-wing/climate sceptic group is more densely con- nected internally than the left-wing/environmentalist group (Student’s t-test applied to average clustering coefficients of all left-wing and right-wing communities in Figs 3 and 7, with clus- tering coefficients calculated in Gephi using Latapy’s algorithm [53], p = 0.019). This finding echoes the observation by Colleoni et al. [54] that in the US, active Republicans exhibit greater levels of homophily than Democrats in their patterns of interaction on Twitter. The temporal analysis presented above gives confidence that the polarised network struc- ture is robust over time, despite turnover in the user populations, sets of shared articles and news sources. We studied similarity between the sets of users, URLs and domains across the seven weeks, as well as the proportion of users, URLs and domains appearing only in the exceptional Week 4 (in which climate change was the subject of a mainstream news event when President Trump announced the US withdrawal from the Paris Agreement). This major event substantially increased the volume of social media messages related to climate change and drew in an increased number of new users, who shared a greater number of articles from a

greater number of online information sources. Thus we observed network structure and com- position in both ‘normal operation’ as well as in an unusual state of high activity, showing the typical level of week-to-week volatility as well as the substantial change under the influence of the disruptive mainstream news event. At all times, including the disruptive event, the polar- ised network structure remains strong and clearly visible, despite substantial changes in its constituent parts (users and articles). As such, we conclude that the broad topological features visible in Figs 3 and 7 are stable over the seven-week duration of our dataset. Similarly, we find that the association of left-wing/environmentalist views, and of right-wing/sceptic views, are persistent, as shown in Fig 5a and 5b, S6 and S7 Figs in S1 File. Future study may find it fruitful to examine whether such stability exists for other online networks or other politicised discussions. The similarity statistics in Table 3 suggest a general trend for a small proportion of the users, URLs and domains to persist across weeks while others appear only sporadically. In the URL case, the comparatively low levels of similarity are not surprising as news articles are PLOS ONE Ideological biases in social sharing of online information about climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0250656 April 23, 2021 20 / 25