Ideological biases in social sharing of online information about climate change

across all coders to find a single domain score. This process was applied for both political and climate change ideology. Since judgment of climate change ideology relies on some knowledge of climate change (e.g. the scientific consensus) the coders were recruited from the postgradu- ate research community in sciences at the University of Exeter. Full results for the coding of each domain are presented in S2 Fig in S1 File. To test the reli- ability of our coding exercise we compare our grades to the domain assessments by Media Bias/Fact Check [49] (MBFC), a fact-checking and media bias site. We translate the MBFC rat- ings to our own scale and apply the same z-score normalisation, then calculated the correlation of the political bias scores by our coders against those from MBFC for those domains which have been rated by MBFC. We find strong positive correlation (Pearson’s r = 0.8, p < 10−6, n = 49) demonstrating that the political coding process is consistent with the expert assess- ments provide by MBFC. No equivalent external assessments are available for climate change ideology, but the robust performance of our coding process for political bias gives confidence in the methodology. See Fig 2a and 2b for the average results across domains and coders for political and climate change bias respectively. Comparisons over time To measure the change in patterns of sharing climate media over time, the pairwise similarity of the sets of users, articles (URLs) and source domains was calculated for the seven weeks in

the study period. An asymmetric similarity measure is defined to give an indication of persis- tence between weeks and identify influxes of new participants or content. This expression compares weeks A and B by the fraction of the users, u, links, l, or domains, d, that appeared in week A that also appeared in week B. In order to account for the repeated usage common in online social networks, a measure of how many times a user, URL or domain appears in each of the two weeks is included by using multisets for u, l and d. Su A;B ¼ juA \ uBj juAj ; Sl A;B ¼ jlA \ lBj jlAj ; Sd A;B ¼ jdA \ dBj jdAj : ð3Þ Values of the similarity measure fall within the range [0, 1]. Values approaching 0 signify that very few of the users (respectively URLs, domains) in week A are also present in week B, whereas values approaching 1 signify that nearly all of the users (respectively URLs, domains) in week A are also present in week B. Note that by design SA,B 6 ¼ SB,A for A 6 ¼ B in general. Results This section divides our results into three main findings. In the first part we focus on Week 4 of the study period, in which the US withdrawal from the Paris Agreement was announced, to demonstrate the broad ideological polarisation observed in the information-sharing networks. Secondly, we characterise the sub-communities that make up the network, showing that sev- eral linked left-wing/environmentalist communities co-exist with a single right-wing/sceptic community. Finally, we look at all seven weeks in the study period to explore how network structure and polarisation changes over time.

Climate media sharing is polarised and politicised We begin by characterising the information-sharing network during the central week of the study period, Week 4, in which the US withdrawal from the Paris Agreement was announced. During this week, 7, 496 URLs were shared by 42, 113 Twitter users. After projection this pro- duces a unipartite network of 7, 496 URLs connected by 107, 304 edges indicating which pairs 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 10 / 25