Ideological biases in social sharing of online information about climate change

number of reasons. Firstly, each URL must be accessible to allow subsequent analysis of con- tent. Secondly, typographical errors by tweet authors are sometimes incorrectly interpreted by the Twitter API as URLs, such as periods followed by alphanumeric characters (hello.world). Thirdly, the validation process handles shortened URLs, which represent the majority of URLs in tweets. Many Twitter users make use of third-party URL shorteners, services which create a redirect to a given long URL from a shorter version that can then be used to reduce the charac- ters needed to embed URL in a tweet (this is unnecessary since Twitter includes its own short- ening service that reduces any URL to 23 characters). URL shorteners are also sometimes used to conceal the target URL from anyone who may potentially click on it. These short links must be traced to their destination both to ensure they point ultimately to a valid resource, and also to be treated alongside other links to the same destination. Each URL is resolved individually, marking as valid only those that return HTTP status code 2XX immediately or through a small number of permanent redirects (301 ! 301 ! . . . ! 2XX status codes). The 2XX status codes indicate various successful outcomes to an HTTP request. Many valid URLs included modifiers inserted for tracking and metadata purposes, without altering the destination page. This can lead to many URLs pointing to the same resource, so to minimise the possibility of considering these as distinct URLs, all schemes and queries are

removed from the resolved URLs. A small number of domains were adversely affected by this. YouTube and other Google services were the most prominent, largely due to the structure of their links. As a result, our process does not distinguish between different YouTube videos within the networks. The validation process results in a mapping between each raw URL (as given in a tweet) and the validated final destination after any acceptable redirect paths. Any tweet containing at least one unvalidated URL is removed from the dataset. Finally, all users who tweeted more than 50 times within a given week are removed; this threshold is found to be a sensible limit to mitigate the impact of automated accounts, especially news aggregator accounts. Across the seven weeks 271 unique accounts were affected, leading to the removal of 66, 892 tweets (approximately 1% of the total dataset). This left 245, 446 tweets by 113, 154 users sharing 54, 462 distinct, validated, URLs across all of the seven one-week windows. Network construction Bipartite networks, containing two node classes representing users and URLs, are constructed by creating an edge i ! j whenever user i shares URL j, illustrated by Fig 1. Multiple shares of URL j by user i increment the edge weight so that the final bipartite edge weight w(i, j) repre- sents the number of times user i shared URL j. Next a unipartite projection is produced from the bipartite network, whereby a network of only URL nodes is created, where edges encode the number of users that shared the connected

pair of URLs. The network construction process is illustrated in Fig 1. The URLs shared by a single user (Fig 1a) form a fully connected clique of nodes representing the sharing pattern for that user (Fig 1b). The whole unipartite network is then constructed by composing the cliques for all users (Fig 1c). The projection allows the use of efficient unipartite network analysis algo- rithms and focuses the analysis on the relationships between URLs. The unipartite network of URLs represents the collective pattern of sharing online content of the targeted Twitter user population. Statistics for the seven one-week networks created are given in S1 Table in S1 File. We restrict our analyses to the giant component to avoid the issue of network fragmenta- tion and limit its impact on community detection. In all cases the giant component consists of around two thirds of all URL nodes in the network. One important design decision made in 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 6 / 25