Content analysis of tweets produced by alternative media sites in the UK: data

This dataset is based on a content analysis of 14,807 Tweets concentrated on the main Twitter accounts run by nine alternative media sites between 2015 and 2018. Our sample was drawn from four periods: 6–25 October 2015; 9–29 October 2016; 30 April–7 June 2017 (the UK general election); and 8–28 October 2018. 

We chose 2015–2018 as this would provide some insight into the patterns of Twitter use and behaviour either side of a general election (in June 2017). 

Tweets were collected using Twitter’s Full Archive Search AP, which we accessed using Twurl to collect JSON files, and subsequently converted into Excel files ready for manual coding. Our sample, therefore, represents the “full” content from each account, excluding deleted content, but including all Tweet types. In total, there were 9,284 standard Tweets, 634 quote Tweets, 1443 reply Tweets, and 3446 Retweets.  Besides quantifying Tweets, we coded each type and share metrics (as of August 2019). we examined the purpose of tweets in the following of ways: 

• to share content e.g., links to articles, videos, images produced by the outlet who is Tweeting;

• to share content from other media publications;

• to share opinion, conjecture, speculation, viewpoints, hypothesis, predictions;

• to share information e.g., a fact, figure, report, announcement, event;

• to share hominem, dismissive, inflammatory, sarcastic, insulting content aimed at others;

• Other purposes, including promoting individuals or organisations, appeals for subscribers, running polls, etc.

In practice, coding was straightforward, since the limit of characters naturally restricts Tweets from performing many functions simultaneously. Accordingly, there was no double coding, and where there was a decision to make, the more dominant Tweet function was chosen. If an opinion, for example, was in any way inflammatory and specifically directed, then this was coded as “attack”, rather than the sharing of less contentious and less targeted opinions. Most often, Tweets simply share online content, and this was straightforward to code.

Political reference and sentiment. Where a Tweet referred to a UK political party, politician, representative or general references to the “left”, “right” or “the government”. We determined “positive” as anything supportive of a party or associated ideology, including the validity of its policies or the behaviours of those representing it. We coded “Negative” for anything interpreted as critical, such as a policy failing or suggestions of poor practice, or corruption. Whenever there was no evaluative judgement, we coded “neutral”. Manual coding enabled us to capture nuanced versions of these categories, including sarcasm or more oblique references that nonetheless could clearly be assigned as either “positive” or “negative”.

Media Reference and sentiment. Where a Tweet referred to the BBC, a UK media outlet or journalist, the “mainstream media”, or other alternative media outlets. As before, we coded “positive” for anything supportive of legacy media, such as the quality of their journalism and whether they were performing well, for example. “Negative” was anything interpreted as critical of legacy media bands, perhaps pointing to “biased” coverage or no coverage of a particular issue at all. As before, “neutral” was coded in the absence of any evaluation.

The data was analysed by three coders and intercoder reliability tests were performed on 1,485 Tweets (10% of the sample). Levels of agreement for all variables were between 89.7% and 94.3%, and the more intuitive Krippendorf Alpha scores ranged between 0.83 and 0.91, indicating a robust and repeatable framework and a reliable coding process.