Skip to content

Influence of fake news in Twitter during the 2016 US presidential election

Authors: Alexandre Bovet, Hernán A. Makse Venue: Nature Communications, Volume 10, Article 1234 (2019) — DOI

TL;DR

Using 171 million tweets from the five months preceding the 2016 U.S. election, this study found that 25% of news-sharing tweets contained fake or extremely biased content. Network analysis revealed distinct clusters of fake, biased, and mainstream news spreaders. Causal modeling showed that Trump supporters' activity drives fake news dynamics, while Clinton supporters primarily engage with traditional and left-leaning outlets—a reversal of the typical influence pattern for mainstream news.

Contributions

  • Large-scale quantification of fake and extremely biased news spread on Twitter during the 2016 election: 30.7 million tweets with URLs to news outlets, 2.2 billion individual users, 25% reaching fake/extremely biased sources.
  • Classification of news outlets into six categories (fake news, extremely biased right, right, center, left, extremely biased left) based on manual expert review and automated fact-checking databases.
  • Network analysis identifying separate clusters for each media category with distinct connectivity patterns: fake and extremely biased networks are denser with more heterogeneous degree distributions than center/left leaning outlets.
  • Causal network reconstruction using multivariate STL (Seasonal-Trend decomposition using Loess) and Granger causality to uncover directional influence relationships between media categories and voter behavior.
  • Demonstration that Trump supporters drive fake news dynamics while Clinton supporters' activity influences center/left outlets, contrary to typical patterns where traditional news outlets dominate influence.

Method

Data collection: Continuously harvested tweets from June 1 to November 8, 2016, using Twitter's streaming API. Filtered to 171 million tweets from 11 million users by using hashtag/Boolean operators targeting the two presidential candidates (Trump and Clinton). Extracted URLs from tweets to identify links to news outlets. Classified URLs into media categories using OpenSources.co database and manual expert review by a Merrimack College research team.

News outlet classification: Categorized 276 news outlets into six types: fake news (entirely fabricated information); conspiracy theory (well-known conspiracy promoters); extremely biased (strong advocacy/propaganda); right, center, left leaning; mainstream fact-based outlets. Classification based on Media Bias/Fact Check ratings and OpenSources.co.

Network analysis: Built retweet networks showing the top 100 news spreaders for each media category. Computed network properties: in-degree, out-degree, average clustering coefficient, and heterogeneity. Analyzed the spread patterns and influence structure among spreaders using collective influence algorithm.

Causal analysis: Applied STL filtering to remove seasonal and trend components from time series of tweeting activity for each news category. Used Granger causality testing with residuals to identify causal links between top news spreaders' activity and supporters' engagement with different media categories. Performed iterative conditional independence testing to distinguish direct from indirect causal effects.

Results

Prevalence and distribution: 25% of tweets linking to news outlets spread either fake or extremely biased news. Among 30.7 million tweets with URLs to news outlets: 2.99 million (9.7%) to fake news, 3.97 million (12.9%) to extremely biased right, 4.03 million (13.1%) to right leaning, 6.32 million (20.6%) to center, 7.49 million (24.4%) to left leaning, 6.0 million (19.5%) to extremely biased left.

Network structure: Fake and extremely biased right news retweet networks showed higher clustering and denser connectivity than center/left networks. The average degree for fake news spreaders (k̄ ≈ 6.5) was comparable to extremely biased networks but with more heterogeneous distribution. Right leaning top spreaders had larger out-degree than center/left spreaders, consistent with automated posting via bots.

Top spreaders: Fake news spreaders were primarily journalists with verified Twitter accounts and many unverified accounts; some accounts (e.g., @PrisonPlanet) showed signs of bot activity. Extremely biased right spreaders included major conservative outlets and personalities. Center/left networks featured mainstream news organizations. Trump supporters dominated top fake news spreader networks while Clinton supporters dominated center/left networks.

Causal dynamics: Causal network reconstruction revealed asymmetric patterns: Trump supporters' activity causally influenced fake news dynamics, while Clinton supporters' activity influenced center/left outlets. The activity of top fake news spreaders had no significant causal effect on Trump supporters' engagement—suggesting that top spreaders follow rather than drive the dynamics. Center and left leaning news spreaders showed strongest causation from supporters' activity, indicating they are the primary information source for that demographic.

Connections

Notes

Strengths: Large representative sample covering the full 2016 election period; multiple definitions and independent classification methods for media outlets; novel application of causal inference to untangle influence direction between media categories and voters; detailed network structure analysis revealing qualitative differences between media types; replication-ready with code availability.

Limitations: Limited to Twitter—findings may not generalize to Facebook, cable TV, or other platforms which dominated news consumption. Causal analysis limited to top 100 spreaders for each category; smaller influence networks not examined. Exposure measured as potential (based on who spread URLs), not actual viewing. Cannot distinguish whether causal effects operate through direct influence or reflect homophily in follower networks. Debate about whether top spreaders drive or follow audience demand.

Open questions: Mechanisms driving the reversal of causality between Trump and Clinton supporter influence patterns. Role of algorithmic amplification in shaping the dynamics. Generalization to other elections and platforms. Whether observed patterns reflect selection by supporters or active recruitment/direction by spreaders.