Guess, A., Nagler, J., & Tucker, J. (2019). Less than you think: Prevalence and predictors of fake news dissemination on Facebook. Science Advances, 5, eaau4586. https://doi.org/10.1126/sciadv.aau4586
Summary
This paper offers one of the first behavioral portraits of who actually shared fake news on Facebook during the 2016 U.S. presidential campaign, moving beyond self-report to observed platform activity. By linking a three-wave YouGov panel survey to respondents’ consented Facebook profile data, the authors document that sharing articles from fake news domains was a rare behavior — over 90% of respondents shared none. Where sharing did occur, it was concentrated among conservatives and, most robustly, among older Americans: users over 65 shared far more than any other group, an effect that survives controls for ideology, partisanship, education, and overall posting volume. The paper argues that age deserves to be treated not as a nuisance control but as a central explanatory factor for misinformation dissemination.
Key Contributions
- Provides an individual-level, trace-based (rather than self-reported) account of who actually shares fake news, overcoming self-report bias.
- Offers a corrective to alarmist post-2016 narratives by showing fake news sharing was empirically rare.
- Identifies the over-65 cohort as a robust, politically-independent predictor of misinformation sharing.
- Demonstrates a reusable methodological template for linking survey data to platform behavioral traces.
- Opens research agendas connecting digital literacy, cognitive aging, and news-feed exposure to misinformation spread.
Methods
The authors fielded a three-wave online panel survey (wave 1 N=3500) via YouGov with sample-matching weights, then linked respondents to their private Facebook profile data (timeline posts, external links, page likes) obtained through a consented web app — successfully matching 1191 respondents (~44% of Facebook users in the sample). Shared external links were cross-referenced against journalist- and academic-curated fake news domain lists, principally Craig Silverman’s BuzzFeed-based list (reduced to 21 mostly pro-Trump domains after filtering hard-news domains), with the Allcott and Gentzkow list and others as robustness checks. Individual-level share counts were modeled with Poisson/quasi-Poisson regression (negative binomial and OLS as robustness), including sociodemographic and political predictors, supported by extensive supplementary robustness testing.
Findings
- 91.5% of respondents shared zero fake news articles; only 8.5% shared at least one.
- Low sharing is not attributable to inactivity — 61.3% shared 100–1000 links overall.
- Republicans shared more than Democrats (18.1% vs. 3.5% sharing at least one); independents shared about as much as Republicans.
- Very conservative respondents shared the most (~1.0 articles on average).
- Users over 65 shared nearly seven times as many fake news articles as the youngest group (18–29), and ~2.3 times the next-oldest group, in multivariate models.
- The over-65 age effect is significant (P<0.01) and robust across alternative measures and controls.
- People who share the most links overall are not the ones sharing fake news, undermining a “shares anything” explanation.
- Applied to hard news, the same model yields varied predictors that do not include age.
Connections
This is a foundational empirical study in the misinformation literature and connects closely to work on the prevalence and reach of fake news exposure such as Guess2020-rr, Grinberg2019-ua, and Allcott2017-yz, as well as to research on why people share false content, including Pennycook2021-jq, Osmundsen2021-et, and Vosoughi2018-at. Its focus on the modest real-world footprint of fake news echoes broader efforts to right-size alarmist narratives, as in Allen2020-nj and Guess2021-ym, and its account of individual-level sharing determinants relates to platform-exposure studies like Gonzalez-Bailon2024-rq and Bakshy-style feed research; among recent work, it links to studies of misinformation sharing and recalibration such as DeVerna2025-dl and Lyons2026-ca.