Eady, G., Paskhalis, T., Zilinsky, J., Bonneau, R., Nagler, J., & Tucker, J. A. (2023). Exposure to the Russian Internet Research Agency foreign influence campaign on Twitter in the 2016 US election and its relationship to attitudes and voting behavior. Nature Communications, 14, 62. https://doi.org/10.1038/s41467-022-35576-9
Summary
This paper asks whether ordinary Americans’ exposure to the Russian Internet Research Agency (IRA) foreign influence campaign on Twitter during the 2016 US presidential election bore any relationship to their political attitudes, polarization, or voting behavior. By linking a three-wave longitudinal YouGov survey of 1,496 US Twitter users to their actual reconstructed Twitter timelines (roughly 1.2 billion tweets from followed accounts), the authors measure real, as-it-happened exposure rather than post-hoc or content-only proxies. Their central finding is deflationary: exposure was extraordinarily concentrated among a small group of strongly Republican users already predisposed toward Trump, was overshadowed by domestic news and politician content by more than an order of magnitude, and showed no meaningful relationship to changes in issue positions, polarization, or vote choice. The paper thus locates foreign influence campaigns within the political-science tradition of minimal persuasion effects.
Key Contributions
- First study to link longitudinal survey data collected during the 2016 campaign to respondents’ actual Twitter timelines, measuring exposure to the IRA campaign as it occurred.
- Quantifies both the scale and the concentration of exposure, and benchmarks it against domestic news media and politician content — context prior work lacked.
- Directly estimates the relationship between exposure and changes in attitudes, polarization, and voting, addressing the gap left by structural/content-focused analyses and earlier one-month post-hoc data.
- Uses equivalence testing (TOST) to substantiate null findings rather than merely reporting non-significance.
- Offers a corrective to alarmist narratives about foreign manipulation, while flagging caveats about second-order and cross-platform effects.
Methods
The authors combined a three-wave YouGov panel (April–October 2016 plus post-election vote validation) with respondents’ linked Twitter accounts, reconstructing potential timeline exposure from all posts by followed accounts in the eight months before the election. Twitter’s released state-sponsored account datasets (Russia, China, Iran, Venezuela) identified foreign-influence content in timelines. Concentration was analyzed via empirical cumulative distributions and benchmarked against domestic content. Exposure was modeled with OLS (log exposures, including a squared party-ID term to test U-shaped effects) and with within-subject/panel fixed-effects regressions relating changes in issue positions, perceived candidate polarization, and vote choice to exposure. Equivalence tests bounded the null effects, counterfactual simulations estimated vote-choice changes, and a back-of-the-envelope calculation scaled exposure to the national population.
Findings
- 70% of respondents saw at least one foreign-influence post; Russian accounts made up 86% of such exposures.
- Exposure was overwhelmingly incidental via retweets from ordinary followed accounts (75–80% of pathways by election day), not direct following.
- Extreme concentration: 1% of respondents accounted for 70% of Russian exposures; 1% of Russian accounts produced 89% of the content seen — far more concentrated than domestic news (1% = 24%) or politicians (1% = 37%).
- In the final month, respondents saw ~4 Russian posts/day versus ~106 from national news and ~35 from politicians (25x and 9x more); median weekly Russian exposure was zero.
- Strong Republicans were exposed to ~9x as many Russian posts as Democrats or Independents; the relationship was monotonic, not U-shaped.
- No significant or substantive relationship between exposure and changes in eight issue positions, ideology, or perceived polarization (magnitudes ~0.05–0.06 SD, bounded below 0.2 SD by equivalence tests).
- Voting relationships were near zero, non-significant, and signed against benefiting Trump; simulated median change in Trump vote was about −0.18 to −0.4 points, well below the 0.77-point Wisconsin margin.
- A back-of-the-envelope estimate suggests ~32 million Americans were potentially exposed to IRA content over the eight months before the election.
Connections
This paper is a companion to the authors’ broader work on measuring real exposure through survey-linked behavioral data (see Eady2025-vm) and directly engages the earlier IRA-exposure study by Bail et al. (Bail2018-fk), whose one-month 2017 data it critiques and extends. Its concentration findings — that a tiny fraction of users absorb the vast majority of dubious content — echo work on the limited reach and skewed distribution of misinformation exposure such as Grinberg2019-ua, Guess2020-rr, and Allen2020-nj, while its minimal-effects conclusion resonates with persuasion-effect research including Guess2023-ai and Nyhan2023-gb.