Coordinating a Multi-Platform Disinformation Campaign: Internet Research Agency Activity on Three U.S. Social Media Platforms, 2015 to 2017¶

Venue: Political Communication, Vol. 37:2, pp. 238–255, 2019 — DOI

TL;DR¶

This paper analyzes Russian Internet Research Agency activity across Facebook, Twitter, and Reddit from 2015–2017 using vector autoregression models on 3,126 Facebook ads, 1.9M tweets, and 12,603 Reddit posts. Key finding: Reddit activity influenced subsequent Twitter activity (a "trial balloon" effect), but the IRA's Twitter engagement responded to Trump approval ratings, suggesting platform-aware, temporally-coordinated disinformation strategy.

Contributions¶

First empirical test of whether state-sponsored disinformation campaigns coordinate strategies across multiple platforms.
Documents temporal dynamics: Reddit as testing ground for messages before broader Twitter distribution.
Shows IRA activity on Twitter was sensitive to external political factors (Trump approval), suggesting real-time adaptive strategy.
Demonstrates asymmetric cross-platform effects using Granger causality and impulse response functions.

Method¶

The study collected three datasets: (1) Facebook ads purchased by IRA accounts (released by Facebook to Congress); (2) tweets from 9 IRA-associated accounts via Twitter's archive released October 2018; (3) Reddit posts from 944 IRA-identified accounts (preserved by Reddit in April 2018).

The unit of analysis was weekly aggregation from July 2, 2015 to May 31, 2017. Six time series were constructed: three measuring IRA activity per platform (tweets, Facebook ads, Reddit posts) and three measuring external factors (US-Russia military/diplomatic interactions, Trump approval rating).

Time series were tested for stationarity using ADF and KPSS tests. Non-stationary series were differenced. A VAR(2) model was fit with Granger causality tests to assess lead-lag relationships between platforms. Impulse response functions analyzed the magnitude and duration of shocks from one platform to another. External variables (Trump approval, geopolitical events) were included as exogenous regressors.

Results¶

Volume: IRA produced 3,126 Facebook advertisements, 1,886,919 tweets, and 12,603 Reddit posts during the study period.
Content themes: IRA content targeted U.S. politics (2016 election, Trump, travel ban), focusing on far-right and social justice narratives to deepen polarization.
Platform dynamics: Facebook ads and Twitter tweets increased around the election. Reddit activity spiked mid-campaign then declined post-election.
Granger causality: Reddit → Twitter relationship was statistically significant (χ² = 18.38, p < .001), but no other cross-platform lead-lag relationships found.
Impulse response: A shock of 100 Reddit posts produced a peak Twitter response of ~20 tweets two weeks later, then dissipated.
External influence: Trump approval rating had a strong positive relationship with Twitter activity (coeff = 899.89, p < .01), but not with Facebook or Reddit.
H1 and H3B (geopolitical influence) were rejected: US-Russia relations did not significantly predict IRA activity on any platform.

Connections¶

Related to Troll Factories: Manufacturing Specialized Disinformation on Twitter via shared focus on IRA operations and bot-driven amplification.
Extends Coordinated inauthentic behavior framework to temporal cross-platform dynamics.
Complements It takes a village to manipulate the media: coordinated link sharing behavior during 2018 and 2019 Italian elections on coordinated campaigns' multi-platform nature.
Methodologically similar to Election interference and information warfare studies using time series analysis.
Cited alongside Fake news on Twitter during the 2016 U.S. presidential election for 2016 election disinformation accounts.

Notes¶

Strengths: This is a rigorous empirical study with strong data provenance (official releases from Facebook, Twitter, Reddit). The VAR framework is appropriate for testing temporal dependencies across platforms—a novel research question at the time. The impulse response functions provide intuitive interpretation of cross-platform effects. Open data and reproducible analysis enhance credibility.

Limitations: The study acknowledges that platform identification of IRA accounts may contain inaccuracies, and it does not examine which message types drove cross-platform coordination. Content analysis is descriptive rather than systematic. The study only covers three platforms and may not generalize to other disinformation actors or campaigns. External event variables (Trump approval, geopolitical crises) are crude proxies and may not capture state-level strategic intent.

Follow-up questions: What message characteristics predict successful trial-balloon campaigns on Reddit? How did the IRA's strategy evolve after the 2016 election? Do other state actors use similar Reddit-first testing strategies?