The rise of social media platforms initially fostered optimism about their democratizing potential, enabling collective action and social movements. However, this optimism waned as governments began using these platforms for domestic and foreign policy objectives, including undermining social movements and interfering in elections. The 2016 US election saw a high-profile example of this with the alleged interference by the IRA, a Russian government-linked organization. The IRA is accused of using fake accounts to polarize the electorate and influence voters. This study addresses the gap in understanding the relationship between exposure to such campaigns and actual political behavior. Previous research faced limitations in data availability, timeframes, and account removal. This study leverages unique longitudinal survey data linked to Twitter feeds to overcome these limitations and assess the influence of the IRA campaign on US voters' attitudes, polarization, and voting behavior. The research explores both the potential for success and failure of foreign influence campaigns, considering their scale and reach. While the IRA campaign reached millions, the research questions whether this scale outweighed the impact of other political content and if exposure was concentrated among those already inclined towards a specific candidate.

Existing research on foreign influence campaigns has focused on two main areas: the impact of these campaigns on attitudes and political behavior, and the structure and content of the campaigns themselves. Studies examining the former often suffer from limitations in data, such as short timeframes, data collected after account removals, and a lack of longitudinal perspective. Other work has analyzed the structure and content of campaigns but not the impact of exposure on user behavior. This study aims to bridge this gap by using unique linked data to assess the relationship between IRA account exposure and changes in political attitudes, polarization, and vote choice among US Twitter users during the 2016 election.

This study employs a three-wave longitudinal survey of 1496 US Twitter users conducted by YouGov. Respondents consented to share their Twitter data for research, allowing the researchers to link survey responses with their Twitter feeds. The survey collected data on socio-demographic characteristics, political attitudes, issue positions, and candidate preferences at multiple time points (April, October 2016, and post-election). The researchers collected tweets from accounts followed by respondents in the eight months preceding the election, totaling 1.2 billion posts. Using data released by Twitter, they identified posts originating from foreign influence campaigns (IRA and smaller campaigns from China, Iran, and Venezuela). The researchers define exposure as tweets and retweets respondents were potentially exposed to in their timelines, acknowledging that this doesn't capture actual views. They analyze exposure levels, concentration, and partisan profiles. To analyze the relationship between exposure and political outcomes, they use within-respondent changes in attitudes, polarization, and vote choice, comparing pre-campaign preferences to post-election votes and survey-based measures. Regression models assess the relationships, with exposure measured as both a count and a binary variable. Equivalence testing is employed to determine the magnitude of relationships, while accounting for multiple comparisons.

The study's key findings reveal a highly concentrated exposure pattern. 70% of respondents (n=1042) were exposed to at least one foreign influence campaign post, with the IRA campaign accounting for 86% of exposures. However, 1% of respondents accounted for 70% of the total exposures, and almost all exposure was concentrated within 10% of respondents. This concentration was significantly higher than exposure to domestic news media or political figures. Exposure was strongly correlated with strong Republican identification, with “Strong Republicans” exposed to roughly nine times more IRA posts than Democrats or Independents. Despite a back-of-the-envelope calculation suggesting 32 million Americans were potentially exposed, the researchers emphasize the relative insignificance of this number compared to exposure to domestic sources. Regression analyses found no statistically significant relationship between exposure to IRA posts and changes in issue positions, political ideology, or perceived polarization between the candidates. Equivalence testing confirmed that the effect sizes were negligible, with relationships not exceeding 0.2 standard deviations for all but one outcome. Similarly, analyses of voting behavior found no statistically significant relationship between IRA exposure and shifts in voting preference or changes in behavior favoring Trump, even when considering broader measures like abstention or third-party votes. In fact, the coefficients in these models suggested the opposite: a negative (albeit statistically insignificant) relationship with voting for Trump.

The findings challenge the notion that foreign influence campaigns on social media readily manipulate individual attitudes and voting behavior. The concentrated nature of exposure, primarily amongst strong Republicans already inclined to support Trump, explains the lack of significant impact. The sheer volume of domestic political content dwarfed the influence of the IRA campaign. The absence of a substantial relationship aligns with previous research on the minimal effects of political campaigns. This research highlights the limitations of observational data, emphasizing that this study isn't a randomized controlled experiment. However, the use of within-respondent comparisons helps to mitigate selection bias. While no direct impact on individual voting behavior was found, the researchers acknowledge the potential for indirect, second-order effects, like the enduring debate surrounding election legitimacy and mistrust.

This study demonstrates the limited impact of the IRA's Twitter campaign on individual-level attitudes and voting behavior during the 2016 US election. Exposure was highly concentrated, outweighed by domestic sources, and showed no significant relationship with changes in political opinions or vote choice. While acknowledging limitations of observational data and potential second-order effects, the findings suggest a need for caution when assessing the potential of such campaigns to manipulate voters. Further research might explore campaign adaptation strategies and potentially changing contexts in the digital sphere.

The study acknowledges several limitations. First, reliance on Twitter's identification of IRA accounts prevents independent validation. Second, as an observational, not experimental, study, the non-random assignment of exposure is a potential limitation, although the panel nature of the data and within-respondent comparisons mitigate some of these concerns. Third, the analysis focuses only on Twitter, excluding potential effects from other platforms or methods of interference (hacking, etc). Finally, the study measures potential, not actual, exposure to posts.