Contributions
- Demonstrates that out-group language is the strongest predictor of social media engagement across Facebook and Twitter, surpassing established predictors (emotional language, moral contagion).
- Shows that posts about political opponents elicit "angry" reactions (the most popular emoji reaction), while in-group posts elicit "love" reactions, reflecting out-group derogation vs. in-group favoritism.
- Validates that the effect is not driven by specific politicians (e.g., Trump) but holds across different categories of political language (identity terms, famous politicians, congressional members).
- Provides evidence that the effect is stronger among political leaders than media accounts, suggesting elite discourse may amplify partisan animosity.
Method
The authors analyzed datasets from two platforms:
Study 1 (Media outlets): 599,999 Facebook posts and 227,229 tweets from liberal (e.g., MSNBC, NYT) and conservative (e.g., Fox News, Breitbart) news media accounts, classified using All Sides Media Bias ratings.
Study 2 (Congress): 1,078,562 tweets and 825,424 Facebook posts from Democratic and Republican US congressional members.
For each post, they counted occurrences of:
- Democratic political language (politician names, congressional member names, identity terms like "liberal," "Democrat," "left-wing")
- Republican political language (same categories, conservative-partisan terms)
- Negative affect words, positive affect words, and moral-emotional language (using validated LIWC dictionaries)
They fit ordinary least squares regression models predicting retweet and share counts, controlling for account followers, URLs, media attachments, and retweet status. For Facebook, they also analyzed reactions (like, love, haha, sad, wow, angry) as dependent variables. Results were meta-analyzed across all eight datasets to estimate average effect sizes.
Results
Across all datasets (meta-analyzed):
- Each additional out-group word increased odds of sharing by 67% (exp(b) = 1.67, 95% CI [1.43, 1.69])
- Posts with at least one out-group term were 2.3× more likely to be shared than posts with in-group terms only
- Out-group language was ~4.8× as predictive as negative affect language and ~6.7× as predictive as moral-emotional language
Emotional responses:
- Out-group language strongly predicted "angry" reactions (exp(b) = 2.19 across datasets)
- In-group language strongly predicted "love" reactions (exp(b) = 1.57)
- The angry reaction was the single most popular emoji reaction across datasets
Political orientation and platform: No moderation by ideology or platform (Twitter vs. Facebook), but effect was significantly stronger among politicians (99% increase per word) than news media (41% increase per word).
Notes
Strengths:
- Large, naturalistic dataset across two major platforms with high ecological validity
- Robust across multiple operationalizations of "out-group" (identity terms, famous politicians, all congressional members)
- Clear evidence that effect is not driven by a single divisive figure
- Direct measurement of engagement (shares/retweets) rather than self-report
- Mechanism analysis via Facebook reactions clarifies emotional drivers
Limitations:
- Correlational; cannot establish causation or clarify why out-group content engages audiences
- Data reflect 2016–2020 period (peak US polarization); unclear if findings generalize to other eras or nations
- No access to Twitter/Facebook algorithmic ranking; cannot isolate platform amplification from user choice
- Congressional accounts may have more explicitly partisan audiences, limiting generalizability
Practical significance: The paper argues this creates a "perverse incentive" on social media: people report not wanting partisan animosity, yet such content receives the most engagement. When platform business models prioritize engagement, they may inadvertently reward divisive speech. Facebook's own internal research (cited) showed posts rated "bad for the world" received more engagement; when the company tested down-ranking such posts, engagement fell and the feature was not deployed.
Relevance to misinformation: While not the paper's focus, the authors note that false political news is often highly negative about the out-group, suggesting that out-group animosity may partly explain false news virality.