David RandContributions
- Introduces a methodology for measuring users' exposure to misinformation originating from political elites, moving beyond prior focus on what users share or believe to what they are exposed to
- Develops falsity scores for 816 elite Twitter accounts based on professional fact-checks from PolitiFact, creating a quantifiable measure of elite-level information quality
- Validates the measure by showing misinformation-exposure scores predict quality of news users share (both by professional fact-checkers and politically balanced crowds), controlling for ideology
- Analyzes co-follower, co-share, and co-retweet networks of 5,000 Twitter users, identifying echo chambers of both political and misinformation content
- Demonstrates a significant ideological asymmetry: estimated ideological extremity correlates more strongly with misinformation exposure for conservative users than liberal users
- Releases open-source R package and API enabling any researcher to calculate elite misinformation-exposure scores for arbitrary sets of Twitter users
Method
The study identifies 1,005 elites (politicians, bureaucrats, media figures, organizations) fact-checked at least three times by PolitiFact as of October 28, 2020. For each elite, a falsity score is calculated by assigning veracity ratings to each fact-check (True=1, Mostly True=0.8, Half True=0.6, Mostly False=0.4, False=0.2, Pants on Fire=0), averaging across all fact-checks for that elite, and subtracting from 1 to create a falsity score from 0 (entirely truthful) to 1 (entirely false).
The authors identify 950 Twitter accounts associated with 816 of these 1,005 elites and retrieve all followers of these accounts (N = 122.5M unique users). They restrict to users following at least three elites (N = 38.3M) and randomly sample 5,000 users. For each user, a misinformation-exposure score is calculated by averaging the falsity scores of all rated elite accounts they follow, weighted by account tweet frequency as a proxy for exposure intensity (robustness checks show results hold without weighting).
Political ideology is estimated using accounts followed, adapted from prior work: users are scored on ideology based on the partisan leanings of elite accounts they follow (liberal = −2.5 to 0, conservative = 0 to 2.5).
News quality is measured using two methods: (1) average trustworthiness ratings from eight professional fact-checkers for 60 news websites, and (2) politically balanced crowd ratings from 970 demographically representative US respondents. For each user, links in their 3,200 most recent tweets (collected July 23, 2021) are extracted and matched against these domain ratings. Only ~5% of links match the available domain list, representing a limitation.
Results
Exposure and content quality: Users' misinformation-exposure scores are negatively correlated with the quality of news they share, using both professional fact-checker ratings (b = −0.728, 95% CI [−0.753, −0.704], p < 0.001) and crowd ratings (b = −0.540, 95% CI [−0.570, −0.510], p < 0.001). This relationship remains significant controlling for estimated ideology (b = −0.712 and b = −0.565 respectively), with misinformation-exposure scores explaining 53% of variance in fact-checker-rated quality and 29% in crowd-rated quality.
Ideology: Misinformation-exposure scores are positively correlated with conservative estimated ideology (b = 0.747, 95% CI [0.727, 0.767], p < 0.001). Critically, when misinformation-exposure score is included in the model, the relationship between ideology and content quality becomes insignificant (p = 0.265 using fact-checker ratings, p = 0.191 using crowd ratings), suggesting that misinformation exposure mediates much of the ideological effect on sharing behavior.
Toxicity and moral outrage: Misinformation-exposure scores are positively correlated with language toxicity (b = 0.129, p < 0.001; b = 0.319 controlling for ideology, p < 0.001) and moral-outrage expressions (b = 0.107, p < 0.001; b = 0.329 controlling for ideology, p < 0.001).
Co-share network: Community detection on the co-share network of 1,798 domains reveals three clusters: liberal-leaning domains (average misinformation-exposure 0.389, ideology −0.470), center-left domains (0.404, 0.038), and conservative domains (0.506, 1.22). Average misinformation-exposure and toxicity scores differ significantly across clusters even controlling for ideology.
Ideological extremity: Estimated ideological extremity (absolute value of ideology) is more strongly associated with misinformation exposure for conservative users (b = 0.825, p < 0.001) than liberal users (b = 0.160, p < 0.001). This asymmetry holds when ideology is estimated using media sharing instead of accounts followed (conservative: b = 0.567, p < 0.001; liberal: b = 0.111, p < 0.001).
Notes
The paper makes a methodological contribution by measuring exposure rather than sharing or belief, capturing an intermediate but understudied part of the misinformation pipeline. The finding that misinformation-exposure scores explain much of the ideological effect on sharing quality is notable and suggests that exposure (choice of whom to follow) may be a more fundamental driver of information-sharing patterns than ideology itself.
A key limitation is reliance on PolitiFact's fact-checking selections and evaluations. PolitiFact may exhibit bias in which claims it fact-checks or how it evaluates them; the authors acknowledge this but note that the methodology can be extended to other fact-checking sources. The study is also limited to users following at least three elites, potentially skewing toward more politically engaged users. Coverage of news domains in the sharing analysis is low (~5% of links), and the measure relies on users' immediate follower network without capturing exposure through retweets or algorithm recommendations. Finally, the study focuses on Twitter, which is not nationally representative.
The ideological asymmetry (extremity mattering more for conservatives) aligns with prior work on partisan divergence in misinformation sharing but warrants further investigation into whether it reflects genuine differences in exposure strategies, algorithmic amplification patterns, or elite messaging behavior.