Quantifying the impact of misinformation and vaccine-skeptical content on Facebook

Contributions

• Framework for measuring impact at scale: Proposes impact = exposure × persuasive influence. Combines experimental treatment effects with observational Facebook data to estimate causal effects for 13,206 URLs using crowdsourcing and machine learning.
• Distinction between veracity and persuasiveness: Shows that content's predicted harm to vaccine confidence predicts negative persuasive effects (−0.69 pp per scale point), not whether the content is factually false. Factually accurate headlines can strongly reduce vaccination intentions if they imply health risks.
• Empirical finding on exposure asymmetry: Flagged misinformation received only 0.3% of vaccine-related URL views (8.7 million of 2.7 billion views). A single Chicago Tribune article ("healthy doctor dies 2 weeks after vaccine") reached >50 million users—more than 6x the total reach of all flagged misinformation combined.
• Mainstream media as primary vector: Despite low-credibility domains posting more hesitancy-inducing content (66%), mainstream outlets like NYT, Fox News, and local news generated far more views and thus greater total estimated impact than fringe sources.

Method

Study 1 (N=8,603): Online survey experiment on Lucid. Participants randomly assigned to control or one of 40 fact-checked vaccine-misinformation headlines. Outcome: self-reported vaccination intentions (0–100 scale) before/after exposure. Result: exposure to misinformation reduced intentions by 1.5 pp (p<0.00001), but effect varied by 3pp across items (bottom decile), suggesting factors beyond falsity matter.

Study 2 (N=10,122): Representative sample of 90 highly-shared vaccine-related articles from Facebook (balanced by topic and domain quality). Same procedure. Crowdsourced raters labeled all 130 items (studies 1+2 combined) on: surprising, plausible, political lean, familiar, harmful-to-health. Random-effects meta-regression predicting observed treatment effects. Result: only "harmful-to-health" rating predicted effect (−0.69 pp per scale point; p=0.0003). Falsity was associated with stronger negative effect but did not remain significant when harmful-to-health was included as a predictor.

Exposure analysis: Used Facebook Social Science One URL Shares dataset covering Jan–Mar 2021 (vaccine rollout period). Identified 13,206 public URLs shared >100 times. Labeled as flagged if rated false/misleading/mixture by professional fact-checkers; all others as unflagged. Aggregated views by fact-check status, domain credibility, and individual story.

Scaling predictions: Recruited crowd raters to predict which of the 13,206 URLs would increase/decrease vaccination intentions, and to rate them on harmful-to-health and accuracy. Trained COVID-Twitter-BERT NLP model on 1,139 crowdsourced-labeled URLs (oversampled high-reach and flagged content). Model achieved 86% accuracy (±0.5 scale points) predicting crowdsourced scores on holdout test set; 97% AUC on binary hesitancy-inducing classification. Passed crowdsourced predictions through meta-regression model from studies 1+2 to predict each URL's treatment effect.

Impact calculation: Multiplied each URL's predicted treatment effect by its observed view count, summed across URL categories (flagged vs. unflagged), and normalized by ~233 million US Facebook users.

Results

• Persuasive effects: On average, flagged misinformation URLs predicted to reduce intentions by 1.36 pp (95% CI: [−1.91, −0.73]); median unflagged URL: −0.3 pp (95% CI: [−0.92, 0.34]). Difference highly significant (t=−40.2, p<0.00001).
• Exposure: Flagged misinformation received 8.7M views (0.3% of total vaccine-related views). Low-credibility domains: 5.1% of views. Most-viewed URL (Chicago Tribune "healthy doctor" story): 67.8M views (>20% of US Facebook user base).
• Total impact (weighting by exposure): Among 3,711 hesitancy-inducing URLs: unflagged vaccine-skeptical content lowered predicted vaccination intentions by −2.28 pp (95% CI: [−3.4, −0.99]) per US Facebook user. Flagged misinformation: −0.05 pp (95% CI: [−0.07, −0.02]). Ratio: 46-fold difference. At 60% intention-to-behavior conversion, implies ~3 million additional vaccinated Americans had vaccine-skeptical content been prevented.
• Source of hesitancy-inducing content: 98% of total view-weighted negative impact came from unflagged content. Although 66% of low-credibility-domain content was hesitancy-inducing, it accounted for only 9.3% of total estimated negative impact; mainstream outlets were responsible for 90.7% due to greater volume and reach.
• Primary narratives: Coverage of rare deaths shortly after vaccination (without contextualization of rarity or vaccine's causal role) achieved disproportionate reach and estimated impact.

Notes

Strengths:

• First large-scale causal study combining experiments, observational data, and machine learning to measure actual impact of vaccine (mis)information on Facebook during early COVID-19 vaccine rollout.
• Rigorous distinction between content's persuasiveness (conditional on exposure) and its real-world impact (which also depends on reach). Most prior work focuses on one or the other.
• Crowdsourced prediction approach is generalizable to other topics and platforms; scalable alternative to labor-intensive manual content review.
• Policy-relevant: findings directly inform interventions. Suggests that focusing only on veracity/source credibility misses the largest source of vaccine hesitancy (mainstream, factually accurate but misleading reporting).

Limitations:

• Survey measure of intentions (not actual vaccination behavior). Authors note Athey et al. found ~0.6x conversion, so actual impact ~40% smaller. Intent-behavior gap is smaller for avoidance (refusing vaccine) than for action-taking, but still a limitation.
• Experiments conducted in mid-2022; Facebook data from Q1 2021. Some robustness checks with contemporaneous data show similar patterns, but temporal gap remains.
• Analysis covers only URL link content; excludes native video, photos, text-only posts on Facebook. Misinformation may be proportionally higher in non-link formats (lower bound on actual vaccine-skeptical impact).
• Sample of crowdsourced raters may differ from actual Facebook users in attention, political beliefs, or vaccine baseline attitudes. Authors find no significant heterogeneity in treatment effects by demographic, but cannot rule out selective exposure to anti-vaccine content among users with preexisting hesitancy.
• Cannot definitively separate causation from correlation in Facebook data. Vaccine-hesitant users may have actively sought out skeptical content rather than being passively exposed.

Significance:

This is a landmark study in the misinformation measurement literature. It moves beyond correlational claims to estimate causal effects at scale, and crucially, reveals that focusing public health and platform interventions solely on "flagged" misinformation ignores ~46x more impact from factually accurate but misleading content. Has already influenced policy discussions around fact-checking effectiveness and the need for platform oversight of "gray area" content. The method (experiments + observational exposure data + crowdsourced predictions) is a replicable template for future impact quantification studies.