Effects of Antivaccine Tweets on COVID-19 Vaccinations, Cases, and Deaths

Despite the wide availability of COVID-19 vaccines in the United States and their effectiveness in reducing hospitalizations and mortality during the pandemic, a majority of Americans chose not to be vaccinated during 2021. Recent work shows that vaccine misinformation affects intentions in controlled settings, but does not link it to real-world vaccination rates. Here, we present observational evidence of a causal relationship between exposure to antivaccine content and vaccination rates, and estimate the size of this effect. We present a compartmental epidemic model that includes vaccination, vaccine hesitancy, and exposure to antivaccine content. We fit the model to data to determine that a geographical pattern of exposure to online antivaccine content across US counties explains reduced vaccine uptake in the same counties. We find observational evidence that exposure to antivaccine content on Twitter caused about 14,000 people to refuse vaccination between February and August 2021 in the US, resulting in at least 545 additional cases and 8 additional deaths. This work provides a methodology for linking online speech with offline epidemic outcomes. Our findings should inform social media moderation policy as well as public health interventions.

💡 Research Summary

The paper investigates whether exposure to anti‑vaccine tweets on Twitter causally reduced COVID‑19 vaccination uptake in the United States and, consequently, increased cases and deaths during the first half of 2021. The authors combine large‑scale social‑media data with an extended compartmental epidemic model to quantify this effect.

Data collection: From February to August 2021 the authors gathered 26 million geolocated tweets mentioning COVID‑19, of which 2.2 million were classified as anti‑vaccine using a supervised text classifier. About 10 % of all tweets could be assigned to a specific county. To capture not only the raw volume but also the diffusion of misinformation across counties, they constructed a county‑level exposure metric (E) that weights each county’s anti‑vaccine tweet count by the retweet network linking counties.

Epidemiological model: They extend the classic SIR framework by adding a Vaccinated compartment (V) and splitting the Susceptible pool into willing (S′) and unwilling (A) individuals, yielding the SIR‑VA model. The key new parameter is γ_e, the rate at which exposure to anti‑vaccine content converts willing susceptibles into the hesitant group per unit of exposure. The total conversion rate γ = γ_e·E + γ_p includes a baseline drift term γ_p that captures all non‑Twitter influences on hesitancy.

Parameter inference: Using county‑level time series of reported cases, vaccinations, and the exposure metric, the authors fit the model with Bayesian Markov‑Chain Monte Carlo. The posterior mean of γ_e is ≈0.18 (95 % credible interval 0.15–0.22), significantly greater than zero (p = 0.0002). This indicates that higher exposure to anti‑vaccine tweets predicts a measurable increase in the hesitant population, which in turn depresses vaccination rates.

Causal analysis: To rule out reverse causality (hesitant individuals seeking more anti‑vaccine content) and confounding platform‑wide trends, they embed the exposure‑hesitancy relationship in a causal graph that includes a feedback edge from A to E. Sensitivity tests—shuffling exposure across counties and comparing with other social‑network indices—destroyed the observed effect, supporting the claim that the relationship is specific to Twitter’s diffusion network.

Impact estimation: The authors define an average treatment effect (ATE) as the change in vaccinations per capita per unit of exposure, finding ATE = −3.2 × 10⁻⁴. Applying this nationwide yields an estimated 14,086 fewer vaccinations attributable to anti‑vaccine tweets during the study period. Using published vaccine efficacy against infection, they compute a lower bound of 545 additional COVID‑19 cases and eight additional deaths that would not have occurred without this misinformation‑driven hesitancy.

Model validation: A leave‑one‑out cross‑validation comparing the SIR‑VA model to a simpler SIR‑V model (which omits the hesitant compartment) shows the former has a substantially better Bayesian fit score (three standard errors improvement). This demonstrates that explicitly modeling hesitancy improves predictive accuracy.

Limitations: The analysis relies on CDC data that contain reporting lags and imputed values, and on Twitter geolocation data that cover only a fraction of users, potentially biasing exposure estimates. The classifier may miss some anti‑vaccine content, and the model assumes uniform γ_e across all counties, ignoring possible heterogeneity (e.g., political leanings). Moreover, the model does not account for infections among vaccinated individuals, vaccinations of hesitant people, or time lags between exposure and infection.

Implications: The findings suggest that platform‑specific misinformation can have tangible public‑health consequences. Moderation strategies—such as reducing the visibility of anti‑vaccine posts, promoting fact‑checked information, or inoculating users with accurate messages—could measurably increase vaccination rates and save lives, not only for COVID‑19 but for future vaccine‑preventable diseases. Future work should incorporate data from other platforms (e.g., Instagram, TikTok) and explore county‑level heterogeneity in the effect of misinformation.