Bollenbacher, J., Menczer, F., & Bryden, J. (2026). Effects of antivaccine tweets on COVID-19 vaccinations, cases, and deaths. EPJ Data Science, 15, 12. https://doi.org/10.1140/epjds/s13688-025-00606-1
Summary
This paper asks whether exposure to antivaccine content on Twitter/X causally reduced COVID-19 vaccination uptake in US counties between February and August 2021 — and, downstream, whether it produced additional cases and deaths. The authors bridge the gap between controlled misinformation experiments (which measure intentions) and correlational studies (which cannot establish causation) by building a mechanistic epidemic model coupled to causal inference. Their central estimate is that antivaccine tweets caused roughly 14,000 vaccine refusals, translating to a lower bound of ~545 additional cases and ~8 additional deaths over the six-month window. The work argues that platform-level speech can be linked to offline public health outcomes with an interpretable, reproducible pipeline.
Key Contributions
- Observational causal evidence tying online antivaccine exposure to real-world vaccination behavior, cases, and deaths — going beyond intention-based lab studies and purely correlational work.
- Introduces SIRVA, a compartmental epidemic model extending SIR with Vaccinated and hesitant/Antivaccine compartments, where the hesitancy conversion rate is driven partly by exogenous information exposure.
- Combines Bayesian epidemic modeling with causal graphical modeling (do-calculus) to derive an interpretable Average Treatment Effect of exposure.
- Offers an open, reproducible methodology for connecting platform speech data to offline epidemic dynamics — with implications for moderation policy and public health communication.
Methods
The authors trained a RoBERTa classifier to detect antivaccine tweets (F1 ≈ 0.74), then constructed a county-to-county COVID retweet network to define per-capita antivaccine exposure propagated across counties and normalized by population. The SIRVA model adds a Vaccinated compartment and an Antivaccine/hesitant compartment (A = αS), with a conversion rate γ = γ_p + γ_e·E partly driven by exposure. Parameters were inferred per county and globally via Bayesian MCMC (NumPyro/NUTS), fitted to CDC county-level case, death, and vaccination data plus CoVaxxy tweets (Feb 6–Aug 9, 2021). Causal graphical modeling yielded the ATE. Model comparison used PSIS-LOO against SIRV and static variants; robustness checks included county-shuffle nulls and comparison against Meta’s Social Connectedness Index.
Findings
- The exposure-to-hesitancy coefficient γ_e had posterior mean ≈ 0.18 (95% CI: 0.15–0.22), significantly greater than zero (p = 0.0002).
- Estimated ATE of exposure on vaccination rate ≈ −3.2×10⁻⁴ vaccinations per daily antivaccine tweet per capita.
- An estimated 14,086 people (95% CI: 11,414–16,759) refused vaccination due to Twitter exposure, against ~27 million who became hesitant overall.
- Roughly 545 additional cases and 8 additional deaths attributed to these Twitter-induced refusals (a lower bound).
- SIRVA outperformed SIRV in ELPD-LOO by about three standard errors, indicating better out-of-sample fit.
- Shuffling exposure across counties nullified the effect, and the COVID retweet network was uncorrelated with Meta’s Social Connectedness Index — supporting a platform-specific causal interpretation.
Connections
This is one of several works in the corpus focused specifically on COVID-19 vaccine misinformation on Twitter/X; it complements Pierri2025-hm, whose correlational findings on online misinformation and hesitancy this paper explicitly extends toward causal estimation, and DeVerna2025-dl, which shares the CoVaxxy-style approach to studying vaccine discourse. More broadly it sits within the health-misinformation-networks literature that models how misinformation propagates and affects offline behavior.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify · Apple Podcasts