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

TL;DR

This study addresses how antivaccine misinformation on Twitter can translate into offline COVID-19 outcomes by linking county-level exposure to vaccination uptake and disease metrics. It introduces the SIRVA model, an SIR-based epidemic framework augmented with vaccine hesitancy and a measurable exposure channel, and estimates parameters using Bayesian MCMC on geolocated county data and a classifier for antivaccine tweets. The authors find that exposure to antivaccine content increases hesitancy (\gamma_e \approx 0.18) and causally reduces vaccination uptake (ATE \approx -3.2\times 10^{-4} vaccinations per daily tweet), corresponding to about 14,086 vaccinations prevented, ~545 cases, and ~8 deaths among the unvaccinated from February to August 2021. The work demonstrates a methodology to connect online speech to offline epidemic outcomes and discusses policy implications for social media moderation and targeted public health interventions.

Abstract

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.

Figures (4)

• Figure 1: Antivaccine tweets. (a) Number of geolocated antivaccine tweets for each day of the observation period. (b) Antivaccine tweets per capita per day, geolocated in each US county during the observation period. Grey coloring denotes counties where we have insufficient geolocated Twitter data.
• Figure 2: SIRVA model. (a) Compartmental model diagram (see Methods). Note that $A=\alpha S$ and $S'=(1-\alpha)S$, where $\alpha$ is the vaccine hesitancy ratio, i.e., the fraction of susceptibles who are unwilling to be vaccinated at time $t$. $E$ is the magnitude of exposure to antivaccine tweets and $\gamma_e$ is the rate at which people become vaccine-hesitant due to this exposure. (b) Posterior distribution of $\gamma_e$, with 95% high-density interval. The posterior mean value is $\gamma_e \approx 0.18$. (c) Posterior distributions of other global model parameters. Note that there are multiple curves for $\nu$ and $\beta$ because we used different values of these parameters during different time periods to account for changing infectivity and national vaccine availability (24-day periods for $\beta$, 48-day periods for $\nu$); these are ordered from earliest to latest. (d) Vaccinated population (V), and population who are both susceptible and vaccine-hesitant (A).
• Figure S3: Steps to construct the causal graphical model from the SIRVA compartmental model.
• Figure S4: Causal graphical model (CGM) corresponding to the SIRVA model. The diagram shows the causal antecedents of the model variables at time $t$. Here the notation $\dot{X}$ denotes the quantity $\frac{dX}{dt}$, and each arrow represents a causal relationship, where the source variable (on the left) causes the target variable (on the right). The dotted arrows denote an additional possible confounding relationship from $\alpha$ to $E$ in the model; this relationship is assumed to exist in the CGM, and our Average Treatment Effect accounts for its presence. Note that the causal graph is a lattice that can be extended backward in time to an initial state (e.g., $\alpha_0$).