Source-linked AI summary
Online misinformation is linked to early COVID-19 vaccination hesitancy and refusal
Francesco Pierri, Brea Perry, Matthew R. DeVerna, Kai-Cheng Yang, Alessandro Flammini, Filippo Menczer, John Bryden
TL;DR
Early COVID-19 vaccine uptake varied across U.S. states, and the relationship between online vaccine misinformation and hesitancy was unclear. The paper examines these associations using social-media misinformation, vaccination and survey data, regression models, and Granger causality analysis, finding evidence that misinformation is linked to lower uptake and higher hesitancy, with a directional relationship toward hesitancy.
Problem
The study addresses limited evidence on how online vaccine misinformation relates to uneven vaccination uptake and vaccine hesitancy during the early COVID-19 vaccination program.
Method
The authors analyze roughly 55 million English-language vaccine-related Twitter posts alongside vaccination and survey data, using multivariate regression adjusted for six confounders and Granger causality analysis.
Results
Online misinformation is negatively associated with vaccination uptake and positively associated with hesitancy, with evidence of a directional relationship from misinformation to vaccine hesitancy.
Takeaways & Limitations
The findings support interventions addressing vaccine misbeliefs to help individuals make better-informed health decisions.
Takeaways & Limitations
The analyses alone do not demonstrate a causal relationship between online misinformation and vaccine hesitancy.
Abstract
from arXiv · showhide
Widespread uptake of vaccines is necessary to achieve herd immunity. However, uptake rates have varied across U.S. states during the first six months of the COVID-19 vaccination program. Misbeliefs may play an important role in vaccine hesitancy, and there is a need to understand relationships between misinformation, beliefs, behaviors, and health outcomes. Here we investigate the extent to which COVID-19 vaccination rates and vaccine hesitancy are associated with levels of online misinformation about vaccines. We also look for evidence of directionality from online misinformation to vaccine hesitancy. We find a negative relationship between misinformation and vaccination uptake rates. Online misinformation is also correlated with vaccine hesitancy rates taken from survey data. Associations between vaccine outcomes and misinformation remain significant when accounting for political as well as demographic and socioeconomic factors. While vaccine hesitancy is strongly associated with Republican vote share, we observe that the effect of online misinformation on hesitancy is strongest across Democratic rather than Republican counties. Granger causality analysis shows evidence for a directional relationship from online misinformation to vaccine hesitancy. Our results support a need for interventions that address misbeliefs, allowing individuals to make better-informed health decisions.
3 Observatory on Social Media, Indiana University, Bloomington, Indiana, USA
The section concludes by emphasizing the importance of enabling individuals to make better-informed health decisions.
- The findings support interventions that address misbeliefs so individuals can make better-informed health decisions.
Introduction
The introduction frames COVID-19 vaccine hesitancy and uneven uptake as public-health problems and motivates studying their relationships with online misinformation. Using geographically resolved social-media, CDC, and survey data, the paper reports negative associations between misinformation and uptake, positive associations with hesitancy, and evidence that misinformation precedes hesitancy.
- Motivation: Around 40-47% of American adults were hesitant to take the COVID-19 vaccine, while herd immunity requires 60-70% vaccination.Uneven vaccination distributions could create geographical clusters of non-vaccinated people.
- Study design: The study examines relationships among vaccine uptake, vaccine hesitancy, and online misinformation across U.S. states and counties.It combines Twitter, Facebook, and CDC data and uses Granger Causality analysis to test directional association.
- Key findings: State-level vaccine hesitancy was positively associated with misinformation (b = 6.88, p = .007) after adjusting for confounding factors.Vaccine uptake and hesitancy were strongly negatively correlated across states (Pearson R = –0.71, p < .001).
- Key findings: At the county level, Democratic and Republican counties converged toward the same vaccine-hesitancy level as misinformation increased.The reported ceiling effect was around 30% of residents being vaccine hesitant, with Republican counties already near that ceiling.
- Directionality: Granger analysis found that misinformation helped forecast vaccine hesitancy weakly at state level (p = .0519) and strongly at county level (p < .001).Significant lagged coefficients indicated an approximately 2-6-day lag from county misinformation posts to increased hesitancy in the same county.
Discussion
The discussion links online misinformation with lower COVID-19 vaccine uptake and higher hesitancy, while emphasizing that causal uncertainty remains. It calls for countering misinformation and improving information moderation to protect public health.
- Causal interpretation: Although analyses show directional evidence, they do not establish causality because ecological design and unmeasured confounding remain possible.Further investigation is warranted, including individual-level and ecological evidence concerning misinformation and vaccine hesitancy.
- Information diffusion: Vaccine opinions appear to spread geographically at a local scale, providing context for how online misinformation may produce broader offline effects.The authors note that social media users are not representative of the general public and that their findings offer insight into information diffusion, behavior, and infectious-disease spread.
- Limitations: Interpretation is limited by measuring sharing activity rather than regional exposure, averaging data geographically, detecting misinformation incompletely, and observing only a short period.Other influences include vaccine accessibility, changing COVID-19 infection and death rates, and legitimate vaccine-safety reports.
- Implications: The study supports better moderation of the information ecosystem so individuals can access information that benefits public health.The discussion frames associations between online misinformation and detrimental offline effects as a rationale for intervention.
Competing interests
The authors declare no competing interests. The study’s measurements and scripts are publicly available, although Twitter’s terms restrict release to tweet IDs that can be reconstructed through the API.
- The authors declare no competing interests.
- Vaccine uptake, hesitancy, socioeconomic, political, demographic, and aggregated online-misinformation measurements are publicly available in the associated repository.
- State- and county-level results can be fully reproduced with provided STATA scripts, while Twitter terms permit release only of tweet IDs reconstructable through the Twitter API.
Supplemental Information for: Online misinformation is linked
The supplemental information describes the study’s data sources and operational measures for vaccine-related misinformation, hesitancy, vaccination uptake, and geographic covariates. It also reports robustness procedures and a county-level tradeoff between measurement error and coverage.
- Data collection and sources: Around 55 M English-language vaccine-related Twitter posts were collected from January 4th to March 25th, 2021, using the Twitter POST statuses/filter v1.1 API.The keyword list was developed through snowball sampling and contained almost 80 keywords; a restricted four-keyword set covered almost 95% of geolocated tweets, with equivalent results.
- Data collection and sources: Around 1.67 M users were matched to 50 US states, while a subset of 1.15 M users was matched to over 1,300 US counties.Locations were identified from Twitter profiles and assigned using the carmen Python library.
- Data collection and sources: Misinformation prevalence was the regional average of each geolocated account’s proportion of vaccine-related tweets linking to low-credibility news websites.The source list contained 674 low-credibility sources, including sites rated “Very Low” or “Low” in factual reporting and those classified as “Questionable” or “Conspiracy-Pseudoscience”.
- Sensitivity analyses: The main county-level analyses used a minimum threshold of 100 geolocated Twitter accounts, while sensitivity analyses included counties with at least 10 and 50 accounts.The larger threshold likely contains less error but omits more counties.
Additional correlation results
Additional analyses examine correlations among vaccine demand, vaccine hesitancy, political partisanship, and online misinformation at state and county levels.
- Additional correlation results: Figures S1 and S2 report additional correlations involving vaccine demand, vaccine hesitancy, political partisanship, and online misinformation.The analyses are presented at both state and county levels.
- Additional correlation results: The reported correlation results span state-level and county-level comparisons.These comparisons address vaccine-related outcomes, political partisanship, and online misinformation.
Main findings from regression analysis
Regression analyses found that misinformation and Republican vote share explained substantial variation in state-level vaccine hesitancy and vaccination rates. These predictors remained significantly associated with vaccination rates after controlling for additional covariates.
- Regression results: The analyses used weighted and ordinary least-squares regressions of state-level vaccine hesitancy and vaccination rate on covariates.Models 1 and 2 were weighted, whereas Models 3 and 4 were ordinary least-squares regressions.
- Regression results: Misinformation and Republican vote share explain nearly 80% of state-level variation in vaccine hesitancy in Model 1.Both predictors remain significant after multiple control variables are added in Model 2.
- Regression results: Misinformation and Republican vote percentage explain nearly half of the variation in state-level vaccination rate in Model 3.They remain significantly associated with vaccination rate net of controls in Model 4.
- Measures: State-level analyses measured vaccine demand as daily vaccinations per million during March 19–25, 2021, and vaccine hesitancy using Facebook surveys from January 4 to March 25, 2021.Partisanship was the percentage of Republican voters in the 2020 presidential election, while misinformation was vaccine-related content shared on Twitter during January 4 to March 25, 2021.
Sensitivity analyses · Supplementary Bibliography
Sensitivity analyses show that the main findings are robust across alternative transformations, model specifications, misinformation definitions, county inclusion thresholds, and controls. They also support moderation by political context and a strong relationship between misinformation and GOP vote share after adjustment for covariates.
- Sensitivity analyses: Alternative state-level specifications produced results consistent with the main findings, while the untransformed misinformation variable had the better model fit.The logged model addressed positive skewness; lower BIC favored the untransformed variable.
- Sensitivity analyses: There was no evidence of an interaction between misinformation and GOP voters at the state level.The interaction test evaluated whether misinformation effects depended on political partisanship.
- Sensitivity analyses: Restricted-keyword definitions of vaccine misinformation yielded findings consistent and robust with the main analyses at the state and county levels.These checks used narrower sets of keywords to identify vaccine-related misinformation shared by Twitter users.
- Sensitivity analyses: County-level results remained significant with untransformed misinformation and were very similar under a polynomial model capturing nonlinear associations with vaccine hesitancy.The log-transformed county measure was retained for the main findings because it provided the best model fit.
- Sensitivity analyses: Political context moderated the county-level association between misinformation and vaccine hesitancy, with Republican versus Democratic state composition affecting the relationship.A scatterplot of Republican- and Democratic-leaning counties confirmed this moderation finding.
- Sensitivity analyses: Adding county tweet volume as a control did not affect results, and lowering the inclusion threshold from 100 to 50 or 10 Twitter accounts produced similar findings.These analyses addressed variation in Twitter activity and tested robustness to county inclusion criteria.
- Sensitivity analyses: A negative binomial regression confirmed a strong county-level relationship between misinformation and GOP vote share net of potential confounding variables.The model predicted mean percent information using percent GOP vote and control variables.
- Supplementary Bibliography: Supplementary materials document the covariates and provide OLS regression results for Granger causality analyses at the county and state levels.Tables S9–S11 describe covariates and report lagged-variate regressions for 610 counties and 50 states.