Estimating COVID-19 vaccine hesitancy

Interact with the vaccine hesitancy data visualization.

Data source: The Delphi Group at Carnegie Mellon University U.S. COVID-19 Trends and Impact Survey, in partnership with Facebook

We estimate vaccine hesitancy for 32,963 ZIP Code Tabulation Areas (ZCTAs) and 3,141 counties and county equivalents across the United States from the beginning of 2021 onward based on responses in the CMU/Facebook COVID-19 Symptom Survey. Specifically, we look at the following questions:

Have you had a COVID-19 (coronavirus) vaccination?

1. Yes
2. No
3. I don’t know

If a vaccine to prevent COVID-19 (coronavirus) were offered to you today, would you choose to get vaccinated?

1. Yes I would definitely choose to get vaccinated
2. Yes I would probably choose to get vaccinated
3. No I would probably not choose to get vaccinated
4. No I would definitely not choose to get vaccinated

Where if an individual answers that they have not had a vaccination, they are asked the second question regarding sentiments toward receiving a vaccine. We consider two metrics of “hesitance” – a “somewhat” variable, which is formulated as the proportion of total respondents (including those who have already been vaccinated)* who answered, “Yes I would probably choose to get vaccinated” or “No I would probably not choose to get vaccinated”; and an “all” variable that also includes “No I would definitely not choose to get vaccinated.” The data are tabulated based on the provided survey weights^† and stratified in this analysis by ZCTA and week. In order to borrow strength across local geography and capture change in hesitancy over time, we incorporate these data into a cascading spline Poisson model (using week as the independent variable) for each state that first fits to all available data in the state, then each county within that state separately, then lastly each ZCTA that overlaps with a given county. Each step reduces the model dataset to only be inclusive of a more granular geographic unit, while passing on statistical information from the “parent” model to inform the estimation through Bayesian priors. We present results based on the ZCTA models, including county results based on population weighted average of the constituent ZCTA estimates.

*For “somewhat,” the proportion is calculated as the number of respondents who answered (2) or (3) to the second question divided by the total number of respondents who gave any answer to the first question. For “all,” the proportion is the number of respondents who answered (2), (3), or (4) to the second question divided by the total number of respondents who gave any answer to the first question.

†https://cmu-delphi.github.io/delphi-epidata/symptom-survey/symptom-survey-weights.pdf