Factors Associated With US Adults’ Likelihood of Accepting COVID-19 Vaccination

Key Points Question What factors are associated with US adults’ choice of and willingness to accept a hypothetical COVID-19 vaccine? Findings In this survey study of a national sample of 1971 US adults, vaccine-related attributes (eg, vaccine efficacy, adverse effects, and protection duration) and political factors (eg, US Food and Drug Administration approval process, national origin of vaccine, and endorsements) were associated with preferences for choosing a hypothetical COVID-19 vaccine. Health care attitudes and practices, political partisanship, and demographic characteristics, including age, sex, and race/ethnicity, were also associated with willingness to receive a vaccination. Meaning The results of this survey study may help inform public health campaigns to address vaccine hesitancy.


Additional Analyses
The experiment employed two approaches to measure the associations between vaccine characteristics and public willingness to vaccinate. Each subject completed five choice tasks. In each choice task, subjects evaluated two hypothetical COVID-19 vaccines with randomly assigned attributes as summarized in Table 1 of the text. After reading about each vaccine, subjects were first asked to choose whether they would get vaccine A or vaccine B, or whether they would choose not to get either vaccine. After making a discrete choice between the two vaccines, subjects were also asked to indicate how likely or unlikely they would be to get each vaccine individually on a 7point Likert scale from "extremely unlikely" to "extremely likely." These questions afford a different test of the associations between each attribute and subjects' willingness to receive vaccination. From these questions, we created a dichotomous dependent variable coded 1 if the subject was "slightly," "moderately," or "extremely" likely to receive each hypothetical vaccine profile and 0 if not.
In the text, we employ two different analytical strategies. For the discrete choice question, we estimated a benchmark OLS regression model and plotted average marginal component effect sizes (AMCEs) in the top panel of Figure 1. For the second dependent variable measuring subjects' willingness to receive each hypothetical vaccine individually, we also estimated a benchmark OLS regression model and presented marginal means in the bottom panel of Figure 1. The two measures are closely related. In a fully randomized survey-experimental design, such as the one employed, AMCEs are the differences between marginal means of a given attribute-level and that attribute's baseline level, all else equal. Each approach has advantages for substantive interpretation. AMCEs estimate the association between each attribute-level and vaccination preferences compared to those at the attribute's baseline, whereas marginal means offer information about vaccination likelihood at all feature levels.
To complement the analyses presented in Figure 1 in the text, eFigure 2 presents the marginal means for the discrete choice question (Model 1 of Table 3) and the AMCEs for the individual vaccine likelihood dependent variable (Model 3 of Table 3). The results are substantively similar to those presented in Figure 1. The marginal means are lower in the discrete choice question because subjects could choose between three options here: vaccine A, vaccine B, or neither. However, the pattern of estimated effects is very similar. The AMCEs for the model of responses to the individual vaccine evaluation question are the differences in marginal means of each attribute-level from its baseline. Marginal means for all attribute-levels in both models are presented in eTable 2.
Finally, as robustness checks we re-estimated the analysis of the individual vaccine evaluation (Table 3, Model 3) using different operationalizations of the dependent variable. Model 1 of eTable 3 again uses a dependent variable coded 1 if the subject was "slightly," "moderately," or "extremely" likely to take each hypothetical vaccine profile and 0 if not. This model is identical to that reported in the text. Model 2 of eTable 3 uses a dependent variable coded 1 if the subject was only "moderately" or "extremely" likely to take the presented vaccine and 0 if not. Model 3 analyzes unwillingness to vaccinate, and uses a dependent variable coded 1 for subjects who report being "extremely," "moderately," or "slightly" unlikely to take the presented vaccine, and 0 if not. Finally, Model 4 uses an ordinal dependent variable reporting the full 7-point Likert scale. The results are substantively very similar across operationalizations of the dependent variable. The most important exception is that the coefficient for minor side effects is only statistically significant (i.e. P<.05, two-tailed test) in two of the four models. Protection duration and the Biden endorsement coefficients are significant in three of four models; the coefficients for all other vaccine attributes are statistically significant and in the expected direction for all four models.

Other Factors Associated with Vaccine Acceptance
Models 2 and 4 of Table 3 include a range of additional variables to examine the relationships between vaccine acceptance and subjects' health background and preferences; personal exposure to those affected by COVID-19; beliefs about the future course of the pandemic; political partisanship; and demographics, including gender, age, educational attainment, religious affiliations, and race.
We measure partisanship with two indicator variables identifying subjects who self-identify as a Democrat or a Republican; both measures include those who lean toward one party or the other. Educational attainment is measured on an 8-point scale from less than high school (2% of our sample) to professional degree (4%). Past frequency of flu vaccination is measured on a 4-point scale: never; once or twice; most years; every year. Favorability toward the pharmaceutical industry was measured on a 5-point scale from "very negative" to "very positive." Religious affiliation was measured with three dummy variables: non-Evangelical Christians (defined, following Pew, as Catholics, Protestants, Orthodox Christians, and Mormons who did not self-identify as Evangelical in a follow-up question); Evangelical Christians; and those who identified as atheists, agnostics, or "nothing in particular. The omitted baseline category captures subjects who identified as Jewish, Muslim., Buddhist, or Hindu (just over 10% of our sample, combined). Finally, the models included indicator variables for those identifying as Black or Latino. Note: Regression coefficients and 95% confidence intervals. Model 1 uses a binary dependent variable coded 1 for subjects who report being "slightly", "moderately", or "extremely" likely to take the vaccine. Model 2 uses a binary dependent variable coded 1 for subjects who report being "moderately" or "extremely" likely to take the vaccine. Model 3analyzing unwillingness to vaccinate --uses a binary dependent variable coded 1 if the subject reported being "extremely", "moderately", or "slightly" unlikely to take the vaccine. Model 4 uses the full seven-point ordinal scale ranging from extremely unlikely to extremely likely to take the vaccine presented. The results are very similar across operationalizations of the dependent variable. P-values for coefficients indicated with an † are below the adjusted target P value (adjusting for α = 0.05) calculated via the Benjamani-Hochberg correction controlling for the false discovery rate in multiple comparisons.

eFigure 1: Sample Discrete Choice Set
As you may know, scientists around the world are working to develop a vaccine for COVID-19. Please consider the hypothetical vaccines described in the How likely or unlikely would you be to get Vaccine A described above (7-point scale)?
How likely or unlikely would you be to get Vaccine B described above (7-point scale)?