Visual policy narrative messaging and COVID-19 vaccine uptake

Our research team conducted a three-wave online panel survey through QualtricsXM of individuals in 50 U.S. states and Washington D.C. in 2021. We contracted with QualtricsXM to administer the survey using their proprietary panels of online survey respondents. These panel participants are recruited from a variety of sources, including website intercept recruitment, targeted email lists, gaming sites, customer loyalty web portals, and social media. Here, we present data from T1 (n = 3,900; launched between January 11 and February 3, 2021) and T2 (n = 2,268; launched between March 22 and April 9, 2021). In these survey waves, we conducted an experiment wherein respondents were randomly assigned to one of four experimental conditions in T1, comprised of three visual policy narrative treatment conditions (“protect yourself,” n = 986, 25%; “protect your circle,” n = 974, 25%; “protect your community,” n = 955, 25%) and a control condition (“get the vaccine”, n = 969, 25%). We conduct two main analyses. Analysis 1 assessed the overall effect of narrative risk messages (Xi) on vaccination (Y1), moderated by political ideology (W), controlling for covariates risk perception, COVID-19 experience, flu vaccine behavior, and demographics. Analysis 2, we test (i) the mediation effects of message conditions (Xi) through affective response (M1) and motivation to vaccinate (M2) on vaccine behavior at T2 (Y1), both as individual mediators and as serial mediators, and (ii) the moderation effects of political beliefs (W) for these mediation effects, as well as for the direct effect of message conditions on affective response (M1), motivation to vaccinate (M2), and on vaccine behavior (Y1), all controlling for covariates risk perception, COVID-19 experience, flu vaccine behavior, and demographics. For the analysis, we use Hayes’ PROCESS models, a regression-based moderated mediation model, the regression coefficients of which are estimated using OLS regression (when M1 or M2 are the outcome variables) and logit regression analysis (when Y1 is the dichotomous outcome variable). For all models, we use 95% confidence intervals, and 5000 bootstrap samples to generate bias-corrected confidence intervals. Given that the Xi variable is a multicategorical variable, both Analysis 1 used two kinds of coding to represent the experimental conditions. The first is indicator coding, which compares each treatment condition to the control. X1 = protect yourself compared to control X2 = protect the circle compared to control X3 = protect the community compared to control The second is Helmert contrast coding, which is a successive comparison of conditions. X4 = the average of all three narrative message conditions compared to control X5 = “protect the circle” plus “protect the community” vs. “protect yourself” X6 = “protect the community” compared to “protect the circle” Thus, in the full statistical model, each path from Xi to other variables includes three different paths, each representing the path from Xi to other variables. Similarly, each moderating effect of W includes interaction terms with Xi.