Survey population

We used the panel data from CVACS 1 (2021) and CVACS 2 (2022), which were implemented by the Southern Africa Labour and Development Research Unit (SALDRU). Study participants were interviewed by telephone in both surveys. Unvaccinated CVACS 1 participants were 18+ years old (n = 3510). A follow-up survey, CVACS 2 (n=3608), was conducted in February–March 2022. Of these, 1386 of them participated and were re-interviewed in CVAC 2. In addition, CVACS 2 also included another 2222 unvaccinated individuals who were interviewed in CVACS 2 the first time.

Measurements

Surveys were linked by their study numbers. We considered the following variables: age, employment, education, residential area, household income, receiving government grants, medical aid/insurance (yes/no), and presence of chronic conditions (e.g., HIV, diabetes, heart conditions). The following questions were asked in both surveys: (1) I will not get infected with COVID-19; (2) Vaccine will prevent death from covid? (3) Vaccine is safe (4) Would you get vaccinated by February/March (in CVACS 1) and May 2022 (in CVACS 2) (5) Would you get vaccinated as soon as possible versus not.

Primary outcome variables

The primary outcomes were: (1) the proportion of survey participants that were vaccine-hesitant, which was derived from the question: “ (a) I will get vaccinated as soon as possible (yes versus no) are categorized as “vaccine hesitant” and (2) Lack of trust in government was derived from the question “How much do you trust vaccine information provided by the government”: (a) a lot, (b) little, (c) not at all, where groups (b) and (c) were combined.

Geographical components: latitude and longitude

We used “district-level” geographical information, which was linked to latitude and longitude using the Maps tools in R-software package.

Statistical analysis

Geospatial investigation of the data was conducted using the geoadditive models, which fit a smoothed bivariate function of latitude and longitude using the logistic regression setting (Kammann and Wand 2003). For example, geospatial variations in an outcome of “vaccine hesitant” vs “not” after adjusting for income, sex, and ethnicity:

$$\log it\lbrack Pt)}_i/1-Pt)}_i\rbrack=f(latitude_i,\;longitude_i)+\beta_1(income)+\beta_2(male)+\beta_3(ethinicity)+error_i,$$

i = 1, …, n for ith individual; in this setting, f is a smoothed bivariate function of latitude and longitude, βi,i = 1, 2, 3 are regression coefficients for the covariates considered in the model; errori ~Normal(0, σ2). Models produced degrees of freedoms (e, d, f) for f, which can be considered the total influence of all observations. In semiparametric regression setting, e, d, f>2 and the corresponding significant p-value (>0.5) provide evidence for the sub-geographical heterogeneity of the outcome of interest where the null hypothesis states that the outcome is uniformly distributed. We used data-visualization techniques to present the intensity of prevalence of an outcome of interest across the region (i.e., image plots), which were created using “Multivariate Generalized Cross Validation” (mgcv) (Wood et al. 2016). To make our methodology available, the scripts were written using the R-software package (3.5.3) (http://cran.r-project.org/) (Appendix).

Population Attributable Risk Percentage

In province specific-analysis, the population-level contributions of the vaccine-related beliefs and attitudes were estimated after accounting for their correlation structure of the variables considered in a multivariable modeling setting. This part of the analysis combined the adjusted odds ratios from a standard logistic regression model and the prevalence of exposure(s) and estimated the population-attributable risk (PAR%) (Wand and Ramjee 2012):

$$PAR=\frac=1-\frac^2_s_s}$$

Where ORs is the odds ratio from a multivariable logistic regression model, and s is the binary response variable of interest. When there is more than one exposure:

$$PAR=\frac^S_s\left(_s-1\right)}_^S_s\left(_s-1\right)+1}=1-\frac^S_s_s}$$

where ORs and Ps, s = 1, ⋯, S, (proportion of the individuals exposed to the sth covariate). Methodology, SAS codes (SAS 9.0), and the Macro (R-3.34) are presented in the Appendix.