Exploratory study of the global intent to accept COVID-19 vaccinations.

Background: As the world begins the rollout of multiple COVID-19 vaccines, pandemic exit strategies hinge on widespread acceptance of these vaccines. In this study, we perform a large-scale global exploratory study to examine the levels of COVID-19 vaccine acceptance and explore sociodemographic determinants of acceptance.
Methods: Between October 31, 2020 and December 15, 2020, 26,759 individuals were surveyed across 32 countries via nationally representative survey designs. Bayesian methods are used to estimate COVID-19 vaccination acceptance and explore the sociodemographic determinants of uptake, as well as the link between self-reported health and faith in the government's handling of the pandemic and acceptance.
Results: Here we show that intent to accept a COVID-19 vaccine is low in Lebanon, France, Croatia, and Serbia and there is population-level polarisation in acceptance in Poland and Pakistan. Averaged across all countries, being male, over 65, having a high level of education, and believing that the government is handling the pandemic well are associated with increased stated acceptance, but there are country-specific deviations. A belief that the government is handling the pandemic well in Brazil and the United States is associated with lower vaccination intent. In the United Kingdom, we find that approval of the first COVID-19 vaccine in December 2020 did not appear to have an impact on the UK's vaccine acceptance, though as rollout has continued into 2021, the UK's uptake exceeds stated intent in large-scale surveys conducted before rollout. Conclusions: Identifying factors that may modulate uptake of novel COVID-19 vaccines can inform effective immunisation programmes and policies. Differential stated intent to accept vaccines between socio-demographic groups may yield insights into the specific causes of low confidence and may suggest and inform targeted communication policies to boost confidence.

Introduction

The rollout of vaccines against the novel coronavirus disease (COVID-19) has begun to populations around the world. While large-scale vaccination manufacture, storage, supply, and delivery all present noteworthy logistical challenges for successful vaccination campaigns, addressing acceptance barriers to COVID-19 vaccines needs equal attention.

Prepandemic evidence from the Vaccine Confidence Project’s (www.vaccineconfidence.org) multinational studies of confidence in vaccines suggests that older age groups are generally among the most confident in vaccines[1-4]. The initial rollout of vaccines to (predominately older age) high-risk groups across many settings is unlikely to meet substantial resistance. Emerging evidence from Public Health England and the Israeli Ministry of Health suggest suggest higher rate of breakthrough infections than was expected via clinical vaccine trials[5,6]; therefore, herd immunity may be “vaccine-assisted” in the sense that vaccines substantially reduce morbidity and mortality while population-level herd immunity builds up through natural infection. However, pandemic exit strategies are widely viewed as relying on achieving vaccination levels that exceed those required for herd/community immunity, which will require uptake from younger age groups, who are among the least likely to state willingness to accept a COVID-19 vaccine in a variety of settings, including France[7], Germany[8], Sweden[8], United States[9,10], and the United Kingdom[11,12] (though there is also evidence that younger age groups are more likely to accept the vaccine in Mexico[8]). In addition, other factors such as gender, education level, income, and ethnicity have been found to be associated with intent to vaccinate in a range of large studies[13]. To date, two multicountry studies exploring national intent to accept a COVID-19 vaccine appear in the published literature, exploring national-level differences in acceptance in June 2020 in 19 countries[14] and from March to September 2020 in 12 countries[8].

This study builds on the growing literature that explores potential uptake of a novel COVID-19 vaccine by widening the number of countries surveyed and by providing a more recent appraisal of vaccine acceptance in many settings. This study also includes countries that were surveyed after the Pfizer–BioNTech vaccine had been submitted for emergency use authorisation[15] and, in the United Kingdom, after this vaccine had both been approved for use[16] and the first patient vaccinated[17].

In this study, intent to accept a COVID-19 vaccine is explored for 26,759 individuals across 32 countries between October 21 and December 15, 2020, using data from the Worldwide Independent Network of Market Research (WIN) World Survey. As of 25 January 2021, these countries represent 73% of the total global mortality burden and among countries with recent and historic vaccine confidence issues such as France[18], Nigeria[19], Pakistan[20], Poland[21], United States[22], and the United Kingdom[23]. A range of possible drivers of COVID-19 vaccine acceptance are considered, including sociodemographic status (sex, age, highest education achieved, employment status, and income); self-reported health (overall health and stress); and perceptions around government handling of the pandemic. Country-level variables are used to explore trends at the national level. A comparison is made between the WIN World Survey data and 10,822 individuals surveyed in June 2020 in 15 of the same countries as surveyed here[14]. Time-varying trends in intent to accept a COVID-19 vaccine are assessed in the United Kingdom where previous survey data[11,24] allow a temporal comparison before and after the approval and introduction of the Pfizer–BioNTech vaccine in the United Kingdom.

We show that intent to accept a COVID-19 vaccine is relatively low in Croatia, France, Lebanon, and Serbia, and highest in Vietnam, India, China, and Denmark. We also find that being male, older, and having a high level of education is associated with higher intent to accept a COVID-19 vaccine, but there are country-level variations around these global trends. Individuals who believe that the government is handling the pandemic well tend to be more inclined to vaccinate, except in Brazil and the United States, where a belief that the government is handling the pandemic well is associated with lower intent to vaccinate.

Methods

Data

Between October 31, 2020, and December 15, 2020, a total of 26,759 individuals aged 18 years or older across 32 countries were surveyed as part of WIN World Survey: Argentina, Brazil, Canada, Chile, China, Croatia, Denmark, Ecuador, Finland, France, Germany, Hong Kong, India, Indonesia, Ireland, Italy, Japan, Lebanon, Malaysia, Mexico, Nigeria, Pakistan, Paraguay, Peru, Poland, Serbia, Slovenia, South Korea, Spain, United Kingdom, United States, and Vietnam. The WIN World Survey is an international survey carried out by WIN every year to measure people’s thoughts, expectations, and perceptions toward relevant topics for society[25]. Surveys are collected via online surveys, computer-assisted telephone surveys, and face-to-face interviews, depending on, for example, countries’ internet penetration and COVID-19 limitations see supplementary Table 1 for further details on survey methodologies, sample sizes, and fieldwork dates. In the survey, respondents are asked “When a vaccine for the coronavirus becomes available, will you get vaccinated?” There are four possible responses (henceforth denoted Y and modelled as an ordinal four-vector): “definitely will get vaccinated” (4), “unsure but probably will get vaccinated” (3), “unsure but probably will not get vaccinated” (2), and “definitely will not get vaccinated” (1). In addition, individual-level sociodemographic data on respondents’ sex, age, highest education level, work status, and income quintile are collected as well as individuals’ self-reported health (“how do you consider your overall health in general?”) and stress (“how often would you say that you suffer from stress?”). Individuals’ perceptions on how their government has handled the pandemic (“please rate […] the way your government [has] handled the [coronavirus] crisis”) are also measured. To allow the statistical investigation into factors that may affect country-wide variation in vaccination intent, country-level data were collected on: the cumulative SARS-CoV-2 deaths per 100,000 population since the start of the pandemic until the beginning of survey fieldwork in each country (deaths, rather than cases, are used as deaths are likely a more robust measure of the state of an epidemic as testing capacities may vary more substantially across countries) and within the most recent two weeks before fieldwork[26]; the Human Development Index (HDI)[27]; GDP per capita (GDP)[28]; and confidence in the importance of vaccines for children[4]. Estimates of country-level confidence in the importance of vaccines for children[4] are used as a proxy for a country’s overall vaccine confidence as they show a strong univariate association with parental acceptance of vaccines for their children (and because there is a high collinearity between national-level confidence in the importance of vaccines for children and in other national-level confidence measures for the safety and effectiveness of vaccines: collinearity may lead to inflated confidence intervals around parameter estimates[29]). The survey questionnaire is provided in the supplementary information. We note that some variables collected were not of interest to us in this exploratory study, for example, public views on the health-system capacities and self-reported rates of smoking or drinking.

During data collection, quotas aligning with national demographic distributions for age, sex, and subnational region were set with survey weights calculated when quotas are not met so that proportions of these demographic traits match national-level distributions. All surveys—with the exception of Ecuador and Vietnam—are therefore nationally representative according to national-level sex, age, and region demographic distributions. In Ecuador and Vietnam, surveys took place in Quito-Guayaquil (Ecuador) and Ha Noi and Ho Chi Minh City (Vietnam) and quotas (and associated weights) were set to align with these subnational regions. Data are collected using online surveys (25 countries, n = 20855 respondents), computer assisted telephone interviews (four, n = 2803), telephone-assisted web interview (one, n = 600), and face-to-face interviews (two, n = 2500), see supplementary Table 1. Informed consent was obtained by WIN for all survey participants.

Statistical analysis

National-level intent to accept COVID-19 vaccines is estimated using a categorical distribution with a Dirichlet prior. The associations between individuals’ uptake intent and individual- and country-level determinants are obtained via Bayesian multilevel regressions[30]. Intent to accept COVID-19 vaccines is related to sociodemographic data (sex, age, highest education level, work status, and income quintile) and self-reported health and stress, as well as perceptions on how their government has handled the pandemic and the country-level factors described in Data. Intent to accept a COVID-19 vaccine is related to these individual- and national-level covariate data via Bayesian ordinal multilevel logistic regressions (detailed below). A total of 95 respondents responded that they “do not know” or provided no response as to whether they would get a COVID-19 vaccine. To avoid the loss of missing data, these 95 responses are recoded to “unsure but probably will not get vaccinated” as they demonstrate some hesitancy about vaccinating, but no strong intent to reject the vaccine. (A sensitivity analysis is performed to test the robustness of this classification against other classification methods, and this does not impact our findings: see supplementary methods for further information).

There were five missing values for vaccination intent, these responses (all from Vietnam) were removed from the analysis as were seventeen unspecified values for covariate data (10 in Germany for income and seven for age in Pakistan). These missing data represent 0.09% of all responses. For all individual-level covariates, a “do not know/ no response” category was created for all covariates to again avoid the loss of missing data. The reference group for individual-level covariates is an employed, female, aged 18–24, with secondary education, in the middle-income group, and who self-reports as healthy, not often stressed, and thinks their government is handling the pandemic badly (see Table 1). For most of these groups (and under our exploratory analysis), there is no strong statistical reason for the selection of the baseline category (e.g., males versus females). However, with regard to employment, we selected the most commonly selected employment option as the baseline group. For income and education, it was of interest to examine the behaviour of the extreme groups with regard to an “average” sociodemographic group. As some country-level covariates are on a different order of magnitude from each other (e.g., GDP and HDI), all level-2 covariates are scaled to the unit interval using to aid prior specification and model convergence.

Table 1

Data summary.

			Definitely will get vaccinated		Probably will get vaccinated		Probably will not get vaccinated		Definitely will not get vaccinated		Do not know /No response
		count	n	%	n	%	n	%	n	%	n	%
Sex	Female	13673	4404	32.2	4973	36.4	2464	18.0	1783	13.0	49	0.4
	Male	13090	4952	37.8	4720	36.1	1949	14.9	1425	10.9	43	0.3
Age	18–24	3593	1322	36.8	1302	36.2	554	15.4	403	11.2	13	0.3
	25–34	5623	1906	33.9	2090	37.2	841	14.9	756	13.4	29	0.5
	35–44	5034	1580	31.4	1821	36.2	907	1.08	700	13.9	25	0.5
	45–54	5041	1654	32.8	1831	36.3	934	18.5	607	12.1	14	0.3
	55–64	4131	1470	35.6	1477	35.8	721	17.4	459	11.1	4	0.1
	65+	3337	1422	42.6	1170	35.1	454	13.6	283	8.5	7	0.2
Employment status	Employed	16445	5786	35.2	5952	36.2	2706	16.4	1943	11.8	58	0.3
	Unemployed	2955	856	29.0	1076	36.4	541	18.3	481	16.3	0	0.0
	Housewife	2464	776	31.5	903	36.7	405	16.4	357	14.5	23	0.9
	Retired/disabled	2838	1233	43.4	972	34.2	402	14.2	225	8.0	6	0.2
	Student	1826	655	35.8	713	39.0	307	16.8	147	8.0	5	0.3
	Refused or do not know (empl.)	234	50	21.4	77	32.9	52	22.2	55	23.5	0	0.0
Highest education	Secondary	11489	3871	33.7	4209	36.6	1986	17.3	1379	12.0	43	0.4
	Primary	1990	662	33.3	692	34.8	289	14.5	332	16.7	14	0.7
	Higher	12536	4603	36.7	4593	36.6	2037	16.2	1272	10.2	31	0.2
	None/other	553	171	31.0	132	23.8	60	10.8	187	33.7	4	0.7
	Refused or do not know (educ.)	195	48	24.5	67	34.5	42	21.3	39	19.7	0	0.0
Income	Medium/low	15343	5234	34.1	5629	36.7	2530	16.5	1895	12.3	56	0.4
	High	8722	3391	38.9	3158	36.2	1306	15.0	844	9.7	24	0.3
	Refused or do not know (inc.)	2688	730	27.1	902	33.5	574	21.4	469	17.5	13	0.5
Health	Unhealthy	5400	1890	35.0	1963	36.4	953	17.6	570	10.6	24	0.4
	healthy	21136	7415	35.1	7656	36.2	3417	16.2	2579	12.2	69	0.3
	do not know (health)	227	51	22.3	75	33.0	43	18.9	59	25.9	0	0.0
Stress	No	3481	1424	40.9	985	28.3	455	13.1	588	16.9	28	0.8
	Yes	23076	7873	34.1	8651	37.5	3902	16.9	2588	11.2	63	0.3
	Do not know (stress)	206	59	28.7	57	27.7	56	27.2	33	15.8	1	0.6
Gov’t handling	Badly	12002	3326	27.7	3950	32.9	2541	21.2	2166	18.0	19	0.2
	Well	13918	5906	42.4	5393	38.8	1670	12.0	885	6.4	63	0.4
	Do not know (gov’t handling)	843	123	14.6	350	41.5	203	24.0	158	18.7	10	1.2

A summary of study factors and raw and breakdown of (weighted) vaccination intent by each covariate used in the study. Note that the total number of weighted responses (26,763) exceeds the total number of respondents (26,759) by four.

Vaccination intent is modelled as , where is vaccination intent for an individual in country/territory ; is an (indicator) matrix of individual-level covariates (which are provided by a binary—or “one hot”—vector according to the sociodemographic status of the individual); are individual-level parameters; and . We use the ordered logistic distribution for ordered outcomes specified by , where and is the standard logistic sigmoid function. This definition is equivalent to the proportional-odds assumption, wherein the difference in the log of cumulative odds ratios between successive categories is independent of the slope β. Individual-level covariates are modelled as for , where country-level covariates are specified in the matrix (Q is the number of country-level covariates); and is the matrix of fixed-effect parameters. A t-distribution is used to allow robust regression of country-specific covariates. Semi-informative normal prior distributions are used for all fixed-effect parameters . Half-normal hyperpriors are placed over variance parameters, . These prior widths (specified as variances) place the vast majority of prior mass over plausible parameter values. All individual- and country-level covariates and their recoding are provided in Table 2. To establish the most parsimonious model for vaccine-uptake intent, three subsets of the hierarchical model specified above are fit: (1) an intercept-and-slopes-as-outcomes model (all model parameters, the “full” model); (2) intercepts-as-outcomes (model 1, with if and ); and (3) the “null” model (no level-1 or level-2 covariates)[31]. The model with the lowest deviance-information criterion[32] is determined to be the most parsimonious model.

Table 2

Study data for Bayesian ordinal multilevel regressions to establish the determinants of intent to accept COVID-19 vaccines.

	Survey item	Values (recode in parenthesis)	Regression baseline
Response variable	COVID-19 vaccination intent (response)
	When a vaccine for the coronavirus becomes available, will you get vaccinated?	Definitely will get vaccinated (4), probably will get vaccinated (3), probably will not get vaccinated (2), definitely will not get vaccinated (1); do not know/no response* (2)	Not applicable (variable is the response)
Individual-level covariates	Socio-demographic
	Sex	Male and female	Female
	Age	18–24, 25–34, 35–44, 45–54, 55–64, 65+	18–24
	Highest educational attainment	None or only basic education (none/other), completed primary school (primary), completed secondary school (secondary), completed high level of education (higher), completed higher level of education, e.g., master/PhD (higher), other educational level (none/other), refused/do not know/no response (do not know or refused, education)	Secondary
	Work status	Working part-time (employed), retired/disabled, student, working full-time (employed), housewife, unemployed, refused/ do not know/ no response (do not know or refused, work status)	Employed
	Income	Low (low/middle), medium low (low/middle), medium (low/middle), medium high (high), high (high), refused/ do not know/no response (do not know or refused, income)	Low/middle
	Self-reported health
	How do you consider your overall health in general?	Very healthy (healthy), healthy (healthy), somewhat healthy (healthy), unhealthy, refused/do not know/no response (do not know or refused, health)	Healthy
	How often would you say that you suffer from stress?	Very often (often), fairly often (often), sometimes (often), occasionally (not often), never (not often), refused/do not know/ no response (do not know or refused, stress)	Not often
	Government handling
		Very badly (badly), rather badly (badly), pretty well (well), very well (well), do not know	Badly
	National statistics
Country-level covariates	COVID-19 mortality (total deaths per 100,000 population preceding fieldwork)	Continuous variable scaled to the range [0, 1]	n/a
	COVID-19 mortality (total deaths per 100,000 population in two weeks preceding fieldwork)	Continuous variable scaled to the range [0, 1]	n/a
	Human development index (HDI) 2019	Continuous variable scaled to the range [0, 1]	n/a
	Gross domestic product per capita (GDP)	Continuous variable scaled to the range [0, 1]	n/a
	Vaccine confidence	Continuous variable scaled to the range [0, 1]	n/a

Questionnaire items from WIN World Survey are shown with possible responses and their recodes (individual-level covariates). Country-level covariate data definitions are also shown (country-level covariates). The baseline for the hierarchical ordinal regression (see “Methods”) is shown for all covariates. (* A sensitivity analysis is performed to assess the impact that recoding “no response / do not know” to “probably will not get vaccinated” has on our findings, see “sensitivity analysis” and supplementary Figs. 1 and 2 in the supplementary information for further details).

A Bayesian linear correlation is used to assess the relationship between vaccination intent in the WIN World Survey and 10,822 individuals surveyed from 15 countries in June 2020 from Lazarus[14] (Brazil, Canada, China, Ecuador, France, Germany, India, Italy, Mexico, Nigeria, Poland, South Korea, Spain, United Kingdom, and United States)[14,33]. The survey question in that study differs from the analysis here; respondents were asked to reply to (on a scale from strongly agree to strongly disagree), “you would accept a [COVID-19] vaccine if it were recommended by your employer and was approved safe and effective by the government.” The aim of this analysis is to show consistency in estimates for national-level intent to vaccinate, despite variation in survey wording. To show trends in the UK’s acceptance of a COVID-19 vaccine, uptake estimates from previous surveys (see Data) are presented with their corresponding credible intervals.

Temporal trends in intent to accept a COVID-19 vaccine are assessed in the United Kingdom before and after the first person was vaccinated with the Pfizer–BioNTech vaccine in the United Kingdom[34] using similar survey data conducted in September[24] (n = 1000) and October 2020[11] (n = 16820).

All inference is performed via Gibbs sampling, with models implemented in JAGS (using R version 4.0.3). In total, 2000 burn-in iterations and 20,000 iterations were sufficient for parameter convergence for all models. Individual-level parameters are reported as odds ratios (OR) or log odds ratios. All parameters are reported with the corresponding 95% highest posterior-density intervals (95% HPDIs). The 95% HPDI is the smallest interval of the posterior distribution that contains 95% of the probability mass. A breakdown of responses to vaccination intent across all individual covariates for all countries can be found in the supplementary data 1 and 2. Individual-level data and their variable recodings are shown in Table 2. No ethical approval for Lazarus[14] data or WIN’s World Survey data was sought as these datasets are in the public domain. Ethical approval for the UK study data was obtained by the Imperial College ethics committee on 15 June 2020 with reference 22130. In the UK dataset, informed consent was obtained from all respondents before they participated in the survey.