Is There Broad-Based Support in High-Income Countries for COVID-19 Vaccine Donation? Evidence from Seven Countries

The data were collected as part of the first wave of the CANDOUR study.⁵ Overall, this first wave involved 15,536 individuals from 13 countries. Respondents completed an anonymous survey between 24 November and 28 December 2020 using Qualtrics web-based software. Quota sampling (and in five countries additional weighting) was used to obtain a sample that reflected the distribution of age, education, gender and region in each country.⁶

As in Clarke et al. [12], the present study used information from 8,209 individuals aged 18 years or more from the subset of HIC in the sample (Australia, Canada, France, Italy, Spain, the UK and the USA).⁷ As described in [12], we adapted a question previously used in the context of H1N1 [9] and asked whether respondents supported donating some COVID-19 vaccine doses for distribution to poor countries with insufficient resources to buy their own vaccines. The exact wording was:

Overall proportions of each response category for each country, including 95% confidence intervals (CIs), were reported by Clarke et al. [12]. In this study, we hypothesised that, in each country, the responses favouring a greater emphasis on vaccine provision for poorer countries would vary by levels of altruism, income, political ideology and education, and be greater among:

In exploratory analyses, we also aimed to test whether the responses were heterogenous in subgroups defined by:

The more ‘altruistic’ half of the sample were identified according to responses to a question that asked:

Imagine the following situation: Today you unexpectedly received $700 [US version: amendments by country to currency and amount] . How much of this amount would you donate to a good cause?

There were some missing data (including “do not know” and “prefer not to say”) for education (1.5%), income (24.3%), political ideology (15.9%), altruism (39.9%) and willingness to take risks (7.0%). We wished to avoid excluding these responses from the analysis, particularly in light of the fact that unanswered responses to these questions might group people with similar attitudes to donating vaccines. To avoid excluding these responses, we included ‘unanswered’ as a separate category.

The household income data were equivalised for household composition using the Modified OECD Equivalence scale. A three-category variable was created for “high income” (on or above the median), “low income” (below the median) and “unanswered”.

We created a three-category variable for altruism based on the donation amount. The first category combined “unanswered”, “do not know” and “prefer not to answer”. The next two categories were “meagre”, which was below the median donation amount for that country, and “generous”, which was on or above the median donation amount for that country. Stratifying by country for this variable was necessary because each country had different amounts and units of currency.

Political ideology was identified by asking respondents to indicate where their political views lay on a scale from 0 (“left”) to 10 (“right”). The centre of the distribution on this scale varied greatly between countries. To make a more comparable variable for this multi-country analysis we split each country’s results into tertiles to give a relative measure of political leaning within countries. So this variable had four categories: left, centre, right and unanswered.

For education we created a variable with four categories: primary education or less, secondary, university degree and unanswered.

Willingness to take risks with health was assessed based on a widely used risk-preference indicator originating from the German Socio-Economic Panel (SOEP) survey. Respondents were again asked to use a scale from 0 (“completely unwilling to take risks”) to 10 (“very willing to take risks”). We divided this into two groups defining those with a score of 0–5 as “unwilling” and those with a score of 6–10 as “willing”. We added a third group of “unanswered”.

For gender, we looked at male gender versus all other responses. For age, the categories were up to age 39, 40–59 and 60+ years.

In our main analysis, we used a Bayesian ordered logistic model with the following four ordered categories:

Answers of “Prefer not to say” or “Do not know” were excluded for this analysis.

Our ordinal logistic regression model assumed a latent normal distribution for donation willingness with three cut-points that defined the observed ordinal response [16]. The model was formulated as a cumulative logit model and we assumed that the odds ratio for a one unit change in the predictors was the same across the ordinal responses. The odds ratios are interpreted as the odds of moving to a higher category (greater willingness to donate). The ordered cut-points can be unequally spaced to allow for differences in the proportions of responses in the four categories, and we allowed the cut-points to vary by country, permitting differences between countries in the proportions supporting donation.

We examined the association with this ordinal outcome and the independent variables of age (up to 39, 40–59, 60+ years), gender (male, not male), education (primary or less, secondary, university degree, unanswered), altruism (meagre, generous, unanswered), income (high, low, unanswered), political ideology (left, centre, right, unanswered) and willingness to take risks with health (unwilling, willing, unanswered). We fitted each variable independently and then used all seven in a multiple variable model.

Each model was fitted as a Bayesian ordinal logistic regression model using random effects by country. We allowed the effect of the independent variables to vary by country as we had a strong expectation of differences between countries. For example, the education level “university” is not fixed over countries but instead varies by country around an overall effect. Estimates were therefore made both at country level and overall. We estimated the mean odds ratio and a Bayesian 95% credible interval.

To test whether the independent variables did vary by country, we fitted an alternative fixed-effects model and compared the model fit using the deviance information criterion (DIC), which balances model complexity and fit [17, 18]. We compared the more complex model containing a random effect in each country with the simpler model containing a fixed effect. All models used a random intercept in each country to account for overall differences between countries in the willingness to donate vaccines.

The Bayesian models were fitted in WinBUGS (version 1.4.3) and the plots were made using R (version 4.1.1). We used two chains thinned by three with a burn-in and sample of 4,000. We visually checked the convergence and mixing of the chains. We used vague normal priors for the mean parameters in the ordinal regression model and vague gamma priors for the inverse variances. The Bayesian models and R code are freely available on github (https://​github.​com/​agbarnett/​donate).

We used Bayesian posterior probabilities (BPPs) to compare the groups. The BPPs examine the odds ratios for a group relative to the reference category, for example, oldest age group relative to youngest. The BPP is the estimated probability that the odds ratio is equal to one (the null hypothesis).

It could be argued that there is a case for interpreting responses of “do not know” or “prefer not to say” as being less supportive of donation than a response that clearly expresses some support. As a sensitivity analysis, we therefore repeated the above analysis including the unanswered donation responses, with the following five ordered categories:

As above, we used an ordinal logistic regression model assuming a latent normal distribution for donation willingness, this time with four cut-points that defined the observed ordinal response [16].

The results of the deviance information criteria (DIC) tests indicated that, for every variable, the random model was preferred (Electronic Supplementary Material (ESM), Appendix A). Results from the univariate Bayesian ordered logistic regression models, all with random country effects as indicated by DIC, are plotted in Fig. 2 (see ESM, Appendix B, Tables B1–B7 for all odds ratio estimates and 95% credible intervals.) In most cases, across the full-sample of countries, the odds ratios were not significantly different from 1, as indicated in Table 2 by the 95% credible intervals not containing 1 (i.e., BPP < 0.05). This suggests only limited heterogeneity in willingness to donate COVID-19 vaccines. An exception was political ideology. As hypothesised, those on the left were significantly more likely to support vaccine donation than those on the right. They were also more likely to support donation that those in the centre or those with unanswered responses for political leaning. This trend was consistent in all countries (Fig. 2 and ESM, Appendix B, Table B2). The only other exception was altruism. As also hypothesised, those who were categorised “generous” were more likely to support vaccine donation than those categorised “meagre”. This trend was consistent in all countries (Fig. 2 and ESM, Appendix B, Table B5).

While the odds ratios for university education versus primary or less were in the direction we hypothesised (> 1) in all countries, and were significant in Australia and the UK (Fig. 2 and ESM, Appendix B, Table B1), the 95% credible interval for the overall odds ratio estimate (1.43) contained 1. While the odds ratios for high income versus low income were also in the direction we hypothesised (> 1) in most countries and overall (1.08), all credible intervals contained 1.

Figure 3 displays the estimated odds ratios from a multiple variable ordered logistic regression model with independent variables of age, gender, education, political ideology, altruism, income and willingness to take risks with health (see ESM, Appendix B, Table B8 for all odds ratio estimates and 95% credible intervals).

The results are broadly very similar to the single variable models. Those on the left were again significantly more likely to support vaccine donation than those on the right or centre, both overall and in each country. They were also more likely overall, and in most individual countries, to support donation than those with unanswered political ideology. With regard to altruism, those who were categorised “generous” were again more likely to support vaccine donation than those categorised “meagre,” both overall and in all countries.

The odds ratios for university education versus primary or less were again in the direction we hypothesised (> 1) in all countries, and were significant in Australia, the UK and the USA, but the 95% credible interval for the overall odds ratio estimate again contained 1. As in the single variable models, the odds ratios for high income versus low income were also again in the direction we hypothesised (> 1) in most countries and overall, but all credible intervals contained 1.

The analogous results for the single variable and multiple variable ordinal logistic regression models including unanswered responses to the donation question are reported in ESM, Appendix B (Tables B9 and B10) and Appendix C Figs. C1 and C2. These results were qualitatively very similar to the main analyses. Those on the left were significantly more likely to support vaccine donation than those on the right or centre or with unanswered responses, both overall and in each country. Those who were categorised “generous” altruistically were more likely to support vaccine donation than those categorised “meagre”, both overall and in all countries.

There are two main limitations to the study. First, as is inherent to all online studies, the need for access to a computer or personal device means that the online samples in each country may differ from their populations in important respects. The CANDOUR study country samples are broadly representative of their populations at least on the observable measures we can benchmark against, including gender, education, income and region [15]. However, as discussed by Duch et al. [15], it is possible that the online samples may differ from their populations on important unobservable variables, such as political engagement and general knowledge. This could potentially reduce the generalisability of our results.

Second, the views captured by our survey are a snapshot of public opinion at a particular period of time (November–December 2020). Notably, at that time, COVID-19 vaccines had been found to be effective and the distribution of COVID-19 vaccines was no longer a hypothetical question. While public opinion may well have changed since this time, there are reasons to conjecture that current opinion in HICs may be more strongly supportive of vaccine donation. First, increasingly large proportions of respondents in HIC have now been vaccinated, which may reduce the public’s concerns that donating doses to other countries would mean that they would need to wait longer to be vaccinated. Second, as of November–December 2020, there had been very limited public discussion of the risk that future variants of COVID-19 might emerge that could be more transmissible and, crucially, more resistant to current vaccines. This is no longer the case, and the public may now be more likely to see a link between large numbers of infections overseas and the risk of variants emerging that are more resistant to our vaccines.