The links between COVID-19 vaccine acceptance and non-pharmaceutical interventions

Abstract

The information set from which individuals make their decision on vaccination includes signals from trusted agents, such as governments, community leaders and the media. By implementing restrictions, or by relaxing them, governments can provide a signal about the underlying risk of the pandemic and indirectly affect vaccination take-up. Rather than focusing on measures specifically designed to increase vaccine acceptance, this paper studies how governments' non-pharmaceutical policy responses to the pandemic can modify the degree of preventive health behavior, including vaccination. To do so, we use repeated waves of a global survey on COVID-19 Beliefs, Behaviors and Norms covering 18 countries from October 2020 to March 2021. Controlling for the usual determinants, we explore how individuals’ willingness to get vaccinated is affected by changes in government restriction measures (as measured by the Oxford Stringency Index). This relationship is mediated by individual characteristics, social norms (social pressure to conform with what most people do), and trust in government institutions. Our results point to a complex picture as the implementation of restrictions is associated with increased acceptance in some contexts and decreased acceptance in others. The stringency of government restrictions has significant positive correlations with vaccine acceptance in contexts of weak social norms of vaccine acceptance and lower trust in government. In countries or communities with tighter social norms and high trust in health authorities, vaccine acceptance is high but less sensitive to changes in policies. These results suggest that the effect of government policy stringency is stronger among individuals who report lower trust and weaker social norms of vaccine acceptance.

1. Introduction

Vaccine hesitancy, or the reluctance to get vaccinated despite availability can undermine governments’ effort to reduce the burden of diseases and has been a key issue during the COVID-19 pandemic. Even with highly effective vaccines, the level of vaccine take-up in the population has to be very high for the pandemic to be contained and avoid breakdowns of the health systems. Understanding the drivers of vaccine acceptance, and the role that government action plays in it, is thus crucial.

The primary research aim of this paper is to examine the link between vaccine acceptance and governments' non-pharmaceutical policies (lockdowns and other restrictions to mobility). These policies provide information signals about the virus and the risks it poses. And these signals, mediated by individual and social factors, can impact the final choice to vaccinate. Controlling for the usual determinants of vaccine acceptance – such as education, age, other individual characteristics, as well as country level prevalence and mortality rates – we assess the strength of the link between vaccine acceptance and stringency of governments' pandemic restrictions. In this assessment, we also consider the role of social norms, trust in authorities, and individuals’ experience of the pandemic.

Once we control for individual characteristics – like higher education and age which are, predictably, positively correlated with vaccine acceptance, and individual perceptions of risk due to personal knowledge of a COVID-positive cases which are more relevant than country-wide COVID-19 death rates – our main empirical finding is on the strength of the link between vaccine acceptance and non-pharmaceutical policies. We find that a unit increase in the stringency of government policies is associated with a 0.4 percentage point increase in vaccine acceptance. In other terms, an increase from the 25th percentile to the 75th percentile of the sample distribution of policy stringency is associated with a 7.8 percentage point increase in vaccine acceptance, a magnitude larger than the difference in vaccine acceptance between individuals with primary education and those with tertiary education, and about the same size of the effect of knowing someone who has been infected by COVID-19.

This main result has important qualifiers. First, the correlations between policy stringency and vaccine acceptance are significantly stronger in societies with weaker social norms about vaccine acceptance, i.e. in societies where individuals do not feel social pressure to get vaccinated due to their perception that the majority of their peers (or other relevant reference group) will not accept a vaccine. Second, these stronger correlations are also found where trust in government is low. Lockdowns and stringent restrictions plausibly convey a stronger informational signal of COVID-19 risk in these societies, rather than in societies with tighter social norms and with populations trusting their government more. Higher vaccine acceptance in these societies tend to pre-exist the implementation of stringent non-pharmaceutical interventions. These findings are in line with evidence from other studies (Blair et al., 2022; Lazarus et al., 2021; Jabar et al., 2022; Mannan and Farhana, 2020; Kerr et al., 2021).

In addition to offer empirical evidence on the under-researched link between non-pharmaceutical policies and vaccine acceptance, our paper contributes to three branches of the literature on vaccine uptake and adoption of other preventive health measures. Firstly, a large body of work exists on individual drivers or correlates of vaccine take-up. In developing countries, these studies mainly focus on child immunization rates and find that mothers’ education and household socio-economic status are significantly correlated to the probability of immunization (Devasenapathy et al., 2016). Education continues to play a role for adult vaccination, along with acceptance of vaccines for other diseases. Maurer et al. (2009) show that the intention to vaccinate against the novel H1N1 virus is strongly associated with uptake of seasonal influenza vaccinations in the US. The association between individual characteristics and vaccine take-up is often non-monotonic. For example, in a study conducted in Indonesia, illiterate women and women with secondary education are more likely to get their children vaccinated than women with primary education (Streatfield et al., 1990). The authors argue that this is possibly due to higher social compliance with local authorities among women with no education, and greater awareness of the protective function of vaccines among women with higher education.

In addition to their socio-economic characteristics, individuals’ beliefs and experiences are linked with their probability of being vaccinated. Perceptions of the risk of infection influence the decision to get vaccinated. Using data collected at the onset of the pandemic in the United States, Bundorf et al. (2021) find that individual beliefs about COVID-19 risks were related to actual preventive behavior. Those who considered themselves to be at high risk to infection significantly reduced activities which may expose them to the virus. Personal past experiences of disease positively affect the likelihood of getting vaccinated in the future. Jin and Koch (2018) show that this “learning by suffering” effect works both ways and find a negative impact on flu vaccination rate for individuals who get sick in the previous year despite having taken the shot.

The second strand of research deals with socio-cultural factors that influence individual health behavior. Social norms have increasingly been shown to play a role in vaccine acceptance and other preventive health measures (Moehring et al., 2021; Allen IV et al., 2021). The direction of the effect is, however, dependent on the context. Social norms can have a conformity or peer effect such that individuals align themselves with others in adopting or rejecting preventive health measures (Goldberg et al., 2020). On the other hand, due to the positive externalities associated with vaccines, norms can generate a free-riding effect so that individuals benefit from others’ immunity without incurring the associated costs themselves. Using a study in Mozambique, Allen et al. (2021) show that the dominating effect depends on the level of local infection rates. When COVID-19 infection rates are low, individuals tend to free-ride and reduce protective measures but as infection increases, the perceived risk pushes people to conform and take greater precautions by social distancing. The dominance of free-riding or pro-social behavior has also been explained by long-standing cultural values. Gelfand et al. (2021) suggests that countries with greater “cultural tightness”, i.e. stricter adherence to norms and punishments for deviance are better able to control the pandemic. As our empirical analysis shows, the levels of vaccine acceptance in a society are not static. Since individual uptake of vaccines depends on external factors like access to information, perception of risk and social norms, there is a vital role for policy to “nudge” or influence individual behavior.

The third body of work is on effective policy action. The urgency and severity of the pandemic has spurred several studies focusing on informational interventions about the risks of exposure to COVID-19, effectiveness of preventive measures like masks (Abaluck et al., 2021), social distancing and vaccines, and updating people about prevailing social norms (Moehring et al., 2021). Interventions in the form of mandates or lockdowns also convey information about the threat of the virus.

Glaeser et al. (2021) find that when local governments lift mobility restrictions, individuals interpret this as a signal that the underlying risk of contracting COVID-19 is low and therefore indulge in more risky behavior – for instance, visiting an indoor restaurant. If the information signal is inaccurately interpreted – that is, if the level of virus circulation is not as low as individuals expect it to be – the lifting of restrictions can result in increased COVID-19 infections or lower rate of vaccinations. Further, government mandates serve to reduce negative externalities arising from the relationship between subjective risk perceptions and preventive behaviors (Betsch et al., 2015; Bundorf et al., 2021). If government action alters these risk perceptions, it has also the potential to affect the take up of COVID-19 vaccines – a topic which, so far, remains mostly unexplored. Recent evidence on the increase in interest for vaccination following the United States’ CDC recommendation for vaccinated users not to wear a face mask (see CNN, 2021) suggests that government action has indeed a role to play.

Apart from government mandates and mobility restrictions, other policy interventions which have seen some success in increasing vaccine take-up include behavioral nudges such as reminders to get vaccinated (Dai, et al., 2021), messaging campaigns about its benefits and social acceptance (Argote Tironi et al., 2021) and tackling misinformation (Loomba et al., 2021). According to evidence in Sweden however, monetary incentives to get vaccinated have more bite than information campaigns (Campos-Mercade et al., 2021). If we consider that estimates of the value of a cure for COVID-19 range between 5 and 15 percent of total global wealth (Acharya et al., 2020), the use of monetary incentives or behavioral nudges may well be cost-effective (Costa-Font et al., 2021).

The rest of this paper is organized as follows: the next section discusses the theory behind the relationship between vaccine acceptance, government stringency and other mediating factors. Section 3 describes the data and empirical methods. Section 4, starting from descriptive statistics and trends of the data, presents the main results. Section 5 concludes with some reflections on limitations and further research.

2. Theory

The intention or decision to vaccinate is the result of a complex process where macro-level objective variables such as health risks and government policy are filtered by individual characteristics and experiences. Betsch et al. (2015) summarize this process in graphical form in Fig. 1 . This figure highlights that vaccine acceptance depends on perceptions of risk (of both the disease and side effects of the vaccine) and a set of other factors including individual characteristics, such as age, gender and education levels, and contextual characteristics, such as social norms, attitudes, identity, and barriers. This relation can be represented as:

$$A_{ict}=f\left(P_{ict},X_{ict},Z_{ct}\right)$$ (1)

where $A_{ict}$ is the vaccine acceptance of individual i, in country c, at time t, and $P_{ict}$ is the corresponding perception of risk, while $X_{ict}$ and $Z_{ct}$ are the individual level and macro (country or community level) determinants. Note that in this simple relation, some individuals may have the same perception of risk, but their decision to vaccinate may differ because they are living in communities with more or less restrictive social norms, or simply because of their different trust in authorities, or different age or gender, or even because governments implement vaccination campaigns which affect the costs, time and effort to vaccination, identified as barriers in Fig. 1.

Fig. 1.

Framework for empirical analysis.

Source: Betsch et al. (2015).

Perceptions of risk, in turn, are based on objective available information and individual characteristics, as follows:

$$P_{ict}=g\left({\mathit{Inf}}_{ict},X_{ict}\right)$$ (2)

The set of variables ${\mathit{Inf}}_{ict}$ represents objective information about both the severity of the disease and possible side-effects of the vaccine. This information set includes several pieces: the individual's own experience of the disease, either in her past (as, for example, in the “learning from pain” of Jin and Koch, 2018) or among her social network; the overall community level of morbidity/mortality from the disease; other facts about the disease and vaccine disseminated by health authorities; and, key for the purpose of this paper, government non-pharmaceutical restrictions which, as highlighted by Glaeser et al. (2021), convey relevant information about risks.

The simple theory framework of Fig. 5 embedded in equations (1), (2) is flexible enough to include what Betsch et al. (2015) label the “four C” model. This model postulates that people do not vaccinate because of these four motives: complacency, convenience, confidence, and calculation. More in detail, individuals may not engage in preventive behaviors, such as vaccination, if they perceive the risks from the infectious disease as too low. Alternatively, when there are barriers to vaccination because of scarce availability, low geographical accessibility, complexity of the process (for example multi-shot vaccine), or other difficulties, there may be willingness to vaccinate but avoidance to overcome these barriers results in non-compliance. In the third case, individuals may have a degree of mistrust in health authorities, the science behind the vaccine, or towards motivation of policy makers mandating the vaccine. In this case, and differently from the first two, individuals actively decide against vaccination. Finally, the calculators may carefully consider pros and cons but may end up in a situation where benefits from vaccination do not clearly outweigh the costs and decide to wait and see, potentially free riding on other people vaccinating. Other theoretical contributions deepen our understanding of vaccine acceptance. Cucciniello et al. (2022) focus on the importance of altruism. They show that by providing clearer information on the externalities of taking the vaccine (when a person vaccinates, she provides benefits for herself but also for others, and especially those who cannot vaccinate because of age or other health issues), vaccination rates increase. Martinelli and Veltri (2021) focus on the cognitive approach, i.e. how people's thinking style – whether intuitive or analytic – impact the way in which information about vaccine and disease is processed.

Fig. 5.

Community norms, trust in government and vaccine acceptance.

Source: Authors calculations using the Facebook COVID-19 Beliefs, Behaviors & Norms Survey (2021). Note: Data are weighted using sample survey weights. Dotted line indicates linear fitted values. All 67 countries in the survey (including the 18 repeated cross-section countries which our core analysis sample) are included in this figure. Values correspond to the average across the period July 2020–March 2021.

In our empirical analysis, by adopting the flexible theoretical framework just described, we control for these potential mechanisms and thus we can assess the specific association between non-pharmaceutical government policies – those related to mobility and activity restrictions (a.k.a. “lockdown policies”) – and vaccine acceptance.

3. Data and methods

3.1. Survey details

We use data from a global survey on COVID-19 Beliefs, Behaviors and Norms jointly conducted by researchers at the Massachusetts Institute of Technology, Facebook, Johns Hopkins University (JHU), the World Health Organization (WHO), and the Global Outbreak Alert and Response Network (GOARN) (see https://covidsurvey.mit.edu/and Collis et al., 2022). The survey was administered via Facebook to a sample of its users in 67 countries from August 2020 to March 2021. We restrict our main regression analysis to 18 countries (Argentina, Bangladesh, Brazil, Colombia, Egypt, India, Indonesia, Japan, Malaysia, Mexico, Nigeria, Pakistan, Philippines, Romania, Thailand, Turkey, United States, Vietnam) for which data is available over multiple 2-week long waves so that we are able to control for country and time fixed effects in a panel setting. Specifically, waves 7-17 administered between October 2020–March 2021 are used with the full set of variables relevant to our analysis. While survey respondents are mostly urban, aged between 20 and 50 years and with tertiary education, the data contain weights to represent either each country's adult or its adult Internet-using population, rather than just Facebook users, and to reduce bias due to nonresponse. The survey uses non-response modelling and post-stratification techniques to design how Facebook users were sampled and weighted. Using regularized regressions, the probability of responding to the survey is modeled as a function of user attributes and the non-response weights are calculated as the inverse of their response probabilities. Next, the survey sample is compared with demographic data available from national censuses to adjust weights such that the sample is representative of each country's adult population age and gender composition and finally, trimmed for outliers (refer to Collis et al., 2022 for more details).

The survey contains information on willingness to be vaccinated, perception of community norms around preventive behaviors, access to information about the virus and trust in media and the government. It also includes data on individual characteristics and personal experience of the pandemic, with regard to health and employment. Online appendix 4 contains a detail of the survey questions and variable definitions used in our study. Globally, 69 percent of the sampled individuals report that they will accept the vaccine with some regional variation ranging from slightly above 50 percent in North America to relatively higher acceptance in South Asia at around 75 percent (Fig. 2 ). The vaccine acceptance rate from the COVID-19 Beliefs, Behaviors & Norms Survey is similar to rates found in other studies, e.g. 71.5 percent of participants in a study of 19 countries by Lazarus et al. (2021) reported that they would accept a vaccine. Higher acceptance of around 80.3 percent is found in Low- and Middle-Income Countries relative to only 64.6 percent in the US (Arce et al., 2021).

Fig. 2.

Vaccine acceptance around the world.

Source: Authors calculations using the COVID-19 Beliefs, Behaviors & Norms Survey. Data are weighted using sample survey weights. All 67 countries in the survey (including the 18 repeated cross-section countries which our core analysis sample) are included in this figure. Values correspond to the average across the period July 2020–March 2021.

We complement the individual level survey data with the Stringency Index collected by the Blavatnik School of Government at Oxford University. This index measures how stringent are government's non-pharmaceutical interventions during the pandemic. The index is collected at the national level (with some sub-national data for selected countries) with a daily frequency and covers almost all countries and regions of the world. The underlying information used to calculate the Stringency Index is collated from publicly available information on government responses and containment measures such as school closures, travel bans and other restrictions in movement (Hale et al., 2021). The process of compiling and aggregating information to build the stringency index may induce measurement error of the actual government intervention but it is currently the most reliable and globally comparable measure available. More details on the implications of measurement error on our estimations is discussed below. Lastly, we use data on the incidence of the virus from the COVID-19 Data Repository by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University (Dong et al., 2020). Summary statistics of the variables used in our study for the full analysis sample are presented in Table 1 . Online appendix Table A1 contains country-level summary statistics of some key variables for the 18 countries included in our analysis sample.

Table 1.

Descriptive statistics for the analysis sample.

	Mean	Standard deviation	Minimum	Maximum	Non-missing observations
Individual characteristics
Age 20–30 years	0.311	0.463	0	1	349,237
Age 31–40 years	0.220	0.414	0	1	349,237
Age 41–50 years	0.173	0.379	0	1	349,237
Age 51–60 years	0.148	0.355	0	1	349,237
Age 61–70 years	0.102	0.303	0	1	349,237
Age 71–80 years	0.039	0.194	0	1	349,237
Age 80+ years	0.007	0.083	0	1	349,237
Male	0.529	0.499	0	1	346,927
Urban	0.808	0.394	0	1	349,020
Primary education or less	0.049	0.215	0	1	348,285
Secondary education	0.211	0.408	0	1	348,285
Tertiary education	0.740	0.439	0	1	348,285
In good health (self-assessment)*	0.786	0.410	0	1	347,284
Experiences and subjective beliefs
Will accept vaccine	0.671	0.470	0	1	345,392
Perception of high risk of infection*	0.399	0.490	0	1	344,758
Knows positive case	0.617	0.486	0	1	338,870
Trusts govt. health authorities*	0.438	0.496	0	1	343,906
No. people out of 100 who will vaccinate*	65.813	28.603	0	100	346,937
Country-level circumstances
Stringency index	64.295	9.021	30	88	349,550
Days with stringency index above 80	63.917	54.589	0	300	349,550
Vaccination campaign underway	0.258	0.437	0	1	349,550
Daily COVID deaths per million (7-day avg.)	1.457	2.394	0	17	349,550

Note: Sample weights are applied. * indicates variables with imputed values. For these variables, the statistics correspond to averages over 10 imputations. See online appendix 2 for more details on the treatment of missing data.

3.2. Empirical methodology

The first assessment of the relationship between vaccine acceptance and non-pharmaceutical policies uses a reduced form equation that is derived by combining equations (1), (2) of our theoretical framework. The following regression specification shows vaccine acceptance as a function of information (which in equation (2) was determining risk perceptions), individual and aggregate characteristics:

$$A_{ict}=\alpha +\beta {NPIPolicy}_{ct}+\gamma X_i+\delta {COVID}_{ct}+\pi {VaxAccess}_{ct}{+\mu }_t+{\theta }_c+{\epsilon }_{ict}$$ (3)

More specifically, $A_{ict}$ is an indicator of whether individual i residing in country c and observed in the survey period t is willing to get vaccinated. We only observe individuals once, as we do not have a panel sample but rather a series of repeated cross-sectional samples for every country. $NPI{Policy}_{ct}$ is the government's non-pharmaceutical interventions at the time of the survey period as measured by the Oxford Stringency Index. The vector $X_i$ includes individual-level covariates such as education, employment and demographic characteristics. Individuals' information set includes also the objective epidemiological situation (COVID-19 prevalence and death rates at the country level during the survey period), and personal experience related to the pandemic (measured at the individual level). Access to the vaccine is expected to have a direct effect on vaccine acceptance, and is captured by a dummy variable indicating whether a vaccination campaign against COVID-19 was underway at the time of the survey in the respondent's location. Country and time fixed effects are included by ${\theta }_c$ and ${\mu }_t$ and ${\epsilon }_{ict}$ is the idiosyncratic error term. In terms of the estimation method, we will use Ordinary Least Squares (OLS). The estimated coefficients of equation (3) should be understood as partial correlations and not causal relationships. An analysis of causal effects would require a different empirical setup than the one we use in this study.

An aspect of our analysis to be aware of is that our measure of government pandemic response -the stringency index-is a de jure indicator, in the sense that it measures the policies as they were designed in paper. The implementation of these policies, however, may have differed from their design. The indicator we use could then introduce some degree of measurement error with the respect to the actual implementation of policies. This would induce an attenuation bias in our estimates of the association between government stringency and vaccine acceptance. Our empirical estimates can thus be understood as a lower bound of the actual effect.

In addition to the reduced form of specification (3), we also estimate separately equations (1), (2) of our theoretical framework, i.e. the relation between information and subjective perceptions of risk (equation (2)), and between perception of risk and vaccine acceptance (equation (1)). We also assess how vaccine acceptance is influenced when we interact independent variables, such as information and perceptions of risk with social norms and trust in public authorities.

To control for country level and month fixed effects, our sample is restricted to the 18 countries where the survey was carried out repeatedly, at the pace of one country-level wave per month. The variables on trust and social norms were only included in the survey starting in October 2020, therefore we restrict our sample to the period October 2020–March 2021.

As there is a substantial number of observations with missing data on key covariates -such as the level of trust in government health authorities, the beliefs about community vaccine acceptance, the perceptions of infection risk, and the self-assessment of health status-we perform a multiple imputation analysis which we detail in the online appendix section 2 following the recommendations of Sidi and Harel (2018). The main assumption of this analysis is that data in these variables is missing at random conditional on observable covariates. The results using multiple imputation analysis are our baseline results, and the results using complete case analysis are included in the online appendix.

Lastly, it is important to assess the statistical power of our analysis. Our empirical analysis consists of a multivariate regression analysis in an observational setting – that is, we are not manipulating the treatment variable in an experimental setting. Moreover, the sample set up is complex as it consists of a series of cross-sectional samples of different countries. We have thus carried out a simulation exercise in order to assess whether the sample size at hand allows to detect the presence of a statistically significant association between government stringency and vaccine acceptance. This exercise is detailed in online appendix section 3. The analysis shows that the sample size that we are using in this study allows to detect the presence of an effect even if smaller than the one found in comparable studies.

4. Results

4.1. Descriptive statistics

Vaccine acceptance is heterogeneous across countries -ranging during the sample period from as low as 55% in Egypt to as high as 80% in Bangladesh- and is not static over time (see online appendix table A1 for country level averages of vaccine acceptance, trust, social norms and other variables included in the analysis). The two panels of Fig. 3 present the evolution of attitudes, or personal beliefs, towards the vaccine and the evolution of social descriptive norms about the vaccine. More in detail, panel a presents the evolution of the share of individuals willing to accept a vaccine in each country in the panel sample over the period November 2020–March 2021. In countries like the United States, attitudes towards vaccine acceptance increased from about 40% of the sample in early November 2020 to around 70% in late January 2021. In Brazil, they increased from around 65%–85% in the same period, while in India they remained relatively stable around or slightly above 70%. Panel b presents the social expectations about vaccine acceptance in the community – a descriptive social norm about vaccine acceptance measured by “the number of people out of 100 that the respondent believes will take the vaccine”.

Fig. 3.

Trends in vaccine acceptance (panel a) and beliefs about community acceptance (panel b).

Source: Authors calculations using the Facebook COVID-19 Beliefs, Behaviors & Norms Survey. Data are weighted using sample survey weights. The variable in panel b is missing for around 23% of the respondents. The values plotted in this panel correspond to complete cases only. See Online appendix 2 for more detail on the treatment of missing data.

Compared to individual attitudes towards the vaccine, the variation over time is smaller for social norms. This is expected as social norms do not tend to change swiftly. However, some shifts are also recorded for social norms: in the United States, on average individuals thought that around 55 people out of 100 in their community would accept a vaccine in November 2020. This figure increased to around 65 people out of 100 by March 2021. In Brazil the increase in believed vaccine acceptance moved from slightly below 70 people out of 100 in the community in November 2020 to around 75 people out of 100 in March 2021. In India, a small but consistent decrease in believed vaccine acceptance in the community was observed during the same period – from slightly above 70 people out of 100 to around 65 people out of 100. The fact that individual vaccine acceptance shows greater variation over time than social norms, which tend to be stickier, suggests that factors other than social norms may be explaining the time variation in vaccine acceptance. Cross-country differences in vaccine acceptance, however, could potentially be explained by differences in norms.

The main time variant factor which this paper is interested in is government policy. When correlating the evolution of vaccine acceptance and government policy response – as measured by the Oxford Government Stringency Index, different patterns emerge. In countries like Brazil, the evolution of vaccine acceptance mirrors the evolution of government restrictions: the more stringent measures are, the higher the level of vaccine acceptance (Fig. 4 , panel a). In other countries, like Pakistan, vaccine acceptance and the stringency of government response appear to be uncorrelated over time (Fig. 4, panel b). Lastly, in some other countries like Argentina (panel c) or Turkey (panel d), it is a particular government action -the start of the vaccination campaign-which appears to increase vaccine acceptance.

Fig. 4.

Evolution of vaccine acceptance and government stringency.

Source: Authors calculations using the Facebook COVID-19 Beliefs, Behaviors & Norms Survey (2021). Data on vaccine acceptance are weighted using sample survey weights.

The different patterns in the relationship between government stringency and vaccine acceptance suggest that factors like trust in government institutions, social norms and the evolution of the pandemic itself may be relevant mediators in how individuals react to government action. Country level correlations show that individual vaccine acceptance and descriptive norms about vaccine acceptance in the community are positively correlated but show considerable variability (Fig. 5 , panel a); a similar pattern is observed in the correlation between vaccine acceptance and trust in government health authorities (Fig. 5, panel b).

4.2. The role of government policies and individual characteristics

4.2.1. Individual characteristics and vaccine acceptance

The results in the baseline specification (equation (3)) using multiple imputation analysis across 10 different imputations (Table 2 , column 2) indicate that higher vaccine acceptance appears to be associated to being male, on average 7.3 percentage points more willing to take a vaccine if offered than women; being of older age, those 20–50 years old are on average 10 percentage points less willing to take a vaccine than those 80 years and older; and those having tertiary education, around 5 percentage points more willing on average than those having primary education only and 3 percentage points more willing than those having secondary education only. The correlations with age and gender are consistent with the mortality risk by COVID-19, a fact that suggests that risk perceptions appear to be a significant factor driving individual vaccine acceptance. The positive correlation with education levels – higher vaccine acceptance for more educated people – does not seem to reflect risk perceptions, as more educated people tend to have lower COVID-19 related mortality (Hawkins et al., 2020) but, rather, may be related to differences in information and knowledge about the pandemic and the vaccines. Lastly, individuals’ assessment of their own health does not appear to be correlated with increased or decreased acceptance.

Table 2.

Main specification (OLS).

Dependent variable: Willing to vaccinate
	Complete cases only	Multiple imputation analysis
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Stringency index	0.0030***	0.0029***	0.0029***	0.0033***	0.0035***	0.0049	0.0030***
	(0.0008)	(0.0008)	(0.0008)	(0.0009)	(0.0010)	(0.0044)	(0.0008)
1-day change in stringency index (absolute)			−0.0028**
			(0.0012)
7-day change in stringency index (absolute)				−0.0017*
				(0.0009)
14-day change in stringency index (absolute)					−0.0016*
					(0.0008)
Stringency index^2						−0.0000
						(0.0000)
Days above stringency score 80							−0.0011***
							(0.0003)
Vaccination campaign dummy	0.1037***	0.1057***	0.1056***	0.1062***	0.1048***	0.1061***	0.1051***
	(0.0206)	(0.0207)	(0.0206)	(0.0199)	(0.0194)	(0.0206)	(0.0200)
Daily deaths per million (7 day avg)	0.0075	0.0075	0.0074	0.0072	0.0071	0.0078	0.0061
	(0.0070)	(0.0070)	(0.0070)	(0.0069)	(0.0068)	(0.0072)	(0.0068)
Knows positive case	0.0857***	0.0851***	0.0851***	0.0849***	0.0848***	0.0851***	0.0847***
	(0.0102)	(0.0099)	(0.0098)	(0.0098)	(0.0097)	(0.0099)	(0.0098)
Individual attributes
male	0.0723***	0.0727***	0.0727***	0.0727***	0.0727***	0.0727***	0.0728***
	(0.0137)	(0.0135)	(0.0135)	(0.0135)	(0.0135)	(0.0135)	(0.0135)
urban	0.0136	0.0180	0.0180	0.0181	0.0180	0.0180	0.0180
	(0.0101)	(0.0110)	(0.0110)	(0.0110)	(0.0110)	(0.0110)	(0.0110)
age 20-30	−0.0964***	−0.0971***	−0.0970***	−0.0968***	−0.0970***	−0.0972***	−0.0971***
	(0.0270)	(0.0297)	(0.0296)	(0.0296)	(0.0296)	(0.0296)	(0.0296)
age 31-40	−0.1026***	−0.1055***	−0.1054***	−0.1052***	−0.1054***	−0.1056***	−0.1054***
	(0.0272)	(0.0293)	(0.0292)	(0.0293)	(0.0293)	(0.0292)	(0.0292)
age 41-50	−0.0931***	−0.0997***	−0.0996***	−0.0994***	−0.0996***	−0.0997***	−0.0995***
	(0.0272)	(0.0294)	(0.0294)	(0.0295)	(0.0295)	(0.0294)	(0.0293)
age 51-60	−0.0743***	−0.0826***	−0.0826***	−0.0825***	−0.0826***	−0.0826***	−0.0824***
	(0.0228)	(0.0246)	(0.0246)	(0.0247)	(0.0247)	(0.0246)	(0.0245)
age 61-70	−0.0303	−0.0438*	−0.0438*	−0.0437*	−0.0439*	−0.0438*	−0.0436*
	(0.0223)	(0.0215)	(0.0215)	(0.0215)	(0.0215)	(0.0215)	(0.0215)
age 71-80	0.0046	−0.0056	−0.0056	−0.0054	−0.0056	−0.0056	−0.0055
	(0.0148)	(0.0141)	(0.0141)	(0.0141)	(0.0141)	(0.0141)	(0.0140)
Primary education or less	−0.0498***	−0.0543***	−0.0543***	−0.0544***	−0.0545***	−0.0543***	−0.0542***
	(0.0166)	(0.0181)	(0.0181)	(0.0181)	(0.0181)	(0.0181)	(0.0181)
Secondary education	−0.0305***	−0.0325***	−0.0325***	−0.0325***	−0.0325***	−0.0325***	−0.0326***
	(0.0079)	(0.0085)	(0.0085)	(0.0085)	(0.0085)	(0.0085)	(0.0085)
In good health	−0.0034	−0.0038	−0.0038	−0.0038	−0.0038	−0.0038	−0.0037
	(0.0053)	(0.0042)	(0.0042)	(0.0042)	(0.0041)	(0.0042)	(0.0042)
Constant	0.3880***	0.3950***	0.3907***	0.3690***	0.3551***	0.3324**	0.4721***
	(0.0476)	(0.0488)	(0.0498)	(0.0573)	(0.0607)	(0.1473)	(0.0633)
Observations	232,292	336,471	336,471	336,471	336,471	336,471	336,471
R-squared	0.058	0.057	0.057	0.057	0.057	0.057	0.058

Standard errors in parentheses clustered at country level. All estimations include country and month fixed effects.

Reference category for age is age 81–90 and reference category for education is tertiary education.

*p < 0.05 **p < 0.01 ***p < 0.001.

4.2.2. Epidemiological situation

Personally knowing someone who tested positive for COVID is strongly associated with increased vaccine acceptance: those that know a positive case are on average 8.5 percentage points (95% CI 6.42–10.61) more willing to take a vaccine if offered than those that don't know one. However, the country level death rate in the week before the individual was interviewed bears no correlation with vaccine acceptance. This suggests the importance of the information set the individual uses to assess the risk of COVID-19 and to decide about vaccination. Rather than the overall epidemiological situation, it is the immediate experience with the pandemic that is associated with increased willingness to get vaccinated.

4.2.3. Government policy response

The stringency of government policy response to the pandemic appears to be positively associated with vaccine acceptance: the more the stringent the measures are, the higher the willingness of individuals to take the vaccine. Each additional unit in the Oxford Stringency Index is associated to a 0.29 percentage point higher vaccine acceptance (Table 2, column 2; 95% CI 0.12–0.47). While this effect is larger than the effect of government stringency on other individual variables found in other studies (Aknin et al. (2022) find an effect equivalent to 0.07 percentage points in mental health, and Götz et al. (2021) find an effect equivalent to 0.09 percentage points in sheltering-in-place), it is comparable to the effect of policy and contextual changes on vaccine acceptance. Moving from percentile 25 in the distribution of the index within our sample (a value of 55.09) to percentile 75 (a value of 74.54) would be associated to a 5.4 percentage point increase (95% CI 2.33–9.14) in individual vaccine acceptance, a magnitude that is close to two thirds of the one associated to knowing personally a positive case. The effect is also comparable to the 4.6 percentage point increase in vaccine acceptance associated to an informational treatment on the need to achieve herd immunity like the one implemented by Argote Tironi et al. (2021) in 8 countries of Latin America. Another useful reference for the magnitude of the effect are the findings of Karaivanov et al. (2022) on the effect of vaccination mandates on vaccine uptake. They find that, across Canadian provinces, vaccination mandates increased cumulated vaccine uptake by up to 5 percentage points 13 weeks after the provincial mandate announcements.

A change in lockdown policies, however, would not be associated with an immediate increase in vaccine acceptance. The nature of the data at hand - including the exact day when the survey was carried out for each individual – allows us to assess the effect of actual changes in the stringency index, which is also measured daily. Columns 3, 4 and 5 of Table 2 include as additional regressors the 1-day change, 7-day change and 14-day of the stringency index respectively. The negative coefficients associated to the change variables indicate that when government policy changes, the effect on vaccine acceptance is not immediate. In particular, the coefficients associated to the 7-day and 14-day change suggest that during that period the overall effect of an increase of one unit in the stringency index is 0.15 and 0.19 percentage points respectively – lower than the 0.29 percentage point result in the longer term. A non-linear effect cannot be accurately estimated as the quadratic polynominal of stringency is not statistically significant (column 6). Further, prolonged and very stringent measures seem to have an adverse effect on vaccine acceptance (column 7). Moving from the 25th percentile in the distribution of days above a stringency score of 80 (equal to 0 days in our estimation sample) to the 75th percentile (equal to 80 days, incidentally) is associated with a 9.1 percentage point decrease (95% CI -14.20 to −4.05) in individual vaccine acceptance. This phenomenon, or diminishing effect of non-pharmaceutical interventions, is witnessed also with respect to actual spread of the disease as well and has been termed as lockdown fatigue in other work (Goldstein et al., 2021).

An important aspect of government policy response to the pandemic that is not captured by the stringency index -which is focused on non-pharmaceutical interventions-is the deployment of a vaccination campaign. The first country to launch a COVID-19 vaccination campaign was the United Kingdom on December 8, 2020. In the weeks that followed, several countries -including the 18 countries in our sample-started their vaccination campaigns. The different specifications of Table 2 all show that, as expected, the presence of an active vaccination campaign is associated with an increase in vaccine acceptance of around 10 percentage points (95% CI 6.16–14.97 in the specification of Table 2, column 2), a magnitude higher than the effect of personally knowing a COVID-19 positive case or almost double the effect of an increase in the stringency index from the 25th to the 75th percentile of the distribution. In any case, the fact that changes in the stringency index are associated to changes in vaccine acceptance even when accounting for vaccine access indicates that both policies -pharmaceutical and non-pharmaceutical- can have an effect on individuals' willingness to accept the vaccine. This result is robust to the use of an alternative indicator of vaccine access, such as the availability of a vaccine for the respondent's age group (online appendix table A.5).

4.3. The role of risk perceptions

Individuals form their risk perceptions using the information set available to them and their individual circumstances, like their age, gender or education. As our analytical framework indicates with equations (1), (2), these risk perceptions, mediated by other factors such as trust and social norms, drive vaccine behavior. While it is out of the scope of this paper to provide a detailed analysis of the formation of risk perceptions, the data allows to have a preliminary look at the relationship between vaccine acceptance and risk perceptions about COVID-19 infection. The survey we use includes a question on the likelihood that a person of the same age and in the same community as the respondent becomes sick from COVID-19. We create a dummy variable that indicates whether a respondent stated that this event was “very” or “extremely likely”. We call this variable the “perception of high risk of infection”. Columns 1 and 2 of Table 3 correspond to equation (2) of our framework and look at the correlations between risk perceptions and information and individual characteristics. Men, people living in rural areas, and younger individuals have all a perception of a lower risk of becoming sick from COVID-19. People in good health perceive their risk to be low as well. Interestingly, individual beliefs appear to be associated to risk perceptions in an unusual direction: individuals who have higher trust in government health authorities and those who believe that many individuals in their community will accept the vaccine also report higher infection risk perceptions (Table 3, Column 2). While this correlational analysis does not allow to establish the exact causal nature of this relationship, it suggests that individuals who do not trust health authorities and who do not believe that the community in which they live will accept the vaccine may systematically underestimate the actual spread of the disease and the severity of the pandemic. Misinformation could be a potential explanation, as Roozenbeek et al. (2020) find that individuals who distrust scientists are more susceptible to misinformation about the pandemic.

Table 3.

The role of risk perceptions (OLS, multiple imputation analysis).

	Dependent variable
	Perception of high risk of infection		Willing to vaccinate
	(1)	(2)	(3)	(4)
Stringency index	0.0008*	0.0007*	0.0028***	0.0026***
	(0.0004)	(0.0004)	(0.0008)	(0.0008)
Stringency index X Perception of high risk of infection				0.0004
				(0.0004)
Perception of high risk of infection			0.1047***	0.0778***
			(0.0055)	(0.0240)
Vaccination campaign dummy	−0.0002	−0.0040	0.1057***	0.1056***
	(0.0061)	(0.0061)	(0.0206)	(0.0206)
Daily deaths per million (7 day avg)	0.0039**	0.0037**	0.0071	0.0070
	(0.0014)	(0.0013)	(0.0069)	(0.0069)
Know positive case	0.0629***	0.0617***	0.0786***	0.0786***
	(0.0059)	(0.0057)	(0.0093)	(0.0093)
+Mediating factors
No. people out of 100 who will vaccinate		0.0006***
		(0.0001)
Trusts govt. health authorities		0.0442***
		(0.0050)
Individual attributes
male	−0.0420***	−0.0433***	0.0771***	0.0771***
	(0.0060)	(0.0058)	(0.0133)	(0.0133)
urban	0.0587***	0.0590***	0.0119	0.0119
	(0.0058)	(0.0054)	(0.0106)	(0.0106)
age 20-30	−0.0843**	−0.0774**	−0.0883***	−0.0882***
	(0.0262)	(0.0254)	(0.0283)	(0.0283)
age 31-40	−0.0240	−0.0181	−0.1030***	−0.1030***
	(0.0253)	(0.0247)	(0.0280)	(0.0280)
age 41-50	0.0002	0.0036	−0.0997***	−0.0996***
	(0.0248)	(0.0241)	(0.0281)	(0.0281)
age 51-60	−0.0111	−0.0096	−0.0814***	−0.0814***
	(0.0245)	(0.0240)	(0.0234)	(0.0234)
age 61-70	−0.0226	−0.0221	−0.0414*	−0.0414*
	(0.0219)	(0.0218)	(0.0207)	(0.0207)
age 71-80	−0.0345	−0.0341	−0.0020	−0.0020
	(0.0192)	(0.0194)	(0.0138)	(0.0138)
Primary education or less	−0.0178*	−0.0199**	−0.0524**	−0.0524**
	(0.0085)	(0.0082)	(0.0179)	(0.0179)
Secondary education	−0.0115**	−0.0118**	−0.0313***	−0.0313***
	(0.0049)	(0.0048)	(0.0083)	(0.0084)
In good health	−0.0719***	−0.0751***	0.0037	0.0037
	(0.0049)	(0.0050)	(0.0040)	(0.0040)
Constant	0.4408***	0.3899***	0.3488***	0.3605***
	(0.0315)	(0.0296)	(0.0486)	(0.0476)
Observations	337,653	336,612	336,471	336,471
R-squared	0.181	0.184	0.067	0.067

Standard errors in parentheses clustered at country level. All estimations include country and month fixed effects.

Reference category for age is age 81–90 and reference category for education is tertiary education.

*p < 0.05 **p < 0.01 ***p < 0.001.

When introduced as a separate, independent regressor in equation (3), this “perception of high risk of infection” variable is associated with a 10.5 percentage points (95% CI 9.26–11.68) higher probability of accepting the vaccine (Table 3, column 3). The coefficients associated to the stringency index and the personal knowledge of a positive COVID-19 case in that same specification remain significant but decrease slightly in magnitude when compared to the baseline estimates (Table 2, column 2). A similar pattern is observed if risk perceptions are interacted with the stringency index (Table 3, column 4).

These results confirm the relevance of risk perceptions as highlighted by our analytical framework. Government policies and individual experience of the pandemic have an effect on vaccine acceptance through risk perceptions. However, the fact that both variables still have a considerable explanatory power suggests that this particular measure of risk perception does not capture fully the risk perceptions that individuals have in mind. Other dimensions of risks beyond individual infection may be driving vaccine acceptance.

4.4. Trust and social descriptive norms

Information and individual circumstances form risk perceptions, which are in turn mediated by trust and social norms as individuals decide on accepting the vaccine or not. The different specifications of Table 4 explore the role that mediating factors such as social descriptive norms (the number of people in the community that the individual believes are going to get the vaccine) or trust in government health authorities play in individual vaccine acceptance. Column 1 presents a simple specification in which both variables are included as independent regressors. Both are strongly and positively correlated with vaccine acceptance. The more individuals believe vaccine acceptance to be the prevalent social norm in their community, the higher their own willingness to get the vaccine. For every additional percentage point of believed acceptance in the community, individual vaccine acceptance increases by 0.6 percentage points (95% CI 0.50–0.62). Moving from percentile 25 in the distribution of believed vaccine acceptance in the community (which represents a value of about 50 persons out of 100 accepting the vaccine in the community) to percentile 75 (a value of about 87 persons out of 100) is associated with a 15 percentage points (95% CI 13.63–16.85) increase in individual vaccine acceptance.

Table 4.

Mediating effects: trust and descriptive norms (OLS, multiple imputation analysis).

Dependent variable: Willing to vaccinate
	Whole sample		Trust in govt health authorities		Whole sample	Belief about community acceptance of vaccine
			Low	High		Low	High
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Stringency index	0.0023***	0.0028***	0.0031***	0.0023***	0.0036**	0.0031***	0.0019***
	(0.0007)	(0.0008)	(0.0009)	(0.0006)	(0.0017)	(0.0008)	(0.0006)
Stringency index X Trust in govt health authorities		0.0001
		(0.0007)
Stringency index X No. People out of 100 who will vaccinate					−0.0000
					(0.0000)
Vaccination campaign dummy	0.0808***	0.1018***	0.1033***	0.0977***	0.0825***	0.0996***	0.0794***
	(0.0168)	(0.0198)	(0.0196)	(0.0211)	(0.0170)	(0.0250)	(0.0143)
Daily deaths per million (7 day avg)	0.0062	0.0072	0.0079	0.0055	0.0063	0.0081	0.0059
	(0.0053)	(0.0066)	(0.0065)	(0.0071)	(0.0056)	(0.0073)	(0.0048)
Know positive case	0.0778***	0.0836***	0.0924***	0.0699***	0.0786***	0.1030***	0.0656***
	(0.0083)	(0.0094)	(0.0117)	(0.0083)	(0.0084)	(0.0142)	(0.0059)
Mediating factors
No. people out of 100 who will vaccinate	0.0056***				0.0073***
	(0.0003)				(0.0017)
Trusts govt. health authorities	0.1087***	0.1480***
	(0.0068)	(0.0444)
Individual attributes
male	0.0594***	0.0736***	0.0789***	0.0649***	0.0582***	0.0551***	0.0654***
	(0.0112)	(0.0130)	(0.0142)	(0.0130)	(0.0114)	(0.0139)	(0.0127)
urban	0.0124	0.0216*	0.0259*	0.0158	0.0096	0.0102	0.0122
	(0.0077)	(0.0103)	(0.0126)	(0.0091)	(0.0081)	(0.0107)	(0.0077)
age 20-30	−0.0526**	−0.0887***	−0.1091***	−0.0495*	−0.0569***	−0.0789**	−0.0534**
	(0.0185)	(0.0291)	(0.0367)	(0.0265)	(0.0187)	(0.0327)	(0.0230)
age 31-40	−0.0642***	−0.0993***	−0.1237***	−0.0538*	−0.0669***	−0.1001***	−0.0566**
	(0.0174)	(0.0288)	(0.0360)	(0.0270)	(0.0175)	(0.0307)	(0.0227)
age 41-50	−0.0708***	−0.0987***	−0.1208***	−0.0563**	−0.0700***	−0.1078***	−0.0565**
	(0.0172)	(0.0289)	(0.0368)	(0.0246)	(0.0172)	(0.0294)	(0.0235)
age 51-60	−0.0660***	−0.0847***	−0.1031***	−0.0481**	−0.0634***	−0.0980***	−0.0512**
	(0.0143)	(0.0240)	(0.0319)	(0.0213)	(0.0145)	(0.0275)	(0.0202)
age 61-70	−0.0347**	−0.0453**	−0.0563*	−0.0217	−0.0329**	−0.0716**	−0.0215
	(0.0143)	(0.0211)	(0.0280)	(0.0225)	(0.0145)	(0.0298)	(0.0177)
age 71-80	−0.0018	−0.0066	−0.0104	0.0009	−0.0007	−0.0400	0.0054
	(0.0125)	(0.0137)	(0.0198)	(0.0220)	(0.0129)	(0.0284)	(0.0136)
Primary education or less	−0.0591***	−0.0605***	−0.0695***	−0.0466**	−0.0548***	−0.0443*	−0.0516***
	(0.0162)	(0.0170)	(0.0182)	(0.0170)	(0.0169)	(0.0247)	(0.0148)
Secondary education	−0.0332***	−0.0326***	−0.0365***	−0.0259**	−0.0334***	−0.0258**	−0.0309***
	(0.0070)	(0.0083)	(0.0088)	(0.0095)	(0.0071)	(0.0115)	(0.0061)
In good health	−0.0206***	−0.0103**	−0.0144**	−0.0024	−0.0167***	−0.0254***	−0.0044
	(0.0039)	(0.0043)	(0.0048)	(0.0058)	(0.0038)	(0.0063)	(0.0031)
Constant	0.0495	0.3426***	0.3179***	0.5165***	−0.0095	0.2097***	0.5606***
	(0.0370)	(0.0502)	(0.0641)	(0.0338)	(0.1042)	(0.0567)	(0.0409)
Observations	336,471	336,471	204,253a	132,218a	336,471	117,850a	218,621a
R-squared	0.178	0.080	0.062	0.042	0.167	0.063	0.044

Standard errors in parentheses clustered at country level. All estimations include country and month fixed effects.

Reference category for age is age 81–90 and reference category for education is tertiary education.

*p < 0.05 **p < 0.01 ***p < 0.001.

^a– average sample size across 10 imputations.

Similarly, trust in government health authorities is associated with, on average, almost 11 percentage points (95% CI 9.41–12.32) higher vaccine acceptance. It is noteworthy the fact that the effect of age on the degree of vaccine acceptance is diminished when social norms and trust are included as explanatory variables – the magnitude of the partial correlation coefficients associated to the different age group dummies is almost halved when comparing Table 2, column 2 to Table 4, column 1. This suggests that a substantial part of the difference in vaccine acceptance across age groups can be attributed to differences in beliefs across age.

The conceptual framework we use suggests, however, that beliefs such as trust and social descriptive norms are not independent factors in the vaccine acceptance decision process, but they rather act as “filters” of the effects brought by differences in information or personal and social characteristics. Columns 2 and 5 in Table 4 present the results of an alternative specification in which trust in health authorities and beliefs on vaccine acceptance in the community are treated as interacting factors of government stringency. These interaction models do not show results that are qualitatively different from the specification in Column 1. Columns 3–4 and 6–7 present an alternative analysis by which the sample is split in two groups for each variable – individuals with low trust in government health authorities (column 3) or high trust in government health authorities (column 4); and individuals whose social context is one of loose social norms about vaccination (i.e. individuals believing that 50 or fewer people out of 100 in the community would take the vaccine, column 6) or those experiencing stringent social norms (i.e. believing that 50 or more out of 100 in the community would take the vaccine, column 7).

The results of the subsample analyses show that the effects of government stringency and the personal acquaintance with a positive COVID-19 case on vaccine acceptance are of lower magnitude among those who have high trust in government health authorities compared to those with low trust, while the average level of vaccine acceptance is substantially higher (columns 3–4). Individuals who don't trust health authorities appear to be more susceptible to change their minds with respect to vaccine acceptance when governments implement more strict non-pharmaceutical interventions or when they personally know someone who got infected. The age effect on vaccine acceptance -by which younger individuals are less willing to accept the vaccine than older ones- and the education effect -by which more educated individuals are more willing to take the vaccine-are considerably larger in magnitude among those who distrust government health authorities than among those who trust them. In this sense, trust in government health authorities seems to be a powerful mediating factor: changes in the information set of individuals, either triggered by government policies or by personal experiences, and differences in individual characteristics have a sizeable effect on vaccine acceptance only among individuals who do not trust health authorities. Individuals who trust them, instead, appear to have a uniformly higher willingness to take the vaccine, irrespective of individual circumstances. Part of this could be a mechanical effect, as individuals with higher vaccine acceptance may have less “room” to improve their acceptance for any given contextual change (the marginal effect can only be smaller the closer one is to 1 in a binary variable such as the decision to accept or not the vaccine). However, this result could also be associated to a smaller informational effect of contextual changes in this subgroup of the population. In fact, Blair et al. (2022) find that, while individuals who trust in government tend to comply more with public health measures, they are not at the same time more knowledgeable about COVID-19 – if anything, their knowledge about the disease may be even more limited than those who distrust government.

A similar pattern is observed in the subsample analysis of the descriptive social norm on vaccine acceptance. The effects of government stringency, knowledge of a positive COVID-19 case and even the existence of a vaccine campaign are of a lower magnitude among those who believe community acceptance of the vaccine is high than among who believe that is low (column 7 versus column 6). Just as in the case of trust in health authorities, the age effect on vaccine acceptance is only present among those who believe community acceptance is low. The education effect is present in both subsamples. These results show that descriptive social norms are also a powerful mediating factor: when individuals believe that social norms about vaccination are loose, their decision to take the vaccine will be explained by differences in the information and individual circumstances which form their risk perceptions. When individuals believe that social norms about vaccination are stricter, these variables are less relevant in explaining individual vaccine behavior. These individuals also have a high vaccine acceptance conditional on their information and characteristics.

A graphic depiction of the mediating effect of trust and social norms in vaccine acceptance is presented in Fig. 6 . Panel “a” shows the predicted probability of accepting the vaccine for individuals with low trust in government health authorities under two scenarios – in the first one, government stringency is low (at the 25th percentile of the sample distribution) and the respondent doesn't know personally any COVID-19 positive case; in the second scenario, government stringency is high (at the 75th percentile of the sample distribution) and the respondent knows personally a COVID-19 positive case. In the first scenario the probability of accepting the vaccine is 50.0% (95% CI 47.9%–52.0%), while in the second scenario the probability is 65.3% (95% CI 62.9%–67.6%) – a difference of 15 percentage points. Panel “b” presents the same two scenarios for individuals that have high trust in government health authorities. In this case, the difference is smaller – about 11.5 percentage points, between 69.6% probability of accepting the vaccine in the first scenario to 81.1% in the second scenario.

Fig. 6.

Trust, norms and changes in stringency and personal experience

Note: this graph plots the predicted probability (including the 95% confidence interval) of accepting the vaccine for different groups under two scenarios. In the first scenario (circle markers), the stringency index takes a value equal to that of the 25th percentile of the sample distribution (55.09) and the dummy variable indicating whether the respondent knows a COVID-19 positive case takes a value of zero; the remaining covariates are set at their sample means. In the second scenario (square markers), the stringency index takes a value equal to that of the 75th percentile of the sample distribution (74.54) and the dummy variable indicating whether the respondent knows a COVID-19 positive case takes a value of one; the remaining covariates are set at their sample means. Panel a. presents the predicted probability for the two scenarios for the subsample of individuals who report high trust in government health authorities. The estimation corresponds to the model in column 3 of Table 4. Panel b. presents the predicted probability for the two scenarios for the subsample of individuals who report low trust in government health authorities. The estimation corresponds to the model in column 4 of Table 4. Panel c. presents the predicted probability for the two scenarios for the subsample of individuals who believe vaccine acceptance in their community is low (defined as believing that 50 or fewer people out of 100 in their community will accept the vaccine). The estimation corresponds to the model in column 6 of Table 4. Panel d. presents the predicted probability for the two scenarios for the subsample of individuals who believe vaccine acceptance in their community is low (defined as believing that more than 50 people out of 100 in their community will accept the vaccine). The estimation corresponds to the model in column 7 of Table 4.

Panels “c” and “d” repeat the same exercise, but this time distinguishing individuals between those who believe community acceptance of the vaccine is low (loose social norms) and those who believe community acceptance is high (stringent social norms). For the first group, the difference between the scenario of low stringency and no personal knowledge of a positive COVID-19 case and the scenario of high stringency and personal knowledge of a positive COVID-19 case is 16.4 percentage points – between a 36.6% probability of accepting the vaccine in the first scenario and 53.0% probability in the second scenario. For the group who believe community acceptance is high, the difference between both scenarios is smaller – about 10 percentage points, between 70.2% probability in the first scenario and 80.5% in the second scenario.

4.5. Heterogeneous effects across countries

Government management of the pandemic and public reception to lockdown policies have varied across the globe. While some countries witnessed backlash or protests to lockdowns and mask-mandates, others have not. To examine heterogeneity of impact across countries, we estimate Equation (3) (using the specification of Table 2, Column 3) including an interaction of the stringency index with country fixed effects. Fig. 7 plots the marginal effects of stringency across countries. The average effect of stringency across countries shows quite some variation: while in all countries the effect appears to be positive, in some countries the estimate is not statistically significant – suggesting a rather weak relationship between stringency and vaccine acceptance. This is the case of Argentina, Egypt, Indonesia, Romania, Turkey, and the United States. In the opposite side of the spectrum are Bangladesh, Brazil, India, Malaysia, Mexico, Nigeria, Pakistan, Thailand and Vietnam, where the association between stringency and vaccine acceptance is statistically stronger and of a larger magnitude.

Fig. 7.

Heterogeneous effects of stringency on vaccine acceptance across countries

Note: this graph plots the average marginal effect on vaccine acceptance associated to an increase of one unit of the stringency index for different countries. The average marginal effect is obtained from the estimation of the baseline specification (equation (1)) including an additional interaction term between the stringency index and the country dummy variables. The upper and lower bounds indicate the 95% confidence interval.

The mediating role of trust in government health authorities and beliefs about community acceptance also shows heterogeneity across countries. In Fig. 8 , we present the marginal effect of the stringency index on individual vaccine acceptance across countries distinguishing between individuals who have low or high trust in government health authorities (panel a) and between individuals who believe vaccine acceptance in the community is low or high (panel b). While on average the degree of vaccine acceptance of individuals who have high trust in government health authorities is less sensitive to changes in stringency index than the vaccine acceptance of those individuals who have low trust, this pattern is reversed in some countries. For instance, in Egypt, Romania, Turkey, and the United States, the individuals who have high trust in government health authorities react more to changes in stringency. In some countries like Colombia, Japan, Mexico, and Pakistan, there is no difference between the sensitivity of vaccine acceptance to stringency index across both groups of individuals. When looking at differences across groups of individuals based on the social norms prevalent in their communities, in most countries the point estimate of the effect of stringency on vaccine acceptance is higher for those who believe community acceptance is low than for those who believe community acceptance is higher, although in most cases the difference is not statistically significant. In some countries, like Thailand and Vietnam, the differences between the two groups – those experiencing tight norms versus those experiencing loose norms – is significantly large. In other countries, like Egypt, Japan, Indonesia, Romania, Turkey, and the United States, individuals who believe community acceptance is low see their individual vaccine acceptance decrease when the stringency index increases – the opposite reaction vis-à-vis their peers in the rest of the world.

Fig. 8.

The mediating role of trust and social norms across countries

Note: these graphs plot the average marginal effect on vaccine acceptance associated to an increase of one unit of the stringency index for different countries. Panel a provides an estimate for the effect among individuals with low trust in government health authorities and among individuals with high trust in government health authorities. Low trust in government health authorities include individuals who declare that they “do not trust” or “somewhat trust” government health authorities as a source of COVID-19 news and information. High trust in government health authorities include individuals who declare that they “trust” government health authorities for the same purpose. The average marginal effect is obtained from the estimation of the baseline specification (equation (1)) including additional interaction terms between the stringency index, the country dummy variables, and a dummy variable indicating whether the individual belongs to the low trust or high trust group. Panel b provides an estimate for the effect among those who believe vaccine acceptance in their community is low and among individuals who believe vaccine acceptance in their community is low. The average marginal effect is obtained from the estimation of the baseline specification (equation (1)) including additional interaction terms between the stringency index, the country dummy variables, and the variable indicating the number of people out of 100 in the community that the respondent believes will accept the vaccine. The values plotted in the graphs correspond to the average marginal effect when the respondent believes 25 out of 100 in their community will accept the vaccine (i.e. they believe community acceptance is low) and to the average marginal effect when the respondent believes 75 out of 100 in their community will accept the vaccine (i.e. they believe community acceptance is high). In both panels the upper and lower bounds of the point estimates correspond to the 95% confidence intervals.

5. Conclusion

In this paper we have analyzed the relationship between government policies, personal experience and vaccine acceptance, exploring the mediating factors of trust and social norms. We find that an increase in the stringency of government restrictions and a personal experience related to the pandemic – knowing personally someone who tested positive for COVID-19 – is associated with increased vaccine acceptance. Government policies, just like personal encounters with the disease, change the information set of individuals and thus their risk perceptions about the pandemic and the vaccine. By signaling the salience of the spread of the disease, the implementation of more stringent policies may induce individuals to be more willing to take a vaccine. On average, the change in vaccine acceptance associated with an increase in government stringency from the percentile 25th to the percentile 75th of the sample distribution is about 7.8 percentage points, a magnitude larger than the difference in vaccine acceptance between individuals with primary education and those with tertiary education, and about the same size of the effect of knowing a positive case of COVID-19. The effect of government stringency on vaccine acceptance is not immediate, as it takes at least two weeks for an increase in stringency to translate fully into increased vaccine acceptance and, also, the effect becomes smaller the higher the initial stringency index is and the longer it has been at very high levels, suggesting a certain degree of “pandemic fatigue”.

The effect of government policy stringency on vaccine acceptance is remarkably stronger among individuals who distrust health authorities or who live in community with loose social norms about vaccination. These individuals appear to be more sensitive to changes in government policies or to the personal experience of getting to know a positive COVID-19 case. Individuals more trustful of government health authorities or who believe community acceptance is high are less sensitive to changes in government policies and their personal experience. Our results point to the importance of government policy stringency on vaccine acceptance, particularly in contexts where trust in government health authorities is low or where the prevailing social norm is for vaccine acceptance to be low. The implementation of stringent restrictions, beyond helping to curtail the spread of the disease by restricting mobility may also contribute indirectly to reducing the circulation of the virus in another way – by increase vaccine acceptance where it is low. More generally, our findings underline the importance of the effects of non-pharmaceutical policies on health behaviors. Many government policy actions, while not targeted directly to induce a given health behavior, may end up having an effect on it by way of modifying the informational environment with which individuals interact.

In closing, we highlight a few limitations and avenues for future research. An important caveat is that our data focuses on the intention to vaccinate and not on actual vaccine uptake. There is thus possibility that the strength of the link between non-pharmaceutical policies and immunization is weaker if there is a gap between intentions and actual choices to vaccinate. The second limitation is that the role of social norms may also be not precisely assessed. Empirical measurement of social norms is in its infancy, and the data we have on social are better than using simple personal attitudes, but there is still margin for improvement. Finally, a key limitation is the impossibility to establish causality with the current data; however, experimental data (see for example Banerjee et al., 2021) on a large immunization program offers some causal evidence on the relationships we examine in this paper. In terms of future research, in addition to remove the mentioned limitations, we can point out that the theoretical framework and the empirical analysis while focused on vaccine acceptance can be easily adapted to other preventive actions (such as mask wearing or simply screening exams). Also, the relationship between non-pharmaceutical measures and the acceptance of preventive health measures can be used to study communicable diseases beyond COVID-19.

Credit author statement

Maurizio Bussolo, Nayantara Sarma, Iván Torre: Conceptualization, Investigation, Methodology, Data curation, Writing/draft preparation, Reviewing, and Editing,

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Handling Editor: Social Epidemiology Editorial Office

Appendix A. Supplementary data

The following is the Supplementary data to this article.

Data availability

The authors do not have permission to share data.