Introduction

In January 2020, the World Health Organisation (WHO) declared coronavirus disease 19 (COVID-19) a pandemic. Since then, the world experienced numerous epidemic waves. Around 7 million deaths have been reported as of May 2023 when the WHO declared that COVID-19 “no longer constitutes a public health emergency”1. The true death toll is even higher. COVID-19 vaccines were developed at record speed in response to the emergence of COVID-192, with vaccination campaigns having been rolled out worldwide since late 2020 and early 2021. However, despite strong evidence on the effectiveness of current COVID-19 vaccines for preventing severe illness, hospitalisation, and deaths associated with COVID-193,4,5, vaccine hesitancy remains widespread and an important barrier to high population uptake of the vaccines6.

COVID-19 vaccine hesitancy is complicated and shaped by an array of psychosocial and contextual factors as well as their interactions21, existing studies on COVID-19 vaccine hesitancy or acceptance mainly included individual/interpersonal and vaccine-specific factors29,30,31. Contextual determinants of vaccination uptake received increasing attention but were usually studied separately with other levels of determinants15,20,27. It is suggested that individual and vaccine-specific factors are interconnected to co-shape people’s vaccination decision32. There remains limited understanding about how the contextual factors interact with individual and vaccine-specific factors. In this study, following the framework proposed by WHO SAGE on Immunisation Working Group11, we adopted a network approach to map individual, interpersonal, vaccine-specific and contextual determinants of vaccine hesitancy simultaneously33. Compared with conventional statistical models such as multivariable (linear or logistic) regression models, a mixed graphical network approach has the strengths of incorporating various data formats and allowing mutual interactions amongst variables and hence can reveal complex pathways lying between multilevel factors and vaccine uptake33.

Vaccine hesitancy is also dynamic, changing with the evolving situation13. First, public perceptions of the disease risk versus vaccine risk can change due to the rapid evolution of the pandemic situation and policy response34. Second, media focus may shift from disease risk to vaccine risk as disease incidence in the community declines and vaccines become widely available35. Third, people’s experiences with the new vaccines also evolve as more people initiate their vaccination, which in turn reshapes their attitudes toward the vaccines and intention to take a second or third dose of the vaccines36,37. Furthermore, policies or official recommendations for COVID-19 vaccination have been constantly changed as more evidence on the characteristics of diseases and the efficacy of vaccines became available38. One key change lies in the shift of the initial recommendation for completing two doses of COVID-19 vaccines (the primary doses) to the uptake of a third dose (booster dose) 6 months after receiving the primary doses39 due to observing a decline in neutralising antibodies among vaccinated individuals40. Strategies for promoting the uptake of the primary and booster vaccinations should be different41, highlighting the importance of a tailored approach for vaccination promotion by stage of the vaccination campaign. Overall, it indicates that it is essential to monitor and track the dynamics of determinants of vaccine uptake throughout the vaccination campaign. A few large-scale population-based studies have been conducted to tackle the temporal changes of determinants on vaccination. These studies concluded that the key determinants of vaccine uptake varied by stage of the pandemic and vaccination campaign6,48, our findings provide more nuanced insights into the changing importance of vaccine confidence attitudes for vaccination uptake with the evolution of the vaccination programme. We found that at the initial stage when COVID-19 vaccines became available to the public, vaccine safety was the most important determinant of primary vaccine doses uptake. Concerns about vaccine safety were mainly driven by the novelty of the vaccine technology and its rapid development49, and further intensified by the negative news regarding COVID-19 vaccines and people’s conspiracy beliefs28,50. As uptake of the primary doses became high, people gained more confidence about the safety of the vaccine. However, the resurgence of COVID-19 outbreaks due to the emergence of new virus variants and waning immunity51 could dampen people’s confidence in the vaccine effectiveness. In other words, people’s attention shifted from the potential harms of the vaccine to its expected benefits in the later stage of the vaccination programme.

Our network analysis also consistently identified that persons with chronic disease status were less likely to take the vaccinations across the whole vaccination programme. Relatedly, older age was found to be negatively associated with vaccine uptake in P2, but the association shifted to be positive in P3. Participants’ verbal reasons for vaccine hesitancy revealed that concern about chronic disease status was the main reason for refusing the primary vaccine doses uptake in older adults (Supplementary Table 2). This explains why uptake of COVID-19 vaccine was low among older people before P3 (the Omicron wave), with around 35%, 50%, and 80% of people aged 60–69, 70–79, and 80 years or above, respectively, had not received any dose of COVID-19 vaccines44, resulting in high daily COVID-19 mortality47. The low vaccine uptake among older adults could be attributed to the special contexts in Hong Kong17,52. First, older people in Hong Kong particularly those with chronic diseases received insufficient information and no explicit advice from healthcare workers and their family members on COVID-19 vaccination14. This may link to the facts of lacking continuity in health care shaped by the existing healthcare system53 and that most older people lived alone or with their older partners in Hong Kong54. Second, Hong Kong Chinese older adults generally have lower educational attainment, which limits their ability to utilise health information for making a medical decision. There were only 11% of Hong Kong older adults obtained the tertiary education in 202155, compared to 47.1%, 32.1%, and 38.8%, respectively, in Japan, New Zealand and  the United Kingdom56 where a higher willingness to take a COVID-19 vaccine was reported among older adults19,57,58. However, older age was positively associated with uptake of COVID-19 vaccine during P3, which may be attributable to the joint effect of intensive media attention on deaths among older people due to COVID-19 and the implementation of door-to-door mobilisation of vaccination for older people during this period44.

Previous studies have shown that merely reporting daily case numbers has little impact on behavioural change, including vaccination59,60. Our study found that reporting the number of COVID-19 cases had a negative association with the primary vaccine doses uptake in P1 and P2, and the association with uptake of the booster dose disappeared in P3 and P4. The pattern at the very beginning could be linked to people’s avoidance of public places (such as vaccination centres) to protect themselves against infection. However, the number of reported cases became insignificant in the later stage. This may be attributed to the growing resilience and familiarity of the prolonged public health crisis61 and the fact that vaccination no longer aimed at preventing infection62. Contrary to previous observational studies concluding that more adoption of non-pharmaceutical preventive behaviours was associated with higher vaccine uptake50,63, we observed a consistent negative association between the adoption of alternative protective behaviour (i.e., “avoiding going out”) and vaccination uptake. In addition, adoption of an array of non-pharmaceutical preventive behaviours was all negatively associated with vaccination uptake at P1, when the vaccine was initially rolled out. Several reasons may explain this. First, people who avoided going out may perceive a higher risk of infection by visiting the vaccination venues. Second, people who avoided/were able to avoid going out may perceive a low risk of exposure to the viruses and hence perceived a low need for vaccination. Third, the negative associations between different non-pharmaceutical preventive behaviours and vaccine uptake at P1 indicate that people tended to adopt alternative behaviours that they may perceive to be safer when evaluating the pandemic risk against the vaccine risk. Fourth, the booster dose was promoted through the implementation of vaccine pass which required people to take a booster dose for proof to access certain premises. This may induce psychological reactance, particularly among people with higher vaccine hesitancy64. Although there may be reverse causality that people who had received the vaccine tended to be more relaxed about taking the non-pharmaceutical measures, we ruled out this possibility by running additional chi-square tests between vaccine uptake status and adoption of non-pharmaceutical preventive behaviours. We found that the negative association was mainly driven by the greater proportion of adopting non-pharmaceutical preventive behaviours in the vaccination non-uptake group (see Supplementary Fig. 6). Overall, this finding suggests that the non-pharmaceutical preventive measures remained important for people who would like to avoid any vaccine side effects or other adverse effects14. Studies consistently found that people’s positive traditional Chinese medicine (TCM) value can induce negative attitudes toward western biomedicine including vaccination14,52,65. In Hong Kong, some individuals, particularly older adults14, are more familiar with TCM and perceive that TCM is less invasive compared with western biomedicine66. Adoption of non-pharmaceutical measures is important for pandemic control at the initial stage when vaccines are not available12. However, it also induces complacency psychology and illusory optimism that vaccines are no longer needed67, which was found to be the main barrier for booster dose uptake in our qualitative analysis of participants’ verbal reasons for vaccine hesitancy. In the later stage of a pandemic, strategies should focus on mitigation rather than containment, during which stringent social distancing measures may induce tremendous societal costs and thereby vaccination is of paramount importance particularly for individuals at higher risk of severe disease47. In the later stage, an illusory belief about the effectiveness of non-pharmaceutical preventive behaviours may be detrimental to promoting vaccination uptake68.

We also identified potential innovation adopters that could be targeted at the early stage of a vaccination campaign. Our results suggest that people with higher educational attainment and reported married status were more likely to take the primary vaccine doses in P1 and a booster dose in P3 when the new recommendation or policy for vaccination was introduced. People with higher education have a greater ability to comprehend new interventions and policies and are thereby usually the early innovation adopters69,70. While married people’s early adoption of novel intervention is likely to be driven by prosocial motivation, to take the risk and endure the uncertainty of vaccine safety to protect their loved ones17. Prior studies found no clear associations between educational attainment and marital status with COVID-19 vaccination acceptance in their survey timeVaccine specific factors

Although vaccine accessibility plays an important role in determining vaccination behaviour, it was not applicable in Hong Kong because Hong Kong had procured sufficient vaccine doses for all its residents and made it easily accessible through setting multiple community vaccine centres. Therefore, we collected vaccine confidence variables (perceptions of vaccine safety, efficacy, importance and value alignment)48 as indicators of vaccine-sepcific factors.

Individual factors

This included: (1) risk perception of COVID-19 (perceived personal vulnerability, severity of the diseases and worry about infection)48. A recent study suggested that intention cannot always reflect actual COVID-19 vaccination behaviour85. Our analyses indicated that the self-reported vaccine uptake rates based on our survey were highly consistent with the actual vaccine uptake rates amongst eligible population reported by the Hong Kong government, suggesting that self-reported vaccination uptake was a good proxy for actual vaccination behaviour (see Supplementary Table 7).

Statistical methods

Descriptive analyses and data pre-processing

We first provided distributions of participants’ sex, age educational attainment, and employment status across the four study periods (P1-P4) and compared that with the most recent census data55. Proportions and the corresponding 95% confidence intervals of each dose of vaccination uptake across P1-P4 were weighted to the population’s age and sex using the 2021 census data, respectively. For contextual data—the number of COVID-19 cases and death numbers in the community, were obtained from the official source and plotted the number of report cases from 23 Jan 2020 to 31 Dec 2022. To account for the potential impact of contextual determinants on COVID-19 vaccine uptake, we used a 14-day time window in our main analyses. We also conducted sensitivity checks on 7-day time window, the results remained robust (Supplementary Fig. 5). The number of COVID-19 report cases and deaths were averaged over a 14-day period. Besides, before running the network analysis, we checked the proportion of missing values in all the selected variables. We removed two variables in the network models that had a high proportion of missing values (monthly household income, missing rate 20.8%; one vaccine confidence item, missing rate 28.7%; Supplementary Fig. 7). Then multiple imputation was conducted in R using mice package86 to replace a small proportion (no more than 8%) of missing values in other variables, using predictive mean matching approach for five iterations with 20 imputations.

Network graphs

Data analysis was carried in R version 4.2.3. All models were visualised as network graphs, with ‘nodes’ representing variables and ‘edges’ representing the conditional dependency between the variables33,87. The edge width was intuitively interpreted as the strength magnitude between the variables. The network layout was based on the algorithm of Fruchterman and Reingold from qgraph R-package88, it generates plots with the most strongly associated nodes being placed at the centre of the graph and weakly associated ones at the periphery.

Network estimates

A mixed graphical model (MGM) was applied given that our model involved variables of categorical, continuous and count data33. One advantage of the MGM model is that that it does not require an a priori commitment to any particular data-generating mechanism89. MGM can handle various types of data without unnecessary data transformation. In addition, unlike other models that focus on one specific variable or outcome at a time, MGMs can explore the relationships between all variables simultaneously33,90. In MGM, edges are parameterised as regression coefficients from generalised linear regression models. We used R-package mgm to estimate the pairwise weighted adjacency matrix amongst variables33, then qgraph package was applied to visualise such edge-weights matrix as a network88. We adopted the penalty approach, namely, the least absolute shrinkage and selection operator (LASSO), to obtain a more conservative network estimation91. The LASSO approach shrinks edge weights by setting smaller edges to zero, thereby reducing the chance of getting false-positive findings. For current study, we used the Extended Bayesian Information Criterion and set its hyperparameter to 0.50 to obtain a succinct network92. Codes for replicating all the results are available on the Open Science Framework: https://osf.io/r58e7/.

Relative important determinants of vaccine uptake

As mentioned above, the edge weights can be interpreted as the strength of conditional dependency. Therefore, by assessing the edge weights associated with the vaccine uptake node we can provide intuitive interpretations of the relative importance of various determinants on vaccine uptake in the networks. We ranked the absolute edge weights of all variables that had an edge with the node of vaccine uptake across P1–P4 to assess the dynamic change of the relative importance of determinants associated with vaccine uptake. We extracted the sign (direction of the association) and the relative importance of continuous variables from the weighted adjacency matrix. However, for categorical variables with more than one level (i.e., one parameter for each of the two or more categorical levels), the weight of the edge for categorical variables was measured by the mean of the absolute value of all parameters. Because the signs for categorical variables cannot be directly shown in the graph (they were manifested as grey lines), we further extracted the specific interaction and their direction from the parameterised multinomial regression in glmnet, which models the probability of each level of the  categorical predicted variable with the first category of the predictor variable as the reference category33. The reference category for categorical variables can be found in Supplementary Table 5.

Predictability estimates

Predictability in network analysis refers to how well a node is predicated by all other nodes in the network93. The predictability for categorical variables was estimated using the normalised accuracy, this indicator was computed by the probability of the node that can be predicted by all other nodes in the network after removing the influence that is achieved by the trivial prediction, while that for continuous variables was indicated by the proportion of explained variance (R2). Results were visualised using a grey pie chart surrounding the nodes in the network, with a larger shaded area indicating greater predictability.

Stability

Post-hoc stability analyses for edge-weight parameters were conducted to assess the reliability of the network estimation. The R-package bootnet was used for the stability check91. The accuracy of edge weight estimates was inferred by calculating the 95% confidence interval (95% CI) of the weight estimates using the non-parametric bootstrap. Wider 95% CI indicates more unstable estimate.

Coding of the verbal reasons for vaccine hesitancy

To reduce redundancy, similar reasons for being hesitant or resistant to take a COVID-19 vaccine were combined to represent one reason category. For instance, “I am afraid of vaccine quality” and “I worry about vaccine safety” were coded as “Concerned about vaccine safety”. Two coders (J.Y. and Y.X.) coded the reasons independently and any disagreements were resolved before finalising the coding scheme. Percentage of each reason category was calculated by age groups and by the primary doses and booster doses of vaccine uptake.