COVID-19 Vaccination Willingness and Reasons for Vaccine Refusal

Key Points Question What were the reasons for COVID-19 vaccine refusal in Hong Kong, and what policy measures may be associated with higher vaccine uptake? Findings This cohort study used data from 28 007 interviews over 20 waves, including 1114 participants in the latest wave in 2022, found that 75.0% of vaccine refusal could be attributable to mistrust in health authorities, low vaccine confidence, misconceptions, and political views. The vaccine pass policy was associated with an increase in vaccination appointments. Meaning These findings suggest that building trust in health authorities, promoting vaccine confidence, and countering misinformation may be fundamental for better preparedness and response to future pandemics.

Participants in Singapore reported their trust in COVID-19 information sources (i.e.WHO, government health authorities, academics, and traditional and social media platforms) on a scale of 1-10 (1=complete mistrust, 10=complete trust) in the baseline (May 2020 to June 2021).A score of ≤5 was considered as mistrust.
The three major misconceptions about COVID-19 vaccines were assessed using the following statements in Hong Kong: "Older individuals have a greater need for COVID-19 vaccination", "Individuals with more chronic diseases have a greater need for COVID-19 vaccination", and "COVID-19 vaccines are more harmful than COVID-19 infection" in waves 15-20.Disagreement with the first two statements and agreement with the last statement were considered as vaccine misconceptions.Opposition to COVID-19 vaccination in adults aged 80 years was assessed in waves 17-20.We further assessed the primary information source for these beliefs in wave 19.We also assessed if participants opposed COVID-19 vaccination for adults aged 80 years (waves 17-20).
The three major misconceptions about COVID-19 vaccines were assessed using the following statements in Singapore: "Individuals aged 60 years and above should receive COVID-19 vaccines", "Individuals with more chronic diseases have a greater need for COVID-19 vaccines", and "COVID-19 vaccines are more harmful than COVID-19 infection" in the follow-up (October 2021 to January 2022).Disagreement with the first two statements and agreement with the last statement were considered as vaccine misconceptions.
COVID-19 vaccine confidence was assessed in Hong Kong (waves 11-18, 20) with statements adapted from the Vaccine Confidence Index on a 5-point Likert scale ranging from strongly disagree to strongly agree: "I think COVID-19 vaccines are effective", "I think COVID-19 vaccines are safe", and "I think COVID-19 vaccines are important for children to have" 3 .Strongly agree and agree were grouped together versus the other options 4 .

Assessment of co-variables.
Media reports on Adverse Events Following COVID-19 Immunisation (AEFIs) were extracted from WiseNews, a dataset of news clippings from Hong Kong traditional newspaper publishers and online news media, between August 1 st , 2020 and May 30 th , 2022 10 .Examples of AEFIs recognised by Drug Office, Department of Health, of the Government of Hong Kong Special Administration Region include anaphylaxis, Bell's palsy, and encephalomyelitis 11 .The news articles must include the words "COVID" and "vaccines(s)".The articles must also include any of the following words or their Chinese equivalents: "side effects", "adverse events", "Advisory Panel on COVID-19 Vaccines", "Expert Committee on Clinical Events Assessment", "loss of function", "anaphylaxis", "anaphylactoid reaction", "acute peripheral facial paralysis", "Bell's palsy", "encephalopathy", "myocarditis", "inflammation of the heart muscle", "pericarditis", "inflammation of the lining outside the heart", "thrombocytopenia", "thrombosis with thrombocytopenia syndrome", and "transverse myelitis".

Post-stratification weighting and raking and calculation of Cohen's w.
Post-stratification weighting was estimated by calculating censoring weights.For Hong Kong, it was defined as the inverse probability of participating in the study after wave 2 and estimated using logistic regression with sociodemographic characteristics at wave 2 (i.e.age, sex, education attainment, marital status, employment status, household income and housing type).Raking was then applied for the sample to be representative of the 2016 Population By-census of Hong Kong using age, sex, education attainment, marital status, monthly household income and housing type.
For Singapore, it was defined as the inverse probability of participating in the follow-up after the baseline and estimated using logistic regression with sociodemographic characteristics at baseline (i.e.age, sex, education attainment, marital status and ethnicity).Raking was then applied for the baseline sample to be representative of the 2020 Population By-census of Singapore, and the follow-up sample to be representative of the 2021 population statistics from the Singapore Department of Statistics using age, sex, education attainment, marital status, and ethnicity.
Cohen's w is a measure to assess association between two nominal variables, and is equivalent to the effect size for a chi-square test of association 12 .We calculated Cohen's w using the "ES.w1" function from the "pwr" package in R to compare the distribution of each sociodemographic between our sample and the 2016 Population By-census of Hong Kong.The effect size measured by Cohen's w is considered small for values close to 0.1, medium for around 0.3, large for around 0.5.

Interrupted time-series analyses.
We obtained the official daily number of COVID-19 vaccination appointments in Hong Kong from February 23 rd , 2021 to May 30 th , 2022 to assess the impact of public policies and interventions on vaccine uptake.These included: 1) lottery-based incentives (e.g.residential apartment) announced from May 26 th , 2021; 2) vaccine mandates for civil servants and staff at schools, care homes, and public hospitals announced on August 2 nd , 2021; 3) vaccine pass (restricting access to premises such as restaurants, public buildings, and leisure facilities) announced on December 31 st , 2021 and enacted on February 24 th , 2022, and 4) reopening of premises under the vaccine pass announced on April 14 th , 2022 and enacted on April 21 st , 2022 [5][6][7][8][9] .These measures were examined in the interrupted time-series analyses to examine their impact on daily COVID-19 vaccination appointments.Given the high public awareness of lotterybased incentives, a new equilibrium was assumed to reach soon after them being announced and sustained up till the next measure (i.e.workplace vaccine mandates).We therefore used step functions to model each of the four interventions.In addition, Lunar New Year (January 31 st -February 2 nd , 2022) and surge of Omicron wave (February 7 th , 2022-April 14 th , 2022) were taken into account.
Autoregressive Integrated Moving Average Exogenous Variable Model (ARIMAX) was used to incorporate exogenous covariates (i.e. the four interventions, Lunar New Year and surge of Omicron wave) into the time series model, and daily vaccination appointments were log transformed to stabilise the variance of time series 13 .Due to model parsimony and the lack of autocorrelation in the residuals, we selected an ARIMAX model of (0,1,2)*(0,1,1) to fit the data.Percent change in the vaccine appointment numbers after the implementation of any intervention was estimated by (exponentiate of the coefficient-1)*100.

Generalised estimating equations between political views in 2014 and COVID-19 vaccine refusal.
We used generalised estimating equations with an independent correlation matrix to examine the association between political views during the 2014 Occupy Central (waves 3-4) and COVID-19 vaccine refusal (waves 11-17).

Causal mediation analyses.
We conducted causal mediation analyses using the CMAverse packages in R 14 .The proposed causal diagram for the relationships between political views, trust in vaccine information sources, vaccine misconceptions, vaccination confidence, and vaccine refusal is shown in Fig. 5, with sociodemographics (i.e.age, sex, education attainment, marital status, employment status and monthly household income) as baseline confounders.Exposure, mediators, and outcome were all dichotomised: political views (non-establishment view and others); trust in COVID-19 vaccine information sources (mistrust and trust); vaccine misconceptions (yes and no); vaccine confidence (not agree and agree); and vaccine refusal (yes and no).
Mediator and outcome were modelled using Poisson distribution with log link function.A regressionbased approach was adopted, and natural direct and indirect effects were estimated through direct counterfactual imputation estimation.Multiple imputation was performed to handle missing values (number of imputed dataset=20).Standard errors of causal effects were estimated through bootstrapping (number of bootstrapping=200).In interpreting our results, assumptions included no unmeasured confounding between exposure and outcome, mediators and outcome, and exposure and mediators after adjusting for sociodemographics; no intermediate confounding between mediators and outcome induced by exposure; no interactions between exposure and mediators; no measurement error; correct parametric specification of the models; no interference, and causal consistency 14,15 .
We examined the indirect effects of political views on vaccine refusal via mistrust in vaccine information sources, vaccine misconceptions, and low vaccine confidence in tandem (eTable 8).Proportion mediated was reported.We also calculated adjusted incidence rate ratios for the exposure-mediator, and mediator-outcome relationships in each mediation model.Adjusted incidence rate ratios were calculated by extracting the coefficients and standard errors from mediation models built for each imputed dataset that were stored in CMAverse and combining them using Rubin's rule 16 .

Estimation of population attributable fractions of vaccine refusal and absolute reduction in vaccine refusal if a risk factor was eliminated.
We examined the determinants of vaccine refusal at wave 17 as this timepoint preceded Hong Kong's major surge in COVID-19 mortality.Although vaccinations increased during the Omicron surge (waves 18-19), it was often too late for older adults receiving their first dose to be protected against severe disease or death during the Omicron wave [17][18][19] .We estimated the sequential and average population attributable fractions (PAFs) of vaccine refusal for each determinant as well as their joint contribution, utilising the "averisk" packages in R 20,21 .Vaccine refusal was modelled using logistic regression, and the determinants were modelled simultaneously (i.e.political views, trust in COVID-19 vaccine information sources from WHO, government health authorities and academics, COVID-19 vaccine misconceptions about older adults, chronic diseases and safety, COVID-19 vaccine confidence in its effectiveness, safety and importance for children to have) while adjusting for sociodemographics.To handle the missing data, 20 datasets were imputed.Point estimates and standard errors of each imputed dataset were extracted from the "averisk" package and combined using Rubin's rule 16 .
To estimate the individual contribution of determinants to vaccine refusal, we used the following estimation procedures: First, we applied robust Poisson regression to model the associations of political views during 2019 social unrest, trust in COVID-19 vaccine information sources, COVID-19 vaccine misconceptions, and COVID-19 vaccine confidence at wave 16 with vaccine refusal at wave 17 in separate models (hereinafter "MPoisson"), adjusting for sociodemographics.Political views were additionally adjusted when modelling the association of trust in vaccine information sources, vaccine misconceptions and vaccine confidence with vaccine refusal.Second, a predicted probability (Po) of being vaccine-hesitant was obtained for each participant under their observed level of covariables in the model.To simulate the counterfactual condition, the following were shifted: 1) political views from non-establishment to proestablishment/neutral, 2) mistrust to trust for COVID-19 vaccine information sources, 3) vaccine misconception to no vaccine misconception, and 4) from not agreeing with statements regarding vaccine confidence to agreeing.A simulated probability (Ps) of being vaccine-hesitant under the counterfactual condition was then obtained using the same set of estimates from MPoisson 22 .
The absolute reduction in vaccine refusal in counterfactual scenarios was estimated by averaging the difference between predicted probability and simulated probability (Po-Ps) across all participants.Population attributable fractions were calculated by averaging the fraction of the absolute reduction divided by Po across all participants.Missing data were imputed using multiple imputation by chained equation, and results were combined from 20 imputed datasets using Rubin's rule 23 .Confidence intervals for population attributable fraction and absolute reduction of vaccine refusal were estimated following the MI Boot (Pooled Sample) algorithm 24 .Specifically, 20 datasets were imputed, and then 500 bootstraps were implemented for each imputed dataset (bootstrap for complex survey design using bs4rw package in Stata was applied to incorporate weights) and all bootstrapping results were stored 25 .Finally, 20*500 estimates were ranked and the 2.5 th and 97.5 th percentiles were extracted as the lower and upper limit of the confidence interval, respectively.

eResults Long-term association between political views and COVID-19 vaccine refusal.
Political participation during the 2014 Occupy Central was also associated with COVID-19 vaccine refusal eight years later (eTable 5 and eFigure 4).

Potential gains in vaccination willingness.
In adults aged 18-59 years, building vaccine confidence with regard to the safety, effectiveness, and importance of COVID-19 vaccines could change vaccine refusal by 8.8% (6.2-11.9),6.7% (4.6-9.2), and 5.8% (3.0-9.1),respectively.If mistrust in WHO, government health authorities, and academics were addressed, the absolute changes in vaccine refusal could be 5.0% (1.8-8.4),4.9% (1.1-8.8), and 2.8% (0.4-5.4), respectively.If the three major vaccine misconceptions were removed, this could have reduced vaccine refusal by 4.0% (1.0-7.2) (eFigure 11).Shifting political views from non-establishment to neutral or pro-establishment may change vaccine refusal by 7.0% (2.9-11.0)and 5.0% (0.9-9.1) in adults aged 18-59 years and 60 years, respectively.Sequential and average population attributable fraction of each determinant was calculated by modelling all determinants simultaneously.This may generate statistically non-significant negative associations between determinants and vaccine refusal, and consequently negative values of population attributable fraction.As such, values in the figure may not add up to 100%. ©

eFigure 3 .
Sampling and retention of participants in 20 waves of longitudinal data in a population-based cohort, 2009-2022.

eFigure 4 .
Vaccination willingness and trust in COVID-19 vaccine information sources in Hong Kong by participation in the 2014 Occupy Central.A B (A) Trend in vaccination willingness from 2020 to 2022, overlaid on media reports from WiseNews on Adverse Events Following COVID-19 Immunisation (AEFIs).Measures to enhance vaccine uptake included announcements of vaccine lotteries, vaccine mandates for civil servants, staff at schools, care homes, and public hospitals, implementation of vaccine pass (e.g.restricting access to premises such as restaurants, public buildings, and leisure facilities), and reopening of premises under the vaccine pass.(B) Trust in COVID-19 vaccine information sources in the adult population in June-July 2021.Error bars indicate the 95% confidence intervals.Political views were assessed in waves 3-4 during the 2014 Occupy Central.Participants who participated or visited the protest sites were classified as non-establishment.Participants that did not visit the protest sites or participate in the movement were classified as pro-establishment or neutral.

eFigure 5 .
Residual check of selected ARIMAX model for interrupted time-series analyses.The residual plots of an ARIMAX model of (0,1,2)*(0,1,1) [7] are shown above.The upper plot indicates that the residuals were white noise with no obvious pattern.The lower left plot indicates there was no auto-correlation among lagged residuals, which was also tested by the Ljung-Box test.The lower right plot shows the distribution of residuals and compares it with the normal distribution.Those residual plots indicate an adequate fit of our data using the selected model.© 2023 Lun P et al.JAMA Network Open.

eFigure 9 .
Trends in vaccine misconceptions in Hong Kong, 2021-2022.Trends in COVID-19 vaccine misconceptions in Hong Kong, 2021-2022.The three misconceptions about COVID-19 vaccines were assessed using the following statements: "Older individuals have a greater need for COVID-19 vaccination", "Individuals with more chronic diseases have a greater need for COVID-19 vaccination", and "COVID-19 vaccines are more harmful than COVID-19 infection.Disagreement with the first two statements and agreement with the last statement were considered as vaccine misconceptions.Error bars indicate the 95% confidence intervals.older adults isconception a out c ronic diseases isconception a out accine sa ety © 2023 Lun P et al.JAMA Network Open.

eFigure 10 .
Population attributable fractions for factors, assessed in June-July 2021, associated with vaccine refusal and absolute reduction of vaccine refusal in November 2021.Error bars indicate the 95% confidence intervals.Vaccine isconception a out sa ety Vaccine isconception a out c ronic diseases Vaccine isconception a out older indi iduals o trust in accine in or ation ro acade ics Any a or accine isconception o trust in accine in or ation ro go ern ent ealt aut orities id not agree t at OV 1 accines are i portant o trust in accine in or ation ro W accine isconceptions ig trust in accine in or ation ro go ern ent ealt aut orities Agreed t at OV 1 accines are i portant ig trust in accine in or ation ro t e W P et al.JAMA Network Open.

eFigure 11 .
Population attributable fractions for factors, assessed in June-July 2021, associated with vaccine refusal and absolute reduction of vaccine refusal in November 2021.

Binary 20 © 2023 20 © 2023
of 1 to 10, how much do you trust World Health Organization (WHO) on COVID-19 vaccine related information?Range: 1(having no trust at all)-10(having complete trust) Recoding: Mistrust: ≤5, Trust: >6 of 1 to 10, how much do you trust Government departments and related institutions like FHB, HA, and CHP on COVID-19 vaccine related information?Range: 1(having no trust at all)-10(having complete trust) of 1 to 10, how much do you trust your physician on COVID-19 vaccine related information?Range: 1(having no trust at all)-10(having complete trust) Recoding: Mistrust: ≤5, Trust: >6 15-Lun P et al.JAMA Network Open. of 1 to 10, how much do you trust local public health or infectious disease academics on COVID-19 vaccine related information?Range: 1(having no trust at all)-10(having complete trust) of 1 to 10, how much do you trust the traditional media platforms (e.g.television news, newspaper, radio stations) on COVID-19 vaccine related information?Range: 1(having no trust at all)-10(having complete trust) of 1 to 10, how much do you trust the social media (e.g.Facebook, WhatsApp, Instagram, Telegram, WeChat, Twitter, Weibo, HKGolden, LIHKG, YouTube) on COVID-19 vaccine related information?Range: 1(having no trust at all)-10(having complete trust) Recoding: Mistrust: ≤5, Trust: >6 of 1 to 10, how much do you trust your family/friends on COVID-19 vaccine related information?Range: 1(having no trust at all)-10(having complete trust) what extent do you agree with the following statement pertaining to COVID-19 vaccines?"I think COVID-19 vaccines are effective."Ordinal: Strongly agree; Somewhat agree; Neither agree nor disagree; Somewhat disagree; do you agree with the following statement pertaining to COVID-19 vaccines?"I think COVID-19 vaccines are safe."Ordinal: Strongly agree, Somewhat agree, Neither agree nor disagree, Somewhat disagree, Lun P et al.JAMA Network Open.
COVID-19 vaccine uptake among older adults before the first documented local Omicron transmission.Data were collected from Hong Kong Vaccination Dashboard, The State Council of the People's Republic of China, Australian Immunisation Register, Public Health Agency of Canada, Prime Minister's Office of Japan, Ministry of Health of New Zealand, Ministry of Health in Singapore, Central Disaster Management Headquarters of South Korean, National Health Service of England, and National Center for Immunization and Respiratory Diseases in the United States 26-34 .a Fully vaccinated refers to having completed the primary series of COVID-19 vaccines.b Data included older adults aged ≥75 years.List of outcome and exposures used in Hong Kong.
Demographic composition of wave 20 compared to 2016 Population By-census of Hong Kong.Interrupted time-series analyses of intervention measures and daily COVID-19 vaccination appointments.Abbreviations: MA = moving average.Coefficient refers to the percent change in vaccine appointment numbers associated with corresponding intervention measures; 95% confident intervals of the percent change are presented in brackets.Bolded are statistically significant (p<0.05).Associations of political views, trust in information sources, vaccine misconceptions, and vaccine confidence in June-July, 2021 with vaccine refusal in the general population (≥18 years) and older adults (≥60 years) in November 2021.Abbreviations: aIRR=adjusted incidence rate ratio; ref= reference level; 95% CI=95% confidence interval.Notes: a Robust Poisson regression models adjusted for sociodemographics; b Robust Poisson regression models adjusted for sociodemographics and political views.Bolded are statistically significant (p<0.05).eTable 6. Associations of political views during 2014 Occupy Central with vaccine refusal over COVID-19 pandemic.Abbreviations: aIRR=adjusted incidence rate ratio; ref= reference level; 95% CI=95% confidence interval.Notes: Generalised estimating equations adjusted for sociodemographics and with an independent correlation matrix.Bolded are statistically significant (p<0.05).Association of political views during 2019 Social Unrest with trust in COVID-19 vaccine information sources, vaccine misconceptions, and vaccine confidence in June 2021.Abbreviations: aIRR = adjusted incidence rate ratio; 95% CI = 95% confidence interval.Robust Poisson regression models adjusted for sociodemographics.Bolded are statistically significant (p<0.05).