We used a stratified random sampling method to conduct a multicenter, observational study of vaccine hesitancy in Chinese households. The study was reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines (Checklist 1). Four cities were selected from China’s 4 major economic regions: Eastern (Changzhou, Jiangsu), Central (Zhengzhou, Henan), Western (Xining, Qinghai), and Northeastern (Mudanjiang, Heilongjiang). These regions differ in economic development, health care resources, and cultural norms. The minimum sample sizes for each region were determined based on the population proportions of China’s Seventh National Population Census (2020). Within each city, at least 2 urban and 2 rural areas were randomly selected for sampling, followed by a random selection of households within each identified area. All eligible members aged 18 years and above were invited to participate, as vaccination decisions for minors are generally made by parents or guardians [20]. This survey was conducted in March 2023. Participants completed either an online questionnaire administered through the Wenjuanxing platform or a paper questionnaire with the assistance of survey staff. To reduce selection and information bias, trained staff thoroughly reviewed all questionnaires based on the exclusion criteria. Finally, a total of 5891 participants aged 18 years and above from 2387 households were enrolled.

Following rigorous data cleaning to exclude invalid questionnaires, 5780 participants were included in this analysis. We further excluded participants with missing data on current and future CBVH, yielding a final analytical sample of 5462 participants for the analysis of current CBVH and 5437 participants for the analysis of future CBVH. Among them, 5407 participants responded to both questions and were subsequently included in the concurrent CBVH analysis. The detailed participant flow is shown in Figure 1.

This study was conducted in accordance with the Declaration of Helsinki. This survey project was approved by the Life Science Ethics Review Committee of Zhengzhou University (approval number: 2021-01-12-05), and all participants provided written informed consent. All data were handled confidentially, and no compensation was provided.

Descriptive analyses were conducted to characterize the study sample. For continuous variables, means and SDs were calculated; for categorical variables, frequencies and percentages were reported. Chi-square and 2-tailed independent-samples t tests were used to examine differences between participants with and without CBVH. Poststratification weighting was used to adjust the sample according to the age, sex, and region distributions reported in the Seventh National Population Census of China (2020) [21]. Participants with missing outcome data were excluded from the corresponding outcome-specific analysis. No covariate data were missing after data cleaning.

A 3-level logistic regression modeling approach was used due to the hierarchical structure of our data. Specifically, individuals were nested within households, and households within communities. Fixed effects were presented as adjusted odds ratios (aORs) with corresponding 95% CIs. Random effects were assessed using random-intercept variances, the variance partition coefficient (VPC), the intraclass correlation coefficient (ICC), the median odds ratio (MOR), and proportional changes in variance (PCV). The community- and household-level VPCs indicate the proportions of overall variation attributable to differences between communities and households, respectively. The combined ICC represents clustering among individuals in the same household and includes both household- and community-level variances [22]. The MOR represents the median odds ratio between higher- and lower-risk clusters and quantifies unexplained heterogeneity on the odds-ratio scale, whereas the PCV quantifies the proportional change in the corresponding variance relative to model 0 [23]. The individual-level residual variance was fixed at π²/3, and 95% CIs were estimated using the delta method. Four models were fitted sequentially. First, a null model with only household- and community-level random intercepts was fitted to partition the variance. Subsequently, 3 multivariable models were built: model 1 adjusted for individual-level factors, including age, sex, ethnicity, religion, marital status, educational level, employment status, coresidence status, smoking status, drinking status, chronic disease, history of allergies, prior COVID-19 vaccination status, source of COVID-19 vaccine information, knowledge of COVID-19 vaccine, perceived severity, perceived susceptibility, perceived benefits, perceived barriers, self-efficacy, cues to action, trust in doctors, and trust in vaccine developers; model 2 further adjusted for household-level factors, including number of children, number of older adults, and annual household income; and model 3 further adjusted for community-level factors, including region, distance to vaccination point, and time to vaccination point. Model fit was evaluated using the Akaike information criterion (AIC), Bayesian information criterion (BIC), and log-likelihood (LL), with better fit indicated by lower AIC and BIC values and higher log-likelihood.

Then, network analysis was conducted to visualize and quantify complex interrelationships among factors that were statistically significantly associated with CBVH within multilevel models. The current and future CBVH networks were estimated using mixed graphical models, which allow continuous, binary, and multicategorical variables within a single analytical framework [24]. Only pairwise interactions were estimated (k=2). Least absolute shrinkage and selection operator regularization [25], combined with extended Bayesian information criterion model selection (γ=0.50), were applied to obtain conservative and sparse network estimates [26]. Nodes represented the included variables, whereas edges represented pairwise conditional associations after controlling for all other variables in the network. Positive and negative associations were represented by green and red edges, respectively, with thicker edges indicating stronger associations. For multicategorical variables, the edge weight was summarized using the mean absolute magnitude of category-specific parameters and was displayed in gray because a single direction could not be assigned. Node predictability indicates the extent to which a node could be predicted by its adjacent nodes [27]. It was quantified using the proportion of explained variance (R²) for continuous variables and the proportion of correct classifications (CC) for categorical variables. Centrality indices, including strength, betweenness, and closeness, were calculated to identify the most influential nodes within the network. Edge-weight accuracy was evaluated using nonparametric bootstrapping with 1000 replicates, from which 95% CIs were obtained. The correlation stability coefficient (CS-C) was reported, with a threshold of 0.25 or higher, and ideally above 0.50 [28]. We also ranked the edges directly connected to the CBVH node and displayed them in a heatmap [29]. Signs were retained for continuous and binary variables, while multicategorical edges were summarized using mean absolute parameter values and compared descriptively.

To evaluate the robustness of the primary findings, 3 sensitivity analyses were performed. Because religious beliefs [30] and allergy history [31] have been identified as important determinants of vaccine hesitancy, the analyses were repeated after separately excluding participants reporting religious beliefs and those with a history of allergies. Participants from the eastern region were excluded to assess the robustness of the results with respect to the uneven regional sample distribution. All statistical analyses were performed using Stata (version 18.0; StataCorp LLC) and R (version 4.4.2; R Foundation for Statistical Computing), with a 2-tailed P value of <.05 considered statistically significant.

The unweighted prevalence of current and future CBVH was 28.4% (1553/5462) and 30.1% (1636/5437), respectively. After poststratification, the corresponding prevalence estimates were 30.88% and 32.55%, respectively (Figure S1 in Multimedia Appendix 1). Among the 5407 participants with complete data for both outcomes, 1461 (27.0%) exhibited concurrent CBVH. In the current CBVH analytical sample (n=5462), participants were distributed across all age groups, and 46.5% (n=2538) were male participants. Most participants were married (n=4832, 88.5%) and employed (n=5038, 92.2%). The future CBVH analytical sample (n=5437) showed a similar distribution. For both groups, we found that age, educational level, employment, region, number of children, annual household income, smoking status, drinking status, history of allergies, chronic disease, prior COVID-19 vaccination status, source of COVID-19 vaccine information, knowledge of the COVID-19 vaccine, perceived severity, perceived susceptibility, perceived benefits, perceived barriers, self-efficacy, cues to action, trust in doctors, trust in vaccine developers, distance to the vaccination point, and time to the vaccination point all differed between participants with and without CBVH (all P<.05). The detailed characteristics are presented in Table 1.

Table 2 presents the variance components, clustering measures, and model fit statistics for the 4 models. In model 0, the combined household- and community-level ICCs were 77.73% (95% CI 72.40%‐82.29%) for current CBVH and 78.84% (95% CI 73.57%‐83.30%) for future CBVH, supporting the use of 3-level models. After controlling for all variables, the ICCs remained high at 71.94% (95% CI 66.30%‐76.96%) and 72.52% (95% CI 66.97%‐77.45%), respectively. From model 0 to model 3, community-level variance decreased substantially (current CBVH: 3.89-1.36 and future CBVH: 4.29-1.49), whereas household-level variance changed modestly (current CBVH: 7.60-7.07 and future CBVH: 7.97-7.19). In model 3, the MORs for current and future CBVH were 15.96 (95% CI 11.32‐23.62) and 16.62 (95% CI 11.75‐24.70), respectively. This indicates a median 16-fold difference in the odds of hesitancy when comparing 2 individuals with identical covariates from different household-community contexts. Overall, relative to model 0, model 3 accounted for 26.57% and 29.17% of the higher-level variance for current and future CBVH, with most of the reduction occurring at the community level.

As shown in Figure 2, in the current CBVH analytical sample, participants with level 3 self-efficacy (aOR 0.01, 95% CI 0.01‐0.02) and level 3 perceived benefits (aOR 0.45, 95% CI 0.25‐0.81) had lower odds of CBVH than those at level 1. Conversely, participants with level 4 perceived barriers (aOR 4.81, 95% CI 2.41‐9.64) and level 4 perceived severity (aOR 2.82, 95% CI 1.74‐4.57) had higher odds than those at level 1. Compared with receiving cues to action from multiple channels, cues derived from social pressure (aOR 2.93, 95% CI 1.08‐7.94) were associated with higher odds of CBVH. Regarding trust in the medical system, higher trust in doctors (aOR 0.97, 95% CI 0.94‐0.99) and in vaccine developers (aOR 0.89, 95% CI 0.84‐0.94) were both associated with lower odds of CBVH. Participants with level 3 vaccine knowledge also had lower odds than those at level 1 (aOR 0.48, 95% CI 0.30‐0.76). Several sociodemographic and health characteristics were also associated with CBVH. Compared with those living with others, individuals living alone had lower odds of CBVH (aOR 0.28, 95% CI 0.10‐0.78). Compared with residents in eastern China, those living in the central (aOR 0.12, 95% CI 0.03‐0.47) and northeastern regions (aOR 0.15, 95% CI 0.03‐0.89) exhibited lower odds of CBVH. Prior COVID-19 vaccination was also associated with lower odds of CBVH (aOR 0.12, 95% CI 0.06‐0.23). Conversely, higher annual household income was associated with increased odds of CBVH (aOR 1.01, 95% CI 1.00‐1.01). Additionally, having chronic diseases (aOR 1.48, 95% CI 1.04‐2.12) or a history of allergies (aOR 1.66, 95% CI 1.04‐2.66) was associated with increased odds of CBVH. Similar associations were observed in the future CBVH analytical sample.

The network analysis included 14 nodes and is presented in Figure 3. The current and future CBVH networks retained 43 and 41 of the 91 possible edges, corresponding to network densities of 47.3% (43/91) and 45.1% (41/91), respectively. The CC of the CBVH node was 79.6% (4350/5462) in the current network and 79.3% (4311/5437) in the future network. Overall, the 2 networks showed similar configurations.

The magnitudes of conditional associations between study factors and CBVH are presented in Figure 4. In both networks, self-efficacy (current CBVH: −0.548 and future CBVH: −0.533) and prior COVID-19 vaccination status (current CBVH: −0.476 and future CBVH: −0.457) showed the largest conditional associations with CBVH. Region (current CBVH: 0.406 and future CBVH: 0.430) and cues to action (current CBVH: 0.176 and future CBVH: 0.207) also showed large edge weights. Because these are multicategorical variables, the values represent absolute edge magnitudes without a single direction. Edge weights for perceived barriers, perceived severity, chronic disease, and allergy history varied little between networks. Trust in doctors had a larger absolute association with current CBVH (−0.088 vs −0.052), whereas trust in vaccine developers (−0.069 vs −0.099) and perceived benefits (−0.035 vs −0.063) had larger absolute associations with future CBVH. Beyond the CBVH-related edges, the strongest edge weight in both networks was observed between perceived barriers and trust in vaccine developers (weight=0.593).

Centrality indices and bootstrap results are provided in Figures S2 to S4 in Multimedia Appendix 1. The case-dropping bootstrap analysis indicated stability for strength (CS-C: current CBVH=0.750 and future CBVH=0.672) and betweenness (CS-C: current CBVH=0.672 and future CBVH=0.672). Nonparametric bootstrap analysis supported the accuracy of the strongest edge estimates.

Three sensitivity analyses were performed, separately excluding participants with religious beliefs, those with a history of allergies, and those from the eastern region (Tables S2-S4 in Multimedia Appendix 1). The factors associated with CBVH were largely consistent with those in the primary analysis. Additionally, the combined household- and community-level ICCs ranged from 64.96% to 74.15% across the 3 analyses, remaining comparable to the primary estimates of 71.94% for current and 72.52% for future CBVH.

This study investigated factors associated with CBVH in a sample from 4 provinces in China, finding that 27.0% (1461/5407) exhibited concurrent CBVH, while the poststratification prevalence of current and future CBVH was 30.88% and 32.55%, respectively. Our findings revealed substantial clustering of CBVH at both the household and community levels (77.73% for current and 78.84% for future). Additionally, network analysis showed that self-efficacy and prior COVID-19 vaccination had the strongest conditional associations with CBVH. The strongest association between non-CBVH nodes was between perceived barriers and trust in vaccine developers. The current and future networks showed broadly similar configurations, supporting the prioritization of factors consistently associated with both outcomes. In general, our results provide insights for developing targeted interventions at the household and community levels to enhance vaccination coverage and strengthen pandemic preparedness.

The observed clustering aligns with social norms theory, which posits that behavior is not only driven by personal motivation or legal enforcement but also influenced by perceptions of others’ behaviors and expectations [32]. At the community level, 2 mechanisms may explain the clustering. First, social selection may produce residential homogeneity. People living in the same community often share similar sociodemographic characteristics, health literacy, and cultural values, which may lead to similar vaccine attitudes [33]. Residents may also face shared structural barriers, including distance to vaccination sites, limited transportation options, unequal access to digital systems, and uneven implementation of public health measures [34,35]. Second, social contagion may spread vaccine-related information and attitudes through neighbors, local groups, and social networks. Vaccine hesitancy may cluster geographically when hesitant views are widespread within a community [36,37]. At the household level, the unique relational structure of the family may make these mechanisms more pronounced. Families often operate under a centralized decision-making hierarchy, whereby one member (eg, parents or adult children) assumes primary responsibility for health decisions [38]. A previous study found that when the family decision-maker was hesitant, 65.1% of other household members tended to follow suit [13]. Notably, household-level variance decreased only modestly after adjustment for measured household characteristics, including household size and income. Future studies should examine other household-level factors, such as decision-making patterns, shared values, and family communication [13,39].

Our findings highlight several key factors associated with CBVH. Higher vaccine knowledge was associated with lower hesitancy, consistent with studies showing that limited vaccine understanding increases hesitancy [40,41]. Self-perception factors, including self-efficacy, perceived benefits, and perceived barriers, were also associated with CBVH [9,42]. Therefore, future interventions may prioritize these modifiable self-perception factors. For instance, motivational interviewing [43] or targeted health education [44] may strengthen vaccination self-efficacy and perceived benefits, while transparent communication regarding vaccine safety and efficacy may help improve perceived benefits [45]. Additionally, we observed a positive association between perceived severity and CBVH. One possible explanation is that during severe outbreaks, individuals may rely more on nonvaccine protective behaviors, such as improved hygiene, reduced mobility, and timely testing, which may be associated with a reduced perceived need for vaccination [46]. Higher trust in doctors and vaccine developers was associated with lower CBVH, underscoring the importance of credible health communication [47]. By contrast, chronic diseases and a history of allergies were associated with higher hesitancy, possibly because these participants were more concerned about their health conditions and adverse reactions [48,49]. Similarly, cues to action were associated with CBVH and have been verified in previous studies [50,51]. Compared with cues from multiple channels, cues based mainly on personal experience or social pressure were linked to higher hesitancy. Living alone was associated with lower CBVH in our study. However, the subgroup of participants living alone was small, and future studies with larger samples are needed to confirm this association. Notably, residents in central and northeastern China, as well as lower-income households, exhibited lower CBVH rates. This may be attributable to the free provision of COVID-19 booster vaccines in China [52], which alleviates financial barriers. This pattern is also consistent with the inverse equity hypothesis, which suggests that hesitancy may emerge first among more socioeconomically advantaged groups [53]. Given the markedly higher CBVH observed in eastern China, we repeated the analysis after excluding this region. Substantial clustering persisted, suggesting that the observed clustering was not driven solely by regional composition.

The network analysis illustrated interrelationships among factors associated with CBVH. Among the factors directly linked to CBVH, self-efficacy showed the strongest association. This finding is consistent with health behavior frameworks in which self-efficacy has been associated with preventive behaviors [54]. Prior COVID-19 vaccination status showed the second strongest direct association. This accords with evidence that vaccination history is associated with booster acceptance and subsequent uptake [55]. Region and cues to action also showed large edge weights. Although these multicategorical estimates did not indicate a uniform direction, they suggest that regional and communication-related factors warrant consideration alongside psychological factors. Beyond the edges directly connected to CBVH, perceived barriers and trust in vaccine developers had the strongest edge in both networks, suggesting that safety and effectiveness may be closely linked to confidence in vaccine producers. A prior study also showed that individuals with less concern about infection severity and lower trust in vaccine effectiveness were more likely to be hesitant [56]. The current and future networks were similar. These findings highlight factors that may warrant further evaluation in intervention studies.

Nevertheless, several limitations should be acknowledged. First, the cross-sectional design limits causal inference, and longitudinal studies are needed to confirm the observed relationships. Second, because CBVH was not a rare outcome, the odds ratios do not approximate prevalence ratios and should be interpreted as measures of association. Third, CBVH was assessed via self-report, which may have introduced reporting bias or social desirability effects. Finally, although our sampling strategy ensured geographic diversity across 4 major provinces, substantial regional heterogeneity may limit the generalizability of the findings.

In this multicenter household-based study, we found that vaccine hesitancy clustered substantially within households and communities, highlighting the importance of household- and community-level interventions. Self-efficacy and prior vaccination were most directly associated with hesitancy, while perceived barriers and trust in vaccine developers showed the strongest associations within the network. Effective strategies such as regular family-centered vaccine education delivered through community platforms, community-based vaccine counseling services, and transparent dissemination of vaccine-related information and policies may help address vaccine hesitancy. These insights offer a framework for strengthening vaccination strategies and preparedness in future infectious disease outbreaks.