The Wuxi Center for Disease Control and Prevention Ethics Committee approved this study (2020No10).

From May 24 to June 15, 2021, we conducted surveys in 31 provinces in mainland China using web-based questionnaires. We used convenience sampling. The link to the questionnaire was created through Wen Juan Xing, which is a platform dedicated to the creation and dissemination of questionnaires, and forwarded by the WeChat platform of the Jiangsu Provincial Center for Disease Control and Prevention. Written information was provided as a statement that could be read, which assured participants that the study was conducted on a voluntary basis and for research purposes only. All surveys were conducted in Chinese. (The surveys were translated into English by 2 researchers only for the purpose of this paper.) Questionnaires answered by people under 18 years and over 60 years were not analyzed. To prevent repeated submission of questionnaires, WeChat real name verification was required when using the link to fill out the questionnaire, and an IP address could only be used to submit a questionnaire once. Questionnaires completed in less than 60 seconds were automatically discarded. Questionnaires with selections at the same level of the Likert scale were also considered invalid.

The questionnaire (Multimedia Appendix 1) was used to collect demographic information (age, gender, ethnicity, usual place of residence, education level, annual household income, and whether respondents worked in the health care industry, residential status), chronic disease history, and information on self-assessment of health status.

The theory of planned behavior is widely used to study intentions and behaviors. In this model, intentions are considered the most direct predictor of behaviors and are weighted based on attitudes, which is the degree to which behaviors are positively or negatively evaluated, and subjective norms, which is the pressure society places on implementing or not implementing behaviors. The effects of attitudes and subjective norms are mediated by perceived behavioral control, which is people's perceptions of their ability to perform a given behavior. When perceived behavioral control is accurate, it acts as a proxy for actual behavioral control, that is, the extent to which a person has the ability, resources, and other conditions required to perform the behavior.

In this study, the variable attitude represented people's positive or negative perceptions of the vaccine. The variable subjective norms referred to the expectations of family, friends, and physicians. The variable perceived behavioral control represented people's beliefs about barriers to vaccination (such as the time, cost, and side effects caused by vaccination). Each item was assessed using a 5-point Likert scale (Table 1).

Before structural equation modeling, confirmatory factor analysis was conducted to assess the reliability and validity of the constructs. Reliability was assessed by calculating the squared multiple correlation [24] and composite reliability [25]. We also examined parameter estimates and their associated t values, factor loadings, and the average variance extracted [26]. We established discriminant validity by calculating the square root of the average variance extracted for each latent variable. The error variances and modification indices of items were estimated.

The goodness-of-fit index (GFI), adjusted GFI, comparative fit index (CFI), and the root mean square error of approximation (RMSEA) were used to evaluate the model fit [27]. For GFI, CFI, and adjusted GFI, values closer to 1 are better, and values greater than 0.95 indicate relatively good fit; RMSEA values less than 0.06 indicate relatively good fit [27,28].

We used bootstrapping (5000 trials) to test mediator effects [29]. Hierarchical moderator regression was used to test moderation effects, and all variables were standardized to avoid multicollinearity [30]. The control variables were entered in the block 1 (gender, education level, health care occupation, annual household income, main living condition, self-evaluation of health, chronic diseases, past behavior), followed by the standardized value of the main effect (attitude, subjective norms, and perceived behavioral control) in block 2, and finally, the interactions and moderators (subjective norms * perceived behavioral control, attitude * perceived behavioral control) in block 3.

AMOS software (version 23; IBM Corp) was used to estimate the structural equation coefficients between latent variables in the model. Hierarchical moderator regression analysis was performed using SPSS software (version 23; IBM Corp). Statistical significance was set at P<.05.

A total of 9924 participants (male: 5407/9924, 54.5%; female: 4517/9924, 45.5%) were included in this study (Table 2). Most respondents had a college degree or higher (7589/9924, 76.4%), and did not work in health care–related industries (7007/9924, 70.6%).

In this study, the proportion showing vaccine hesitancy (respondents who selected not sure, hesitant, and very hesitant for Q18) was estimated to be about 26.6% (2640/9924). Of the total sample, 77% (7643/9924) reported that they had received COVID-19 vaccinations, and 29.4% (2922/9924) had received the influenza vaccinations in the previous year. Of the women, 22% (992/4517) had received human papillomavirus (HPV) vaccinations. Of the total sample, of the majority believed that they would choose to receive COVID-19 (8614/9924, 86.8%) and influenza (3315/9924, 33.4%), vaccinations this year.

The 18 items were found to be reliable and valid based on each item’s estimated error variance and modification index (Table S1 and Table S2 in Multimedia Appendix 2). Composite reliability values were greater than 0.6 and average variance extracted values were greater than 0.5, except those for the dimension behavior (Table 3).

The square roots of average variance extracted of the dimensions attitude, subjective norms, intention, and perceived behavioral control exceeded the related correlations (Table 4), indicating discriminant validity in the structures in this study [25]. The overall model achieved a good fit (GFI 0.991; CFI 0.992; adjusted GFI 0.987; RMSEA 0.029); therefore, the measurements and structural model were acceptable.

Annual household income, education, gender, health care occupation, chronic diseases, health self-assessment, and past vaccination behavior affected vaccination behavior (∆R²=0.121, P<.001); however, whether participants lived alone or not did not significantly affect vaccination behavior as a control variable (P=.08) (Table S3 in Multimedia Appendix 2). In block 3, the attitude * perceived behavioral control and subjective norms * perceived behavioral control terms significantly changed the model compared with block 2 (∆R²=0.003, P<.001).

In block 3, perceived behavioral control was positively and significantly correlated with vaccination behavior (β= 0.274, P<.001); however, attitude * perceived behavioral control (β=–0.052, P<.001) and subjective norms * perceived behavioral control (β=–0.028, P=.006) had reverse inhibitory effects on vaccination behavior. This not only indicates that perceived behavioral control moderates the impact of attitude on behavior and that of subjective norms on behavior, supporting hypotheses 5a and 5b, it also shows that there is a substitution relationship between attitude, subjective norms, and perceived behavioral control in their influence on vaccination behavior. Overall, perceived behavioral control weakens the positive effects of attitude and subjective norms on vaccination behavior, and when perceived behavioral control is low, the promotion effects of attitude and subjective norms on behavior are more pronounced, but with increases in perceived behavioral control, the positive effects of attitude and subjective norms on behavior gradually decrease. Specifically, the slope describing effect of attitude on behavior will decrease by 0.050 SD and that of subjective norms on behavior will decrease by 0.021 SD when perceived behavioral control increases by 1 SD (Table 5; Table S3 in Multimedia Appendix 2).

Both attitude (direct effect: β=0.493, P<.001) and subjective norms (direct effect: β=0.244, P<.001) showed significant positive associations with behavior. Hence, hypotheses 1 and 2 were confirmed. Intention also positively and significantly affected vaccination behavior (direct effect: β=0.462, P<.001); therefore, hypothesis 3 was confirmed. In addition, bootstrapping indicated that intention was present as a positive and significant mediator between attitude and vaccination behavior (indirect effect: β=0.159, P<.001). Similarly, the mediating effect between subjective norms and vaccination behavior was positive and significant (indirect effect β=0.258, P<.001); therefore, hypotheses 4a and 4b were confirmed (Table 6). Furthermore, the difference between the indirect effect of attitude on behavior and that of subjective norms on behavior was statistically significant (difference=–0.091, P<.001), which indicated that subjective norms had a greater influence on vaccination behavior than attitude.

Our findings support the hypothesis that intentions mediate vaccination behavior, attitudes, and subjective norms and addresses concern about the mediating process of intention in the theory of planned behavior [31], while also justifying initial theoretical claims that distal attitudes and subjective norms can influence behavior, through proximal intention mediators [32]. In particular, perceived behavioral control was found to be a moderator that influences the mediating processes attitude→intention→behavior and subjective norms→intention→behavior.

A large meta-analysis [15] of applications of theory of planned behavior showed that subjective norms were weak predictors of intention and behavior. Recent studies [33,34] that have used the theory of planned behavior empirically for health behavior have suggested that subjective norms are not the most critical predictor. In contrast, some argued that subjective norms are strong predictors [35,36]. Our findings are similar to those in [35,36], and the indirect effects in our study further show that subjective norms have greater impacts on behavior than attitudes. This indicates that decisions about health issues are more likely to be influenced by social surroundings. Debate about whether subjective norms have a strong or weak influence on behavior might due to differences in the types of behaviors that have been targeted. In contrast to high-frequency health behaviors, such as exercise, smoking cessation, when making decisions about vaccination, people expect to be counseled by someone close to them (eg, family members and close friends [37]), or someone they trust (eg, a physician [38]). Additionally, socially desirable responses likely contributed to our finding that attitudes have less explanatory power for behavior. Socially desirable responses are defined as the tendency to give a positive self-description, which is relatively common for potentially sensitive problems. Korn [39] suggested that vaccination is a prosocial behavior and demonstrated that it is part of the social contract by showing that there is significant intergroup bias in vaccinated and unvaccinated cohorts. Therefore, it is reasonable that participants glorified attitudes toward the vaccine and responded positively to the investigation, rather than providing their true thoughts.

Our findings also show that high perceived behavioral control weakens the effect between attitude, subjective norms, and vaccination behavior, using intention as a mediator. Additionally, perceived behavioral control has an alternative relationship with attitude and subjective norms, when present as a moderator in the model. These findings extend those in existing literature on perceived behavioral control. However, perceived behavioral control was considered to be a positive predictor of intentions in many existing studies [40-42]. Our findings (block 2 in the hierarchical regression analysis) also confirm this view. It is important to consider that if a given behavior is seen as positive and implementable, people tend to engage in that behavior. However, the gap between intentions and behavior exists precisely because intention is not sufficient to ensure that a person converts ideas into actual behavior, due to the limitations of actual capabilities. Thus, we find that the mechanisms by which perceived behavioral control affect intention and behavior are different. In other words, the influence of perceived behavioral control on intention is as positive as attitude and subjective norms. However, when using perceived behavioral control to predict behavior, it should be thought of as actual behavior control [23]. Therefore, it is more appropriate to use perceived behavioral control as a moderating variable to influence the mediating effect of intention on behavior.

The moderated mediation model provides evidence for the practical application of theory of planned behavior in vaccination behavior. Future interventions for people who are unwilling to be vaccinated or subjectively postpone vaccination should not only focus on changing their negative attitudes toward the vaccine but should also pay attention to the ideas of their families and close friends about vaccination and intervention. This will facilitate positive intentions to get vaccinated. In a study [43] that explored factory workers' vaccination behavior using the theory of planned behavior, it was also shown that a positive attitude and the support of health care workers, relatives, and friends can contribute to individuals getting vaccinated, and in another study [44], it was found that trusted individuals had a unique influence on young women in encouraging them to get their HPV vaccination. In addition, health authorities may need to take measures such as communicating risk or giving rewards for vaccinations to improve influenza and HPV vaccine uptake.

This study was conducted in the form of a web-based survey; thus, convenience sampling might have caused selection bias. Additionally, we used a cross-sectional survey; thus, information on respondents' vaccine-related behavior at the time of the survey was used in behavior dimension. For this reason, the values of composite reliability and average variance extracted for the behavior dimension were below the ideal value. However, the overall model fit was ideal, with GFI, adjusted GFI, and CFI >0.9 and RMSEA <0.06.

Data from 1 month after the survey showed that the actual COVID-19 vaccination rate (having completed at least one shot), in China was 74% [45], approximating the self-reported behavioral data in this study (8614/9924, 86.8%). Although the self-reported data of this study are credible, cautious interpretation of the sample’s representativeness of vaccination behavior is needed.

The behavior dimension was measured with 3 items—self-reported COVID-19 vaccination behavior and influenza vaccination behavior, as well as overall vaccine hesitancy. This is because according to the schedule of immunization in China [46], there are 5 main vaccines for adults 18 to 59 years of age—COVID-19, influenza, HPV, hepatitis B, and rabies. Yearly influenza and the recent COVID-19 vaccine were the used to examine vaccination behavior because the administration of the other 3 types are limited—hepatitis B vaccinations are valid for a long period, rabies vaccinations are for emergency use after possible exposure, and HPV vaccinations are mainly target the female population in China.

Our findings showed that subjective norms have a stronger influence than attitudes on this particular vaccination practice. Moreover, perceived behavioral control not only is a positive facilitator of intention but also has an alternative relationship with attitudes and subjective norms. As a moderator, perceived behavioral control conversely weakens the positive effects of attitudes and subjective norms on vaccination behaviors. When perceived behavioral control is low, the positive influence of attitudes and subjective norms are evident. However, when perceived behavioral control is high, the positive effects of attitudes and subjective norms gradually decrease. In particular, as a moderator, perceived behavioral control has more predictive power for vaccination behavior.