This cross-sectional study used a causal-comparative design. This design was selected because the groups were predetermined prior to the relationships among the variables of interest being analyzed [16].

Data was collected as part of a large-scale transnational survey where the web-based survey was advertised on social media (Facebook, Instagram, and Google Ads) and the responses were recorded from June 8 to June 29, 2020. Due to budget limitations, we targeted the advertisement only to Taiwan, the United States, Canada, and Japan. Participants were included if they were aged 20 years or older and able to read and understand the selected language (English, Mandarin, or Japanese). Based on the recommendation for the estimation of a sample size for comparative studies, about 59 participants were needed for the high and low engagement groups (the proportion of the 2 groups was estimated to be 10% and 30%) [17].

A web-based survey, designed by the investigator, was used and developed based on a literature review and the HBM (Figure 1). Details regarding the survey content and development process have been published elsewhere [18].

The survey contained 7 subscales (ie, perceived susceptibility, severity, benefits, barriers, self-efficacy, cues to action, and actions) with a total of 35 items that inquired about an individual’s perception of the pandemic and actions of protective measures. All items were rated on a 5-point Likert scale that showed the level of agreement or frequency (eg, always, sometimes, or never). Higher scores represented higher levels of agreement or more frequent adoption of behaviors. The 7 subscales were defined as follows:

Apart from the abovementioned variables, demographic data were also included in the web-based survey. Cronbach α was .71.

The study was approved by the Research Ethics Committee of the National Taiwan University Hospital (202005043RINC). All participants were required to provide digital written consent before the anonymous survey began.

In addition to the descriptive analysis, cluster analysis was applied to group participants based on their level of engagement in all protective behaviors. The scores of the 6 behaviors were first standardized based on the z scores, given that the scales for these behaviors were different. Additionally, k-means clustering was used, and 3 clusters were determined using the NbClust package [29] in the statistical computing software R (R Foundation for Statistical Computing). ANOVA, Fisher exact tests, and chi-square tests were used to further examine the differences among the groups. Posthoc tests, Fisher least significant difference, and Bonferroni correction were applied to further clarify the directions of the aforementioned analyses. Finally, multinomial logistic regression was applied to adjust the relationships among potentially related health belief variables. A 2-sided P<.05 was considered statistically significant.

Among the 629 responses received, 245 (38.95%) were excluded due to duplication (n=1) or incompletion (n=244). Of the remaining 384 participants (age: mean 39.92, SD 14.65 years), 145 (37.8%) were male, 238 (62%) were female, 1 participant did not specify their sex, 106 (27.6%) were health care professionals or students, and 65 (16.9%) had chronic diseases. Nearly all participants (n=352, 91.6%) had completed a college education or higher. For the past 6 months, 258 (67.2%) participants had lived in Taiwan, 86 (22.4%) in Japan, 28 (8%) in North America, 5 (1.3%) in Europe (Switzerland, Germany, and the United Kingdom), 2 (0.52%) in Hong Kong, 1 (0.26%) in China, and 1 (0.26%) in Macau. The protective behaviors that were mostly adopted by the public were avoiding traveling abroad (n=224, 58.3%), practicing good cough etiquette (n=218, 56.8%), wearing facemasks (n=186, 48.4%), handwashing (n=179, 46.6%), cleaning and ventilating (n=128, 33.3%), and maintaining social distance (n=101, 26.3%). Table 1 displays the demographic data and the frequencies of the adopted protective behaviors.

Cluster analysis divided participants into 3 groups: those who adopted protective measures more frequently (cluster 1, high engagement; n=181, 47.1%), less frequently (cluster 2, low engagement; n=34, 8.9%), and those in-between (cluster 3, medium engagement; n=169, 44%) (Table 2). Note that since the values were standardized scores, negative values do not imply that participants did not engage in such behaviors. For instance, cluster 1 had higher standardized scores (z scores: 0.47334-0.67822) for all 6 behaviors than cluster 2 (z scores: –0.81341 to –1.65617) and cluster 3 (z scores: –0.29885 to –0.41468). Higher z scores represented more frequent adoption of protective behaviors.

ANOVA was used to examine whether the variables of the HBM were different among the 3 groups. The 1-way ANOVA showed significant differences in perceived barriers (F_2,381=3.046, P=.049), self-efficacy (F_2,381=23.935, P<.001), cues of action regarding recommendations from informal information sources (F_2,381=21.152, P<.001), and environmental cues (F_2,381=8.396, P<.001) (Table 3).

The least significant difference posthoc test was performed to clarify the direction of the ANOVA results (Table 4). Individuals in cluster 1 perceived significantly fewer barriers than those in cluster 2 (P=.02). No significant differences was identified between cluster 1 and 3 or between cluster 2 and 3. For self-efficacy, cluster 1 had significantly higher scores than cluster 2 (P<.001) and cluster 3 (P<.001), while cluster 3 had significantly higher scores than cluster 2 (P<.001). Regarding recommendations from information sources, cluster 1 followed behavioral instructions recommended by informal sources more often than cluster 2 (P<.001) and cluster 3 (P<.001), while cluster 3 cluster followed the suggested behaviors more often than cluster 2 (P=.001). When making decisions about adopting protective measures, clusters 2 and 3 considered environmental cues more often than cluster 1 (P=.001 and P=.003, respectively). There was no significant difference between clusters 2 and 3 regarding the consideration of environmental cues. In order to clarify if there was an interaction between the 2 health belief variables that are significantly different among the 3 groups, perceived barriers and self-efficacy were included in the multinominal logistic regression analysis. The results showed that after controlling for perceived barriers, self-efficacy was still significantly associated with group differences (P<.001).

Fisher exact and chi-square tests were used to evaluate whether categorical variables were significantly different among the clusters. The 3 clusters were significantly different based on sex (n=383, χ²₂=8.276, P=.02). Bonferroni correction showed that there were significantly more men (13.8%) than women (5.9%) in cluster 2. A Fisher exact test also revealed that the clusters were significantly different based on the place of residence (P<.001). Bonferroni correction showed that there were fewer participants from Taiwan (37.2%) than Japan (68.6%) and North American (71%) in cluster 1, and more from Taiwan (51.6%) than Japan (27.9%) and North American (22.6%) in cluster 3.

This study identified individuals with different levels of engagement in protective behaviors and the significantly different characteristics among them. We categorized individuals into 3 distinct groups: high, low, and medium levels of engagement in protective behaviors. That is, a group of individuals who stuck to all protective behaviors, while another group engaged in them significantly less. While most studies focused only on the adherence to a single behavior, our study was one of the few that addressed a group of protective measures. Observing and categorizing the adherence to behaviors collectively is particularly valuable when identifying possible populations or factors contributing to gaps in outbreak prevention. These results led to the second aim of our study, which addressed the more important question of the factors associated with the different levels of behavioral engagement.

Individuals from each group were significantly different from each other in terms of both intrinsic and extrinsic factors. Intrinsic factors included sex, perceived barriers, and self-efficacy, while extrinsic factors covered cues to action and residency. For the intrinsic factors, some of our results supported the current evidence, and some demonstrated variations compared to previous studies. Similar to other studies, the results of our study confirmed that sex played an important role in behavioral engagement during a pandemic. Specifically, women were more willing to perform protective measures than men [7,10,30-34]. This suggests that men may need more health education or incentives to enhance their self-protective behaviors. Prior studies have indicated that a high level of self-efficacy was strongly associated with self-protective behaviors [9,13,14,33,35], which aligned with our finding that self-efficacy was a strong factor associated with engagement in protective behaviors, even after controlling for perceived barriers, another health belief variable that was significantly associated with group differences. Individuals with higher self-efficacy in performing protective behaviors and preventing infection are more willing to adopt protective behaviors.

Some studies conducted at the beginning of the COVID-19 pandemic showed that perceived risk, perceived susceptibility [11,14,36], perceived severity [8,9,11,35], and perceived benefits [14,30,35] predicted behavioral engagement. In contrast, our study did not find significant differences in the aforementioned variables among the groups. Our results suggested that perceived barriers were a significant contributing factor [14,30,35]. Several reasons may cause this variation, including the data collection time and location. Compared to other studies, we collected data during a relatively later period, approximately 6 months after the first reported COVID-19 case [37]. It is suspected that fear of a disease decreases when the public knows more about it. Thus, the role of fear-related concepts, such as perceived risk, perceived susceptibility, and perceived severity, in the adoption of preventive measures may not be as important as it was when COVID-19 was an unknown disease. Therefore, future studies should explore whether some intrinsic factors (eg, self-efficacy) remain fairly constant in their impact on protective behavior, while other factors may change over time or by disease status. Inconsistent results may also be related to location, as our study was the first to include a large population from Taiwan, which reported lower COVID-19 cases and deaths compared to other countries [38]. However, our results showed that Taiwanese people were a minority within the high engagement cluster. The relatively stable situation in Taiwan may not have triggered the constant urge to implement protective measures for infection prevention.

It is interesting to note that the 2 extrinsic factors, informal recommendations and environmental cues, had contrasting relationships with the adoption of protective measures. While it seems that all participants followed the recommendations from formal sources to a certain extent, our results suggested that individuals who practiced protective measures more often actually followed recommendations from informal sources more frequently. Alternatively, individuals who adopted protective measures at a medium or lower frequency were more likely to make relevant decisions based on environmental cues, such as the behaviors of surrounding people. There is a tendency for individuals with the highest adherence to protective behaviors to grasp all kinds of information and strictly follow the recommendations. However, individuals with lower adherence made their decisions more flexibly based on the changing situation. Future research should investigate whether these differences are affected by decision-making styles. For example, Scott and Bruce identified 5 distinct decision-making styles. Among these styles, the rational decision-making style is characterized as making decisions based on “a search for and logical evaluation of alternatives,” and the dependent style is “according to advice from others [39].”

This study had several limitations. First, we did not follow the behavioral changes and associated factors longitudinally. Future longitudinal studies are needed to understand the more dynamic phenomenon of the adoption of protective behaviors. Second, the web-based data collection method inevitably reached a younger population with a higher educational level. Thus, our results may not be generalizable to younger or older populations or populations with a lower educational level. The number of participants would be more representative of public opinion if the countries of origin and types of occupation were more equal in number. Specifically, while about a quarter of the study participants were health workers or students, their knowledge and training may have affected their health-related beliefs and behaviors. Future studies may explore if a health-related background can affect health beliefs and behaviors. Third, due to the lack of compensation and the length of the questionnaire, respondents’ motivation was weakened, and about 38% of the responses were incomplete. A similar phenomenon was also observed in other studies, which have shown an effective response rate of web-based surveys ranging from 10.2% to 58.6% [40,41].

This study is one of the few that used an HBM-based questionnaire to survey a collection of protective behaviors against COVID-19 and the associated factors across different countries. The results identified 3 groups of people with different levels of behavioral engagement. These individuals were significantly different from each other in terms of a number of the HBM concepts, including demographics, perceived barriers, perceived self-efficacy, and cues to action. Our results are worth considering in future policy-making and research. Specifically, enhancing self-efficacy may be a powerful way to facilitate engagement in protective measures, especially since self-efficacy continuously affects individuals’ adoption of behavior regardless of the stage of the pandemic. Tailored messages targeted at men during stable but ongoing pandemic conditions are important for minimizing the possible ignorance of these protective measures. Future studies are needed to clarify whether the degree of impact of the associating factors on protective behaviors changes over time, and whether decision-making styles contribute to the engagement with protective behaviors.