We used the mobile WeChat app to search the official accounts by the name of the province and the key words “Centers for Disease Control and Prevention” and “CDC.” We found 28 WOAs of provincial CDCs on January 15, 2018, and subscribed to them. Data for this study consisted of WeChat articles found on these WOAs between January 1, 2017, and December 31, 2017.

For each article, we recorded the code of official accounts, push time, and amounts of article reading and liking. We developed a features frame for each article, referring to the research of Kite et al [17]. Specifically, we added article content, title type, and article length, which are features that may affect the effectiveness of the WOAs. Next, we conducted a presurvey on 100 articles. Based on the results of the survey, we have merged some of the less frequently used categories into other and made some modifications during iterative testing to make the features frame more relevant to public health communication. The final features framework with definitions is shown in Table 1. Then we invited six experts in relevant fields to evaluate the content validity of the features frame and calculated the content validity index (CVI). Experts used a 4-point Likert scale to assess the degree of consistency between the content of each item and the corresponding article features. Very unrelated was counted as 1 point, comparatively unrelated was counted as 2 points, comparatively related was counted as 3 points, and very relevant was counted as 4 points. The CVI of the item level was 0.83 to 1.00, and the CVI of the entire framework was 0.98.

We coded the data quantitatively using a coding scheme that assigned numeric values to article features. Title type, article content, article type, communication skills, and article length contained mutually exclusive categories, and they were coded with the corresponding number. However, multiple marketing elements are possible in a single article, so not all categories were mutually exclusive. Each marketing element was coded as 0 or 1 (0 represents not using, 1 represents using). Next we tested the interrater reliability between the two coders. After receiving training and getting familiar with the content, LFH and MJY independently coded the same subset of articles (n=300) from 28 WOAs. Any disagreement between the coders was resolved by discussion. Once interrater reliability reached 80%, LFH and MJY then individually coded half of the WOAs each.

For every WOA, we examined the descriptive characteristics of articles, including the number and percentage of articles pushed in a year, median reading, and median liking per article. Next we generated descriptive statistics for each title type, article content, article type, communication skill, marketing element, and article length. We then investigated associations between the features of the articles and user engagement, operationalized as reading and liking. Since the data distribution is skewed, we categorized the amount of reading and liking by the 75% position, defining an amount of reading and liking less than the 75% position as low-level reading and liking and above the 75% position as high-level reading and liking. Given that the marketing elements were multichoice, we converted the features of marketing elements into the number of marketing elements. Next, we applied logistic regression analyses to assess associations between the features of the articles and the user engagement with the level of reading and the level of liking as the outcome variables and six features as categorical independent variables and used the P value to represent the result of the hypothesis test, which decided whether to reject the null hypothesis (the regression coefficient is equal to 0). We conducted a series of two-category univariate logistic regression analyses to perform the initial screening of variables. Variables that gave P<.10 in the univariate analyses were evaluated further using multivariable logistic regression. For the multivariable regression, P<.05 was considered the statistically significant level. Risks were expressed as adjusted odd ratios with 95% confidence intervals. EpiData 3.1 software (EpiData Association) was used to establish the database; SPSS Statistics version 15.0 (IBM Corp) was used for the statistical analysis.

Overall, 28 provincial CDC WOAs pushed a total of 5976 articles in 2017 (Table 2). Guangdong CDC pushed the most articles (565/5976, 9.45%), and Hainan CDC pushed the fewest articles (4/5976, 0.07%). Shanghai CDC articles attracted the most user engagement (median reading: 3777, median liking: 37), and Ningxia CDC articles attracted the least (median reading: 21, median liking: 0). For all articles, the median reading was 551.5 and the median liking was 10.

As described in Table 3, article titles were usually declarative sentences (2358/5976, 39.45%). Most articles were related to healthy lifestyle (1355/5976, 22.67%), and text with pictures (3686/5976, 61.68%) was the most common article type. The most common communication skill was informative or instructive (3096/5976, 51.81%). Only 60.66% (3625/5976) of the articles contained any marketing elements, and the most articles used only one kind of marketing element (2543/5976, 42.55%). The article length was usually 1000 words or fewer (3615/5976, 60.49%).

As shown in Table 4, univariate logistic regression analysis revealed that title type, article content, article type, communication skills, number of marketing elements, and article length were significantly related to reading level (P<.001). Liking level displayed a similar pattern. All article features were evaluated further using multivariable logistic regression.

Articles for which the title type was an exclamation or emphatic sentence were less likely to obtain a high level of liking than those whose title types were declarative sentences. With regard to article content, articles about infectious diseases, vaccination, food safety and nutrition, and healthy lifestyle were more likely to receive high-level reading and liking than those about other topics. Articles related to health education activities or national health policy conferences were less likely to obtain high-level reading and liking. Environmental and occupational health–related articles were more likely to obtain high-level reading, unlike articles describing research progress, but both showed no effect for level of liking. For article type, articles containing text, pictures, and links were 5.21 times and 5.64 times more likely to receive high-level reading and liking, respectively, than text-only articles. Those containing text and pictures, similar to those containing text, pictures, and videos, were also more likely to receive high-level reading. With regard to communication skills, when compared with informative or instructive articles, those using traits of humor, questioning, fear appeal, and other were more likely to receive high-level reading and liking. Articles with positive emotional appeal were more likely to obtain a high-level liking, but this had no effect observed for level of reading. Articles using three or more kinds of marketing elements were more likely to obtain high-level reading and liking than those using none, while articles using one and two kinds were both less likely to obtain high levels of reading and liking. Compared with articles containing 1000 words or fewer, those containing 1000 to 1500 words and 1500 to 2000 words were more likely to obtain high-level reading and liking.

To our knowledge, this is the first study to investigate the current status of WOAs of CDCs for health information dissemination. Our analysis identified several features of articles that were associated with better information dissemination, which we defined as high-level user engagement. In the field of public health research and practice, WeChat represents a convenient and accessible tool for health education in China, which historically has been difficult using traditional methods [2,19]. Health promotion organizations should be aware of strategies to engage their target audience. The results presented in this paper may provide these organizations with some guidance on how to improve engagement with WeChat users.

According to our results, the content of articles was associated with user engagement. Indeed, article content was identified as an essential factor in determining whether WeChat users forward or share articles with friends [20]. Research on other social media platforms has also revealed that the content of posts seems to have a significant effect on user engagement [16,21]. Our findings further showed that what the public liked to read and praise were articles about infectious diseases; vaccination; food safety and nutrition; and healthy lifestyles, which reflects that they seem to be interested in popular science articles about daily health knowledge. In contrast, the public was relatively less interested in articles related to research progress and health education activities of institutions or national health policy, which may seem to be far away from daily life.

As the results showed, article type was an important covariate of user engagement. This is consistent with previous research. Research found that article type is an important indicator that alters the effectiveness of article dissemination on WOAs [22]. Among studies based on Facebook, the effect of the post type is also displayed. One found higher engagement to be associated with pictures, videos, and links [16]. Another showed that video posts were more attractive post types than picture posts [17]. Our study seems to give another possibility related to the combination of article types. A combination of text, pictures, and links was the most engaging article type, 5 times more popular than text only. Although the combination of text, pictures, and videos and the combination of text and pictures were also more likely to get a high-level reading and liking, the effect was far lower. This seems to indicate that links play a more important role in the combination of article types for increasing the user engagement. The effect of the links can be easily found in the advertising field. Ads may work better if the users can directly access the relevant pages through the links in the social media website pages instead of being forced to visit the external website [23]. And in the field of health promotion using networking platforms, relevant links are also identified as a key strategy for successful user engagement [24]. However, only 26.4% of all articles we coded included the combination of text, pictures, and links and the combination of text and pictures accounted for 61.7%, which suggested that public health organizations were trailing behind marketers in advertising field. The administrators of the CDC official accounts should probably add links to the traditional combination of text and pictures.

The communication skills used were also related to user engagement. Compared with the use of a peaceful tone to convey health knowledge, articles using humor were more likely to receive high levels of reading and liking. This result is similar to the findings of previous research: Klassen et al [25] found that using an optimistic tone was associated with more interactions on Facebook, and posts from health promotion organizations that were more serious in tone had minimal engagement from fans. This may reflect the effect of emotions [26]. Some researchers think that emotion is an important motivator of social sharing, and when participants experience positive emotions viewing a post on social media, they are more likely to engage with that post than are those who do not experience positive emotions [27,28]. In addition, articles using questioning skills were also nearly two times more likely to achieve high-level reading and liking than those using informative or instructive skills. Some surveys have found that many respondents believe that internet health information is not reliable [2,29]. Indeed, there are potential dangers associated with using social media for health communication, such as sharing of misleading or inaccurate information [5]. As professional health organizations, CDCs use questioning expressions to correct this information, which may attract readers’ attention. According to our results, articles with fear appeal were likely to show similar but slightly lower results than those using the above two skills. In the past, fear has been employed by many health promotion organizations to induce behavior change [25]; however, Soames Job [30] and Keller [31] think that fear is only likely to work under particular circumstances. Therefore, the effectiveness of fear appeal may require further research.

The number of marketing elements was associated with the level of user engagement. Articles that used one or two marketing elements were less likely to obtain high-level reading and liking than those that used none. However, articles using three or more kinds of element obtained the opposite result. This showed that some marketing elements were associated with lower levels of engagement and some with higher engagement. Previous research has discussed the impact of some marketing elements on user engagement in other social media. The effect of celebrities is usually positive. Chapman [32] believed that celebrity involvement in public health campaigns delivers long-term benefit. Similarly, Kite [17] found that the use of celebrities and athletes led to higher average engagement. However, the effect of authoritative people is still controversial. One study found that, compared with Facebook posts that have no marketing elements at all, using authoritative people reduced likes, shares, and comments [17]. But Preece et al [21] thought that using charismatic leaders or respected authorities increased the chances of user engagement. Therefore, which marketing methods can play positive roles in increasing the user engagement on articles of WOAs remains to be further discussed.

The results of this study showed that article length was also a covariate of user engagement: articles containing 1000 to 1500 words and 1500 to 2000 words were more likely to obtain high-level reading and liking. A study of WeChat articles considered that text above 1500 words can be considered as a long text [33]. Our study showed that users seem to have a preference for articles of moderate length.

In our study, title type was negatively associated with user engagement. When the title was an exclamation or an emphatic sentence or phrase, the articles were less likely to receive high-level liking compared with titles involving declarative sentences. Zhang [34] analyzed the influence of WeChat article titles on people’s willingness to open the full text and found that the content of the title could affect the dissemination of the article. However, we do not have sufficient evidence to demonstrate what type of title was more popular with readers. Therefore, the link between title type and user engagement seems to present an interesting direction for future research.

Our findings should be interpreted with consideration of some study limitations. First, the study involved a short time frame of data collection. Evaluating a longer period may identify time differences and improve strategies. Second, this study used reading and liking as indicators of user engagement, and we further transformed them into two-category variables: high-level and low-level. We may thus be obscuring the effects of the original variables. Future analyses should expand our findings by quantitatively evaluating the indicators associated with engagement. Last, other factors that we have not focused on in this study may also shape user engagement such as individual-level factors. These individual factors and article features should be analyzed comprehensively in future research.

Social media involves technologies that facilitate opportunities for engaging with the audience [35]. How social media can best be used to achieve health information dissemination and public health outcomes is a topic of much discussion and study in the public health community. Given the lack of related studies based on WeChat or official accounts, we conducted this study and found that the article content, article type, communication skills, the number of marketing elements, and article length were associated with reading level and liking level. However, title type was only associated with liking level. Our results may provide public health and community leaders with insight into the diffusion of important health topics of concern. Because this study was focused primarily on user engagement and article features, future studies might improve our understanding of other factors that contribute to the dissemination of specific key themes. For example, the agency must identify what audience they are trying to reach, how that audience uses WeChat, and what goals and objectives are most appropriate.