In this study, we collected a comprehensive dataset from Facebook (Facebook, Inc), the biggest social media platform. In addition to social networking functions, Facebook allows individuals or organizations to create (public) Pages for users to form communities for various purposes. Facebook Pages have been widely used by companies (including all major e-cigarette manufacturers and retailers) as a platform for marketing and maintaining customer relations [28]. Those Pages also represent active communities for e-cigarette users to discuss topics related to e-cigarettes, including flavors, promotional campaigns, the pros and cons of consuming e-cigarettes, and safety issues. The rich discussions about e-cigarettes in public Pages provide an ideal data source to identify consumers’ perceptions and preferences. and the diffusion of multiple flavors.

Based on keywords generated by domain experts (as Textbox 1 shows), we retrieved a set of e-cigarette–related Facebook Pages through Facebook’s application programming interface (API). We derived the keywords in Textbox 1 from the combination of domain expertise and the published literature [29-31]. For consistency, we manually extracted the Pages related to smoking promotion run by e-cigarette manufacturers and retailers. Finally, we collected the full information of all posts with comments. In total, we collected 7132 e-cigarette–related Facebook Pages with 765,321 posts up to April 24, 2015. Of these posts, 86.68% (663,357/765,321) were generated during 2013 to 2015. A post may receive comments and likes from Facebook users and can be shared by users (to their own Facebook timelines). We collected 2,737,840 comment records and 17,671,614 like records. For each post, we collected the Page identifier (ID), post ID, user ID (who posted the post), time when the post was created, textual content, and the records of comments, likes, and shares. For each comment record, we collected user ID (who posted the comment), time when it was created, and textual content. For each like record, we collected user ID (who clicked the Like button of the original post) and time of clicking the Like button. In total, we identified 1,414,240 unique user IDs. Then, we collected the full public profiles of these users, including their screenname, language, location, and sex. To be consistent, we chose 384,792 posts generated by users with the label “en_US,” indicating they were English-speaking Facebook users located in the United States.

We characterized the influence of different flavors on the information propagation patterns using regression models. In this section, we explain the variables for candidate regression models.

When users browse posts, photos, and other information on Facebook, they can click the Like button for that information, post comments, and share the information to their own timelines. The frequency of a post being shared and liked, and the number of comments received, are explicit proxy measures of information propagation. Figure 1 shows the distributions of these 3 variables (note that we added 1 to each value on the x-axis to avoid the logarithm of zeros on the horizontal axis). In general, we observed a power law–shaped curve in these distributions. This “rich-get-richer” effect indicated that the popularity of a post and the information propagation were unevenly distributed, with most of the posts being seldom shared, commented on, or liked, while a small number of popular posts received a huge number of shares, comments, and likes.

When Facebook user A shares a post published on a Facebook Page, this post then appears in A ’s timeline, as well as on the newsfeed (home page) of A ’s friends. Therefore, the sharing behavior presents the information propagation from the Page to the user and the user’s friends. If one of A ’s friends, B, also shared the same post after reading it from the newsfeed (because A shared it), our data collection also captured this new sharing behavior. It is impossible to differentiate the original shares and subsequent shares caused by specific propagation paths through the newsfeed, because Facebook’s API prohibits the collection of friendship information. On the other hand, 2 additional proxies of information propagation, comments and likes generated by user A, will not be explicitly presented to A ’s friends. Therefore, data on post-sharing behaviors is the most effective and reliable proxy to identify, track, and model information propagation on Facebook [26,32]. In this study, we calculated the frequency of being shared by Facebook users for each post (denoted as Shares) as the dependent variable (representing information propagation) in regression models.

Because of the lack of regulations, manufacturers and retailers do not have a universal flavor classification system. Researchers have used questionnaires and data mining methods to identify a set of the main categories of e-cigarette flavors [12,22,27]. Borrowing and evaluating these categorizations, we identified 11 categories of flavors of e-cigarettes in our dataset: beverage, coffee, sweet, dessert & bakery, fruits, herbs & spices, menthol & mint, nutty, cream, tobacco, and chocolate. We also identified a set of keywords for each category (eg, coke and pepsi are keywords for beverage). It is worth mentioning that the content of a post could contain more than one flavor. Figure 2 shows the distribution of posts that mentioned flavors. Among all of the flavors, fruits was the most popular, followed by sweet and cream.

To characterize the influence of these 11 categories of flavors on information propagation (measured by the frequency of being shared), we introduced 11 binary variables for flavor categories. Each binary variable represented the existence of keywords belonging to the corresponding flavor category.

To avoid bias, we introduced a set of nonflavor-related variables that could have influenced information propagation and correlated with flavor-related variables. In Facebook Pages, manufacturers and retailers often promote their products by offering consumers rewards and gifts by lottery among those users who liked, shared, or commented on the posts. Obviously, such promotional activities would largely increase the appeal of posts to the users. We first identified promotion-related posts based on a set of keywords related to promotions (eg, reward, share, gifts, and free). Then, we added the binary dummy variable promotion to represent the existence of promotion in the corresponding post.

The activeness of a Facebook Page is often associated with its popularity. In general, the more active a Facebook Page is, the more frequently its posts can be viewed by users. To capture this effect, we used the count of posts in a Facebook Page as an independent variable, Posts, to measure the activeness of the Page. In addition, the level of user engagement is diverse because of many unknown factors (eg, the popularity of the brand). To differentiate the influence of flavors and the Page-specific user engagement level, we calculated the average number of shares per post of each Page as a control variable, average share.

The topics conveyed by posts could have a significant influence on information propagation. To capture the potential effect of topics, we employed the commonly used latent Dirichlet allocation, an unsupervised learning model for topic modeling, to extract 3 topics hidden in the text of posts: details about products (product), methods of consuming e-cigarettes (method), and other related discussions (other). Table 1 lists the top 10 most frequent words for each topic. For more details about topic modeling, please refer to Multimedia Appendix 1.

The content of posts often contained URLs and hashtags. URLs provide external information related to the posts. Hashtags are used to help Facebook users label and identify posts with specific topics. Both URLs and hashtags have been found to be associated with the likelihood of information propagation [33]. We introduced 2 control variables, URL mention and Hashtag, to represent the existence of URLs and hashtag labels, respectively. Table 2 summarizes all variables, and Multimedia Appendix 1 summarizes Pearson correlation coefficients.

We evaluated a set of regression models for count data to characterize the influence of flavors on e-cigarette–related information propagation. We used Stata software version 12.0 (StataCorp LLC) to estimate parameters.

The Poisson regression model is the most common method to model count data. It assumes that the mean and variance of the dependent variable are equal; thus, we needed to test the overdispersion effect of the data to confirm the assumption. We used the following z score test to evaluate whether the overdispersion effect in the Poisson regression model was significant enough to violate the fundamental assumption [34]: z=[(y –µ)²– y]/µ√2, where y is the dependent variable and µ is the expectation. We obtained a z score of 1228.886 with a t probability of .04. This indicated that there was a significant overdispersion effect and the Poisson regression model was not ideal for these data. This was also reflected by the poor goodness-of-fit, as indicated by the large value of the Akaike information criterion (AIC).

The negative binomial regression model is widely used to resolve the overdispersion problem by relaxing the Poisson assumption through adding constant dispersion parameter α. However, the negative binomial assumption is difficult to meet when excessive zeros exist in dependent variables.

In our study, we found that 61.1% of observations of the dependent variables were zero. To handle excessive zeros, we used the zero-inflated regression model and the hurdle regression model. In the zero-inflated model, the dependent variable is modeled as a mixture of the count data model (eg, Poisson regression model, negative binomial regression model) and a separate Bernoulli distribution. In the hurdle model, there are 2 components to model the dependent variable: positives are generated by a truncated-at-zero count data model, and zeros are generated by a Bernoulli distribution. Both models can overcome the limit of standard count data models, which assume that zeros and positives are both generated by the same process.

This study was, to our knowledge, the first data-driven research to characterize the influence of categorized flavors in e-cigarette–related information propagation on social media. Surprisingly, we found that flavors did not enhance the popularity of e-cigarette–related information. Certain flavors even reduced the popularity, indicating users’ lack of interest in flavors and potential concern about the safety issues of flavoring additives. For manufacturers and retailers, this study suggests that promoting e-cigarettes with flavors is not an effective marketing approach. For regulatory authorities and policy makers, this study suggests that new policies with updated regulations and restrictions on flavors are needed, for the sake of the health of e-cigarette users.

Our study had several limitations. First, data derived from social media are obviously biased. We only studied English-language content posted by US users. Higher-resolution data with detailed demographic information could improve the practical value of this research significantly. In addition, for consistency, Facebook Pages in our dataset were mainly from parties making or selling e-cigarettes, with a commercial focus. The information propagation patterns on Facebook Pages of nongovernmental organizations and health authorities could be different, thus needing further studies.

Second, information propagation is only one aspect of examining the diffusion of health products. Analyzing the content of information could help us extract users’ opinions and emotions while discussing e-cigarette–related topics. Content analysis could also help us understand the root cause of the lack of interest in flavors revealed in this study.

Third, we evaluated the propagation of information by counting the number of shares of each post. This method measures the scale of propagations well, but could not measure the depth of propagations accurately. The Facebook API prohibited us from retrieving more detailed information about the accurate propagation path because of privacy concerns. There is a need for future research on the depth of information propagation using other data sources (eg, Twitter and Reddit).

Fourth, more data-driven medical research is critically needed to identify the root cause of the lower popularity of certain flavors of e-cigarettes.

This study found that mentions of flavors in posts did not enhance the popularity of e-cigarette–related information. There are several future works that we will pursue. First, we plan to validate the findings of this study using the data of other social media platforms under different cultural and language settings. In addition, we will develop state-of-the-art text mining methodologies to identify social media users’ opinions of flavors and the use of e-cigarettes with different flavors. We will also develop probabilistic topic models to identify various topics related to e-cigarettes for smoking surveillance. This line of social media research has great potential to help e-cigarette manufacturers, retailers, regulatory authorities, and policy makers understand the behaviors and opinions of e-cigarette users. This study demonstrated the potential of using social media data to understand the behaviors of e-cigarette users through an empirical study of flavors, and it calls for more research from other perspectives to fulfill the potential of this valuable big data source.