Twitter data were obtained from Radian6 [33], a leading social media monitoring tool that collects data from more than 400 million sources across social networking sites (eg, Twitter, Facebook, YouTube), forums, blogs, and mainstream news. Radian6 offers historical Twitter data from July 1, 2008, and provides users full access to all available tweets from the Twitter firehose.

To identify tweets about e-cigarettes, we developed a search syntax that included 55 search keywords, including general e-cigarette terms (eg, electronic cigarette, eCig), specific e-cigarette brand names (eg, blu, NJoy, green smoke), and terms about e-cigarette use (eg, vaping). We reviewed the initial search results and amended the syntax to exclude other tobacco or drug terms (eg, marijuana, hookah). The final search syntax was entered into Radian6 to identify relevant tweets from July 1, 2008 to February 28, 2013 (when the search was conducted). The search results were downloaded into Microsoft Excel from Radian6, including the date and time of the tweet, Twitter handle, the entire text of the tweet (including URLs, hashtags), and the number of followers at the time the tweet was posted.

Data processing was done in the Python programming languages. We extracted hashtags and links from the text using regular expressions [34]. For hashtags, we used the regular expression #[a-zA-Z][a-z-A-Z0-9_]. For URLs, we used +:/{2}[-]+(.[-]+)(?:(?:/[^s/])) [34]. Because many links are URL shorteners (eg, bit.ly, t.co), we unshortened the URLs using the application programming interfaces (APIs) provided by unshort.me and Unshorten.It! [35].

We randomly selected 507 tweets from the Radian6 corpus and manually classified them as advertising or not advertising. A tweet was coded as advertising e-cigarettes if it mentioned specific brands or websites and listed a price, promotional offer, and words such as “buy.” By using the scikit-learn module in Python [36], we built a classification algorithm to tag the remaining tweets as advertising or not. We used text content of the tweet and 5 metadata features in our classification algorithm: follower count at time of tweet, following count at time of tweet, number of tweets at time of tweet, a binary feature indicating whether or not the tweet contained a URL, and a binary feature indicating whether or not the tweet was a retweet.

To turn the tweet text into a feature matrix [37], we used scikit-learn’s CountVectorizer method [36]. CountVectorizer creates a feature for each word and, in this case, a feature for every n-gram of length 1 to 4 [37]. For each tweet, the value of the feature for a given n-gram was one if the tweet contained the n-gram and zero otherwise. Initially, we had 15,313 n-gram features along with the 5 metadata features.

Because the number of features was 30 times the number of data points, we performed feature selection before training a classification model [38,39]. We trained an extremely randomized trees model [39] (implemented in scikit-learn as ExtraTreesClassifier) on our data and selected only the features with nonzero feature importance. This resulted in 2167 features. We then fit the reduced data with a random forest model (as implemented by scikit-learn) with 10 trees. We scored our model with a tenfold cross-validation according to the accuracy metric (proportion correctly classified). (Because we were concerned about correct classification in both classes, precision and recall were insufficient metrics here.) The mean accuracy across the 10 cross-validations was 0.907. Because 78% of the tweets in the training data were advertising, the naive classifier that labeled everything as advertising would result in an accuracy of 0.78. Our model was a 16% improvement over this baseline. The classifier had high precision (91%) and recall (93%).

By using a combination of natural language processing and manual classification, we extracted prepositional phrases indicating where people used e-cigarettes or observed others using them. We started with the corpus of tweets that we classified as not advertising.

First, we selected a subset of the tweets containing the words smoking, vaping, smokes, vapes, smoked, and vaped that indicated active usage of e-cigarettes. Second, we cleaned each of these tweets according to a series of rules that were iteratively developed to maximize the number of prepositional phrases extracted. These rules included removing @-mentions and hashtags, removing various interjections (eg, “lol,” “smh,” and various expletives), removing common subordinate prepositional phrases (eg, “in the middle,” “in front”), and correcting common misspellings and slang commonly used in text-based communication (eg, “ur” to “you’re”). Third, we ran each cleaned tweet through the parts-of-speech tagger in the Natural Language Toolkit (NLTK) library [40]. The tagger attempted to label each word in a document (ie, tweet) with its part of speech. Fourth, to extract prepositional phrases from the tweets, we extracted each preposition and all subsequent words before the next verb, preposition, or “wh-” adverb (eg, “whenever,” “where”). Statistical methods for extraction of prepositional phrases, such as the Stanford Parser [41] are available. However, because of the prevalence of poor spelling and bad grammar on Twitter, the Stanford Parser proved ineffective at correctly extracting prepositional phrases. Looking at a sample of 100 tweets containing prepositional phrases, we determined that the 3 primary prepositions indicating the location of e-cigarette use were “in” (62 tweets), “at” (12 tweets), and “on” (11 tweets). The next most frequent prepositions were “around” (3 tweets) and “during” (3 tweets). Thus, we restricted further analysis to tweets containing one of the top 3 prepositions. Finally, we manually compiled a list of nouns that commonly occur in prepositional phrases, but do not refer to a physical location (eg, at “night,” in “life”). Prepositional phrases containing these words as objects were excluded from our analysis. Once prepositional phrases with these objects were excluded, we were left with a list of tweets containing references to e-cigarettes being used in physical locations.

In summary, we found that Twitter conversations about e-cigarettes have increased dramatically in recent years. This pattern is consistent with the recent rise in e-cigarette advertising expenditures [8] and e-cigarette use among youth [46] and adults [47]. It is not surprising that the majority of tweets appear to be advertising-related because in the early introduction of a new product claims are often made that new products are more cost-effective (eg, claims that e-cigarettes are cheaper than tobacco cigarettes and nicotine replacement therapy) [48]. Furthermore, because anyone can create a Twitter account and start posting content, e-cigarette vendors can freely “advertise” their products via tweets to the more than 316 million monthly active users with no advertising costs per se. This also means that anyone—whether a legitimate brand, an online-only vendor, an affiliate marketer, or an e-cigarette user/enthusiast—can post content. Indeed, our look at the most active Twitter handles suggests that actual online vendors and potential affiliate marketers are promoting e-cigarettes online, with some potentially using computer programs to generate and post tweet content automatically. This practice is not unique to e-cigarettes as evidenced by the fact that, for years, Twitter has been attempting to identify spam accounts that are generating automated content from computer programs and not actual individuals. This type of “spamming” suggests that there is an online marketplace for e-cigarettes whereby individuals can earn money by driving customers to visit an e-cigarette vendor website. This also means that the online marketplace is more diverse than traditional brick-and-mortar offerings; the top brands mentioned in tweets, except for blu eCigs, are not the leading brands advertised on other media channels (eg, television) or sold in retail stores [8].

The ease of posting user-generated content online and sharing this information across social media platforms such as Twitter suggests that Twitter users may be exposed to more e-cigarette brands and online vendors than non-Twitter users. Although only approximately 23% of the US adult population is on Twitter, use rates are highest among young adults aged 18 to 29 years [17] and have been increasing among youth in recent years [17,49]. Just because youth are on Twitter does not necessarily mean they are exposed to these tweets. One would have to follow these Twitter accounts or be exposed to the tweets through their social network (ie, followers or those they are following) or via searches. This study was not designed to examine audiences that may have been exposed to these e-cigarette tweets; however, recent studies suggest that e-cigarette users learn about e-cigarettes from sources such as the Internet. Therefore, monitoring how e-cigarettes are discussed on social media platforms such as Twitter is important, especially if frequent tweets about e-cigarettes from multiple Twitter handles may give consumers the false perception that e-cigarettes are readily available and use is more common than it really is. Previous research has shown that tactics of fostering a false sense of consensus have been used to oppose passage of e-cigarette policies [28]. We also know from the literature that youth overestimate the prevalence of youth smoking, so exposure to frequent tweets about e-cigarettes and visual cues of people vaping their e-cigarettes via photos may influence their interest in and perceived social norms about e-cigarette use [50].

Our analysis of patterns of tweets among the most active Twitter users also revealed important patterns (eg, accounts gaining many followers during a short window of time) that suggest some of the most active accounts may be engaging in practices such as purchasing followers. Previous research has shown that abnormal patterns in posting behavior (eg, bursts of posts in a short period of time) are indicative of spamming or fake consumers [51]. In this context, a burst of followers for a particular e-cigarette-related Twitter handle likely suggests spamming. Additional work is needed in this area because distinguishing legitimate e-cigarette companies/vendors from spammers/affiliate marketers will help prioritize and inform future regulatory efforts.

Our analysis of nonadvertising tweets indicates that organic conversations about e-cigarettes are occurring online that can provide insights into consumer use behaviors. Interestingly, we find posts about e-cigarettes being used in public spaces, with class being the top mention. Although some of these tweets indicate that students (and staff in some cases) were using e-cigarettes in class, the majority of tweets were from nonusers who expressed surprise or disdain at seeing others using e-cigarettes in public and were confused as to whether this is allowed. We do not know from the tweets whether these were youth or young adults tweeting about a high school class or a college class, and future studies should examine ways to determine the demographics of tweeters. Regardless, these results suggest that youth and young adults may be exposed to e-cigarette use in their everyday lives, and the lack of action by staff may give them a sense that e-cigarette use is permissible and not as harmful as cigarettes. This is a concern among public health professionals and increasingly local governments have introduced [52] and passed [53] policies banning the use of e-cigarettes in public places and in schools (eg, [54]).

This study has several limitations. Although we attempted to identify the entire population of tweets about e-cigarettes using a comprehensive search strategy, we may have missed some relevant tweets. Additionally, we attempted to characterize the most active Twitter handles, but our analysis was based largely on posted information in their profile, which may not reflect who they truly are or their motivations for tweeting e-cigarette content. This is a general challenge in conducting Twitter research and more advanced computational methods are needed to mine the profile descriptions and tweet content of any given Twitter handle to determine the individuals’ demographic characteristics. Additionally, we do not know the extent to which Twitter users were exposed to e-cigarette tweets. Finally, the data presented in this study may have limited implications today given likely changes in the e-cigarette marketplace, brands’ advertising strategies, e-cigarette policies, and consumer behaviors.

This study builds on prior work assessing e-cigarette marketing on Twitter (eg [29]). A key strength is that we analyzed data across multiple years to characterize the emerging trend of e-cigarette conversations on Twitter. We used a data source with access to the full Twitter sample and historical tweets. In contrast, most studies using Twitter data have examined substantially shorter time periods (eg, months) using freely accessible data from the Twitter API, which only provides 1% of all tweets. The Twitter data and analyses presented here also helped us to understand a key period (2008-2013) for e-cigarette marketing when these devices began to increase in popularity. This study also identifies locations where e-cigarettes are being used and suggests the need for better enforcement of policies restricting e-cigarette use in public places such as schools.

In conclusion, Twitter may provide valuable insights into how information about new products, such as e-cigarettes, are disseminated via social media. Our results suggest that Twitter is being used to promote e-cigarettes and that the online marketplace is more diverse than the offline brick-and-mortar product offerings and advertising strategies. Monitoring and regulating these entities online will be challenging given how easily anyone can set up multiple social media accounts with no verification of their identity before they advertise their products to consumers worldwide. Therefore, it will be important to not only examine the content of e-cigarette advertising tweets, but other characteristics of marketers, such as frequency of tweeting behavior and patterns of acquiring followers, to identify entities that state and federal governments may need to monitor and regulate. For example, enforcing advertising restrictions may have different implications for an actual e-cigarette brand company with brick-and-mortar presence versus an individual who is an affiliate marketer for an online e-cigarette retailer. Our results also highlight that e-cigarettes are being used in public places such as schools and underscore the need for education and enforcement of policies banning e-cigarette use in public places. In summary, Twitter data can provide new insights on a rapidly evolving public health phenomenon such as e-cigarettes to help inform future research, regulations, surveillance, and enforcement efforts.