Tweets were collected using a Twitter firehose application programming interface through Brandwatch, a subscription-based social analytics software. Each tweet was selected for the dataset if (1) it was posted between October 2019 and September 2020, (2) it used at least 1 vaping-related term such as e-cig, vape, and e-liquid (see Table S1 in Multimedia Appendix 1 for full term list), (3) it originated from a handle (username) not identified as an organization or business, and (4) it originated from a geocode within the United States. This led to an initial sample of 63,008 tweets. The selected time period reflects a period of change in the United States, including both the emergence and initial peaks of COVID-19 and new federal restrictions on tobacco products (December 2019: Tobacco 21 and February 2020: federal flavor ban of cartridge-based e-cigarettes).

Similar to prior work, retweets (tweets originally composed by a different Twitter user and reshared by another user) and replies (responses to tweets that may or may not include the original tweet) were retained, as they were assumed to reflect an endorsement of the original post [50].

All tweets containing 1 or more alcohol terms were selected from the overall vaping tweet sample using the grepl() function in R (R Foundation for Statistical Computing) to form the final analytic sample of tweets mentioning both vaping and alcohol (N=6437 tweets). Alcohol-related terms used in prior topic modeling studies were reviewed by the study team, and a final set was selected for the current analysis based on a review of the content analysis literature, team consensus, and team review of initial test searches of the dataset. The final list of terms used was selected to ensure a broad search strategy reflected by tweets, which included use-related (tipsy and drunk), event-related (bar crawl and Thirsty Thursday), and product-related (beer and vodka) terms (see Table S1 in Multimedia Appendix 1 for full list) [46,47,51,52]. A subsample of tweets included in the final dataset was checked by the study team to confirm that they contained at least 1 vaping and 1 alcohol-related search term and that the content of the tweet was relevant to vaping or alcohol.

We used latent Dirichlet allocation (LDA) to evaluate the major topics—or themes—among the sampled tweets. LDA is a machine learning topic modeling approach to identify naturally occurring topics. Model selection was based on the coherence score [53-55]. Specifically, coherence scores were evaluated for each topic count, and the highest score was selected to maximize model fit [56]. Tweets within each topic were then reviewed, and the labels for each topic were created by the research team, reflecting qualitatively emerging themes.

Tweet-level descriptive statistics were generated overall and by topic group. The top 15 most salient words overall are reported. The salience of a word is how informative it was for identifying a distinct topic within the model. More salient terms help differentiate across topics. For each topic, we report the top 15 most relevant terms (relevancy metric λ was set to 0.6 to assist topic interpretation as recommended by Sievert and Shirley [57]), sorted by frequency, along with mock tweets reflecting the topic theme. To protect user anonymity, no tweets were quoted verbatim. Instead, we created mock tweets reflecting the shared perspectives and common themes within each topic group.

Each tweet contains metadata for the tweet itself and the Twitter handle (Twitter’s term for a username). These data are used as descriptive information and for demographic prediction.

We used deep learning techniques such as M3 demographic inference [58] and Bidirectional Encoder Representations From Transformers (BERT) [59] models to predict whether a Twitter handle was likely an individual or an organization, as well as the account user’s gender and age. Using machine learning approaches to predict demographics provides additional contextual data to tweets often missing from user accounts [45]. Given that vaping and drinking vary by age and gender, it is critical to understand potential variations in post activity and content by these key demographics [25,60,61]. A handle was required to have ≥50 tweets for gender and organizational prediction via M3 inference using information from the user’s Twitter bio, handle, username, and profile image (when available) [58]. M3 is a previously validated deep learning system to predict demographic information with an accuracy of 91.8% and 89.8% for gender and organizational status, respectively [58]. For the current dataset, gender was additionally validated against the gender prediction available in Brandwatch, and M3 inference was found to have a 93% agreement rate. Following prior applications of M3 inference, individuals with a probability score ≥0.5 for being a man were categorized as men, while those with scores <0.5 were categorized as women. Users with ≥0.5 probability score for being an organization were categorized as organizations, while those with scores <0.5 were categorized as individuals. Those categorized as organizations were not included in the analytic dataset, as previously noted in the Data Collection and Sample section.

For age prediction (<21 and ≥21 years), we built and trained a BERT [59] model using a dataset of 3000 prelabeled tweets and their last 100 tweets. The training model yielded high accuracy in age prediction (81%). We then used the trained age-BERT model to predict the age (<21 and ≥21 years) of the users using their last 100 tweets (see Multimedia Appendix 1 for additional details on model development).

Tweet sentiment was calculated using Valence Aware Dictionary for Sentiment Reasoning, a lexicon designed for quantitatively analyzing the sentiment of social media text, which has been previously validated and applied to Twitter data [62]. Valence Aware Dictionary for Sentiment Reasoning reviews text and automatically calculates a compound score ranging from –1 (most extreme negative) and +1 (most extreme positive). Scores were used to summarize average tweet sentiment overall and by topic group.

Two federal regulations related to tobacco products were announced during the study period. The federal Tobacco 21 law (December 20, 2019) raised the minimum purchasing age in the United States from 18 to 21 years [48]. On January 2, 2020, a federal ban on flavored cartridge-based e-cigarettes was instituted (with exceptions for tobacco- and menthol-flavored products) [49]. We use the second regulatory change as the regulatory change indicator, given the close proximity of their implementation. Dates from January 2, 2020, and later were considered to have occurred in the post–regulatory change period, while dates prior to this date were coded as the pre–regulatory change period. Variation in the prevalence of each topic across regulatory time periods was tested via chi-square tests, and the resulting P value was reported.

We also examined variations in topic prevalence by whether they occurred before or after the declaration of the COVID-19 pandemic. Specifically, the World Health Organization declared COVID-19 a global pandemic on March 11, 2020. For this study, dates before versus on and after the pandemic declaration were considered pre– versus post–COVID-19 pandemic declaration periods, respectively. Variation in the prevalence of each topic by pandemic time periods was tested via chi-square tests, and the resulting P value was reported.

This study used publicly available Twitter data. Only study staff had access to the full tweets and associated metadata. All data were stored on password-protected devices. No tweets are quoted verbatim to ensure anonymity. There is no path from this manuscript or any supporting material to individual users, usernames, or tweets. The Boston University Institutional Review Board determined that studies using such data did not meet the definition of human participants and was thus exempt.

Roughly 10% of tweets that mentioned e-cigarettes or related terms included alcohol-related terms (6437/63,008, 10.2% after the exclusion of organizational accounts). Among the vaping-alcohol tweets where gender and age could be predicted, 61.9% (n=3448) of tweets were designated as likely originating from men, and 78.9% (n=3708) were likely posted by individuals ≥21 years of age (Table 1). The median follower count was 384 (IQR 141-1146), and the accounts followed were 505 (IQR 218-1170). Average tweet sentiment was largely neutral (80.1%). The majority of the tweets (n=5188, 80.6%) were posted in January 2020 or later, after the Tobacco 21 law and cartridge-based e-cigarette bans went into effect. This is slightly higher than would be expected if tweets were evenly distributed across months (75% of tweets would be from January 2020 or later). A little over half of tweets (n=3773, 58.6%) were posted in March 2020 or later, after the COVID-19 pandemic emerged in the United States. This is roughly what would be expected if tweets were evenly distributed across the time periods.

A 3-topic model was the best fit for the data (coherence=0.39). Including additional groups resulted in less distinct topics, without improving model fit. Vape was the most common among the top 15 salient terms, followed by alcohol, drink, care, smoke, flavor, and store (Figure S1 in Multimedia Appendix 1). While COVID-19–specific terms (ie, pandemic and COVID-19) were mentioned, no specific COVID-19–related themes or terms were frequent or salient across topics. After qualitative examination of relevant terms and full tweets, the following topic themes were identified: (1) flavors and flavor ban (n=3334, 51.8% of tweets; Figure 1), (2) co-use discourse (n=1119, 17.4%), and (3) availability and access regulation (n=1984, 30.8%).

Within the flavors and flavor ban topic (topic 1), the most relevant and frequent terms were vape, flavor, alcohol, adult, tobacco, and product (Figure 1). Tweets tended to include arguments focused on how banning vaping flavors was incongruent with alcohol regulation where flavors are allowed and common. Tweets also often discussed the benefits of flavors for adults who made the switch from combustible tobacco products or espoused the opinion that e-cigarette use was less detrimental to youth’s health and safety than alcohol. This topic had the lowest proportion of tweets from individuals ≥21 years (76% vs 82% in other groups). Mock tweets representing common themes and tone within this group include:

For the co-use topic (topic 2; n=1119, 17.4%), the most frequent words were vape, drink, smoke, alcohol, beer, and teen. Tweets often discussed plans or experiences related to using e-cigarettes and alcohol together in a positive light. This topic included tweets mentioning regulation, but these were less common compared with the other groups. Tweets within this group had a more even gender distribution (51.3% women vs approximately 35% in the other groups; Table 1). They were also more likely (compared to topic 1) to include individuals predicted to be older than 21 years of age (n=610, 82.4%). This topic also had the highest prevalence of both negative (18.3% vs roughly 10% in the other groups) and positive (10% vs roughly 8% in the other groups) sentiment, suggesting more polarization within this topic. Mock tweet examples include:

The availability and access regulations topic (topic 3) represented roughly a third of tweets (n=1984, 30.8%). The most relevant and frequent terms were vape, care, store, liquor, love, and place. Tweets focused on regulation (similar to topic 1) but tended to include discussions of whether and how regulation restricts access. Some tweets supported regulations restricting youth access (ie, Tobacco 21) but expressed concern or anger that stricter regulations would limit access among legal-age adults. Tweets often mentioned age verification or limiting locations of sale as strategies rather than stricter regulations seen as moving toward “prohibition.” Alcohol regulations were held up as a comparison—either as an example of how complete bans of certain vape products may not work or to acknowledge the need to strengthen or enforce age verification processes and limit sale locations similar to alcohol. Statistics surrounding alcohol use and harm were also used for comparison, as were regulations and use prevalence of other substances, including cannabis.

We next examined the distribution of topics by when in the regulatory change period they were posted. While the flavor ban topic (topic 1) represented over half of the tweets across the study period (n=3334, 51.8%, Table 2), flavor-related tweets were more common in the pre–regulatory change period prior to the changes to federal policy at the start of 2020 (Tobacco 21 and ban on flavored cartridge-based e-cigarettes) than in the post–regulatory change period (860/1249, 68.8% of pre–regulatory change period tweets compared to 2474/5188, 47.6% of post–regulatory change period tweets). Co-use discussions were more common in the post–regulatory change period (n=11, <1% of tweets prior to policy change vs n=1108, 21.4% of all tweets in the post–regulatory change period), while availability and access discussions (topic 3) were similar in prevalence across the 2 time periods (roughly 30%). Variations in topic prevalence across time were statistically significant across the 3 topics (P<.001).

Finally, variations in topic prevalence before and after the pandemic declaration were examined. Topic prevalence varied significantly in the pre– and post–COVID-19 pandemic declaration periods (P<.001). The availability and access topic (topic 3) was slightly more common in the post–COVID-19 pandemic period compared to the pre-period (n=1221, 32.4% vs n=763, 28.6%, respectively) as was the co-use topic (topic 2; n=846, 22.4% vs n=273, 10.3%, respectively).

In a sample of tweets mentioning e-cigarettes and alcohol, we identified 3 topics. Most tweets fell into 2 topics focused on vaping regulation while drawing comparisons to alcohol regulations. One additional topic including discussions of both alcohol and vaping (co-use) was also identified. Our findings demonstrated that while there were co-use discussions on Twitter, a large subset of comentions during this time focused on the regulation of these substances both before and after the implementation of federal policies related to tobacco. We also found that topic prevalence varied before and after key regulatory events as well as in the postpandemic period. Specifically, co-use discussions were most common after the pandemic declaration, which may suggest increased stress and substance use during a public health emergency. In addition, while flavor and flavor ban–related discussions were common across the study, they were more prevalent in the time period leading up to Tobacco 21 and the ban on flavored cartridge-based e-cigarettes. This may have reflected increased discussion prior to implementation as a way to express support or concerns surrounding the regulations.

These analyses add to our understanding of the social media discourse surrounding vaping regulation. Prior work examining Twitter discussion in response to the 2016 Deeming Rule by the Food and Drug Administration (FDA) on e-cigarettes found a rise in negative reactions after the announcement of the new regulations [35]. They also found that tweets often debated whether the rule would harm health or focused on how it would impact the new, emerging e-cigarette market [35]. Another more recent analysis that applied LDA topic modeling before and after the FDA’s flavor enforcement policy found similarly negative sentiment after implementation and discussions surrounding purchasing and accessing products after regulatory implementation [63]. Similar to these prior studies, tweets within the current analyses, which were collected during a period of changing policies, often debated the value of regulating e-cigarettes. In particular, how regulations implemented to protect youth would limit adult access to flavors and make purchasing products more challenging. Public health officials and policy makers should be aware of regulatory discourse in order to understand current public knowledge and sentiment as well as to contribute accurate and timely information to address concerns surrounding regulatory shifts. This may be particularly important in situations where posts are generated or elevated by the industry either directly or indirectly (ie, astroturfing) [64].

A novel finding was the use of alcohol as a regulatory comparator when reflecting on vaping regulation. While some tweets acknowledged alcohol regulation was beyond the FDA’s purview, the opinion that vaped nicotine products are more regulated than alcohol was consistently used to argue against vaping regulations. Additionally, use statistics among youth were commonly shared as evidentiary support. Specifically, some reported alcohol as a more common, harmful substance used by youth compared with e-cigarettes.

The prominence of access and regulatory commentary during this time period may have been in response to the new, evolving nature of regulations. One study analyzing tweets responding to an antivaping campaign found frequent objections to regulation as well as tweets touting e-cigarette health benefits [37]. In another study focused on marijuana use mentions—a substance experiencing regulatory change—calls for legalization were common [65]. Recent regulatory changes may have been more salient after the pandemic declaration due to challenges with access from business-related closures during this time period.

The identified co-use topic demonstrates an opportunity for those on these platforms (including youth audiences) to be exposed to content, which may normalize joint use. Common terms defining this subtopic included broad alcohol terms such as “alcohol” and “beer” but also terms that may be associated with heavy drinking such as “frequent” and “binge.” Our search strategy of including both general alcohol terms as well as common excessive drinking terms in the United States such as “blacked out” and “trashed” may have allowed us to more readily identify co-use tweets that would have not been captured with more standard drinking terms. These findings emphasize the importance of understanding cultural terms and specificity when collecting and analyzing social media data.

Counteradvertising efforts may be particularly useful during times of crisis and change as pathways for sharing accurate information in a timely fashion [44]. The unique time period of data collection may have also affected co-use discussions. Interestingly, the co-use topic was almost entirely restricted to the postpolicy period and was most common during the pandemic period. While regulatory shifts may have brought substance use discussions to the forefront, the early stages of the pandemic likely influenced co-use discussions. Over three-quarters of tweets within this topic were more specifically from the period after the COVID-19 pandemic was declared in March 2020. A prior study found alcohol use mentions related to coping during COVID-19 and addiction concerns during this time period [36]. With individuals spending time at home and living their social lives in virtual spaces, individuals during this time period may have increased their likelihood of sharing experiences related to substance use and co-use in particular. As LDA is a data-driven process, it is possible that if we had run the LDA separately across time periods for the regulatory change period or the COVID-19 declaration period, different themes may have emerged. Future research could adopt this approach to explore additional themes that may be missed within a larger dataset.

This study has important limitations. First, analyses were limited to Twitter and to the text of tweets. Findings may not be generalizable to other social networks or more visual content (eg, videos and images), which are an important element of social media discourse. Second, although we included retweets, we did not explore key influencers that may drive content and were unable to examine the demographics of tweet viewers (ie, whether tweets reached underage youth). In addition, we did not distinguish between human and bot accounts. Furthermore, it is unclear whether users expressing frustration over regulations were industry funded or endorsed accounts. An analytic limitation is that LDA uses a “bag of words” method, which ignores word order and context. In addition, our focus on tweets mentioning both vape- and alcohol-related terms and the resulting smaller sample size likely limited the identification of additional topics. Finally, while we reviewed a subsample of tweets to ensure they included the relevant vape and alcohol search terms, we did not examine tweets for whether they were germane to vaping or alcohol use. It is therefore possible that although a term was used, the tweet was not focused on these topics.

This study also has several significant strengths. First, we used a data-driven approach to identify tweets mentioning both vaping and alcohol. We made no a priori assumptions about the number of topics or the thematic groupings and were able to quantify salient and frequent terms mentioned within identified themes. Additionally, we used state-of-the-art machine learning algorithms to predict account user demographics. Finally, we analyzed tweets from a time period, which included many changes in the United States, shedding light on how policy change and major national emergencies may shift social media discourse surrounding substance use.

In conclusion, our study identified 3 distinctive topics across tweets including both vape- and alcohol-related terms. Specifically, a discussion of flavors and pending flavor bans was the largest topic subgroup, followed by a discussion of potential restrictions to e-cigarette availability and adult access resulting from regulatory changes, and finally, a portion of tweets included co-use of alcohol and vape products. While there were co-use mentions, tweets more often mentioned alcohol to make comparisons regarding how the substances were differentially regulated. Our findings demonstrate how the still-evolving landscape of electronic nicotine product policies influences social media discussion and debate. It also signals an opportunity to better understand and prepare for common concerns surrounding proposed regulatory changes—particularly as it relates to limiting youth access. A deeper understanding of social media discussions may inform health communication campaigns and policies focused on restricting underage youth’s access to substances and advancing youth’s health.