The studies included in this review used a variety of analytical methods, including qualitative analysis, quantitative content analysis, machine learning, rule-based, and statistical analysis. The types of analysis include sentiment assessment, thematic analysis, content analysis, named entity recognition, social networks, and geographic analysis. Table S3 in Multimedia Appendix 1 summarizes the analyses.

Themes were identified in 62% (26/42) of the studies. Manual coding of the themes was performed by 69% (18/26) of the studies, either by using pre-existing categories or by observing a sample of the data and generating a codebook [22,23,25,26,28,30,31,37,41,47-49,52,55-58,60]. Of the 26 studies, 2 (7%) studies used the services of social media data analytics companies [42,50].

Of the 26 studies, 4 (15%) studies used topic modeling to infer themes or topics [34,35,46,54]. The algorithm of choice for this task is the latent Dirichlet allocation [66]. The choice of the number of topics was based on intrinsic evaluation metrics (eg, coherence and perplexity) and iterative qualitative analysis informed by prior experience with topic models. Of the 26 studies, 1 (4%) study used temporal topic modeling techniques to study changes in topics over time, with the goal of analyzing how web-based vaping narratives changed during the COVID-19 pandemic [46].

Of the 26 studies, 1 (4%) study identified themes by using rule-based methods. Frequency counts of the most common unigrams and bigrams were generated and formed the basis of the topics [45]. Another study used SAS Text Miner software, a text-topic node algorithm, to discover topics [38].

Socioeconomic and demographic analyses of the study population were performed in 26% (11/42) of the studies. Of the 11 studies, 2 (27%) studies used the provided gender, age, and other user characteristics from user profiles or inferred from posts by users [33,52]. Of the 11 studies, 2 (27%) video-based studies used the perceived age and gender of the subjects after observing the videos [25,26].

Of the 11 studies, 2 (18%) studies that used social media analytics providers obtained age and gender data by using the supplied analysis [42,50]. Of the 11 studies, 2 (18%) of the Twitter-based studies used a commercial tool called DemographicsPro, which uses proprietary algorithms to infer user demographic characteristics [23,28]. Other studies used existing census data [32], demographic information obtained from survey data [37], and a 2-step method based on a gender-name lexicon and a face recognition algorithm applied to users’ profile information to identify the users’ gender [43].

Geolocation data analysis was performed in 40% (17/42) of the studies. User profiles or message metadata were used in 52% (9/17) of the studies [24,29,32,34,36,43,54,55,60]. Of the 17 studies, 2 (12%) studies used information provided by social media analytics companies [38,50]. The DemographicsPro tool was used in 5% (1/17) of studies [28]. Of the 17 studies, 3 (17%) studies used location information provided by Google Trends [40,44,51]. Another (1/17, 5%) study collected geographical information from survey data [37]. Of the 17 studies,1 (5%) video-based study used the geographic location of video channels [26].

An individual’s perception of a topic can be characterized as having a positive, negative, or neutral sentiment. The analysis of these sentiments is often performed using automated language tools and is named “sentiment analysis” [67].

Out of the 12 studies that performed sentiment analysis, 5 (42%) used automated methods. Of the 12 studies, 1 (8%) study trained a binary Naive Bayes classifier on a sample of 1000 “marijuana” related tweets to classify posts into 2 opinion polarities, positive and negative or neutral [32]. Another study used sentiment analysis provided by a social media analytics company [50]. Of the 12 studies, 3 (25%) studies used Valence Aware Dictionary and Sentiment Reasoner (VADER) [68], a lexicon and rule-based sentiment analysis tool [43,54,62]. The VADER performance was compared with in-house machine learning classifiers trained on 3000 manually coded cannabis-related tweets, which showed a 30% performance improvement over VADER. Although VADER is widely used for general tweet sentiment analysis, its performance suffers in substance-use-related domains where negative words are often used to carry positive sentiments. For example, “I took CBD oil, that stuff was bad” [69]—in this sentence, “bad” actually means good.

For conducting user analysis, 57% (24/42) of the studies examined either the subject of the posts, as from individuals or others (ie, from self, retail, media, or professionals), or who the post was about (self, others, or general) [22,23,25-29,33,37,41-43,45,47-50,52,55,57,58,60-62].

When manual data labeling was performed, the determination of both the poster and subject of the post was part of the labeling process. Self-reporting and self-use were easily determined by observation of videos, as were most texts based on the structure of the language. For example, a study [27] first identified whether the subject of the tweet was about the self, other, or general and then identified whether the tweets were about actual cannabis use. This study included further categories of tone, related behavior, perceived impact, and social context. Automated labeling approaches look for phrases that indicate self-reporting. For example, a study on opioid addiction [42] looked for phrases such as “I am addicted” and “I have been addicted” in the context of opioid mentions. Classifiers were used in another study [59] to separate marketing tweets from nonmarketing tweets; however, their focus was on marketing tweets.

None of the studies used advanced natural language processing techniques to establish subjects and personal mentions. Social media bots are automated accounts that generate artificial activities on social media platforms [18]. Bot detection was used in only 4% (1/24) of studies, which used Twitter as a data source [45].

Of the 42 studies, 2 (5%) studies examined the social networks of contributors to conversations. This allowed the identification of target communities and user interactions [34,43]. Of the 42 studies, 3 (7%) studies examined the impact of governmental cannabis legalization policies on the sentiments and opinions of people or on the volume of social media posts [24,28,54]. Term frequency and count analysis of words and phrases was performed in 12% (5/42) of studies [29,39,50,62,63].

Institutional review boards (or their equivalents) ensured that research using human participants is conducted in an ethical manner [70]. Approval for and overseeing of a study by an institutional review board ensures that researchers adopt an ethically appropriate research protocol that respects the rights and interests of social media users; 62% (26/42) of the studies mentioned an ethics approval review being sought or the study being exempt from ethics requirements. There was no mention of ethics approval in 38% (16/42) of the studies.

The use of standard reporting systems, such as the US Food and Drug Administration reports, helps to assess whether social media research findings can be generalized to real-world data. When a suitable ground-truth data set is not available, validating results against >1 social media platform improves the generalizability and validity of the results. Only a few studies used >1 social media data source or validated their findings against other data sources. Of the 42 studies, 2 (5%) studies used Food and Drug Administration data as an external ground-truth data source to validate their results [36,53]. Of the 42 studies, 1 (2%) study analyzed several web-based forums [31], and 2 (5%) other studies used several social media platforms as their data sources [22,47].

Sources of consumer-generated data for cannabis research used by the reviewed studies include social media platforms, such as Twitter, Reddit, and YouTube; search queries, including Google Trend and Bing query logs; and web-based forums, crowdfunding platforms, blogs, and websites. Twitter was used in most of the studies. One of the studies concluded that, compared with unmoderated platforms, moderated sites focused more on evidence-based information and controlled misleading content [22].

Some studies have used social media analytics companies for some or all of their data collection and processing tasks. Other studies used application program interfaces to interact with Twitter and Reddit. Although Facebook allows researchers to access public posts from public pages through a dedicated platform [71], 2% (1/42) of studies [58] analyzed private Facebook posts—the method used to obtain data was not reported.

Approximately half of the studies used data sets of <8000 records and many of them used 1000 records. These studies either focused on understanding the characteristics and needs of users or the quality of information on the web, or they were directed by an RQ such as “Are individuals using CBD for diagnoseable conditions which have evidence-based therapies?” These analyses play a critical role in understanding the domain but are difficult to replicate and generalize.

More recent neural network–based natural language processing techniques have not been used in the studies in this review. These modern machine learning methods have the advantage that they require minimal data preparation and are characterized by the capacity to learn the nuance of language. However, to function effectively, they typically require high-quality annotated data—a scarce and expensive resource. Textual social media data are highly amenable to these techniques. Creating and sharing deidentified annotated data sets for this purpose should be encouraged within appropriate ethical, regulatory, and legal frameworks [72].

These limitations are mentioned in order of frequency.