Other commonly used smoking-related terms did not exhibit a percentage drop of this magnitude

The purpose of this keyword filtering was to better isolate smoking-related conversations from all other Twitter discussions occurring on college campuses. After tweets were manually reviewed to positively identify smoking-related conversations originating in these college campuses, a snowball sampling design was employed which compared the frequency of all non-keyword terms in “signal” tweets with the frequency of these words in “noise” tweets . This methodology resulted in the identification and querying for ten additional keywords: 420, 818, blunt, bong, cigs, kush, marijuana, roll, smell, and stoge. After isolating a corpus filtered for tobacco-related keywords in areas geolocated for California 4-year universities, four researchers trained in social media content analysis used an inductive coding approach to identify study characteristics of interest by manually annotating all tweets , following an approach also described in prior studies . Annotators had backgrounds in public health and had experience manually annotating social media posts for tobacco behaviors in prior published research projects . Manual annotation included: identifying the type of smoking product discussed ; assessing positive, negative, or neutral sentiment related to smoking behavior ; and identification of whether the tweet included first-person use or second-hand observation of smoking behavior. Table A1 contains further details about topics that were coded as valid and invalid for positive identification as a “signal” tweet. Tweets that did not express sentiment related to smoking were excluded from analysis of signal tweets.

The primary objective of this approach was to conduct exploratory research into what tobacco and smoking products were being discussed by Twitter users at California universities, assess the overall sentiment toward tobacco and smoking by these users,indoor grow shelves and explore whether it was possible to identify self reporting of tobacco use-related behavior . Four authors coded posts independently and achieved a high inter rater reliability for overall coding categories and equally high inter rater reliability for specific sub coding for tobacco , vape , and marijuana specific tweet categories. For inconsistent results and any discrepancies related to coding, all authors convened to discuss, confer, and reach consensus on the correct classification informed by the inductive coding approach outlined in Table A1. Analyses of variation across college campuses were limited to the top twenty colleges by tweet volume, as estimates collected from samples of tweets from other colleges may have been biased due to insufficient volume of tweets collected. A p < 0.10 was considered statistically significant for correlational analyses due to sample size limitation. Point density algorithms were used to visualize and detect geospatial trends. Analysis was conducted in R version 4.0.1 and geospatial visualization of data was done in ArcGIS Desktop version 10.7. This project was part of a broader study to examine college campus smoke-free policies using qualitative focus groups and examining social media data with the qualitative analysis approved by the Institutional Review Board at California State University, Fullerton . Data collection resulted in 83,723,435 geo-identifiable tweets located in the state of California in the 5-year period from 2015 to 2019. From these tweets, 1,381,019 originated from 88 CA 4-year colleges, with the five schools contributing the most tweets including UC Los Angeles , Stanford University , UC Riverside , University of Southern California , and UC Berkeley . Thirty-eight schools contributed over 10,000 tweets each, overall representing 89% of the entire corpus of CA 4- year college geocoded tweets.

Of these tweets, 7,342 contained smoking-related keywords with approximately one third occurring after 2015. In total, smoking-related topics originating from all geocoded tweets in the state made up an extremely small proportion of all topics and tweets specifically geocoded for CA 4-year universities. This low representation of smoking-related twitter topics likely was impacted by the methodology of data collection and its associated limitations . Of the 34 smoking-related keywords used to query the Twitter API, eight returned over 100 tweets from college campuses during this time period: cigarette , dip , joint , njoy , pipe , smoke , smoking , and weed . Upon further examination, it was determined that “njoy” returned tweets with the word “enjoy” in 99.2% of cases, and “dip” also predominantly returned false positives. After excluding these two terms, longitudinal analysis revealed that the rate of “weed” in the corpus decreased through the study time period, with it being found in 28% of tweets in 2015, 15% in 2016, 10% in 2017, 14% in 2018, and 7% in 2019. Conversely, the rate of “pipe” in the corpus increased from 7% in 2015 to 14% in 2016, 16% in 2017, 14% in 2018, and 12% in 2019. Also notable was the rate of “joint,” which increased from 6% in 2015 to 15% in 2016, 19% in 2017, 16% in 2018, and 14% in 2019. The frequency of these keywords may have been impacted by changes in the way users communicate about smoking-related topics, in addition to the potential impact of legalization of adult-use cannabis in 2016. Other terms were comparatively stable, with “smoke” returning the top number of tweets in the corpus for any given year in the study period. From this subset of filtered geocoded data, manual review identified 1,089 “signal” tweets relating directly to smoking topics, with 509 relating to tobacco, 490 relating to marijuana, 79 relating to vaping, and 7 relating to multiple product types in the same tweet . Sixty eight CA colleges were represented in our signal data, though the top 20 accounted for 783 of tweets. Individual colleges exhibited high variation in the proportion of tweets corresponding to each smoking product assessed.

Out of the top twenty colleges by tweet volume, the distribution of tobacco-related tweets ranged from 26.1% for CSU Long Beach to 62.2% for CSU San Jose [median [M] = 43.1%, standard deviation [SD] = 10.7%]. Vaping-related tweets were detected from eighteen of these twenty colleges, ranging from 1.6% for CSU North ridge to 21.1% for CSU Fullerton . Finally, the distribution of marijuana related tweets ranged from 28.6% for CSU San Marcos to 61.9% for the University of Southern California . Positive sentiment about tobacco, marijuana, and vaping was detected from 736 tweets. Out of the top twenty colleges by tweet volume, positive sentiment ranged from 55.0% for CSU Long Beach to 95.8% for CSU Los Angeles . When computed as a proportion of all tweets with smoking-related behavior, including neutral tweets without clear user sentiment, positive sentiment ranged from 47.4% for both CSU Fullerton to 80.9% for CSU Santa Barbara . With the exception of CSU Long Beach and CSU Fullerton , all colleges had at least 50% positive sentiment from tweets about smoking . Across product categories, positive sentiment varied with 58.2% for vaping, 66.1% for tobacco, and 70.7% for marijuana. When calculated as a proportion of all tweets,indoor garden table positive sentiment was 63.9% for vaping, 70.6% for tobacco, and 85.5% for marijuana. The majority of tweets from any product category exhibited either positive or negative sentiment, with only 8.9% of tweets about vaping, 6.3% about tobacco, and 17.3% about marijuana having neutral sentiment. Therefore, while the majority of tweets about any product exhibited either positive or negative sentiment, the data suggests that tweets about tobacco or vaping were much more opinionated than marijuana, which had the highest proportion of neutral sentiment tweets.There were also 502 tweets denoting first-person product use or second-hand observation of another person’s use of smoking products. These reports also ranged by product type, with 40.8% for marijuana, 47.4% for tobacco, and 10.0% for vaping. Out of all tweets, 40.3% of those about marijuana indicated first-person use or second-hand observation, whereas this applied to 48.6% of tobacco-related tweets and 63.3% of vaping-related tweets. Across the top twenty colleges by tweet volume, first-person smoking product use or second-hand observation of another product user ranged from 31.8% from UC Berkeley to 73.7% for CSU San Jose .

As the UC system and CSU Fullerton had smoke-free policies in 2015, and the remaining 22 schools in the CSU system did not have smoke-free policies, these tweets were assessed for evidence relating to campus policy violation. Out of 486 tweets in 2015 indicating first-person smoking or second-hand observation of smoking, 146 were from schools with smoke-free policies. It should be noted that the content of these 146 tweets indicated smoking behavior on campus . As we captured 11 schools with smoke-free policies in 2015 and 19 schools without smoke-free policies , the number of these tweets per school was approximately the same among schools with smoke free policies and those without smoke free policies , potentially indicating a muted effect regarding the implementation of smoke-free policies, at that time, on these college campus populations and their compliance behaviors. Geospatial analysis revealed a distribution of tweets that approximately followed California’s population distribution, with a cluster in the San Francisco Bay Area and a cluster in Southern California, which was dominated by the Los Angeles Basin. However, comparatively fewer smoking-related tweets were captured from colleges in California’s Central Valley region. This distribution may have also been impacted by a low volume of tweets collected and sample bias for higher-population demographic areas based on the data collection process. Based on our use of tweets specifically geolocated for CA 4- year universities combined with a data filtering process to isolate tweets containing smoking-related keywords, 7,342 tweets were obtained for analysis that discussed smoking and also originated from California universities between 2015 and 2019. Within this corpus of social media messages, rates for use of the term “weed” decreased over time, changing from 28% in 2015 to 7% in 2019.The mechanisms underscoring the observed decrease in social media messages with this keyword are not clear but may result from evolving word choices to describe marijuana, decreased use of marijuana on CA college campuses, social inhibition of posting marijuana related public messages on Twitter, or some combination thereof. Further, it is unclear how passage of legalized adult-use cannabis Proposition 64 may have impacted these conversations, attitudes, and behaviors, particularly as despite state legalization, some college-aged students may not be of legal age and campus smoke free policies still restrict their use. Manual review uncovered 1,089 tweets explicitly related to smoking behavior and posted within the boundaries of California 4-year universities, with the majority of tweets expressing positive sentiment about smoking products and behavior. Five-hundred and-two of these tweets reported first-person use or second hand observation of another person’s smoking behavior, with 146 tweets reporting possible violations of smoke- or tobacco free campus policies that were clearly in place from 2015 but were also in the process of being fully implemented. These tweets indicate early lack of compliance to smoke-free campus policy implementation as self-reported by social media users. For campuses where policies were not in place, tweets also reflect general positive sentiment about smoking and reports of smoking behavior, indicating possible barriers to enacting campus smoke free policies that would occur in 2017, when more smoke free campus policies across the California State University system were enacted. These results provide early indications that smoke free campus policy implementation requires continued attention and sufficient resources to ensure appropriate health promotion, education on policy requirements, and policy enforcement measures in college communities. Overall, our analysis found a higher number of tweets in our corpus identified for tobacco and marijuana products, with comparatively fewer for vaping products geolocated for California university campuses. The majority of geolocated data collected during this study originated in 2015, which may explain the overemphasis on tobacco and marijuana Twitter conversations as vaping products were rising in popularity. Additionally, national debate about marijuana legalization occurred during this time frame, though was not legalized in California for adult recreational use until 2016 and licensure of cannabis retailers was permitted in 2018. As previously stated, national and state discussions relating to marijuana legalization may have influenced the relative social acceptability and volume of marijuana-related Twitter conversations among campus populations. Tweets about vaping had the highest proportion containing first-hand accounts of use or other persons engaged in product use and behavior. The increasing popularity of vaping products throughout the study time period, especially among the college aged population, may partly explain why college students posted about themselves or other people using vaping products in this context, despite having an overall lower volume than other smoking products .