Geolocated Twitter posts from 2015, a year after which the UC system had enacted its systemwide tobacco-free policy and prior to enactment in the CSU system, were used to conduct a retrospective analysis of smoking-themed tweets geofenced from UC and CSU campuses (one CSU campus—Fullerton—enacted its own smokefree policy effective August 2013). Data collection was conducted using the Twitter public application programming interface in combination with cloud computing services on Amazon Web Services. Data were filtered for messages that included geospatial coordinates enabled by users. These data were collected in JSON format and stored in a relational database, with information on the date and time of the post, hyperlink to the original tweet, text of the tweet, and geospatial coordinates.

Posts were selected for further analysis if they included any of the following keywords associated with tobacco products and behavior: bidis, cigarette, cigarettes, cigarillos, cigars, ciggie, class, dip, e-cig, hookah, huqqa, joint, JUUL, kreteks, Marlboro, Newport, njoy, pipe, roll-up, shag, smoke, smoking, snuff, snus, tobacco, vape, vaped, vapejuice, vaper, vapes, vapine, vapor, waterpipe, waxpen, or weed. Keywords were selected and adopted on the basis of prior studies [15,16] and because they were related to college life or tobacco products according to manual searches conducted on Twitter. For the sake of parsimony, the top 2 leading brands of cigarette sold in the United States (Marlboro and Newport), which were also the most commonly mentioned brands in tweets related to college life as revealed from manual searches conducted on Twitter, were also included in the list of keywords.

Tweets that included these keywords were then further filtered to identify those physically originating (ie, geofenced) from CSU or UC campuses using base maps of these schools available from the Stanford Prevention Research Center [17], which the authors visually assessed for concurrent validity with school boundaries in satellite imagery via Google Maps. Latitude and longitude coordinates for posts containing keywords were loaded into ArcGIS Desktop (version 10.6) and the clipping function was used to omit tweets outside of CSU or UC campus shape files.

Human annotators manually assessed posts relating to the study theme corresponding to user-generated tobacco products or related behavior, for face validity. Human annotators were trained in tobacco research and have participated in prior research infoveillance research [18,19]. Posts were manually annotated for positive, negative, or neutral sentiment toward smoking. These attributes were selected by 2 authors, with high interrater reliability (Cohen κ=0.96). Discrepancies were resolved through discussion among all authors. Chi-square tests were conducted to determine if there were significant differences in sentiment between campuses that were smoke- or tobacco-free (all UC campuses and CSU, Fullerton) compared to those that were not.

A separate content analysis of tweets with smoking-related sentiment was conducted by the first author to identify central themes [20]. Tweets were imported into Atlas.ti (version 8) [21]. Each tweet was a recording unit and mutually exclusive. All tweets were coded for the smoking policy of the campus from which it originated (tobacco- or smoke-free vs non–tobacco- or –smoke-free). A general inductive approach was then used to develop a coding framework for tweets to assess thematic content [22]. Once the coding scheme was developed, tweets were coded and analyzed for major themes by campus smoking policy type.

Quantitative analysis of tweets originating from campuses of public 4-year universities in California revealed a significant difference in sentiment toward smoking by campus smoking policy status, with a higher proportion of positive sentiment tweets originating from campuses without smoke- or tobacco-free policies, but an overall high percentage of positive sentiment regardless of campus smoking policy type. Greater positive sentiment may have been driven by more tweets containing general comments about one’s own smoking behavior and comments about smoking as a general behavior and a greater variety in types of positive sentiment tweets.

Negative sentiment toward smoking was identified on both campus types, but universities without a smoke- or tobacco-free policy had much higher numbers of positive sentiment tweets in two categories: general comments about one’s own smoking behavior and comments about smoking as a general behavior. These differences may indicate policy passage and implementation at smoke- and tobacco-free campuses may be associated with less positive sentiment toward smoking, consistent with other studies [6,7]. Though efforts such as outreach programs and infrastructure changes may help with policy compliance [23-25], continued effort is needed to change norms around campus smoking and increase policy buy-in. These efforts should be ongoing and consistent, facilitating broad involvement of campus constituents in the policy implementation process. Additionally, twitter posts expressing a desire and intention to smoke as well as self-reported smoking suggest a need to more actively promote cessation on all campuses [24-25]. Future studies should also assess whether sentiment may significantly differ on the basis of different tobacco, e-cigarette, and other smoking product use, particularly in the context of introduction of new products and bans on products (eg, flavored products) that are popular among youth and young adults.

Data collected for this study were limited to Twitter users who enabled geolocation, which may have introduced bias in the volume and types of tweets or Twitter users for whom data were collected. Additionally, this study only reviewed data from a single calendar year to assess differences in sentiment between California universities with and those without smoke- or tobacco-free campus policies, though examining tweets from subsequent years (2016 and 2017) may have yielded additional user sentiment after policy implementation. Another limitation is the constrained contextual information available in some tweets (eg, such as references to “smoking” without mention of a specific product), found mainly among tweets expressing an intention to smoke, which may have referred to marijuana and not tobacco products. We also did not restrict our data set to one tweet per user, though a user expressing negative or positive sentiment may be more likely to have a similar sentiment in other tweets they post. The rationale for this approach included the detection of both second- and first-hand accounts of smoking behavior, which could relate to multiple smoking behavior instances and the possibility that this could lead to possible underestimation of expressed sentiment toward alleged violations of campus smoking bans. Finally, this is an ecological study assessing the relationship between a survey of smoking-related social media posts from college campuses and the policies on those campuses. Associations uncovered in this study may not apply to individuals, and the effect of college policies on sentiment should be further investigated in studies with a breadth of individual participants whose selection is representative of the campus population.

Among Twitter users in California public universities, the overall sentiment toward tobacco products and use is high, although positive sentiment is higher on campuses without smoke- or tobacco-free campus policies. Smoke- and tobacco-free policy initiatives, including implementation, may help reduce positive sentiment toward smoking on college campuses and should be strengthened to maximally increase their impact.