BACKGROUND
Since their entry into the US market, the public health community has debated the health and safety concerns of e-cigarettes. While survey data can provide useful information about public sentiment around e-cigarettes, potential limitations include social desirability bias and response style contamination. Therefore, more organic and innovative methods of assessing sentiment toward behaviors such as e-cigarette use are important to establish. Harvesting sentiment about e-cigarettes from Twitter is one potential avenue.
OBJECTIVE
This study aims to examine Twitter chatter from 2017 through 2020 through use of natural language processing to classify the sentiment of tweets, model sentiment of tweets over time, and identify important predictors of tweet sentiment using the unobserved components model.
METHODS
Human coders annotated a random sample of relevant tweets for sentiment and commercial activity. This was followed by natural language processing using BERTweet to classify the remaining tweets as containing positive, negative, or neutral sentiment and the presence of commercial activity. After classification, an unobserved components time series model evaluated trends in the data and identified predictor variables. Predictors included events listed on the Consumer Advocates for Smoke-free Alternatives Association and Drug Watch timeline and commercial account activity.
RESULTS
Across all years, the proportion of relevant, non-commercial e-cigarette tweets with positive sentiment was 18.5%. Negative and neutral sentiment represented 31.2% and 50.3%, respectively. Significant predictors of positive sentiment included commercial volume of tweets (p<.001) and noteworthy news events (p=.025). An increase of 100 commercial tweets was associated with an increase of approximately 61 positive sentiment e-cigarette tweets, controlling for other predictors in the model. Significant predictors of negative sentiment included JUUL ending its use of Twitter (p=.002) and FDA-related events (p=.004). For neutral sentiment, predictors were JUUL ceasing to use Twitter (p<.001), FDA-related events (p=.032), and commercial volume of tweets (p=.003).
CONCLUSIONS
This combination of methods represents a valuable approach for use of large, publicly available data as an alternative to or augmentation of self-report survey data. Twitter surveillance provides a unique opportunity to capture real-time sentiment changes to important events, because users create posts in direct response to current events. Supporting interventions such as the FDA’s Final Rule as an important intervention in e-cigarette behavior, we found a positive association between positive sentiment and commercial activity.