TL;DR
This study demonstrates that Twitter language can effectively capture regional and cultural differences in excessive alcohol consumption across the US. By analyzing millions of geolocated "drunk" tweets, researchers found correlations between tweet frequency and self-reported drinking rates, and identified distinct linguistic markers associated with drinking in various cultural communities.
What Problem Does This Paper Address?
Excessive alcohol consumption is a major public health concern in the US, leading to significant mortality and economic costs. While traditional surveys like the BRFSS track drinking trends, they often lack the granularity to understand the diverse cultural contexts and specific social practices that contribute to excessive drinking. This limits the effectiveness of targeted public health campaigns. The paper aims to leverage social media, specifically Twitter, as a more granular and culturally sensitive tool to monitor and understand these drinking behaviors.
Methods
The researchers employed a multi-step approach:
- Keyword Filtering: They started with a broad set of drinking-related keywords but ultimately focused on tweets containing the word "drunk" to capture discussions about the act of drinking itself.
- Topic Modeling (LDA): Latent Dirichlet Allocation (LDA) was used to identify 100 distinct "drunk" topics from the filtered tweets. This involved tokenizing tweets, identifying discriminative words, and then clustering these words into thematic topics.
- Correlation Analysis: The frequency of "drunk" tweets at the county, state, and American Communities Project (ACP) community levels was correlated with self-reported excessive drinking data from the BRFSS.
- Differential Language Analysis (DLA): This technique was used to identify language features (topics) that characterized counties with high versus low excessive drinking rates, and to pinpoint drunk language most associated with specific ACP communities.
- Pronoun and Sentiment Analysis: The study also examined the use of personal pronouns ("I" vs. "he/she/they") and tweet sentiment to understand the personal versus impersonal nature of drinking discussions and their valence.
Datasets
- Twitter Data: A massive dataset of approximately 37.6 billion public tweets, filtered down to 19.3 million English-language tweets containing the word "drunk" and geolocated to US counties. This resulted in a final sample of 3.3 million "drunk" tweets spread across 1573 counties.
- Behavioral Risk Factor Surveillance System (BRFSS): County-level prevalence data on self-reported binge and heavy drinking from 2006-2012 was used as the ground truth for excessive alcohol consumption.
- American Communities Project (ACP): This dataset provided a clustering of US counties into 15 community types based on demographic, socio-economic, and cultural indicators, allowing for analysis at a more culturally relevant level than just counties or states.
Key Insights and Ideas
- Twitter as a proxy for public health: The study strongly suggests that Twitter language can serve as a valuable, real-time proxy for monitoring public health behaviors like excessive drinking, complementing traditional survey data.
- Cultural nuances in drinking language: The way people tweet about drinking is not uniform; it reflects distinct regional and cultural contexts, offering insights into the social practices and norms surrounding alcohol consumption.
- Religious communities and drinking: Areas with strong religious identification (e.g., LDS Enclaves, Evangelical Hubs) showed both the lowest rates of excessive drinking and the least amount of "drunk" tweeting, serving as a useful validation point.
- College towns and drinking culture: College towns stood out for their high frequency of "drunk" tweets, often discussing drinking in the context of partying, sexual behavior, and social events.
- Family and drinking: Hispanic communities, in particular, frequently discussed drinking in relation to family members, aligning with higher rates of self-reported drinking in these groups.
Key Results
- Correlation with excessive drinking: The frequency of "drunk" tweets correlated significantly with excessive alcohol consumption at the county level ($r = 0.26$, $p < 0.001$) and state level ($r = 0.45$, $p = 0.002$).
- Topic associations: Excessive alcohol consumption was most correlated with tweets referencing drinking with friends ($r = 0.20$), family ($r = 0.15$), and driving under the influence ($r = 0.14$).
- ACP community differences:
- Religious communities had a high frequency of anti-drunk driving tweets.
- Hispanic centers discussed family members drinking.
- College towns discussed sexual behavior and partying.
- Self vs. Other: The African American South showed a more personal tweeting style about drinking (higher self and other references), while communities like LDS Enclaves and College Towns tended to be more impersonal.
Limitations and Future Work
- Selection Bias: The authors acknowledge that Twitter users may not be representative of the general population, and certain types of individuals might be more inclined to tweet about drinking.
- Twitter Bots: The potential for Twitter bots to influence the data and skew results is noted.
- Aggregated Data: The study relied on aggregated Twitter data and survey data. Future work should ideally use individual-level data with linked self-reports and language analysis.
- Confirmatory Studies: The authors emphasize the need for confirmatory studies to test the identified associations as hypotheses.
- Personalized Messaging: The findings suggest that future public health interventions could benefit from culturally tailored messaging based on these identified regional and community-specific drinking contexts.
Summary updated