Abstract
The language used in tweets from 1,300 different US counties was found to be predictive of the subjective well-being of people living in those counties as measured by representative surveys. Topics, sets of co-occurring words derived from the tweets using LDA, improved accuracy in predicting life satisfaction over and above standard demographic and socio-economic controls (age, gender, ethnicity, income, and education). The LDA topics provide a greater behavioural and conceptual resolution into life satisfaction than the broad socio-economic and demographic variables. For example, tied in with the psychological literature, words relating to outdoor activities, spiritual meaning, exercise, and good jobs correlate with increased life satisfaction, while words signifying disengagement like ’bored’ and ’tired’ show a negative association.
Introduction
Measuring the subjective well-being (SWB) of a population is traditionally a slow and expensive endeavor, relying on massive phone surveys that often act as lagging indicators. This research investigates whether social media can serve as a "fast and cheap" alternative. By analyzing the collective "voice" of a county on Twitter, the authors sought to predict survey-based life satisfaction scores and provide a "behavioral resolution" that explains the underlying factors of regional well-being.
Methodology
The researchers employed a pipeline called Differential Language Analysis (DLA) to process over a billion tweets from 2009 to 2010.
Feature Extraction
Text was converted into data using two primary methods:
- Lexica: Hand-built lists of words categorized by psychological themes.
- LDA Topics: Latent Dirichlet Allocation (LDA) was used to automatically cluster words into 2,000 "topics" based on co-occurrence (e.g., a "gym" topic containing workout, trainer, fitness).
To calculate the probability of a topic within a specific county, the researchers used the following formulation:
Predictive Modeling
The study used LASSO (L1-penalized) linear regression to manage thousands of features across 1,300 counties. To ensure language wasn't just a proxy for wealth or age, the models controlled for socio-economic status (SES) variables like median income, education, and ethnicity.
Key Results
The most significant finding is that language features improved prediction accuracy above and beyond what could be predicted by income and education alone. The model using only Twitter "topics" achieved a Pearson correlation of $r = 0.307$ with survey data. When language features were added to traditional demographic controls, the correlation jumped from $r = 0.435$ to $r = 0.535$.
As illustrated in the maps above, the Twitter-predicted well-being closely mirrored the actual survey-measured well-being across the United States, demonstrating the geographic validity of the linguistic model.
Analysis and Insights
- Engagement vs. Disengagement: Positive well-being was strongly correlated with engagement (topics like meetings, ideas, volunteer) and physical activity (gym, hiking, yoga). Conversely, the strongest negative predictors were words of disengagement (bored, tired, sleepy).
- The "Community Voice" Effect: Interestingly, while Twitter users skew younger and more urban, their collective language still accurately reflected the well-being of the broader representative sample. This suggests that well-being is shared within a geographic culture.
- Professional Identity: High life satisfaction was linked to "high value creation" occupations. Topics involving management, research, and technology were positive predictors, aligning with psychological theories linking accomplishment to well-being.
Limitations and Future Work
The authors acknowledge sampling biases and the noise inherent in geolocating social media users. Additionally, the relationship between language and well-being might evolve over time as slang and platform demographics shift. Despite these challenges, this method offers a path for policymakers to evaluate the impact of local decisions in near-real-time, providing a dynamic supplement to traditional survey methods.
Summary updated