Job Satisfaction Through the Lens of Social Media: Rural--Urban Patterns in the U.S

TL;DR

The paper addresses rural-urban differences in job-related well-being in the U.S. using a novel social-media-based indicator derived from a fine-tuned large language model applied to about 2.6 billion geotagged tweets. It links county-level job-satisfaction sentiment (jobsat) to labor-market conditions from 2013–2023 via logistic regressions with fixed effects and a nested set of controls (UnemploymentRate, LaborForce, ACS5 income, TweetVol, Rural). The findings show unemployment as the strongest predictor of negative job sentiment, with rural counties exhibiting lower sentiment overall, but the rural gap narrows under low unemployment, indicating conditional convergence driven by labor-market tightness. The results imply that policy aimed at sustaining tight labor markets in rural areas could improve perceived job quality and related outcomes, even amid persistent income disparities.

Abstract

We analyze a novel large-scale social-media-based measure of U.S. job satisfaction, constructed by applying a fine-tuned large language model to 2.6 billion georeferenced tweets, and link it to county-level labor market conditions (2013-2023). Logistic regressions show that rural counties consistently report lower job satisfaction sentiment than urban ones, but this gap decreases under tight labor markets. In contrast to widening rural-urban income disparities, perceived job quality converges when unemployment is low, suggesting that labor market slack, not income alone, drives spatial inequality in subjective work-related well-being.

Figures (3)

• Figure 1: Monthly national unemployment rate (%) and labor force (in million) obtained through county-level monthly aggregation compared to the social media indicator jobsat rescaled to national level to account for the missing counties in some years/months.
• Figure 2: County-level averages of job-satisfaction sentiment (rescaled to the interval [0,1]) and unemployment rates in the United States, 2013--2023. The top panel reports the rescaled jobsat indicator based on social-media sentiment, mapped using a perceptually uniform viridis color scale; higher values indicate more positive job-related sentiment. The bottom panel shows the mean county-level unemployment rate over the same period, using an inverted viridis color scale so that brighter areas correspond to lower unemployment. Counties with missing data in specific years contribute only when observed; maps reflect the average over available months for each county.
• Figure 3: Average difference $\Delta P = P(\texttt{jobsat}>0\, |\, \texttt{rural} = r) - P(\texttt{jobsat}>0\, | \,\texttt{rural} = 1)$, between each rurality level ($r=2,\ldots,9$) and the most urban counties ($r=1$), plotted across unemployment rates. Predicted probabilities are obtained from the logit model in equation \ref{['eq:logit']} and averaged over the empirical distribution of year, month, and state fixed effects.