Quantifying generalized trust in individuals and counties using language

Salvatore Giorgi, Jason Jones, Anneke Buffone, Johannes C. Eichstaedt, Patrick Crutchley, David B. Yaden, Jeanette Elstein, Mohammadzaman Zamani, Jennifer Kregor, Laura Smith, Martin E. P. Seligman, Margaret L. Kern, Lyle Ungar, H. Andrew SchwartzFrontiers in Social Psychology 2024 · DOI

Abstract

Trust is predictive of civic cooperation and economic growth. Recently, the U.S. public has demonstrated increased partisan division and a surveyed decline in trust in institutions. There is a need to quantify individual and community levels of trust unobtrusively and at scale. Using observations of language across more than 16,000 Facebook users, along with their self-reported generalized trust score, we develop and evaluate a language-based assessment of generalized trust. We then apply the assessment to more than 1.6 billion geotagged tweets collected between 2009 and 2015 and derive estimates of trust across 2,041 U.S. counties. We find generalized trust was associated with more affiliative words (love, we, and friends) and less angry words (hate and stupid) but only had a weak association with social words primarily driven by strong negative associations with general othering terms ("they" and "people"). At the county level, associations with the Centers for Disease Control and Prevention (CDC) and Gallup surveys suggest that people in high-trust counties were physically healthier and more satisfied with their community and their lives. Our study demonstrates that generalized trust levels can be estimated from language as a low-cost, unobtrusive method to monitor variations in trust in large populations.

Introduction

Generalized trust—the fundamental belief in the basic honesty and reliability of others—serves as the "social glue" for modern society. It is a primary driver of economic growth, civic cooperation, and institutional stability. However, traditional methods for measuring this metric, such as the General Social Survey, are often expensive, infrequent, and lack the geographic granularity required for local policy interventions.

This research addresses these limitations by leveraging "digital breadcrumbs" from social media. By analyzing the massive volume of text data generated online, researchers can create high-resolution, near real-time maps of social cohesion across the United States.

Methodology

The study utilized a two-stage computational linguistics framework to bridge individual psychology with community-level outcomes.

Individual-Level Modeling

The researchers first analyzed a dataset of approximately 16,000 Facebook users who provided both their status updates (3.3 million posts) and self-reported trust scores via the International Personality Item Pool (IPIP). Using Differential Language Analysis (DLA), the team trained a penalized (ridge) regression model. Features included individual words, n-grams (2-to-3-word phrases), and 2,000 semantically related "topics" generated through Latent Dirichlet Allocation (LDA).

Community-Level Application

The resulting model was applied to a corpus of 1.64 billion geotagged tweets (2009–2015). By aggregating predicted trust scores of individuals within specific geographic boundaries, the researchers derived trust estimates for 2,041 U.S. counties.

Validation Benchmarks

To ensure accuracy, language-based estimates were validated against several "gold standard" datasets:

  • Gallup-Sharecare Wellbeing Index: 2.2 million survey responses.
  • CDC/County Health Rankings: Data on obesity, smoking, and mortality.
  • U.S. Census: Socio-economic indicators including income and education.

The Linguistic Signature of Trust

The study revealed distinct linguistic markers for trust and distrust. Trusting individuals frequently use affiliative language (love, we, friends, amazing) and discuss future-oriented social engagements (tomorrow, tonight, conference).

Conversely, distrust is characterized by anger and "othering." High-distrust individuals use aggressive terms (hate, stupid, kill) and dehumanized generalizations. Notably, the word "people" emerged as a strong marker of low trust, suggesting a tendency to view others as a vague, externalized group ("those people") rather than as specific individuals or community members. While general "social" word categories initially showed weak correlations, removing "othering" terms like they and people revealed that core social language (we, us) is highly predictive of trust.

Key Results and Community Impact

The language-based model demonstrated significant predictive power and strong correlations with real-world outcomes:

  • Predictive Accuracy: The individual-level model achieved a correlation of $r = 0.49$ with self-reported scores, comparable to digital models for major personality traits like Extraversion.
  • Community Wellbeing: Language-based trust scores correlated strongly with Gallup’s measures of community satisfaction ($r = 0.59$).
  • Socio-Economic Links: High-trust counties showed higher median household incomes ($r = 0.41$) and high school graduation rates ($r = 0.54$).
  • Public Health: Trust was a significant negative predictor of obesity ($r = -0.57$) and reported mentally unhealthy days ($r = -0.23$).
  • Political Correlation: A shift was observed in the 2016 U.S. Presidential election, where counties with lower generalized trust scores showed significantly higher vote gains for Donald Trump compared to 2012 patterns.

Discussion and Limitations

While this method provides a powerful "dashboard" for monitoring social health, it is not without limitations. The 2009–2015 data reflects a younger demographic, and the "digital divide" may affect representativeness. Furthermore, while the correlations are robust, they do not inherently prove causality—high trust may lead to better health, or healthier environments may foster higher trust.

Future work aims to develop platform-specific models (e.g., for Twitter vs. Facebook) and utilize these tools to monitor how major societal shocks, such as economic shifts or pandemics, impact the fabric of social trust in real-time.

Summary updated

Abstract

Trust is predictive of civic cooperation and economic growth. Recently, the U.S. public has demonstrated increased partisan division and a surveyed decline in trust in institutions. There is a need to quantify individual and community levels of trust unobtrusively and at scale. Using observations of language across more than 16,000 Facebook users, along with their self-reported generalized trust score, we develop and evaluate a language-based assessment of generalized trust. We then apply the assessment to more than 1.6 billion geotagged tweets collected between 2009 and 2015 and derive estimates of trust across 2,041 U.S. counties. We find generalized trust was associated with more affiliative words (<i>love</i>, <i>we</i>, and <i>friends</i>) and less angry words (<i>hate</i> and <i>stupid</i>) but only had a weak association with social words primarily driven by strong negative associations with general othering terms ("they" and "people"). At the county level, associations with the Centers for Disease Control and Prevention (CDC) and Gallup surveys suggest that people in high-trust counties were physically healthier and more satisfied with their community and their lives. Our study demonstrates that generalized trust levels can be estimated from language as a low-cost, unobtrusive method to monitor variations in trust in large populations.
BibTeX
@inproceedings{giorgi2024quantifying,
  author = {Salvatore Giorgi and Jason Jones and Anneke Buffone and Johannes C. Eichstaedt and Patrick Crutchley and David B. Yaden and Jeanette Elstein and Mohammadzaman Zamani and Jennifer Kregor and Laura Smith and Martin E. P. Seligman and Margaret L. Kern and Lyle Ungar and H. Andrew Schwartz},
  title = {{Quantifying generalized trust in individuals and counties using language}},
  booktitle = {Frontiers in Social Psychology},
  year = {2024},
  doi = {10.3389/frsps.2024.1384262},
}
Figure