The assessment of psychological richness, meaning, and happiness with social media text data: Predictive accuracy and distinct behavioral correlates

Cavan Vaux Bonner, Young-Min Cho, Fanyi Zhang, Louis Tay, Lyle Ungar, Sharath Chandra Guntuku

TL;DR

This study explores whether social media text can predict psychological richness, a less-studied dimension of well-being, and compares its predictability to hedonic (happiness) and eudaimonic (meaning) well-being. While language features improved predictions for psychological richness, they offered less added value for happiness and meaning compared to demographic data alone. This suggests that psychological richness might manifest differently in online communication, offering unique insights into individual differences.

What Problem Does This Paper Address?

Traditional methods for assessing well-being, like surveys, are often costly, time-consuming, and suffer from low response rates. Social media text analysis (Language-Based Assessment, LBA) has emerged as a promising alternative for capturing well-being at scale. However, most LBA research has focused on hedonic well-being (happiness, life satisfaction). This paper addresses the gap in understanding whether other important well-being constructs, specifically psychological richness (experiencing a wide range of novel and interesting life experiences) and eudaimonic well-being (purpose and meaning), can be effectively predicted from social media text and how this compares to hedonic well-being. It also investigates the unique linguistic correlates of psychological richness to establish its discriminant validity.

Methods

The researchers employed a multi-faceted approach to predict well-being constructs from social media data. They utilized:

  • Good Life Scale (GLS): A self-report questionnaire to measure psychological richness, meaning, and happiness.
  • Demographic Data: Age, gender, income, and education were used as baseline predictors.
  • Linguistic Inquiry and Word Count (LIWC): A dictionary-based approach that categorizes text based on pre-defined linguistic features (e.g., positive emotion, negative emotion, pronouns).
  • Latent Dirichlet Allocation (LDA): An unsupervised topic modeling technique to identify latent themes or topics within the text data. They used a pre-trained LDA model on a large Facebook corpus to extract 2000 topics.
  • Predictive Modeling: They used 10-fold cross-validation to build models combining demographic data, LIWC features, and LDA topics to predict GLS scores.
  • Incremental Validity Testing: Statistical tests (t-tests and z-tests) were used to determine if adding language features significantly improved prediction accuracy beyond demographic features.
  • Correlational Analysis: To assess discriminant validity, they examined the correlations between language features (LDA topics and LIWC categories) and the different well-being constructs.

Datasets

The study used a dataset of Facebook posts from 2,644 participants recruited through the Qualtrics online survey panel. Participants were required to have at least 500 words across their Facebook posts to ensure sufficient data for analysis. The LDA topics were derived from a pre-existing corpus of approximately 18 million Facebook updates.

Key Insights and Ideas

  • Psychological Richness is Distinct: The study provides evidence that psychological richness, while related to happiness and meaning, has unique linguistic correlates. This suggests it's not just a synonym for positive affect or purpose but a distinct experiential dimension.
  • Social Media as a Window, But With Nuances: While social media text can predict well-being, the type of well-being matters. Psychological richness, which involves diverse and novel experiences, might be less readily expressed or captured through the typical social media interactions compared to more general positive affect or a sense of purpose.
  • Demographics Still Hold Sway: For happiness and meaning, demographic factors alone provided substantial predictive power, and language features didn't significantly add to this. This implies that for these constructs, broader life circumstances might be more influential than specific linguistic expressions in online spaces.
  • Face Validity of Linguistic Cues: The identified language features associated with psychological richness were interpretable and aligned with existing theories, strengthening the validity of using LBA for this construct.

Key Results

  • Predictive Accuracy: Psychological richness had the lowest overall prediction accuracy (r = .21), followed by hedonic well-being (r = .27) and eudaimonic well-being (r = .29).
  • Incremental Validity: Language features (LIWC and LDA) significantly improved the prediction of psychological richness beyond demographics (p = .016). However, they did not significantly improve predictions for meaning (p = .063) or happiness (p = .091).
  • Unique Correlates: Topics related to "Reflection and Insight," "Ideas and Concepts," "Political and Social Issues," and "Negative Affect" were positively associated with psychological richness. Notably, "Negative Affect" topics were positively linked to psychological richness but negatively to happiness and meaning, highlighting discriminant validity. Conversely, "Leisure and Entertainment" topics were negatively correlated with psychological richness but positively with happiness.

Limitations and Future Work

  • Model Fit: The confirmatory factor analysis of the Good Life Scale showed a poor model fit, which the authors acknowledge as a limitation. Future research should focus on developing and validating instruments with better psychometric properties.
  • Observability of Richness: The authors suggest that psychological richness might be less observable in social media text compared to hedonic and eudaimonic well-being, potentially due to the nature of online interactions favoring more common or less complex disclosures.
  • Effect Sizes: While statistically significant, the effect sizes for many linguistic associations were small. The authors emphasize the need for multi-method evidence to verify these findings.
  • Future Directions: The paper calls for future research to explore how improved measurement models and LBA techniques can enhance the accuracy and discriminant validity of assessing distinct well-being components. This could lead to better policy insights and understanding of well-being trends.

Summary updated

Abstract

Assessing well-being with social media text data is a promising method, but besides hedonic well-being, little is known about whether additional well-being dimensions, such as psychological richness and eudaimonic well-being, can be predicted from such data. We compare the predictive accuracy for hedonic well-being, eudaimonic well-being, and the recently proposed construct of psychological richness in a large sample of Facebook users (n = 2,644), and find that the inclusion of language features incrementally improved model prediction accuracy beyond demographic features for psychological richness, but not for hedonic or eudaimonic well-being. Psychological richness had the lowest overall prediction accuracy (r = .21) followed by hedonic well-being (r = .27) and eudeomonic well-being (r = .29). The linguistic features associated with Psychological Richness were face valid, and in many instances the content and direction of the associations were unique to Psychological Richness, which provides discriminant validity evidence.
BibTeX
@inproceedings{bonner2026assessment,
  author = {Cavan Vaux Bonner and Young-Min Cho and Fanyi Zhang and Louis Tay and Lyle Ungar and Sharath Chandra Guntuku},
  title = {{The assessment of psychological richness, meaning, and happiness with social media text data: Predictive accuracy and distinct behavioral correlates}},
  booktitle = {PLoS},
  year = {2026},
  doi = {10.1371/journal.pone.0337649},
}
Figure