Rai et al. · PNAS 2024 · Explore the study
Does the same language signal depression across communities?
In this study, familiar language markers and prediction models related to depression differently for Black and White participants. Explore the findings behind that result.
01 / Table 1 · Model performance
Who a model is tested on matters
The researchers trained models to estimate depression severity from language. Choose a training group and language representation to compare results for the two test groups.
How to read this: Pearson’s r describes how closely predicted and reported depression scores track together. Zero means no linear association; 1 means perfect positive association. These values are correlations, not percentages of correct diagnoses.
See the numbers from Table 1
Reported Pearson correlations and p values from five-fold cross-validation. The paper prints some p values as .000; these are rounded values, not exact zero. These p values concern each correlation and do not test the difference between the two groups.
02 / Figures 1 & 2 · Language patterns
The same words can carry different signals
The paper found that greater depression severity was associated with more “I” language and five negative-emotion topics among White participants, but not Black participants. Select one of these markers to explore the study’s public group averages.
Each point is the average marker use for participants in one racial group with the same PHQ-9 questionnaire score. The score ranges from 0 to 27; higher scores indicate more severe reported symptoms. This view shows the public averages in their original units; the paper’s figures show fitted relationships.
— indicates no group average at that score.
View the chart values
03 / What to take from the study
Evaluate language models across the people they serve
For both language representations, models had weaker correlations with depression scores in the Black test group—even when trained on Black participants. The findings highlight why performance in one group should not stand in for evaluation across communities.
The study matched Black and White English-speaking adults in the United States on age and gender; 76% of participants were women. Participants reported symptoms using PHQ-9 and consented to share Facebook updates. The results concern this sample, platform and set of language features. They do not establish why the differences arose or imply that depression cannot be expressed in language in any group.
This page explains group-level research findings. It does not assess an individual’s mental health.