Fig. 1.

Sentiment ratings predict changes in depression. (A) The x-axis of the graph shows the beta coefficient from a robust linear regression model which included age, gender, and education level as covariates. The y-axis of the graph shows the predictors in the model (covariates not shown). In both studies (Study 1, blue line with a circle; Study 2, orange line with a square), more negative human-rated sentiment scores predicted higher depressive symptom scores (i.e., more depressed) at the three-week follow-up after controlling for initial depression (PHQ-9) scores. (B) For both studies, participants were grouped by whether no change in PHQ-9 was predicted by the model (predicted change between −0.5 and 0.5) or whether increased or decreased symptoms were predicted (Study 1, blue line with a circle; Study 2, orange line with a square). Error bars indicate the SEM. P-values indicate a significant difference from zero change using Wilcoxon signed-rank tests. *P < 0.05; **P < 0.01; ***P < 0.001.