Fig. 4: How academics express themselves online, averaged by author characteristics.

Data reflect n = 99,274 academics (balanced panel) from 2016 to 2022. ‘Egocentrism’ and ‘emotionality’ (ratio of affective to cognitive words) were each measured from the full set of English tweets per user. ‘Toxicity’ (Google’s Perspective API) was computed using up to 10 randomly sampled English tweets per user–year to manage API constraints; see Methods for variable measurement and sampling details. For each subgroup, circles indicate mean metric values; error bars show ±95% confidence intervals (s.e.m.). In each panel, the dashed vertical line represents the overall academic average (the ‘Main’ group mean) for that measure, serving as a baseline for subgroup comparisons. Note that the three measures are on different scales: egocentrism can range from 0 upward, emotionality is typically close to 1, and toxicity is bounded between 0 and 1; they are plotted on their inherent scales, thereby preserving the natural variation in each metric.