Fig. 2.

Predictive performance (mean AUC) for training and test data (Mumsnet), cross-platform test (Reddit), and experimental data; error bars show 95% confidence intervals, the dotted horizontal line indicates no class separation (i.e., “guessing accuracy”)