Extended Data Fig. 5: Results hold for a different toxicity threshold.
From: Persistent interaction patterns across social media platforms and over time
Core analyses presented in the paper repeated employing a lower (0.5) toxicity binary classification threshold. a. Mean fraction of toxic comments in conversations versus conversation size, for each dataset (see Fig. 2). Trends are reported with their 95% confidence interval. b. Pearson’s correlation coefficients between user participation and toxicity trends for each dataset.c. Pearson’s correlation coefficients between users’ participation in toxic and non-toxic thread sets, for each dataset. d. Boxplot of the distribution of toxicity (n = 25, min = −0.016, max = 0.020, lower whisker = −0.005, Q1 = − 0.005, Q2 = 0.004, Q3 = 0.012, upper whisker = 0.020) and participation (n = 25, min = −0.198, max = −0.022, lower whisker = −0.198, Q1 = − 0.109, Q2 = − 0.070, Q3 = − 0.049, upper whisker = −0.022) trend slopes for all datasets, as resulting from linear regression. The results of the relative Mann-Kendall tests for trend assessment are shown in Extended Data Table 5.
