Fig. 6

Thresholds for anticipating drastic community events. a ROC analysis of diagnostic performance. On the two-dimensional surface of detection- and false-detection rates of abrupt community-compositional changes (abruptness > 0.5), area under the ROC curve (AUC) and optimal detection rate (asterisk) were calculated for each warning signals (7-day-ahead forecasting). A high AUC value indicates a high detection rate of abrupt community events with a relatively low false detection rate (maximum AUC value = 1). Note that the low AUC values may be attributed to the small number of points with abruptness > 0.5 in Soil/Medium-B treatment. See Additional file 15: Figure S15 for full results. b Optimal thresholds for anticipating community collapse. For stable-state entropy (top) and local Lyapunov stability (bottom), diagnostic threshold for warning abrupt community changes was obtained based on the Youden index after compiling all the data of Medium-A and Medium-B treatments