Performance of rogue-detection approaches. Changes in consensus tree quality after removing rogue taxa. Rogues were identified using relative bipartition information content (“R,” using RogueNaRok) and shared phylogenetic information content (heuristics H1, H2); or by selecting the same number of leaves at random (-|$r$| suffix). Zero indicates no change in quality from the plenary consensus. a) extra percentage of information retained from tree set: phylogenetic information content (PIC) of retained splits, minus PIC of contradicted splits, normalized against total PIC of all splits; b) extra precision (total PIC of all splits in consensus); c), extra accuracy (PIC of “true” splits minus PIC of “false” splits); d) extra proportion of quartets that are resolved (i.e., precision); e) extra accuracy (“true” quartet statements minus “false” quartet statements), trees with c. 100–200 leaves. Columns in a–d correspond to data sets with c. 100, c. 200, and c. 500 leaves. Crosses denote means; hinges denote interquartile range; whiskers extend to the most extreme data point that is no more than 1 interquartile range from the box. Differences between medians (bars) are significantly different at the |$5{\%}$| level where notches do not overlap (Chambers et al. 1983).