Character stability in 39 data sets

Abstract

Comparisons of the stability of classifications based on phenograms or estimated cladograms have attracted considerable interest and controversy in recent years. Rohlf and Sokal (1981) discussed the general topic while others have addressed specific aspects of stability: OTU stability (Schuh and Farris, 1981; Sokal et al., 1984); character stability tested by congruence of classifications based on different classes of characters (Mickevich, 1978; Rohlf and Sokal, 1980; Mickevich and Farris, 1981; Rohlf et al., 1983a, b); and character stability tested by congruence of classifications derived from random bipartitions of characters (Schuh and Polhemus, 1980; Colless, 1981; Schuh and Farris, 1981; Sokal and Rohlf, 1981a; Sokal, 1983a, c; Sokal et al., 1984). Cladist authors have consistently concluded that cladistic classifications based on estimated cladograms are more stable than those based on phenograms. Pheneticists have contended that the computations carried out by the cladists to support these conclusions were either misleading or erroneous and that, when these same data sets are reanalyzed appropriately, the results frequently show approximately equal stability for phenograms and estimated cladograms or, in some cases, appreciably greater stability of phenograms. In recent studies, Sokal (1983c) and Sokal et al. (1984) reported that the relative stability of phenograms and estimated cladograms was related to the ratio between number of characters and number of OTUs in any given data set. These views are based on findings in four data sets, two 1 Present address: Institute of Zoology, Academia Sinica, Nankang, Taipei, Taiwan, R.O.C. representing real data (the Leptopodomorpha [Schuh and Polhemus, 1980] and house mice [Fitch and Atchley, pers. comm.]) and two artificial data sets (the Caminalcules [Sokal, 1983b] and the simulated phylogenies of Fiala and Sokal [1985])- During a recent investigation of the properties of consensus trees and consensus indices (Shao, 1983), the opportunity arose to test a substantial number of data sets for character stability based on random bipartitions of character sets, and to examine whether the outcomes of these tests are related to the character-number/ OTU-number ratio of these data sets. Our findings are reported and discussed below