28 



Erignathus, Histriophoca, Pagophilus, Phoca spp., and Pusa spp. (hereafter, the "E- 

 Phocinr clade) based on the results of other tests. However, as PAUP cannot produce 

 distributions of truly constrained topologies, distributions were estimated by collapsing the 

 constrained subgroup to its ancestral node. As well, as constrained skewness has not been 

 tested before, reciprocal constraints were analyzed in which the strongly and weakly 

 supported subgroups were alternately collapsed (to nodes 34 and 33 respectively; see 

 Fig.5B) to test whether the collapsing resorted to here had some effect on skewness. If 

 Hillis & Huelsenbeck's (1992) conjecture is accurate, then one would expect the 

 distribution with a collapsed (weaker) "E-Phocini" clade to maintain a significant skew, 

 while that with a collapsed (stronger) "anti-Phocinr clade should possess a non-significant 

 skew. Tests were paired to account for both ACCTRAN and DELTRAN reconstructions 

 of the ancestral node. For the case when the "anti-Phocinr were collapsed only, it was 

 possible to derive the skewness statistic from an exhaustive search of all 135,135 possible 

 trees rather than invoke PAUP's RANDOM TREES subroutine. 



Successive approximations (Farris 1969) 



Successive approximations is an a posteriori weighting technique that seeks to arrive at a 

 more robust solution (i.e., fewer and more resolved equally most parsimonious solutions) 

 by differentially weighting characters in proportion to how well they have performed in 

 a previous analysis. This procedure is typically recursive, and continues until no further 

 change is observed in either tree topology or character weights (Swofford 1993). 

 Although the use of successive approximations has increased resolution and decreased 

 ambiguity when applied to some data sets (Novacek 1993), its use is also somewhat 

 problematic. First and foremost, it is not clear how to determine a character's quality. 

 Typically, one of three goodness-of-fit statistics - CI, RI, or RC (see above) - is used, 

 but there appears to be no reason to favour one over another. As well, characters do not 

 fit equally well to all equally most parsimonious solutions, and a decision must be reached 

 whether to reweight characters according to their maximum, minimum, or average value 

 of the goodness-of-fit statistic chosen (Swofford 1993). Other problems include the 

 tendency of missing data to artificially make their characters less homoplastic (thereby 

 contributing more to future analyses), and the obvious circularity of the procedure as a 

 whole (Novacek 1993). 



Here characters were reweighted (base weight = 1,000) using all combinations of CI, RI, 

 and RC, and their maximum, minimum, and average values. Fractional weights were 

 rounded off to the nearest whole number. All searches used the heuristic search option as 

 detailed above. 



Support analyses (Källersjö et al. 1992) 



The general concept of support tests (also known as decay analyses) is to view trees (or 

 their summaries in the form of consensus trees) of increasingly greater length, and thus 

 homoplasy, so as to determine when a clade of interest disappears or is contradicted. Clades 

 that withstand the intrusion of increasing levels of homoplasy to the greatest extent are 

 judged to have the strongest support (Novacek 1991; Swofford 1993). This basic procedure 

 (termed Bremer support by Källersjö et al. 1992) suffers from being dependent on the 



