23 



Search criteria and summarizing output 



The number of taxa examined here prevented an exact solution from being found (via 

 exhaustive or branch and bound algorithms). Therefore. PAUP's heuristic search option 

 was used, which although highly effective, cannot guarantee an optimal solution (Swofford 

 1993). Unless otherwise indicated, all searches were heuristic and used a random addition 

 sequence (with 25 repetitions), TBR branch-swapping on minimal trees only (with steepest 

 descent on), collapsed zero-length branches, and an unlimited number of MAXTREES. 

 This combination of options seemed to be the most effective in finding an optimal solution, 

 and should minimize the analysis becoming trapped in local optima or on islands of less 

 than optimal trees (Maddison 1991: Swofford 1993). 



In those cases where multiple equally most parsimonious solutions were found, the rival 

 results were summarized through the use of both strict and majority rule consensus trees. 

 These two methods provide different types of information. By retaining only those groups 

 that are found in all rival solutions, strict consensus trees will identify regions with 

 multiple, conflicting solutions as polytomies. However, within these regions, some groups 

 may occur with a greater frequency than others. The majority rule consensus algorithm, 

 by retaining those groups found in greater than 50% of the rival solutions, will tend to 

 preserve these more frequent groups that are ignored by the strict algorithm. 

 Character state assignments for internal nodes (see Character Analysis) were reconstruc- 

 ted using both accelerated and delayed transformation optimization criteria (ACCTRAN 

 and DELTRAN respectively). With no ambiguity in the reconstruction of a character, both 

 methods will yield identical results. For equally parsimonious reconstructions of a 

 homoplastic character, ACCTRAN optimization will tend to favour an early origin of the 

 derived state, followed by a reversal back to the more primitive state, while DELTRAN 

 optimization will tend to favour later parallel derivations of the derived state (but note 

 that these are not hard and fast rules) (Wiley et al. 1991; Swofford 1993). Thus, the 

 repeated claims of a predisposition towards reversals in phocid (and especially phocine) 

 evolution (e.g., Wyss 1988; Berta & Wyss 1994) may reflect the singular use of 

 ACCTRAN optimization (the default choice in PAUP). A third optimization criterion 

 available in PAUP, MINF, was not employed as its output is often identical to that of 

 DELTRAN optimization (Swofford 1993). 



Statistical tests 



One of the more active areas in theoretical cladistics in recent years has been the 

 development, and subsequent dissection, of various statistical tests designed to objectively 

 quantify the robustness of a given cladogram. In this section, each of the tests used in this 

 study are described in turn, including their objectives, their shortcomings and/or criticisms, 

 and how they were implemented here. 



Goodness-of-fit statistics 



The most basic method used to judge the quality of a solution is the use of one or more 

 goodness-of-fit statistics: consistency index (CI), homoplasy index (HI), retention index 

 (RI), and rescaled consistency index (RC) [see Farris (1989), Wiley et al. (1991). and 



