646 



Fishery Bulletin 89(4), 1991 



Similarity analysis was used to classify fish species 

 into distinct trophic groups based on their diets. Two 

 techniques were used. The first, based on the work of 

 Smith (1985), generated a matrix of similarity coeffi- 

 cients and associated variances for each pair of species 

 being compared. Similarity was measured with the pro- 

 portional similarity coefficient (PS) for each pair of fish 

 species being compared. This analysis was based on the 

 average percent composition for each species. 



The second method applied cluster analysis to these 

 data. I used the clustering methods of Nemec and 

 Brinkhurst (1988), which use a bootstrap technique ap- 

 plied to replicate samples to determine the significance 

 of clusters. The Bray-Curtis similarity coefficient was 

 used and, where replicates were available, significance 

 was tested using 50 bootstrap simulations. For each 

 species, a separate cluster analysis was performed to 



determine if there were significant intraspecific sub- 

 groups by size, sex, or depth of habitat (McKenna 1990, 

 app. D). Although the variability in diet of each species 

 was obvious, no significant subgroups were identified. 

 Cluster analysis was then applied to examine the 

 structure of the community. There is virtually no 

 spatial structure to the South Georgia fish community 

 at this season (McKenna 1990:178-251) and, since no 

 significant intraspecific structure was found, each 

 stomach examined was used as a replicate sample of 

 the diet of the predator species (Table 1). These data 

 were used in the SIGTREE cluster analysis program 

 (Nemec and Brinkhurst 1988). Due to a limitation of 

 the program (a maximum of 110 observations [sum of 

 replicates per species] are allowed), the data were 

 pooled into groups of three and averaged to reduce the 

 number of observations from 264 to 97 (Table 1). 



