Heery and Cope: Co-occurrence of bycatch and target species in the groundfish trawl fishery 
39 
This approach added “fake” species (termed “fakies”) 
to the data set that were randomly allocated to each 
tow (i.e., a 0.5 probability of occurring in any tow) and 
were subsequently clustered as members of the full 
data set. The dissimilarity point at which these species 
were grouped (termed the “breakpoint”) represented 
the dissimilarity distance at which group assignments 
were considered to be no better than random place- 
ment. This breakpoint was not affected by the number 
of fakies included in the analyses (Cope and Haltuch, 
2012). Here the results are presented for cases where 
5 fakies were added. 
Results from the HCA and the PA (with the use of 
both the silhouette coefficient and Hubert’s r cluster 
validity diagnostics) were then compared and recon- 
ciled. Reconciliation was performed by looking for con- 
sistently forming groups of co-occurring species in the 
catch that were supported by all clustering methods. 
Instances of groups being supported by 1, but not both, 
cluster approaches were noted. Throughout the pre- 
sentation of our results, we use the term “identifiable” 
clusters to represent clustered species that 1) had an 
average silhouette value >0.25 in each PA (Kauffman 
and Rousseeuw, 2005) and 2) a dissimilarity point that 
was less than that for simulated fake species in HCA 
results (Cope and Haltuch, 2012). 
Species assemblage analyses were completed on 
various subsets of the data to evaluate species co- 
occurrence in the demersal trawl fishery at a variety 
of temporal and spatial scales. These analyses helped 
to resolve fine-scale aspects of species co-occurrence 
with rebuilding species in the commercial catch. As- 
semblages were first evaluated on a coastwide basis 
by applying each clustering method to the data as a 
whole. The next part of the analyses partitioned the 
data by year. Additionally, dominant clusters some- 
times obscured smaller, but nonetheless identifiable 
groupings. To avoid such an outcome, we removed the 
ubiquitous species that had formed clusters when using 
all species combined and then ran all cluster analyses 
again with the remaining species to identify additional 
assemblages. 
Rebuilding species 
The characterization of species assemblages containing 
rebuilding species was an important consideration, yet 
the rebuilding species were some of the rarest of the 
species included in our data set. Thus, it was unlikely 
that they would be well represented in any assem- 
blage. Three approaches were taken to resolve the co- 
occurring relationship of rebuilding species with other 
species in the commercial catch data. With the first 
approach, we compared the proximity of rebuilding 
species with that of the simulated fakies that occurred 
with decreasing frequency. Cluster analyses were ex- 
plored with the occurrence probability of fakies on each 
tow (x) set equal to the frequency of occurrence of each 
rebuilding species (Table 1). This exploration allowed 
evaluation of the level of random assignment which 
best described the presence of rebuilding species in 
clusters. For example, if a species had a 5% frequency 
of occurrence, a probability of assigning a fakie to a 
tow was also set at 5%. A dissimilarity distance equal 
to or greater than the breakpoint of the fakies would 
indicate a randomly occurring, and therefore not a co- 
occurring, rebuilding species. 
With the second approach, we considered species 
co-occurrences only in the rare occasions when a re- 
building species was present on a tow, thus defining 
species assemblages as conditional on the presence of a 
rebuilding species. Using only positive tows for each re- 
building species as data sets, we re-analyzed clusters, 
and species assemblages were identified on a coast- 
wide, year-by-year basis. Fakies were also incorporated 
into this analysis to define clusters. 
With the third approach, we evaluated species as- 
semblages at finer spatial resolutions to identify spa- 
tially explicit co-occurrences with rebuilding species in 
the catch. For each rebuilding species, a tree regression 
was applied to identify a spatial stratification scheme 
on the basis of latitude. Tree regression uses recursive 
partitioning to split data into groups (Clark and Pregi- 
bon, 1992). In this case, the data were split by latitude 
on the basis of the log-transformed catch per tow of 
each rebuilding species and thus identified hot spots of 
species catch. Cluster analysis was then applied within 
each of the resulting latitudinal strata. Additionally, 
data were stratified with 1° latitude intervals, as well 
as on the basis of the departure port recorded by the 
observer. Clustering results from these 3 stratification 
schemes were then compared and summarized. 
All analyses described here were conducted in R soft- 
ware (vers. 2.13.2; R Development Core Team, 20 1 1 2 ). 
Results 
Overall species co-occurrences 
When using observer data from all areas and all 
years, we found 2 strong and consistent clusters: 1) a 
deepwater slope cluster and 2) a shelf cluster (Table 2, 
Fig. 2). The most common components of the slope clus- 
ter were Sablefish, Dover Sole, and Shortspine Thorny- 
head. This group also included Arrowtooth Flounder 
(Atheresthes stomias), Rex Sole ( Glyptocephalus zachi- 
rus), Longnose Skate (Raja rhina), and Pacific Hake, 
depending on the method used to determine clusters. 
The major constituents of the shelf cluster were Eng- 
lish Sole (Parophrys vetulus) and Petrale Sole. Hierar- 
chical clustering analysis also indicated that Lingcod 
( Ophiodon elongatus), Pacific Spiny Dogfish ( Squalus 
suckleyi), and Spotted Ratfish ( Hydrolagus colliei) were 
2 Mention of trade names or commercial companies is for iden- 
tification purposes only and does not imply endorsement by 
the National Marine Fisheries Service, NOAA. 
