Lee and Sampson: Dietary variations for three rockfish species off the Pacific Northwest 
513 
Table 2 
Extrinsic factors and their levels that were examined in the analyses of diet variability in rockfish species. Two data matrices 
were formed for the analyses: quarterly fishery samples (April 1998 to September 1999) and NMFS summer survey samples 
(1980 and 1998). Given in parentheses are the numbers of nonempty stomach samples for each level of the given factor across all 
other factors. 
Data matrix 
Factors 
No. of levels 
Levels 
Quarterly fishery 
Predator 
3 
S. flavidus (139), S. entomelas (194), S. pinniger (65) 
samples 
Season 
6 
Six seasons (81, 101, 51, 23, 97, 45) 
Depth 
2 
<146 m (304), >146 m (94) 
Time of Day 
3 
Morning (before 10 a. m.) (90) 
Midday (10 a. m.-5 p. m.) (230) 
Evening (after 5 p. m.) (78) 
Sex 
2 
Male (162), Female (236) 
Fish size 
3 
<40 cm (136), 40-45 cm (203), >45 cm (59) 
NMFS summer 
Year 
2 
1980(128), 1998(312) 
survey samples 
Latitude 
4 
41°-43°5’ (51), 43°5’-45° (153) 
45°-47° (137), 47°-49° (99) 
Depth 
3 
<110 m (113) 
110-165 m (222) 
>165 m ( 105) 
Time of Day 
3 
Morning (before 10 am) (58) 
Midday (10 am-5 pm) (260) 
Evening (after 5 pm) (122) 
Sex 
2 
Male (241), female (199) 
Fish size 
3 
<40 cm (125), 40-45 cm 
(208), >45 cm (107) 
importance of the factors by variance partitioning in 
a linear model. 
For the analysis of dietary variation, we formed two 
matrices of prey species composition data based on the 
two sampling schemes: a quarterly fishery sample data 
matrix for the three rockfish species collected from 
commercial fishing trips, and a summer survey sample 
data matrix for the S. flavidus stomachs collected dur- 
ing the 1980 and 1998 NMFS summer surveys. Some 
unique extrinsic factors, as well as shared factors, were 
associated with each data matrix. The unique extrinsic 
factors associated with the quarterly fishery sample 
data matrix were predator type and season. We used 
this data matrix to examine differences among the 
three rockfish species and the quarterly patterns in 
their diets but could not explore a latitudinal effect be- 
cause of the limited geographic coverage of the samples. 
The unique extrinsic factors associated with the NMFS 
summer survey data matrix were year (1980 and 1998) 
and latitude. Other factors tested for both data matrices 
were depth, time of day, sex, and fish size. All factors 
were treated as categorical variables. A summary of 
the factors, their levels, and the number of correspond- 
ing stomach samples for each data matrix are given in 
Table 2. 
For the analysis the prey species were grouped into 
seven major prey groups: euphausiids; fishes; salps; 
heteropods; jellyfishes (species other than salps and het- 
eropods in the gelatinous zooplankton group); decapods; 
and miscellaneous prey items. To remove the problem 
of unequal weights across the samples in the PCA, the 
weights of prey groups were standardized to proportions 
based on the total stomach contents weight for each 
individual fish stomach. In the data matrices each row 
represented an individual fish stomach and each column 
represented a prey group. The value of each cell in a 
data matrix is the weight proportion of a particular 
prey group in a particular fish stomach. Before run- 
ning the PCA, we transformed the weight proportions 
( WP ) using the angular transformation, (2/n) arcsine 
( WP 1/2 ), which is considered an appropriate transforma- 
tion for proportion data for improving the assumptions 
of normality and the homogeneity of variance (Sokal 
and Rob If, 1994). 
After running the PCA we fitted a series of general 
linear models (GLMs) to the sample scores extracted 
from the primary PCA axes to relate the diet composi- 
tions to the extrinsic factor. Different strategies are 
available for the selection of independent variables in 
a GLM: forward selection; backward elimination; step- 
wise selection; or use of statistical information criteria 
(e.g., Akaike information criteria) (Ramsey and Schafer, 
2002). For our study, which had limited data, we took a 
parsimonious approach in selecting variables: a forward 
selection strategy with a constraint on the maximum 
level of interactions. We did not consider interactions 
higher than two-way interactions. Given the limited 
number of samples (Table 2) it seemed unlikely that 
