296 
Fishery Bulletin 1 10(3) 
Separation of the effect of Julian day from the effect of 
location on abundance and cannibalism was, therefore, 
not possible and Julian day was not directly included in 
further analyses. 
A step-wise model-selection approach was used for 
GAMs, where all covariates were included in initial 
models and covariates with the least significant P- 
values (P>0.05) were removed one at a time in sub- 
sequent models until all covariates in the model were 
significant. The “mgcv” (Wood, 2000) library in R (R 
Development Core Team, 2010) was used to run the 
GAM models. 
Pollock diet 
For pollock that were subsampled from each trawl haul 
for diet information, lengths and weights were mea- 
sured and stomach contents were removed. Stomach 
contents of pollock were identified to the lowest possible 
taxonomic category (10 prey categories), enumerated, 
and weighed, and in the case of prey fish, measured. 
The 10 prey categories were chaetognaths, euphausiids, 
amphipods, copepods, crab, miscellaneous, and pollock 
(<60 mm standard length [SL], 60-200 mm SL, >200 
mm SL, and unmeasured). If unmeasured (i.e., too 
digested to measure), and prey pollock were accompa- 
nied by measured prey pollock in the same predator 
stomach, the unmeasured prey items were assigned 
the same average length as the measured pollock, 
otherwise, they were not included in the analyses. The 
percent frequency of occurrence, partial fullness, and 
percent body weight of each prey item were calculated 
for each predator pollock and then averaged over preda- 
tor size categories ( PredatorLength ) at each station in 
each year. Predator length categories were 1-19 cm, 
20-29 cm, 30-39 cm, 40-49 cm, 50-59 cm, and >60 
cm fork length (FL). The presence or absence (PA) of 
cannibalism ( l = cannibalism present; 0 = cannibalism 
absent) was recorded. Percent frequency of occurrence 
( %FO ) was calculated as 
n „ 
%FO = — , 
n r 
where n p = the number of predators that consumed prey 
type p; and 
rif - the number of predators with food (f) in their 
stomachs at that station. 
If %FO was zero, a value of zero was assigned to PF. 
Pollock diet models 
GAMs were also used to explore the relationship between 
pollock cannibalism on age-1 pollock (60-200 mm SL) 
and important covariates. There were many zeros in the 
pollock diet database — a feature typical of this type of 
data. To address this issue, the first step was to examine 
whether the presence or absence (PA) of cannibalism 
(binary response variable) was related to the covariates, 
therefore, data from all stations were used (stations 
where cannibalism was and was not found to occur). 
GAMs were used to explore the relationship between the 
occurrence of pollock cannibalism on age-1 pollock (PA) 
and covariates, which were the following: Year, Cold- 
Pool, PredatorLength, Overlap, TempDepth (s [bottom 
temperature, bottom depth]), and Location (s [latitude, 
longitude]). The time of day that pollock were sampled 
was not included in GAMs because it likely would not 
affect evidence of cannibalism as digestion of pollock 
prey takes longer than 24 hours (Dwyer et al., 1987). 
The subsequent step was to determine those factors 
affecting the relative amount of cannibalism (hereafter 
referred to as simply the amount of cannibalism) at 
stations where cannibalism occurred. The amount of 
cannibalism on age-1 pollock was estimated by weight- 
ing %FO, %BW, and PF of prey pollock (60-200 mm 
SL) by the abundance of large predatory pollock (e.g., 
%FOxLgPollock). These weighted estimates were then 
ln + 1 transformed and, hereafter, are referred to as 
%FO LgPolloclr %BW LgPollock’ and PF LgPollock and were the 
response variables assessed. The covariates in these 
GAMs included ColdPool, PredatorLength, Temp- 
Depth, Location, and a smoothed SmallPollock term 
( slSmallPollock ]). The covariate Overlap was not tested 
as a predictor of the amount of cannibalism because it 
was assumed that where cannibalism occurred, preda- 
tor and prey pollock co-occurred. Instead, SmallPollock 
was included as a covariate because the amount of prey 
pollock available may influence the amount of cannibal- 
ism that occurred. The same step-wise model-selection 
approach that was used for the abundance GAMs was 
used for the diet GAMs. 
Results 
Percent body weight ( %BW ) was calculated as 
Pollock abundance 
%BW 
^100 A 
I 
w 
PJ 
BW 
where W pj = the weight of prey type p in predator j; and 
BW = the body weight of predator j. 
Partial fullness (PF) was calculated as 
PF = %FO x %BW. 
During 1982-2006, 2754 bottom trawls were conducted 
on the EBS shelf (Fig. 1). The number of stations sam- 
pled ranged from a low of 13 in 1982 to a high of 170 in 
1993, with an average of 120 stations sampled per year. 
Average CPUE of large pollock (>200 mm SL; LgPollock) 
declined until the late 1990s; whereas, average CPUE of 
small (60-200 mm SL; SmallPollock) pollock was high 
and variable before 1998 and lowest during 2003-2006 
(Fig. 2). CPUE of very small pollock ( < 60 mm SL) was 
rare (<0.06 ha~D in all years. The proportion of sta- 
