164 
Fishery Bulletin 112(2-3) 
Results 
Sea-surface temperature for the study area over the 
period 1979-2009 ranged from 12.7°C to 19.4°C, with 
a mean of 16.2°C. Overall averaged seasonal anoma- 
lies ranged from -1.5°C to 1.1°C around the mean (Fig. 
3). In comparison, seasonal anomalies by grid section 
ranged from -3.8°C to 3.4°C. Years with a strong posi- 
tive PDO (index>l) were 1983, 1987, 1993, 1997, and 
2003, and a strong negative PDO (index<-l) occurred 
in 1999 and 2008 (Fig. 3). Strong positive ENSO years 
were 1982-83, 1987-88, 1991-92, 1997-98, and 2002- 
03, and strong negative ENSO years were 1988-89 and 
1999-2000 (Fig. 3). No long-term trends in SST were 
apparent in our data given the levels of seasonal and 
ENSO variation observed. However, a linear regression 
of PDO anomaly data shows an overall negative trend 
in the last 30 years: coefficient of multiple determina- 
tion (i? 2 )=0.215, P=0.009. This pattern is likely a result 
of the PDO regime switch in the last decade (Overland 
et al., 2008; Hodgkins, 2009). 
The correlation analysis revealed that annual sight- 
ing rates and mean group size were not correlated for 
any of the species examined. This finding indicates 
that, although species may be encountered with vary- 
ing frequency across years, the number of individual 
animals per group does not change in a correlated way. 
For example, if more groups of a given species were 
also was conducted to determine whether group size 
correlated with the number of groups encountered. 
To select predictor variables for inclusion in each 
model, a likelihood-based smoothness selection meth- 
od, instead of a traditional stepwise method, was ap- 
plied with the restricted maximum likelihood (REML) 
criterion (Patterson and Thompson, 1971; Wood, 2006). 
Each predictor variable was tested for inclusion in the 
model with a tensor product approach coupled with 
a smoothing function defined by a cubic regression 
spline with shrinkage. The best model was selected 
on the basis of a combination of the information-the- 
oretic descriptor Akaike’s information criterion (AIC; 
Akaike, 1976) and REML. Next, an interactive term 
selection method was applied to sequentially drop the 
single term with the highest nonsignificant P-value 
and then refit the model until all terms were signifi- 
cant. The best-fit model was therefore one that mini- 
mized AIC and maximized REML and the explained 
deviance and that included only significant predictor 
variables. In addition, the ENSO, PDO, and seasonal 
SST averages, as well as each of the depth metrics, 
were tested for correlation if more than one of them 
was included in a model as a significant predictor. 
These variables were then included together only if 
they were not correlated. If the variables were cor- 
related, then only the most significant variable re- 
mained in the final model. 
