280 
Fishery Bulletin 117(4) 
showed river-specific variability in diet, dietary ontog¬ 
eny, growth, and population structure, we conducted a 
separate CCA for each river (Hilling et al., 2018; Schmitt 
et al., 2019). Each CCA was completed in the package 
vegan (vers. 2.4-4; Oksanen et al., 2017), which is an 
extension of the statistical software R, vers. 3.4.3 (R Core 
Team, 2017). 
Predation models for species of concern 
We used binomial generalized additive models (GAMs), 
which are semiparametric generalizations of logistic 
regression (Hastie and Tibshirani, 1990), to examine 
relationships between the binary occurrence of depleted 
alosines (American shad and river herring), blue crab, and 
American eel in the diet of blue catfish by predictor vari¬ 
ables identified in each CCA. This approach was especially 
useful for identifying when and where predation by blue 
catfish on these species of concern was most likely. Again, 
the GAMs were based on occurrence data because it is the 
best metric for assessing predation and is often more reli¬ 
able than other diet measures (Baker et al., 2014; Buck- 
land et al., 2017). A GAM is flexible because it assumes 
only that functions are additive and relationships are 
smooth (Guisan et al., 2002). A GAM, like a generalized 
linear model, uses a link function to establish a relation¬ 
ship between the mean of the response variable and a 
“smoothed” function of the explanatory variables, making 
it robust to scattered or correlated data (Lin and Zhang, 
1999). Separate GAMs were constructed for American eel, 
depleted Alosa species (American shad and river herring), 
and blue crab. Each GAM included smoothing functions 
for predator length and salinity, yet month and river were 
treated as categorical factors (Wood, 2006). Although 
interannual variation is likely an important driver of 
dietary patterns, stomach contents for each month were 
pooled across years (e.g., contents for June in 2013, 2014, 
2015, and 2016 were combined into the single category for 
June) to increase sample sizes. This pooling was a neces¬ 
sary step because circumstances that were out of our con¬ 
trol (equipment failure and weather) resulted in limited 
sampling during some combinations of month and year 
(e.g., June 2014). 
Each model was created by using the R package mgcv, 
vers. 1.8.28 (Wood, 2017), with default values unless oth¬ 
erwise specified. Each model was constructed as follows: 
logit (P) = P 0 + /i( predator length) 
+ f 2 (salmity) + month + river, 
where logit = the binomial link function; 
P = the probability of a species being consumed; 
P 0 = the model intercept; and 
f\-f 2 = the smoothing functions realized by penalized 
thin plate regression splines (Wood, 2006). 
For each covariate, the mgcv package fits a series of penal¬ 
ized regression splines as smoothing functions and sup¬ 
plies degrees of freedom for smooth terms by minimizing 
generalized cross-validation scores (Wood, 2006). Individ¬ 
ual P-tests were then used to determine which predictors 
contribute significantly to the deviance explained (Wood, 
2006). The probability of encountering each species in 
stomachs of blue catfish was then predicted separately by 
river to elucidate the conditions that lead to higher pre¬ 
dation rates for these species. Overall predictive perfor¬ 
mance of each model was then assessed by using the area 
under the receiver operating curve (ROC) in the package 
ROCR (vers. 1.0-7; Sing et al., 2005) in R. An area under 
the ROC of 0.5 is equivalent to a random guess, a value 
of 1.0 indicates perfect model performance, and a value 
>0.7 indicates adequate model performance (Bewick et al., 
2004; Austin, 2007). 
Results 
Data collection 
During 2013-2016, we collected 14,488 blue catfish stom¬ 
achs at 542 sites on the James, Pamunkey, Mattaponi, and 
Rappahannock Rivers in eastern Virginia (Fig. 1). In addi¬ 
tion, stomachs were collected from several hundred fish that 
were captured from the York River, which forms at the con¬ 
fluence of the Pamunkey and Mattaponi Rivers near West- 
point, Virginia (Fig. 1). For simplicity, fish captured from 
the northern half of the York River were allocated to the 
Mattaponi River sample, and fish captured on the south¬ 
ern half were allocated to the Pamunkey River sample. Of 
the stomachs collected, 7302 contained food items (50%). 
Although stomachs (sample size [/r]=16,110) were collected 
year-round (Schmitt et al., 2019), we limited this study to 
stomachs that were collected by using low-frequency elec¬ 
trofishing following a stratified random sampling protocol 
(April-October) to avoid spatiotemporal biases. 
Major diet drivers for blue catfish 
Diets of blue catfish varied by river, salinity, season, and 
predator TL, and all constraining variables were statisti¬ 
cally significant in the CCA (P<0.001; Table 1). For each 
river, the first 2 CCA axes accounted for a considerable 
amount of variation in the diet of blue catfish: 80.0% in 
the James River, 85.0% in the Rappahannock River, 97.4% 
in the Pamunkey River, and 93.3% in the Mattaponi River. 
Global F-tests on each CCA for each river were highly sig¬ 
nificant (P<0.001 for all), and nearly all constraining vari¬ 
ables significantly affected the diet of blue catfish in each 
river (P<0.001), with the exception of salinity zone in the 
Mattaponi River (P=0.081; Table 1, Fig. 2). 
Results of each CCA indicate several key patterns in the 
diet of blue catfish. First, there were consistent, length- 
related (i.e., ontogenetic) shifts from omnivory to pisciv- 
ory in all rivers. Second, blue catfish more frequently 
preyed on invertebrates or crustaceans during spring than 
during other seasons and began to consume more fish as 
the seasons progressed. Third, the predation of various 
invertebrates is generally associated with lower salinities, 
