278 
Fishery Bulletin 114(3) 
Figure 3 
Oceanographic maps smoothed (with the Kriging method) and created hy Barlow et al. (2009) and used with permission, 
displaying (clockwise from top left) surface temperature, surface salinity, surface chlorophyll, and mixed layer depth 
values in the eastern tropical Pacific (shaded region). Ship track-lines are shown with solid or dashed lines. Solid lines 
indicate sampling was continuous. Dashed lined indicated sampling was conducted at 55-km intervals. Numbers along 
isopleths indicate values for the variable represented in each map. Only variables coinciding with the 32 dipnet stations 
were used in the classification and regression tree (CART) analysis. 
sure of diet diversity ranging from 0 (no diet diver- 
sity) to 1 (high diet diversity) among predicted prey 
categories. A large tree is produced and 10-fold cross- 
validation is used to prune the tree to within one 
standard error of the tree yielding the minimum er- 
ror (i.e., the “1 SE” rule [Breiman et al., 1984; Kuh- 
nert et al., 2012]). Predictions are made by partition- 
ing a new observation down the tree until it resides 
in a terminal node. The prey group with the greatest 
numeric proportion among a suite of prey in the diet 
is displayed at each terminal node. The vector of prey 
proportions, in numbers of prey eaten by an individual 
predator is represented as a univariate categorical re- 
sponse variable of prey type (class), with observation 
(case) weights equal to the proportion of the prey type 
eaten by the predator. Fish with empty stomachs were 
omitted from this analysis because we were interested 
in how predictor variables influenced prey type. Rank- 
ings of variable importance are computed to identify 
which predictor variables are most important in the 
model. In addition, Kuhnert et al. (2012) implemented 
a spatial bootstrapping technique to account for spa- 
tial dependence in the data and to assess uncertainty 
in the predicted diet composition at each node in the 
classification tree. The classification was implemented 
in R software, vers. 3.1.1 (R Core Team, 2014) with 
the ‘rpart’ package (Therneau et al., 2013); further de- 
tails can be found in Kuhnert et al. (2012). 
We used CART analysis to explore the relationship 
among 12 dependent spatial, oceanographic, and bio- 
logical predictor variables (Table 1) and the response 
variable, diet composition. Spatial predictors included 
latitude and longitude, oceanographic predictors con- 
sisted of MLD, SSS, SST, and SCHL concentration, and 
biological predictors contained information on the zoo- 
plankton prey community by using data from the net 
samples and the myctophid predators. Data represent- 
ing the zooplankton community (potential prey) includ- 
ed ostracod, copepod, and euphausiid numeric composi- 
tion and zooplankton displacement volume in the net 
samples (Table 2). Standard fish length was used to 
assess the effects of ontogenetic diet. We used species 
as a predictor variable to assess resource-partitioning 
among species. 
