500 
Fishery Bulletin 11 5(4) 
Table 3 
Classes of generalized linear models (GLMs) and generalized additive models (GAMs) evaluated for goodness of fit in this 
study for the use of acoustic data to predict whether sampling areas in the Gulf of Alaska are trawlable. The best model in 
each category is shown in boldface. An asterisk (*) indicates an interaction term. The parameters include roughness, hard¬ 
ness, first bottom length (length 1), second bottom length (length 2), bottom rise time, depth, maximum volume backscatter 
strength (Sv), kurtosis, and skewness. NS=not significant; poly=polynomial ; S=smooth; d=degree of polynomial; k=number 
of knots. 
Roughness 
Hardness 
Length 1 
Length 2 
Rise time 
Depth 
Max Sv 
Kurtosis 
Skewness 
GLM 
Model 1 
linear 
linear 
linear 
linear 
linear 
linear 
NS 
linear 
linear 
Model 2 
poly (d=3) 
linear 
poly (d=2) 
linear 
linear 
linear 
NS 
linear 
poly (d=2) 
Model 3 
linear 
linear 
*skewness 
linear 
linear 
linear 
linear 
linear 
*length 1 
GAM 
Model 1 
so 
so 
so 
so 
so 
so 
so 
so 
so 
Model 2 
S(k=4) 
S(k=4) 
S(k=4) 
S(k=4) 
S(k=4) 
S(k=4) 
NS 
NS 
S(k=4) 
Model 3 
so 
SO 
*skewness 
so 
so 
so 
so 
so 
*length 1 
linear and interaction terms. When interaction terms 
were introduced, the choice of which terms to include 
in each interaction was based on their correlation. 
Specifically, the correlation between skewness and first 
bottom length, skewness and kurtosis, and between 
hardness and max S v were high at 0.92, 0.97, and 0.82, 
respectively, and were therefore examined for poten¬ 
tial interaction effects. The step function in R, which 
calculates the AIC value of models starting with the 
full model that used all 9 parameters and then in a 
stepwise fashion eliminates 1 parameter at a time, was 
applied to each candidate model to determine whether 
a simpler, reduced model would result in a better fit. 
The best GLM among the 3 classes was then chosen as 
the one with the lowest AIC value. 
Three classes of generalized additive models (GAMs) 
were also evaluated, again with the minimum AIC val¬ 
ue to choose the best-fitting model within each class 
(Table 3). The 3 classes consisted of models with all un¬ 
constrained smoothing terms, models with constrained 
smoothing terms, and models with mixtures of uncon¬ 
strained and bivariate smoothing terms (interaction 
terms). Constraining the smoothing functions consisted 
of setting the maximum number of knots allowed (in 
all cases, 4 knots). Choice of the variables used in the 
interaction terms was again based on the magnitude 
of the correlation between variables. Model selection 
for the GAMs proceeded with a process similar to that 
used for the GLMs, but was done manually because 
a step function is unavailable for GAMs. Instead, we 
used the P-values from the analysis of variance of the 
model to sequentially eliminate nonsignificant terms. 
In the event of more than 1 nonsignificant term, the 
term with the least significance was first eliminated, 
then the reduced model was refitted and further non¬ 
significant terms were sequentially removed. The best 
overall GAM was again chosen among the best models 
within each class on the basis of minimum AIC. 
After selecting the best candidate GLM and GAM, 
we compared them in terms of their ability to correctly 
classify data not used in the model building process. 
The best model (i.e., the one with the highest predic¬ 
tive accuracy) when applied to the training data is not 
necessarily the best choice when applied to new data. 
To assess the relative robustness of the models when 
subjected to new data, we used 33% holdout cross vali¬ 
dation (Arlot and Celisse, 2010; Maunder and Harley, 
2011), which proceeded as follows. Random samples 
of 160 segments, split equally between trawlable and 
untrawlable data, were selected without replacement 
from the pool of 238 acoustic segments. These data 
were then used as a training sample to construct pre¬ 
diction functions by fitting the best GLMs and GAMs. 
The remainder of the sample was used as a proxy for 
new data. Each of the fitted models was then applied 
to the new data to estimate the probability that each 
of the ~50 records within each segment was trawlable 
by using the “predict” function in R. Probabilities >0.5 
was used as a criterion to classify individual records 
as trawlable. Likewise, the criterion used to classify 
entire segments as trawlable was that the proportion of 
records classified as trawlable was also >0.5. After the 
trawlability of all segments was estimated, the values 
were compared with the trawlability classification of 
the sampling cells. The proportion of correctly classi¬ 
fied segments was then calculated. This process was 
repeated 100 times and the average proportion of cor¬ 
rect classification was used as a measure of how well 
each model predicted trawlability. 
Results and discussion 
On the basis of the minimum value of AIC, the best 
GAM, with 7 unconstrained smoothing terms and 1 
interaction term, produced an overall (trawlable and 
