Underwood et al.: Behavior-dependent selectivity of Limcinda ferruginea in the mouth of a bottom trawl 
435 
Table 3 
Summary of the 3 statistical models used for analyses of behavioral responses of yellowtail flounder ( Limanda ferruginea) 
observed in video footage from 5 tows of a bottom trawl in June 2010 on the southern Grand Bank off eastern Newfound- 
land. Initial herding response: initial response=location+swimming direction+length+gait+start density+tow (random factor). 
Change in herding response: change in response=location+swimming direction+length+gait+initial response+residence+start 
density+tow (random factor). Capture outcome: capture outcome=previous gear experience+length+gait+initial 
response+residence+start density+tow (random factor). Variables indicated in bold are significant in the reduced models 
(P< 0.05). Z value is the Wald-Z test. Location 1: port vs. starboard; location 2: port vs. middle; swimming direction 1: port 
vs. starboard; swimming direction 2: port vs. vessel; initial response 1: slope vs. run; initial response 2: slope vs. rise. 
Initial herding response 
Change in 
herding response 
Capture outcome 
Variable 
Z-value 
P(>Z) 
Z-value 
P(>Z) 
Z-value 
P(>Z) 
Intercept 
0.416 
0.68 
-2.384 
0.02 
-2.240 
0.03 
Location 1 
0.939 
0.35 
-0.057 
0.95 
Location 2 
0.570 
0.57 
-0.159 
0.87 
Swimming direction 1 
1.415 
0.16 
0.275 
0.78 
Swimming direction 2 
Previous gear experience 
2.404 
0.02 
-0.607 
0.54 
-2.031 
0.04 
Length 
0.213 
0.83 
0.390 
0.70 
0.278 
0.78 
Gait 
0.616 
0.54 
-1.573 
0.12 
-1.590 
0.11 
Initial response 1 
3.465 
<0.001 
3.366 
<0.001 
Initial response 2 
0.000 
0.99 
Residence 
1.037 
0.30 
-1.117 
0.26 
Start density 
-0.494 
0.62 
0.237 
0.81 
-0.029 
0.98 
sponse=location+swimming direction+length+gait+initial 
response+residence+startdensity+tow [random factor] ), 
and capture outcome (model 4: capture outcome=previous 
gear experience+length+gait+initial response+residence 
+start density+tow [random factor]). 
The influence of fish location in relation to the foot- 
gear on the orientation of 190 individual yellowtail on 
the substrate (previous gear experience, specifically for 
previously herded fish) was tested for uniformity (non- 
randomness) with the Rayleigh test by using Oriana 
software, vers. 3 (Kovach Computing Services, Angle- 
sey, Wales). 
Because we were interested in the effect of fish 
length, along with other covariates, in shaping be- 
havioral responses, 40 individual yellowtail that had 
no length data (i.e., fish that were unmeasured) were 
dropped from the analysis for models 2-4 (initial re- 
sponse, change in response, and capture outcome). For 
the initial response model, we initially attempted a 
multinominal analysis. However, we had zero obser- 
vations for hop and pass under responses and only 9 
observations for the rise response, thereby invalidat- 
ing any further multicategorical analysis. The statisti- 
cal analysis for the initial response model was then 
focused on the herded individuals, and binomial analy- 
sis was used with the initial response variable catego- 
ries of run and slope for 141 observations. The model, 
therefore, was altered and named “initial herding re- 
sponse.” The statistical analysis for the change-in-re- 
sponse model was also focused on only the herded indi- 
viduals (i.e., initial response variable categories of run 
and slope), and binomial analysis was used with the 
change-in-response variable categories of changed and 
continued for 141 observations. The model, therefore, 
was altered and named “change in herding response.” 
However, with the capture outcome model, we exam- 
ined all initial responses (i.e., initial response variable 
categories of run, slope, and rise), using binomial anal- 
ysis with capture outcome variable categories of caught 
and escaped for 150 observations. 
To account for the variance between tows and pseudo 
replication (Millar and Anderson, 2004) in analysis of 
models 2-4 (initial herding response, change in herding 
response, and capture outcome), we used a generalized 
linear mixed model (GLMM) with binomial error, with 
tow as a random factor. Analysis with GLMMs was car- 
ried out with the lme4 package (Bates et al., 2013) in 
R, vers. 3.0.2 (R Core Team, 2013). Explanatory vari- 
ables with more than 2 categories (i.e., location) were 
automatically separated into binomials by R (i.e., port 
location versus starboard location; Table 3). Variables 
in the models were reduced by using backward step- 
wise deletion until only variables that explained a 
significant amount of variation (likelihood ratio test, 
P< 0.05) in the data remained (Crawley, 2007). 
Results 
Catch composition of flatfishes varied with each tow, 
ranging from 84% to 92% for yellowtail and from 8% to 
15% for American plaice. Witch flounder (Glyptocepha- 
lus cynoglossus ) were present in only one tow (Table 
1). The length of yellowtail in the catch ranged from 
