412 
Fishery Bulletin 115(3) \ 
a maximum within the first 5-10 min followed by a 
steady decline in all videos), we chose a 30-min seg- 
ment from each video for counting fish. Once the cam- 
era view was clear of suspended silt and debris, counts 
were made for 2 categories of black sea bass: all black 
sea bass and then separately for those with a nuchal 
hump (i.e., fish that were distinguishable from other 
individuals by their darker body coloration, a nuchal 
hump with the usual blue color appearing grayish- 
white on the videos, and white fin stripes; hereafter 
referred to as nuchal black sea bass ) from each video 
using a variable called MeanCount, which is an alter- 
native counting metric to others commonly reported in 
the literature (e.g., MaxN; maximum number of indi- 
viduals of a particular species present at one time in 
any single point on the video) (Schobernd et al., 2014; 
Bacheler and Shertzer, 2015). MeanCount is the mean 
number of fish counted in a sample of frames from a 
video (Schobernd et al., 2014). In this case, 60 single 
frames were sampled systematically, one every 30 s for 
30 min of videotaping. Counts from the sampled frames 
were then averaged to obtain values of MeanCount. We 
chose MeanCount because, unlike MaxN, it has been 
shown by Schobernd et al. (2014) to be relatively un- 
biased and linearly related to true abundance but has 
similar variation to that of MaxN. 
Fish behavior during the selected 30 min of a video 
was evaluated by noting (in minutes) the time of first 
arrival (TFA) of fish within the camera view followed 
by general observations of behavior around the cam- 
era system. TFA was included as a behavioral measure 
to examine whether faster arrival times to the camera 
view could be related to greater densities of fish in the 
surrounding area. Additionally, we wanted to deter- 
mine whether the presence of bait in the trap would 
result in fish appearing on cameras earlier than when 
bait was not present in the trap. Swimming, resting, 
and habitat-associating behaviors were recorded by se- 
lecting individual fish within the camera field-of-view. 
Swimming fish were followed until they left the camera 
view; no more than 3 fish were followed at any one 
time. Resting and habitat-associating behaviors were 
recorded if, or when, a swimming fish stopped to rest 
on the bottom or near structures such as rocky outcrop- 
pings or boulders. These behaviors were also noted for 
fish already resting on the bottom when the camera 
frame landed. Aggregating behaviors were documented 
for fish resting on the bottom in groups of 2 or more, 
and antagonistic behaviors were noted only when nu- 
chal males were observed chasing smaller non-nuchal 
fish. 
Responses to the trap, including entries through the 
entrance funnel, half entries (entering the entrance 
funnel but backing out), and exits (exiting the trap 
through the entrance funnel or through one of the es- 
cape vents in the parlor) were noted on videos captured 
by the camera facing downward over the top of the trap 
for each deployment (Fig. 2). These data were collected 
on 9 of the 10 sampling days and were used to examine 
the influence of trap soak time on catches of black sea 
bass in fish traps in a complementary manuscript (i.e., 
Cullen and Stevens, 2017). 
Data analysis 
We tested for differences in MeanCount among the 3 
classified habitat types. Because MeanCount is a con- 
tinuous variable and repeated deployments were made 
at a site on each sampling day, we used linear mixed- 
effects models to test for differences in MeanCount for 
the categories of all black sea bass and nuchal black 
sea bass separately among the 3 classified habitat 
types (sand, sand+rock, live bottom). Linear mixed- 
effects models can be used as alternatives to methods 
of repeated-measures analysis of variance (AN OVA) 
when data are unbalanced, and they allow modeling 
of covariance structures (Pinheiro and Bates, 2000). In 
our models, habitat type was treated as a fixed effect; 
however because of limited knowledge of bottom types 
at sampling sites, equal replication of habitats across 
video deployments was not possible a priori. Bait 
method (i.e., baited trap, unbaited trap) was dummy 
coded (i.e., the categorical variable bait method was 
converted to a continuous variable by assigning values 
of 0 for baited trap deployments and values of 1 for 
unbaited trap deployments) and the continuous vari- 
able was included as a covariate in the models to con- 
trol for its possible influence on values of MeanCount. ! 
Sampling site was treated as a random effect because i 
consecutive camera system deployments provided 
multiple, non-independent samples per site (Zurr et 
al., 2009). This method, which was equal to fitting a 
model with a compound symmetrical correlation struc- 
ture, provided a random intercept term for each site, 
and allowed the variance in values of MeanCount 
within sites to be separated from the residual vari- 
ance (Pinheiro and Bates, 2000). Linear mixed-effects 
models with MeanCount as the response variable were 
fitted by using the nlme package, vers. 3.1-129 (Pin- j 
heiro et al., 2017) in the R statistical environment, 
vers. 3.3.2 (R Core Team, 2016). MeanCount data were 
checked for normality and variance homogeneity and 
log-transformed (by taking a natural logarithm of the 
variable+1; i.e., \og e [MeanCount+l], 1 was added to ' 
MeanCount because the data contained some 0 values) 
before analysis to help meet the assumptions of the : 
linear mixed-effects models. Corrected Akaike infor- 
mation criterion (AICc), which is recommended for 
small sample sizes (Burnham and Anderson, 2002) 
was used to compare 3 model types: models with ran- 
dom effect for site, models without the random effect 
for site, and weighted models with the random effect 
for site. The latter models were weighted by using a 
constant variance function (i.e., weights produced with 
the varldent function in the nlme package) to cor- 
rect for heteroscedasticity or different variances for 
MeanCount data among habitat types (Pinheiro et 
al., 2017). The constant variance function in weighted j 
models allowed the variance to differ for each level of j 
habitat type. 
