86 
Fishery Bulletin 1 12(1) 
for analysis (one fish was detected at the study site for 
<3 h and was, therefore, excluded from analysis). 
The activity index was defined as the average num- 
ber of times a fish moved between adjacent acoustic 
stations per hour during a given 3-h time period (dawn, 
day, dusk, and night). This index was calculated for fish 
that moved at least once in a 3-h period, and values for 
this index ranged between 0.33 and 9. Like the move- 
ment index, the activity index indicated movements on 
the scale of about 400 m. Activity indices were calcu- 
lated for each time period between the time of tagging 
and 6 November 2003 because observations on activity 
levels were not temporally consecutive after this date. 
As before, 121 fish contributed information; 11,843 ob- 
servations were available for this index. 
Analysis of movement 
We used a generalized linear model to examine diel 
and seasonal patterns in fish movement and to esti- 
mate the probability that a fish moved as a function of 
sex, water temperature at the bottom of the seafloor, 
and salinity at the bottom of the seafloor. We did not 
consider effects of fish size because preliminary inves- 
tigations indicated that movement did not vary with 
size. Time entered the model as continuous linear, 
quadratic, and cubic effects and corresponded to cal- 
endar day (i.e., time=l on 30 May, and time=174 on 
19 November). Quadratic time effects allow the direc- 
tion of the response to change once (e.g., the response 
increases, reaches a maximum, and then decreases); 
cubic time effects are required when the direction of 
the response changes twice. All time factors were stan- 
dardized to remove effects of collinearity; centering the 
data was not effective. The effects of season (summer 
[30 May-7 Sep] or fall [8 Sep-19 Nov]) and time period 
(dawn, day, dusk, and night) also were considered. 
In addition to these multiscale temporal effects, we 
examined the effect of release group (defined according 
to tagging date: early June, late June, or July) because 
preliminary modeling indicated heterogeneity of move- 
ments for fish tagged at different times. Environmental 
effects were characterized by water temperature and 
salinity measured at the bottom of the seafloor at the 
study site. Specifically, we used mean bottom water 
temperature at station B1 and differences in mean wa- 
ter temperature and salinity at the bottom of the sea- 
floor between stations H7 and Bl; these measures were 
estimated for each day of the study and standardized 
by scaling the data with the standard deviation. These 
environmental factors were selected because tolerance 
estimates for other measures (e.g., mean temperature 
at station H7) indicated strong collinearity with time 
(tolerance values <0.10) that could not be removed 
through standardization or centering (Quinn and Ke- 
ough, 2002). 
Because movement indices (0, 1) were recorded daily 
for each fish during 4 time periods, we analyzed these 
data with a repeated-measures approach to address 
the potential correlation among observations from each 
fish. Furthermore, we identified individual fish nested 
within a time period as the subject for modeling the 
repeated measures. We fitted the following generalized 
linear model to the data: 
Yjklmn = + +4 +ri + 4m+*n +*2 +*£+* + * + « 
+ interactions, 
where Tjiqnm 
M 
bj 
4 
Yi 
the binomial movement index of the i th 
fish of the j th sex of the k th tagging 
group in the 1 th season in the m th time 
period for the n th day (time); 
the expected response (either 0 or 1); 
the effect of the j th sex; 
the effect of the k th tagging group; 
the effect of the 1 th season; 
the effect of the m th time period; 
T n, f n , and T n = the linear, quadratic, and 
cubic effects of the n th day; 
(p = the effect of mean daily temperature at 
station Bl; 
k - the effect of the mean temperature differ- 
ence between stations Bl and H7; 
a - the effect of the mean difference in salin- 
ity between stations Bl and H7; and 
interactions refers to 2- and 3-way interactions 
between the fixed effects. 
We did not include higher-order interactions in the 
generalized linear model because such complexity was 
either unnecessary or severely reduced the precision 
of the estimated parameters. In this model, Tjidmn was 
assumed to be distributed as a binomial. The general- 
ized estimating equation (GEE) method was used to 
estimate model parameters (Liang and Zeger, 1986; 
Littell et ah, 2002) with the GENMOD procedure in 
SAS (vers. 9.3). 
We evaluated several covariance structures to de- 
scribe the potential correlation among the binomial re- 
sponses: independent, compound symmetry, autoregres- 
sive with lag 1, and m-dependent (SAS User’s Guide, 
vers. 9.3). For the m-dependent structure, correlations 
varied with the first m-time intervals, but were equal 
to 0 after the m th interval. Because we fitted a number 
of potential models, we used the information-theoretic 
approach to identify the most suitable model from the 
set of models that we considered (Burnham and An- 
derson, 2002). 
Typically, Akaike’s information criterion (AIC) is 
calculated from model likelihoods, and differences in 
model values of AIC are used to guide model selection; 
the model with the lowest criterion is considered best. 
However, GEE methods do not use likelihoods to com- 
pute model parameters; instead, quasi-likelihoods are 
used. For GEE models, the quasi-likelihood information 
criterion (QIC), which is a modification of the AIC, was 
used to select the covariance structure that best fits 
the data and to guide model selection (Pan, 2004). In 
