188 
Fishery Bulletin 109(2) 
swept was calculated by multiplying the tow distance 
by the net opening width, the latter calculated from an 
empirical relationship between depth and net width. 
To compare differences in salinity and temperature 
among depths, months, and latitude we obtained data 
from the King County Puget Sound Marine Monitor- 
ing Program which monthly monitors water quality 
at two stations in the northern and southern regions 
of the survey area (Fig. 1). Data were collected with 
a conductivity-temperature-density instrument (CTD) 
consisting of an SBE 3 temperature sensor, SBE 4 con- 
ductivity sensor, and SBE 29 pressure sensor (SeaBird 
Electroics Inc., Bellevue, WA) and were binned at 0.5- 
m intervals. CTD sampling occurred within 10 days of 
trawl sampling. 
Statistical analysis 
More than 200 species of fish have been documented in 
Puget Sound, but many of these are rare or sparsely dis- 
tributed such that an intensive sampling effort would be 
required to sufficiently describe the distribution patterns 
of all species. Instead, we focused our research on the 
commonly occurring species that accounted for the bulk 
of the demersal fish biomass and therefore represent the 
most significant fish in the food web. To calculate diver- 
sity metrics for comparisons, rare species that occurred 
in fewer than 10% of the sampled trawls were removed 
from the data set; the exclusion of rare species permitted 
a coarse-scale evaluation of differences in the common 
components of the assemblage. Differences in species 
richness (IV) and diversity (Shannon-Wiener diversity 
index, H'; Krebs, 1989) across depths and among months 
were examined by using two-way analysis of variance 
(ANOVA). Standard two-way ANOVA requires that 
treatment levels be fully replicated across both main fac- 
tors (in this case, depth and month). Because we lacked 
samples from depths of 20 m and 40 m in March, we 
performed two sets of tests. In the first set we included 
samples from all four depths, but only from October and 
July. In the second, we included samples from all three 
months, but only from 40 and 160 m. Initial examination 
of the data indicated normal or near-normal distribu- 
tions, therefore data transformations were not called 
for because ANOVA is robust to minor departures from 
normality (Zar, 1984). In instances where either of the 
main factors was significant, Tukey’s honestly signifi- 
cant difference (HSD) tests were used to identify which 
depth and month levels differed. 
In the above analysis, both N and H' are simple and 
widely used measures of diversity that describe the 
number or relative biomass of species at each sample 
but ignore similarity in species composition among 
samples. In contrast, canonical correspondence analy- 
ses (CCA) are used to explicitly evaluate multivariate 
patterns of species biomass among sample sites. Es- 
sentially, CCA is a multivariate extension of multiple 
regression where species and sites are simultaneously 
ordinated in a manner that maximizes the variance 
related to a set of explanatory (constraining) variables 
(ter Braak, 1986). As with multiple regression, the 
inertia, or variance explained, by a given model can 
be determined and the significance of the explanatory 
variable tested (Legendre and Legendre, 1998). In in- 
stances where two or more sets of explanatory variables 
are of interest (in this case, month and depth), partial 
CCA can be employed to isolate the effect of each vari- 
able (Legendre and Legendre, 1998). The technique is 
analogous to partial regression, where the response 
variable (species biomass) is first constrained by one 
of the explanatory variables (either month or depth, 
expressed as factors with dummy variable coding). The 
resulting residuals are then constrained by the second 
explanatory variable. Effectively, the first explanatory 
variable is treated as a confounding variable and its 
effect is “cleansed” from the data set. The assemblage 
pattern related solely to the second explanatory vari- 
able can then be isolated and explored (Legendre and 
Legendre, 1998). An advantage of CCA over standard 
univariate tests of species biomass (e.g., ANOVA) is 
that the method simultaneously depicts the strength 
and direction of species responses to predictor variables 
by the position and spread of species in ordinate space. 
The approach therefore offers insights into species as- 
sociations that are not readily obtained by univariate 
methods (Legendre and Legendre, 1998). 
We applied partial CCA to the data set alternating 
depth and month as the confounding and explicit ex- 
planatory variables, respectively. We recognized that 
the habitat of most fishes changes with body size (Wer- 
ner and Gilliam, 1984) and therefore we divided species 
with abundant, small size classes (individuals less than 
30% of maximum recorded total length [TL] that oc- 
curred in at least 10% of the sites sampled) into small 
and large size classes (greater than 30% of maximum 
recorded TL were categorized as large) and we treated 
them as distinct species in the analysis. An exception 
was made for spiny dogfish, which possessed a bimodal 
size distribution that separated at approximately 500 
mm or 47% of the maximum recorded TL. 
Owing to the lack of samples from 20- and 80-m 
depths in March, we performed partial CCA (one each 
for depth and month), using 1) samples collected from 
all three months, but from 40 and 160 m only; and 2) 
for all four depths from October and July only for a 
total of four partial CCA tests. We identify the data 
included in each univariate and multivariate analysis 
by labeling tests as “all-months” or “October-t- July,” 
respectively. The analysis was split to avoid ambiguity 
that may have arisen from performing partial CCA on 
data lacking full treatment replication across factor 
levels (Anderson and Gribble, 1998). 
To increase the robustness of each CCA only species 
that occurred in at least 20% of the sites sampled were 
included. The variance explained by a global CCA model 
(month+depth) was used in conjunction with results 
from the partial CCA to identify variance components 
that were uniquely and jointly explained by each predic- 
tor (Borcard et al., 1992; Anderson and Gribble, 1998). 
The significance of each predictor in the global and 
