Harding et at: Regional and seasonal patterns of epipelagic fish assemblages from the central California Current 
265 
data we used the Bray-Curtis coefficient to construct 
resemblance matrices. The variance components and 
degrees of freedom of highly nonsignificant interaction 
terms in the full model were consolidated by sequen- 
tially pooling them with the residuals to generate the 
final reduced model. Regions and seasons were treated 
as fixed effects: the examination and testing of varia- 
tions in community structure between regions and sea- 
sons was the a priori objective of the study. Years were 
treated as random effects because there was no a priori 
reason for the timing or duration of the study: years have 
no particular meaning except to serve as replicates for 
the fixed effects of primary concern. Moreover, interan- 
nual patterns may be complicated by other sources of 
uncontrolled variation such as weather and sea state, 
or different ships and captains. 
To examine community patterns in finer detail (spe- 
cifically, among all four combinations of the two fixed 
factors [2 regionsx2 seasons]), we used a two-way 
crossed PERMANOVA type-III design with hauls aver- 
aged by station and season across all six years. This 
method of cumulating samples provided greater replica- 
tion and more degrees of freedom for each factor than 
was possible in the previous three-way arrangement. 
After this global test, pairwise comparisons were made 
between the two levels of each significant factor. 
We used nonmetric multidimensional scaling (MDS), 
an unconstrained ordination technique, to create graph- 
ical summaries of relationships among samples based 
on the abundance of the various species present and to 
highlight spatial and temporal patterns of community 
structure. Unlike PERMANOVA, MDS operates on the 
rank orders of the elements in the resemblance matrix, 
rather than on the resemblance matrix itself, and con- 
structs a map of the samples in a specified number of 
dimensions. The axes in MDS plots have no meaning 
other than for orientation, and scaling in MDS plots, if 
shown, is arbitrary. A stress value ranging from 0 to 
1.0 is used to measure the reliability of the ordination, 
with zero indicating a perfect fit and all rank orders 
correctly represented by the relative distance between 
all pairs of points in the graph, and with values >0.3 
indicating that points are close to being arbitrarily 
placed in the graph (Clarke and Warwick, 2001). 
Where group differences in community structure 
were found (a<0.05 in PERMANOVA tests), we used 
another exploratory method to identify those species 
most responsible for the difference. For any two groups, 
SIMPER (similarity percentages) calculates the percent 
contribution each species makes to the total between- 
group dissimilarity (Clarke and Warwick, 2001). SIM- 
PER identifies a small subset of species that are more 
consistently present or more abundant in one group 
than another, thus helping to reveal the major con- 
tributors to each group’s biotic identity and simplifying 
the interpretation of community patterns. Because the 
routine incorporates both abundance and frequency of 
occurrence, it allows species at low abundance to be 
major contributors to community patterns if they are 
consistently present in one place or time. 
Water properties: tests and ordinations 
To match the approach used for the community analy- 
sis, we also ran multivariate tests for group differences 
in environmental structure using PERMANOVA and 
graphically summarized the relationships among hydro- 
graphic samples using ordination — in this case principal 
components analysis (PGA). In order to make direct 
comparisons with the biological patterns, oceanographic 
variables were grouped, averaged, and plotted in the 
same arrangements as those used for hauls in the spe- 
cies analysis. Because we were interested in examining 
water properties only as they relate to biotic patterns, 
the starting point for the PERMANOVA environmental 
model was the reduced model from the biotic analysis, 
with all three main effects included and all interac- 
tion terms pooled with residuals. Four variables with 
skewed distributions required transformation before 
PCA, and appropriate transformations were selected by 
using log-likelihood profiles of Box-Cox transformations. 
A square-root transformation was used for DIS and 
PAR, and a log 10 (x) transformation for DEP and CHL. 
Environmental variables were normalized before PCA 
analysis, and Euclidian distance was used to measure 
sample similarity. 
Oceanographic variables were also individually tested 
for regional and seasonal differences by using uni- 
variate analysis of variance (ANOVA) type-II tests. The 
same transformations described above were applied. 
After transformation, the distributions of all environ- 
mental variables met the requirements necessary for 
parametric testing and ANOVA was an appropriate 
choice in this instance. 
The degree of similarity between corresponding spe- 
cies and environmental patterns was measured by us- 
ing a matrix-matching permutation test (the BIO-ENV 
routine in PRIMER). With this procedure, biotic and 
abiotic samples are compared from matching locations 
and a subset of water properties are determined that 
maximize their correlation to the community pattern 
(Clarke and Warwick, 2001). To accomplish this goal, 
the elements of the two corresponding sample similar- 
ity matrices (Bray-Curtis for biotic and normalized 
Euclidian distance for abiotic) are ranked, the ranks 
are ordered by sample number or location, and the 
two matching sets of ordered ranks are compared by 
calculating a correlation coefficient — in this case the 
familiar Spearman coefficient (p s ). The significance 
level of the match is determined by comparing the ob- 
served value of p s to a large set of p s values generated 
by repeated random reassignment of sample labels in 
one of the two similarity matrices. Only samples that 
are jointly present in both matrices are considered in 
the test. 
Results 
We caught a total of 53 different species of fish during 
this study, of which a few common mid-trophic level 
