510 
Fishery Bulletin 95(3), 1 997 
these variables also had the largest rank correlations 
with abundance (Table 1), they were likely to be the 
most important parameters for flathead sole distri- 
bution. The error rates for predicting absence of flat- 
head sole were consistently much lower than those 
for predicting presence. 
Pacific halibut presence or absence could be most 
accurately predicted by using either depth or tem- 
perature with either distance or sand. Halibut abun- 
dance had a higher rank correlation with tempera- 
ture than with depth and a higher correlation with 
sand than with distance from the mouth of the bay 
(Table 1). It is difficult to evaluate the relative im- 
portance of depth and temperature and of sand and 
distance owing to high correlations among these vari- 
ables (Fig. 3). The depth-temperature factor ex- 
plained most of the observed distribution. The error 
rates for predicting presence or absence changed sig- 
nificantly only if both depth and temperature were 
excluded. Error rates for stations where Pacific hali- 
but were present were consistently much lower than 
those for stations where no halibut were found, thus 
this species appears to be strongly associated with 
specific habitat characteristics. 
The three best predictors for yellowfin sole were 
depth and gravel combined with either sand or tem- 
perature. Of these, depth and gravel resulted in the 
lowest total error rates. Only depth was significantly 
correlated with yellowfin sole abundance (Table 1). 
The sediment parameters added very little informa- 
tion because yellowfin sole occurred over a wide range 
of substrate types. Error rates for stations where 
yellowfin sole were present were much lower than 
those for stations where this species was absent, re- 
flecting the restricted depth range within which yel- 
lowfin sole were encountered. Presence and absence 
patterns for all four species are plotted against the 
two best discriminator variables (Fig. 2). 
Regression trees were constructed by using CPUE 
for each species to refine our habitat models. The 
initial trees were allowed to grow, provided the num- 
ber of stations in a node was five or greater. The re- 
sultant regression trees had sizes of 22 terminal 
nodes for rock sole, 16 for flathead sole, 19 for Pa- 
cific halibut, and 18 for yellowfin sole. The total 
deviances for the initial trees were 1.24, 0.57, 0.38, 
and 0.44 respectively, indicating that the model fit- 
ted for rock sole was much poorer than that for the 
other species and that the tree for Pacific halibut 
had the best fit. 
The trees for all species seemed to overfit the data, 
as indicated by cross-validation. Plots of deviance 
against tree size (number of terminal nodes) for the 
four flatfish species indicated that deviance was usu- 
ally at a minimum at very small tree sizes, consist- 
ing of only two or three nodes (Fig. 4). The deviance 
for each species tended to increase steeply at a tree 
size between 4 and 6 nodes, and we chose the largest 
size before a steep increase as optimum size for the 
tree. The initial tree was pruned to six terminal nodes 
for rock sole and halibut and to four terminal nodes 
for flathead sole and yellowfin sole. 
The pruned regression tree for rock sole indicated 
that sediment, depth, and temperature were the best 
predictor variables for rock sole CPUE. The deviance 
of the pruned tree increased to 1.852 from 1.242 for 
the initial tree. This relatively poor fit may again be 
due to the widespread distribution of rock sole, a 
species that does not seem to be limited to any par- 
ticular habitat type. Stations were first separated 
by sediment type into 89 stations on sand or muddy 
sand with a high mean CPUE (18 fish/10-min tow) 
and 80 stations on other sediment types that had a 
much lower mean CPUE (1.6 fish/10-min tow) (Fig. 
5). The highest mean CPUE (25 fish/10-min tow) oc- 
curred at stations on sand or muddy sand which had 
a bottom temperature of more than 8.7°C. The colder 
stations on sand and muddy sand were separated 
into seven low salinity stations with low mean CPUE 
(0.58 fish/10-min tow) and 10 high salinity stations 
with medium to high CPUE (11 fish/10-min tow). 
Most stations on other sediment types, which in- 
cluded gravel, mud, gravelly mud, gravelly sand, 
gravelly muddy sand, gravelly sandy mud, muddy 
gravel, muddy sandy gravel, sandy gravel, and sandy 
mud, had low CPUE values except for a group in 
shallow water (<27.5 m) on gravelly muddy sand, 
sandy gravel, or sandy mud (13 fish/10-min tow). 
Thus, by combining results from the correlation 
analysis, presence and absence patterns, and regres- 
sion trees, rock sole were found to be most common 
on sand or mixed sand substrates and most abun- 
dant in shallow and relatively warm water. 
The regression tree for flathead sole indicated that 
temperature, sediment type, and depth were the best 
predictors of flathead sole abundance. The deviance 
of the pruned tree was 0.774 compared with 0.569 
for the initial tree. Highest CPUE values tended to 
occur at stations where bottom temperature was less 
than 8.9°C (Fig. 6). At warmer stations, mean CPUE 
of flathead sole was very low (0.17 fish/10-min tow) 
if stations were less than 48 m deep, which was the 
case for the majority (n=109) of the stations. Mean 
CPUE at warm stations was higher, however, for the 
six stations located in water deeper than 48 m (4.6 
fish/10-min tow). Stations with bottom temperatures 
below 8.9°C had a low flathead sole CPUE if the sedi- 
ment was categorized as gravel, sand, muddy sandy 
gravel, or sandy gravel (1.6 fish/10-min tow). The 
CPUE was much higher on pure mud or mixed mud 
