206 
Fishery Bulletin 96(2), 1998 
C 
0 
E 
b 
tf shk 
Ar?§&1$ 
rd cj 
bts but 
cr 
bsjf 
sh 
A ^ 18 
Species 
ab - Atlantic bumper 
ac - Atlantic croaker 
acu - Atlantic cutlassfish 
bs - brown shrimp 
bts - blacktip shark 
he - hardhead catfish 
jf - cabbagehead jellyfish 
rd - red drum 
sh- unid. penaeid shrimp 
shk - unid. sharks 
3&14 
-i 
bsk - bull shark 
but - gulf butterfish 
cr - cownose ray 
cj - crevallejack 
gc - gafftopsail catfish 
sps - spotted seatrout 
ss - sand seatrout 
stm - striped mullet 
tf - unid. tonguefish 
Dimension 2 
Figure 7 
Correspondence analysis plot for species-area associations (spring) for 1995 fishing season. 
have ranged from 0.05% (Dunham, 1972) to 3.90% 
(Christmas et al., 1960) by number, and 1.0% 
(Condrey 1 ) to 2.80% (Christmas et al., 1960) by 
weight. However, these values are based on bycatch 
retained in the fish hold. No estimates of the releas- 
able bycatch are available. Based on our analysis, 
releasable bycatch estimates for the U.S. Gulf of 
Mexico menhaden fishery range from 0.033% (me- 
dian) to 0.17% (mean) and reflect the strong posi- 
tively skewed distribution of the bycatch. Values 
based on the winsorized mean are intermediate to 
those of the mean and median and are associated with 
a lower standard deviation than that for the mean. 
As a result of the patchy distribution of menhaden 
bycatch, examination of the relationship of bycatch 
to other factors is made more complex. Even after 
the transformation of our data, gross violations of 
the ANOVA model assumptions made it an unsatis- 
factory technique. Because we could not find a suit- 
able transformation, our solution to examining such 
data would be to convert the variable of interest into 
a categorical variable and to use categorical tech- 
niques in analyzing the data. In our case, the use of 
loglinear models to identify statistically important 
interactions was found to be a useful tool in explor- 
ing such data. This solution can be considered to fall 
between studies that can use ANOVA techniques (e.g. 
Andrew et al., 1995) and those based on the modi- 
fied negative binomial model as used by Perkins and 
Edwards ( 1996). 
Legendre (1987) noted that the responses of living 
organisms to environmental change is nonlinear and 
in instances nonmonotonic. As loglinear models are 
insensitive to the shape of the relationship among 
the variables, Legendre (1987) noted that they are 
well suited for examining nonmonotically related 
variables. A further advantage of this type of analy- 
sis is that because biological variables respond to 
interacting environmental variables, they can be used 
to examine such relationships in detail. By using 
loglinear models we can include a set of potential 
interactions in our saturated model, and through a 
stepwise selection procedure, find those interactions 
that are statistically important. In our study, bycatch 
was the issue of interest and the variable that we 
treated as a response. In effect, we were trying to 
