de Silva and Condrey: Patterns in patchy data discerned from Brevoortia patronus bycatch 
199 
Menhaden catch (std menhaden) 
Figure 2 
Distribution of menhaden catch sampled during the 1995 gulf menhaden fishing season, std = standard 
(1,000 standard menhaden ~ 305 kg). 
The number of by catch observed in each set ranged 
from 0 to 1,600 organisms, with a median, mean, and 
standard deviation of 15, 61, and 153 respectively. 
The winsorized mean and its standard deviation val- 
ues were 53 and 6.9 respectively. The 95% confidence 
interval of the mean was between 41.8 and 79.3. The 
distribution of bycatch organisms was strongly posi- 
tively skewed (5.9) and peaked sharply with a Kur- 
tosis value of 47.2 (Fig. 3). 
The bycatch percentage ranged from 0% to 4% with 
a median, mean, and standard deviation of 0.033%, 
0.168%, and 0.48 respectively. The winsorized mean 
bycatch percentage and its standard deviation were 
0.14% and 0.02, respectively. The 95% confidence in- 
terval of the mean was between 0.11% and 0.22%. 
The distribution of the bycatch percentage was also 
found to be positively skewed and strongly peaked, 
with skewness and kurtosis values of 5.4 and 32.7 
respectively (Fig. 4). 
Analysis of variance with bycatch, bycatch percent- 
age, and their respective transformations, together 
with the arcsine transformation, did not meet model 
assumptions. In all cases the modified Levene’s test 
indicated that the variances were nonhomogeneous, 
and both the residual plots and Shapiro-Wilk test 
indicated that the assumption of normality of residu- 
als were not met. For example, for the response log 
(bycatch percentage +1), the residuals of the model 
were not normally distributed (Shapiro-Wilk 
W=0.678, P<W=0.0001), and the residual plot indi- 
cated nonhomogenous variances. Furthermore, a 
modified Levine’s test indicated that the variances 
were nonhomogeneous (F=6.21, df=ll, df-error=245, 
P>F=0.0001). These characteristics suggest that the 
model assumptions were grossly violated and that 
ANOVA may not be an appropriate form of analysis 
in this case. 
Spatial and temporal patterns in bycatch 
Exploratory analysis using Sogls near models With 
the backward selection procedure, loglinear models 
[SAB SAD DB] and [SAB SAD] (as defined in Table 
2 ) satisfied the criteria for a logit model and had good 
fit (Table 2). The simplest of these models, [SAB SAD], 
was selected; this loglinear model corresponds to the 
logit model with categorical explanatory variables 
of the form 
l 0g ^*^ = a + A A +A s +Af, U) 
^ low\ik 
