Wild Hog Rooting 
177 
Vegetation 
Analyzing the 1989 data with PROC REG and then discarding 
all variables with VIFs above 2.0 eliminated all but 39 of the 
original 97 variables. PROC STEPDISC retained six variables for 
inclusion in the final model. These six variables, in order of entry 
into the model, included percent total herb cover, percent ever- 
greenness, percent canopy cover, percent greater star chickweek, 
percent spring beauty, and percent exposed rock (Table 4). 
The combined 1989 data set failed to meet the assumption of 
homogeneity between within-group covariance matrices (P = 0.0001). 
PROC DISCRIM was programmed to use a pooled covariance ma- 
trix. The equality between within-group covariance matrices assump- 
tion is similar to the equal variance assumption of univariate analysis 
(Green 1979). For the 1989 data set, there was a significant differ- 
ence (P = 0.0001 for each) between the equality of within-group 
covariance matrices. Despite the failure of the data to meet the 
required assumption, we programmed PROC DISCRIM to use a pooled 
covariance matrix. 
Our line of reasoning for this approach was the following. 
First, the test used by PROC DISCRIM is extremely sensitive to 
nonnormality and rejects too often (Dr. Thomas Gerig, Department 
of Statistics, North Carolina State University, personal communica- 
tion). Second, many investigators believe that discriminant function 
analysis is robust and that the assumptions need not be strictly met 
(Johnson 1981, Taylor 1990). Third, we used linear discriminant 
analysis primarily as an exploratory tool and not a confirmatory 
tool (Williams 1983, James and McCulloch 1990). Fourth, our study 
had a relatively large number of observations and a large observa- 
tion to variable ratio. Data from both years consisted of 216 obser- 
vations (one set of observations per trap). During 1989, the ratio of 
observations to variables entered into the stepwise discrimination 
analysis was 30:1. In 1990, the ratio was 15:1. Rexstad et al. (1988) 
found that the median sample size of 28 multivariate studies pub- 
lished between 1985 and 1987 in The Journal of Wildlife Manage- 
ment was 99 observations and 12 variables, or an 8:1 ratio. Taylor 
(1990:188) stated that “relaxation of assumption is most justified 
with large data sets.” However, some authors do caution that the 
assumption is important and should not be dismissed lightly (Wil- 
liams 1983, Rexstad et al. 1990). 
Classification success using this method was fair. Of the 102 
unsuccessful traps, 66.7% were correctly classified as being unsuc- 
cessful. Of the 114 successful traps, 79.0% were correctly classi- 
fied. The Kappa statistic indicated that these classification rates were 
