Getting Started: DISCRIM Procedure
The data in this example are measurements of 159 fish caught in Finland’s lake Laengelmavesi; this data set is available from the Puranen. For each of the seven species (bream, roach, whitefish, parkki, perch, pike, and smelt) the weight, length, height, and width of each fish are tallied. Three different length measurements are recorded: from the nose of the fish to the beginning of its tail, from the nose to the notch of its tail, and from the nose to the end of its tail. The height and width are recorded as percentages of the third length variable. The fish data set is available from the Sashelp library. The goal now is to find a discriminant function based on these six variables that best classifies the fish into species.
First, assume that the data are normally distributed within each group with equal covariances across groups. The following statements use PROC DISCRIM to analyze the Sashelp.Fish data and create Figure 32.1 through Figure 32.5:
title 'Fish Measurement Data';
proc discrim data=sashelp.fish;
class Species;
run;
The DISCRIM procedure begins by displaying summary information about the variables in the analysis (see Figure 32.1). This information includes the number of observations, the number of quantitative variables in the analysis (specified with the VAR statement), and the number of classes in the classification variable (specified with the CLASS statement). The frequency of each class, its weight, the proportion of the total sample, and the prior probability are also displayed. Equal priors are assigned by default.
Figure 32.1
Summary Information
Bream |
34 |
34.0000 |
0.215190 |
0.142857 |
Parkki |
11 |
11.0000 |
0.069620 |
0.142857 |
Perch |
56 |
56.0000 |
0.354430 |
0.142857 |
Pike |
17 |
17.0000 |
0.107595 |
0.142857 |
Roach |
20 |
20.0000 |
0.126582 |
0.142857 |
Smelt |
14 |
14.0000 |
0.088608 |
0.142857 |
Whitefish |
6 |
6.0000 |
0.037975 |
0.142857 |
The natural log of the determinant of the pooled covariance matrix is displayed in Figure 32.2.
Figure 32.2
Pooled Covariance Matrix Information
The squared distances between the classes are shown in Figure 32.3.
Figure 32.3
Squared Distances
The DISCRIM Procedure
0 |
83.32523 |
243.66688 |
310.52333 |
133.06721 |
252.75503 |
132.05820 |
83.32523 |
0 |
57.09760 |
174.20918 |
27.00096 |
60.52076 |
26.54855 |
243.66688 |
57.09760 |
0 |
101.06791 |
29.21632 |
29.26806 |
20.43791 |
310.52333 |
174.20918 |
101.06791 |
0 |
92.40876 |
127.82177 |
99.90673 |
133.06721 |
27.00096 |
29.21632 |
92.40876 |
0 |
33.84280 |
6.31997 |
252.75503 |
60.52076 |
29.26806 |
127.82177 |
33.84280 |
0 |
46.37326 |
132.05820 |
26.54855 |
20.43791 |
99.90673 |
6.31997 |
46.37326 |
0 |
The coefficients of the linear discriminant function are displayed (in Figure 32.4) with the default options METHOD=NORMAL and POOL=YES.
Figure 32.4
Linear Discriminant Function
-185.91682 |
-64.92517 |
-48.68009 |
-148.06402 |
-62.65963 |
-19.70401 |
-67.44603 |
-0.10912 |
-0.09031 |
-0.09418 |
-0.13805 |
-0.09901 |
-0.05778 |
-0.09948 |
-23.02273 |
-13.64180 |
-19.45368 |
-20.92442 |
-14.63635 |
-4.09257 |
-22.57117 |
-26.70692 |
-5.38195 |
17.33061 |
6.19887 |
-7.47195 |
-3.63996 |
3.83450 |
50.55780 |
20.89531 |
5.25993 |
22.94989 |
25.00702 |
10.60171 |
21.12638 |
13.91638 |
8.44567 |
-1.42833 |
-8.99687 |
-0.26083 |
-1.84569 |
0.64957 |
-23.71895 |
-13.38592 |
1.32749 |
-9.13410 |
-3.74542 |
-3.43630 |
-2.52442 |
A summary of how the discriminant function classifies the data used to develop the function is displayed last. In Figure 32.5, you see that only three of the observations are misclassified. The error-count estimates give the proportion of misclassified observations in each group. Since you are classifying the same data that are used to derive the discriminant function, these error-count estimates are biased.
Figure 32.5
Resubstitution Misclassification Summary
The DISCRIM Procedure
Classification Summary for Calibration Data: SASHELP.FISH
Resubstitution Summary using Linear Discriminant Function
0.0000 |
0.0000 |
0.0536 |
0.0000 |
0.0000 |
0.0000 |
0.0000 |
0.0077 |
0.1429 |
0.1429 |
0.1429 |
0.1429 |
0.1429 |
0.1429 |
0.1429 |
|
One way to reduce the bias of the error-count estimates is to split your data into two sets. One set is used to derive the discriminant function, and the other set is used to run validation tests. Example 32.4 shows how to analyze a test data set. Another method of reducing bias is to classify each observation by using a discriminant function computed from all of the other observations; this method is invoked with the CROSSVALIDATE option.