Given a classification variable and several quantitative variables, the STEPDISC procedure performs a stepwise discriminant analysis to select a subset of the quantitative variables for use in discriminating among the classes. The set of variables that make up each class is assumed to be multivariate normal with a common covariance matrix. The STEPDISC procedure can use forward selection, backward elimination, or stepwise selection (Klecka, 1980). The STEPDISC procedure is a useful prelude to further analyses with the CANDISC procedure or the DISCRIM procedure.
With PROC STEPDISC, variables are chosen to enter or leave the model according to one of two criteria:
the significance level of an F test from an analysis of covariance, where the variables already chosen act as covariates and the variable under consideration is the dependent variable
the squared partial correlation for predicting the variable under consideration from the CLASS variable, controlling for the effects of the variables already selected for the model
Forward selection begins with no variables in the model. At each step, PROC STEPDISC enters the variable that contributes most to the discriminatory power of the model as measured by Wilks’ lambda, the likelihood ratio criterion. When none of the unselected variables meet the entry criterion, the forward selection process stops.
Backward elimination begins with all variables in the model except those that are linearly dependent on previous variables in the VAR statement. At each step, the variable that contributes least to the discriminatory power of the model as measured by Wilks’ lambda is removed. When all remaining variables meet the criterion to stay in the model, the backward elimination process stops.
Stepwise selection begins, like forward selection, with no variables in the model. At each step, the model is examined. If the variable in the model that contributes least to the discriminatory power of the model as measured by Wilks’ lambda fails to meet the criterion to stay, then that variable is removed. Otherwise, the variable not in the model that contributes most to the discriminatory power of the model is entered. When all variables in the model meet the criterion to stay and none of the other variables meet the criterion to enter, the stepwise selection process stops. Stepwise selection is the default method of variable selection.
It is important to realize that, in the selection of variables for entry, only one variable can be entered into the model at each step. The selection process does not take into account the relationships between variables that have not yet been selected. Thus, some important variables could be excluded in the process. Also, Wilks’ lambda might not be the best measure of discriminatory power for your application. However, if you use PROC STEPDISC carefully, in combination with your knowledge of the data and careful cross validation, it can be a valuable aid in selecting variables for a discrimination model.
As with any stepwise procedure, it is important to remember that when many significance tests are performed, each at a level of, for example, 5% (0.05), the overall probability of rejecting at least one true null hypothesis is much larger than 5%. If you want to prevent including any variables that do not contribute to the discriminatory power of the model in the population, you should specify a very small significance level. In most applications, all variables considered have some discriminatory power, however small. To choose the model that provides the best discrimination by using the sample estimates, you need only to guard against estimating more parameters than can be reliably estimated with the given sample size.
Costanza and Afifi (1979) use Monte Carlo studies to compare alternative stopping rules that can be used with the forward selection method in the two-group multivariate normal classification problem. Five different numbers of variables, ranging from 10 to 30, are considered in the studies. The comparison is based on conditional and estimated unconditional probabilities of correct classification. They conclude that the use of a moderate significance level, in the range of 10 to 25 percent, often performs better than the use of a much larger or a much smaller significance level.
The significance level and the squared partial correlation criteria select variables in the same order, although they might select different numbers of variables. Increasing the sample size tends to increase the number of variables selected when you are using significance levels, but it has little effect on the number selected by using squared partial correlations.
See Chapter 10: Introduction to Discriminant Procedures, for more information about discriminant analysis.