Use this field to specify a logical expression for subsetting the
predictor class variables
. This is useful when you want to apply a statistical filter to reduce the initial number of candidate predictor variables passed to subsequent stages of
predictive model
building.
Class variables
are coded using 0 and 1 values, so filtering criteria are usually in terms of proportions. The expression must be a valid SAS
WHERE clause
, and it is applied separately to each predictor
class variable
.
For example, to filter only diseased individuals from a data set containing a mixed
population
of diseased (sick) and healthy (healthy) individuals (as indicated in a column named
DiseaseStatus
), you could use the following simple WHERE expression:
Note:
the word
where
has already been entered for you.
You can use one or more of the following statistical keywords in your expression:
CSS CV IQR KURTOSIS MAD MAX MEAN MEDIAN MIN NMISS SKEWNESS STD STDERR SUM USS VAR
.
For example, specifying
MEAN > 7 and VAR > 3
keeps only those predictor variables whose
mean
value across
observations
in the input data set is greater than 7 and whose
variance
is greater than 3.
You can also use the keyword _NAME_ to filter by predictor variable name. For example, specifying
substr(_NAME_,1,3) ne "af_"
removes all predictors whose name begins with
"af_"
.
When you have specified a test data set or when running Cross Validation Model Comparison, the WHERE clause is applied only to the
training data set
, and then all predictors not satisfying the expression are dropped from both the training and test data sets.
Refer to
Definitions of Functions and CALL Routines
for details about the aforementioned statistics.