The HPREDUCE Procedure

REDUCE Statement

REDUCE UNSUPERVISED effects </ reduce-options>;

REDUCE SUPERVISED response <response> = effects </ reduce-options>;

PROC HPREDUCE can be used for both supervised and unsupervised variable selection. In unsupervised case, the REDUCE statement specifies the effects to be considered in the variable selection process. An effect can be an original variable in the input data set or a variable constructed from the original variables. In the supervised case, you need to specify both the effects and the response variables. A response variable can be an original variable in the input data set or a variable constructed from the original variables. For the regression case, you can specify more than one response variable.

Table 2 summarizes the reduce-options, which control the number of variables to be selected.

Table 2: reduce-options

Option	Description
AIC	Performs model selection by using Akaike’s information criterion
AICC	Performs model selection by using the corrected Akaike’s information criterion
BIC	Perform model selection by using Schwarz Bayesian information criterion
MAXSTEPS=	Specifies the maximum number of steps to take; the number must be greater than or equal to 1
MAXEFFECTS=	Specifies the number of effects to select; the number must be greater than or equal to 1.
VARIANCEEXPLAINED \| VAREXP=	Specifies the fraction of the total variance to be explained; the value must be between 0 and 1.
MINVARIANCEINCREMENT \| VARINC=	Specifies the minimum increment of explained variance (in a fraction of the total variance); the value must be between 0 and 1.

The reduce-options determine the number of variables to be selected. You can specify the following reduce-options as stopping criteria for the HPREDUCE procedure. When you specify more than one option, PROC HPREDUCE stops whenever one of the specified options is satisfied, or when the explained variance equals to the total variance. In the latter case, the procedure prints the following message in the log: "Early stop: the proportion of the explained variance to the total variance equals 1."

AIC: stops PROC HPREDUCE if the Akaike’s information criterion (AIC) value fails to decrease in three contiguous steps.
AICC: stops PROC HPREDUCE if the corrected Akaike’s information (AICC) value fails to decrease in three contiguous steps.
BIC: stops PROC HPREDUCE if the Schwarz Bayesian information criterion (BIC) value fails to decrease in three contiguous steps.
MAXSTEPS=n: stops PROC HPREDUCE after it runs n steps.
MAXEFFECTS=n: stops PROC HPREDUCE after n effects have been selected. Because individual levels of one classification variable can be selected in different steps of the variable selection process, PROC HPREDUCE might take more than n steps to select n effects.
VARIANCEEXPLAINED=fraction VAREXP=fraction: stops PROC HPREDUCE when the fraction of the total variance can be explained by the selected variables.
MINVARIANCEINCREMENT=fraction VARINC=fraction: stops PROC HPREDUCE when the minimum increment of the explained variance is less than fraction of the total variance.

Last updated: May 25, 2022