The HPSPLIT Procedure

Observations for which the response variable is missing are omitted from the analysis. The HPSPLIT procedure provides various methods of handling missing values of predictor variables.

By default, observations for which predictor variables are missing are omitted from the analysis. This behavior is common to other statistical modeling procedures in SAS/STAT software. Alternatively, you can use the ASSIGNMISSING= option to request different methods of dealing with missing values of predictor variables.

The ASSIGNMISSING=BRANCH option creates an extra branch for missing values. You determine the maximum number of branches by using the MAXBRANCH= option, which specifies the maximum number of branches per node in the tree. By default, MAXBRANCH=2. The ASSIGNMISSING=BRANCH option has no effect if there are no missing values in the training data set for a particular split. However, even if this is not the case, there could be missing values in the data set that is used for scoring, so this option is used to assign missing values that could be encountered in the future. For more information, see the section Scoring.

The ASSIGNMISSING=POPULAR option assigns missing values to the most popular node of a split. If two or more branches have the same maximum number of observations, then the missing values are assigned to the branch that has the lowest node index.

The ASSIGNMISSING=SIMILAR option assigns missing values to the most similar node. Similarity is calculated using a chi-square test for categorical response variables and an F test for continuous response variables. The ASSIGNMISSING=SIMILAR option has no effect if there are no missing values in the training data set for a certain split. To handle this case, PROC HPSPLIT assigns future missing values to the most popular node of a split.