Modeling — Variable Selection

Gather all variables. If you want to add new variables, do it now.

Run WOE binning. Get WOE value of each bin and Information Value (IV)of each variable.

Check if any variables with IV=0. Excludes them.

Throw the rest variables to Regression with Stepwise Selection.

proc logistic data=Remission DESC;
   model target(event='1')= var1 var2 var3
                / selection=stepwise
                  slentry=0.1
                  slstay=0.05
                  details
                  lackfit;

Here comes the fun part.

A ideal variable set for Scorecard should have 10–20 variables. Not too many.

Find a set and work two ways: Size Up and Size Down.

Size up: add ONE new variable each time and see if it improves the concordance.

Size down: remove variables with large P value and small Chi-squre.

Refer to previous models for inspiration.

Credit: Shan Lin & my dear co-workers