When using the best subset selection method (SELECTION=SCORE option) in PROC LOGISTIC, if the data in the candidate predictors are extremely ill-conditioned then the score chi-square statistics reported in the "Regression Models Selected by Score Criterion" table might be incorrect. The data are ill-conditioned if there are constant variables, if the variables have a very wide variation in scales, or if there are highly-collinear/redundant variables. The problem can be avoided by removing variables causing the ill-conditioning, and then using the SELECTION=SCORE method on the remaining variables. To do this, remove any constant variables, use PROC STANDARD to standardize the variables if there is a wide difference in scales, and remove any variables that are redundant as illustrated below.

This PROC STANDARD step standardizes all numeric candidate variables in data set MYDATA.

      proc standard data=mydata out=stddata m=0 s=1;
        var <all-candidate-variables>;
        run;

Run PROC LOGISTIC on the standardized data using all of the desired candidate variables. Include an ODS OUTPUT statement to save the table of parameter estimates.

      proc logistic data=stddata;
        model <response-variable> = <all-candidate-variables>;
        ods output parameterestimates=pe;
        run;

The following DATA step creates a macro variable (UseTerms) that contains the list of nonconstant and nonredundant candidate predictors.

      %let UseTerms=;
      data _null_; 
        set pe;
        if df=1 and Variable ne "Intercept" then
        call symput("UseTerms",catx(" ",symget("UseTerms"),Variable));
        run;

Using only the standardized variables in the UseTerms macro variable avoids any ill-conditioning in the PROC LOGISTIC step that implements the SELECTION=SCORE model selection method.

      proc logistic data=stddata;
        model <response-variable> = &UseTerms / selection=score;
        run;

Operating System and Release Information