what's the backward selection default stopping rule in proc logistic?

[Xixi97]

If I don't specify the SLENTRY and SLSTAY options in code below, what's the backward selection default stopping rule in SAS?

proc logistic data = dataset;

model Y(event='1')=X /selection = backward;

run; 

[FreelanceReinh]

Hello @Xixi97 and welcome to the SAS Support Communities!

According to the documentation the default criterion is SLSTAY=0.05.

Let's check this:

/* Create sample data for demonstration */

data test;
input g r v n;
cards;
1 0 0 10000
1 0 1 10000
1 1 0 10000
1 1 1 10398
2 0 0 10000
2 0 1 10000
2 1 0 10000
2 1 1 10397
;

/* Fit the full model by group (g) for response r=1 with predictor v */

ods output parameterestimates=est;
proc logistic data=test desc;
by g;
freq n;
model r=v;
run;

proc print data=est;
run;

Result:

                                                                       Prob
Obs    g    Variable     DF    Estimate      StdErr     WaldChiSq     ChiSq    _ESTTYPE_

 1     1    Intercept     1    -1.28E-6      0.0141        0.0000    0.9999       MLE
 2     1    v             1      0.0390      0.0199        3.8446    0.0499       MLE
 3     2    Intercept     1    -1.27E-6      0.0141        0.0000    0.9999       MLE
 4     2    v             1      0.0389      0.0199        3.8256    0.0505       MLE

/* Use backward elimination with the default SLSTAY=0.05 */

ods output parameterestimates=estb;
proc logistic data=test desc;
by g;
freq n;
model r=v / selection=backward;
run;

proc print data=estb;
run;

Result:

                                                                       Prob
Obs    g    Variable     DF    Estimate      StdErr     WaldChiSq     ChiSq    _ESTTYPE_

 1     1    Intercept     1    -1.28E-6      0.0141        0.0000    0.9999       MLE
 2     1    v             1      0.0390      0.0199        3.8446    0.0499       MLE
 3     2    Intercept     1      0.0197     0.00995        3.9014    0.0482       MLE

As expected, only in group g=1 variable v met the selection criterion "ProbChiSq<=SLSTAY=0.05" and thus made it into the final model.