How to do GLMSelect without picking up non-significant variables?

[Davejones]

Hi Experts,

Proc GLMselect model is based on AIC. however, it occasionally picks up non-significant variable in the final Parameter Estimates table. Is a better way to improve the "stepwise" selection method instead of pre-selecting the "p<0.05" variables? 

Thanks

[sbxkoenk]

Hello,

PROC GLMSELECT will occasionally select an input variable whose parameter estimate has a p-value above 5%.
Surely this is not a bad thing at all. If it improves the overall AIC(C), that's OK anyway.

In my opinion, there is no way to avoid this. You can only play (i.e. trial-and-error) with

Criteria Used in Model Selection Methods (the CHOOSE=, SELECT=, and STOP= options in the MODEL).

One or more settings will probably produce a model of which all parameters are significant according to the 5% significance level.

SAS/STAT® 15.3 User's Guide
The GLMSELECT Procedure
Criteria Used in Model Selection Methods

https://go.documentation.sas.com/doc/en/statug/15.3/statug_glmselect_details15.htm

Good luck,

Koen

[Davejones]

Thanks Koen for your help!

As p>n, tried to use both Lasso and elasticnet, and got inconsistent outputs too.

[sbxkoenk]

Hello,

If the number of variables / features is higher than the number of observations / samples, the data are considered high-dimensional and require dimension reduction approaches.

Koen

[Davejones · Accepted solution]

Thanks Koen!