HPGENSELECT/PROC GENMOD - GEE Stepwise

[cxc810]

Does SAS have a stepwise process for the generalized estimating equations (GEE)? Like does HPGENSELECT or PROC GENMOD have a stepwise option for a GEE model?

If not, is there a reason why there is no option for this?

Thanks in advance! 

[SteveDenham]

Read through the examples in the HPGENSELECT documentation.  The first example shows a forward selection method.  The documentation for the SELECTION statement goes through the METHOD= options (none, forward, backward, lasso, stepwise).  So it's out there, you just have to look for it.

SteveDenham

[MichaelL_SAS · Accepted solution]

Unless data splitting is used (that is one set of data are used to select the model and a different set are used to fit the selected model and perform inference), traditional inferential methods will not be valid due to the post-selection inference problem.

On the other hand, the main appeal of GEE models is typically related to inference for longitudinal or clustered data. In particular, GEE models treat the correlation between associated observations as a nuisance parameter and (under appropriate regularity conditions), they will provide consistent parameter estimates even when the working correlation structure is misspecified. Moreover, inference for those parameter estimates can be performed using the empirical/sandwich covariance matrix estimate.

I believe it is because all these (inferential) asymptotic properties of GEE models break down in the presence of post-selection inference that PROC GENMOD does not support a stepwise option for GEE models and PROC HPGENSELECT does not support a REPEATED statement.

[SteveDenham]

Thanks @MichaelL_SAS  - a lot is explained that way.  I hope that helps the OP as much as it taught me.

SteveDenham