BookmarkSubscribeRSS Feed
🔒 This topic is solved and locked. Need further help from the community? Please sign in and ask a new question.
McVeneg
Calcite | Level 5

Hello,

Is it possible to do a stepwise or backward selection model with proc glimmix.

I know it is possible to do that with proc logistic. But I need to have a random effect in the model.

I have more than 50 independent variables to create the model. It will be nice to create this code as a help to create the final model.

Thank you very much for your help

1 ACCEPTED SOLUTION

Accepted Solutions
SteveDenham
Jade | Level 19

Yes, conditional and marginal interpretations are different. The drawback to the GEE approach seems to lie primarily in the area of covariance structure selection, using tools such as change in log likelihood or information criteria.  Because the marginal approach uses pseudo-likelihoods, the "data" are not the same under different covariance structures.  In contrast, the conditional approach uses a quasi-likelihood approach, so that information criteria can be used to select a best fitting covariance structure.

 

Steve Denham

View solution in original post

9 REPLIES 9
SteveDenham
Jade | Level 19

You cannot do model selection with glimmix.  It is something I hope is never added to any of the mixed model methods as it leads to biased results, and one of the biggest things to worry about in generalized mixed models is random specification to avoid bias to begin with.

If you want to do variable reduction, look into PROC VARCLUS or PLS or even GLMSELECT(using LAR or LASSO), and then fit the resulting fixed effects to the mixed model.

Steve Denham

plf515
Lapis Lazuli | Level 10

Stepwise methods don't give correct answers; nor do backwards, nor forwards. This is true even for OLS regression. The problem is probably compounded for GLIMMIX.

I, like Steve, am glad GLIMMIX doesn't let you do this. I wish REG and GLM didn't ....

Using STEPWISE is like telling your boss "please don't give me a raise"

SteveDenham
Jade | Level 19

And today I get an email from SAS advertising PROC HPGENSELECT, so that in SAS/STAT12.3 and 13.1 people can muck about (only selection methods are backwards, forwards and stepwise) in generalized linear models that are generally already biased.

I can hardly wait to see what some brokerage wonks come up with...

Steve Denham

Haris
Lapis Lazuli | Level 10

Could you expand a bit more on the GLM being 'generally biased.'  A reference source would be great!

Thanks,

Haris    

SteveDenham
Jade | Level 19

Read Walt Stroup's Generalized Linear Mixed Models to see how, especially in smaller datasets, marginal models are biased compared to conditional models.  Naive fitting using pseudo-likelihood methods for distributions without a free scale parameter would be the most likely candidates.

Steve Denham

Gokirop
Fluorite | Level 6

Hi Steve,

Would it be a problem for you to offer an intuitive explanation why this happens?

Thanks.

Marin

SteveDenham
Jade | Level 19

This is the way I think of what Stroup presents: Marginal models are averaged over the random effects (or repeated effects).  Consequently, the estimates will regress towards the mean.  Does that help any?

 

Steve Denham

Gokirop
Fluorite | Level 6

Hi Steve, 

Thank you. So you mean that interpretations differ? I mean, if you are looking for a population averaged (not individual) response, GEE should be the way to go, wouldn't it? And I can't really see how this relates to variable selection in the mixed models (to the extent it would be a bigger problem than usual)?

Thanks again.

Marin

SteveDenham
Jade | Level 19

Yes, conditional and marginal interpretations are different. The drawback to the GEE approach seems to lie primarily in the area of covariance structure selection, using tools such as change in log likelihood or information criteria.  Because the marginal approach uses pseudo-likelihoods, the "data" are not the same under different covariance structures.  In contrast, the conditional approach uses a quasi-likelihood approach, so that information criteria can be used to select a best fitting covariance structure.

 

Steve Denham

sas-innovate-2024.png

Join us for SAS Innovate April 16-19 at the Aria in Las Vegas. Bring the team and save big with our group pricing for a limited time only.

Pre-conference courses and tutorials are filling up fast and are always a sellout. Register today to reserve your seat.

 

Register now!

What is ANOVA?

ANOVA, or Analysis Of Variance, is used to compare the averages or means of two or more populations to better understand how they differ. Watch this tutorial for more.

Find more tutorials on the SAS Users YouTube channel.

Discussion stats
  • 9 replies
  • 10432 views
  • 7 likes
  • 5 in conversation