covariates selection and multicollinearity in repeated measures applying proc mixed

[fatemeh]

Hello all,

I really appreciate you to help me find a way to check for covariates selection and multicollinearity like VIF or any other way in repeated measures applying proc mixed with fixed effects and two repeated effects in sas9.4.

[STAT_Kathleen]

There is no VIF option in PROC GLIMMIX or PROC MIXED for multicollinearity diagnostics. There is an INFULENCE option in PROC MIXED for influence observations diagnostics, but no multicollinearity diagnostics options.

[fatemeh]

Thank you for your reply. I found this article related to this topic, "User-friendly SAS® Macro Application for performing all possible mixed model selection - An update", but i can not download the ALLMIXED2 .SAS macro-call file from the authors website at "http://www.ag.unr.edu/gf". Any help will be appreciated to help me find access to ALLMIXED2 in sas 9.4.

[PaigeMiller]

But influence diagnostics are not the same as multicollinearity diagnostics. The original poster needs to run the model through PROC REG to get the multicollinearity (VIF) diagnostics, then (assuming there are no serious problems indicated by the VIF) run PROC MIXED for the actual analysis.

[Ksharp]

You could try CORRB option of MODEL .
CORRB Displays correlation matrix of fixed-effects parameter estimates

if correlation is very big like >0.8 ,that stand for these two variables is high collinearity .

[PaigeMiller]

Yes, CORRB will detect certain types of multicollinearity. However other types of multicollinearity will not be detected via CORRB.

[SteveDenham]

To add to @PaigeMiller 's response, CORRB would enable you to say something about the collinearity between variable X1 and variable X2, but not about the collinearity between X1 and a linear combination of the other X's.

The point I would add is that variable selection is a tricky subject, even for a fixed effects model.  For a mixed model, it is even more problematic.  What you might consider is to use a LASSO based method, treating all factors as fixed during the selection, and then denoting as random those that appropriately define a broader inference space.

SteveDenham