Collinearity diagnostics for (mixed) multinomial models

Have you tried the check_collinearity() function in the performance
package? It's supposed to work on brms models, but whether it will work on
a multinomial model I don't know.  It works well on mixed models generated
by glmmTMB().

John Willoughby


On Fri, Feb 25, 2022 at 3:01 AM <r-sig-mixed-models-request at r-project.org>
wrote:

Send R-sig-mixed-models mailing list submissions to
        r-sig-mixed-models at r-project.org

To subscribe or unsubscribe via the World Wide Web, visit

would

like to see some (multi)collinearity diagnostics on the fixed effects, of
which there are over 30. My models are fit using the Bayesian *brms*
package because I know of no frequentist packages with multinomial GLMM
compatibility.

With continuous or dichotomous outcomes, my go-to function for

calculating

multicollinearity diagnostics is of course *vif()* from the *car*

package.

As expected, however, this function does not report sensible diagnostics
for multinomial models -- not even for standard ones fit by the *nnet*
package's *multinom()* function. The reason, I presume, is because a
multinomial model is not really one but C-1 regression models  (where C

is

the number of response categories) and the *vif()* function is not

designed

to deal with this scenario.

Therefore, in order to obtain meaningful collinearity metrics, my present
plan is to write a simple helper function that uses *vif() *to calculate
and present (generalized) variance inflation metrics for the C-1
sub-datasets to which the C-1 component binomial models of the overall
multinomial model are fit. In other words, it will partition the data

into

those C-1 subsets, and then apply *vif()* to as many linear regressions
using a made-up continuous response and the fixed effects of interest.

Does this seem like a sensible approach?

Best,

Juho

        [[alternative HTML version deleted]]