Skip to content
Prev 19701 / 20628 Next

Collinearity diagnostics for (mixed) multinomial models

Dear John (Fox, and other list members),

Could we achieve the invariance Professor Fox refers to by altering my
initial approach in the following ways:

   1. Instead of fitting a *linear* model with a mock continuous response
   to each subdataset and applying *vif() *to that, we fit to each
   subdataset an actual binary GLM with a real pairing of two response
   categories as LHS.
   2. Instead of fitting only C-1 binary sub-GLMs for which to calculate
   GVIF diagnostics, we fit *choose(C, 2) *such submodels, i.e. *one for
   every possible pairing of response categories*.
   3. Finally, for each coefficient, we average the GVIF statistic
over the *choose(C,
   2)* submodels in order to obtain a summary statistic.

What do you gentlemen think?

Best,

Juho



pe 25. helmik. 2022 klo 19.40 John Fox (jfox at mcmaster.ca) kirjoitti:

  
  
Message-ID: <CAG_dBVe+Ovcae-RoDWRz8cDJLTQx3aLZ9ZZ8Debkz0V3xud5Zw@mail.gmail.com>
In-Reply-To: <63b44aed-cfcf-cfc3-99ba-fc7e7c85a0fb@mcmaster.ca>