www.cmm.bristol.ac.uk/forum

Hi all,
I am bothering you because I have a question about the way you interpret outcomes for ordinal multilevel models.
If this is not the right place to ask I would appreciate it if you could at least give me some references I could have a look at.

So, in my data I have the response variable which is ordinal (no care, occasional care, and regular care). I have data clustered in countries.

I initially estimate the null model. Then I include person-level characteristics (and check whether they meet the parallel line assumptions or not). Finally, I add the contextual country-level variables, for which I relax the parallel line assumption. I fit data in MLwiN using MCMC methods.
When I do this exact same procedure (minus the parallel assumption checks) for a binary outcome (say 'any care', i.e. occasional + regular, vs 'no care') I notice that the country-level variance decreases slightly when I include person-level characteristics, and decreases dramatically when I include country-level variables. This is somehow what I would expect.

However, when I extend the analysis to the ordinal data I get weird results.
To start with, when I include country-level indicators the country-variance increases. How is this possible? And is this ok?
Secondly, I only get one estimate of the variance. I would however expect to have a measurement of the variance for country-level differences between "no care vs occasional or regular care" and one for "no care and occasional vs regular care". Am I completely lost?

Thanks for your help!!
Best wishes,
giodje12

Hi giodje12,

See our free online multilevel modelling course for comprehensive explanation and examples of multilevel ordinal response models.

http://www.bristol.ac.uk/cmm/learning/course.html

The following is a nice article using multilevel ordinal response models where they relax parallel lines assumption

http://www.uic.edu/classes/bstt/bstt513/Addict2000.pdf


You say

"To start with, when I include country-level indicators the country-variance increases. How is this possible? And is this ok?"

I have to admit, this does sound counter intuitive. Note you should fit all these models by MCMC as quasilikelihood results are only approximate. So do check that this results is not simply due to estimation method. Otherwise all I can suggest is that you simulate some data along the lines of your application to check that you can recover the data generating parameters.

You say

" I only get one estimate of the variance. I would however expect to have a measurement of the variance for country-level differences between "no care vs occasional or regular care" and one for "no care and occasional vs regular care". Am I completely lost?"

You get this in a multinomial response multilevel model, so you may want to read up on that topic as well to see what works best for your data.


Best wishes

George