GLIMMIX for repeated measures and multinomial (ordered) response

[toneill]

Looking for some interpretation help with a GLIMMIX output (as most the documentation is from logistic models).

The very basic bivariable model is as follows:

proc glimmix data=have noclprint noitprint method=quad gradient ;

  class study_id ;

  model score=race / solution dist=mult link=clogit ;

  random intercept / type=un subject=study_id ;

      covtest 'Need Random Intercept?' 0 ;

  where race ^in(.,88,99) ;

run ;

I have 5500 observations, with a total of 1795 individuals (study_id) observed over time. Participants have multiple scores (range: 0-4) over time. I have made no assumption about the distance between adjacent levels, but they are ordered (i.e. 0=not present, 4=very bad levels present). Race is ethnic race (i.e. caucasian, African-American etc) with 5 levels. I am modeling the probabilities of levels of score having lower ordered values in the response profile table.

Once run, I have 4 intercepts  (0-3), and 5 estimates for race. Numbers below are just made up.

Effect           Score     Race     Estimate

Intercept      0                                  0.5

Intercept      1                                   0.6

Intercept      2                                   0.7

Intercept      3                                   0.8

Race                        Race1               0.09

Race                         Race 2              0.1

Race                          Race3               0.11

Race                          Race4               0.12

Race                          Race5               0.13

I would like to be able to interpret this (eventually) in a multivariable model as well, then present it in a logical and meaningful way. Any direction would be appreciated.    

[SteveDenham]

The key will be to translate the estimates into odds ratios that describe what is going on, and that isn't always easy.  This is a case where ESTIMATE statements are critical to calculating customized odds ratios.  Do a search of this site on 'glimmix multinomial estimate' as I am sure there are answers out there by people who are better at this than I am.

Steve Denham