Interpreting Estimates/Solution for Fixed Effects in GLIMMIX

[KMun]

Good afternoon,

I am new to multi-level modeling. I was first taught standard OLS for continuous variables. I am not working with an ordinal variable, so I am running a cumulutive logit model with the glimmix procedure, which my book says is appropriate for ordinal outcomes (mine is a five item likert scale).

In the Output, under Solution for Fixed Effects, I get what appear to be coefficient estimates. How are these interpreted for a five item ordinal variable? Not like slopes, right? When doing this in non-multilevel analyses, I always get odds-ratios. Do I want [(exp(reported estimate)+/-1)? Or is that not what is being reported?

I have tried including OR or oddratio in the code. It gives me multiple responses per independent variable. It is not at all clear how to interpret this.

proc glimmix noclpint;
class region;
model generous (descending) = age gender education income minority relig mclrs mcpdis mceu
region_eu/S dist=multi link=clogit ddfm=bw; weight popw;
random int relig/subject=region type=chol;
nloptions tech=nrridg;
covtest 'H: var[random int] =0' 0;
covtest 'H: var[random cov] =0' . 0;
covtest 'H: var[random slope]=0' . . 0;
run;

[ChrisNZ]

Moved to stats forum.

[SteveDenham · Accepted solution]

The multiple ORs per variable are the result of calculating the OR for each level of the dependent variable (cumulative logit link implies this).  Also, several of your independent variables are likely to be categorical (I can't imagine relig as a continuous variable).  So there are two cases:

Variable is continuous: You should get 4 OR for 2v1, 3 v 1+2, 4 v 1+2+3, 5 v 1+2+3+4 (levels of the Likert scale).  These are calculated for an increase of 1 in the continuous variable.

Variable is categorical: The same 4 OR, but these are now calculated for each level of the categorical variable relative to a reference level, which would be defined to have an odds ratio of 1 when compared to itself..

The latter case may be handled better by use of the OR option in an LSMEANS statement.  For multiple continuous variables, OR from LSMEANS for one variable is calculated at the mean of all the other continuous variables, again with a default increase of 1 unit.  Other values for unit can be specified (for instance, suppose unit=10 for age.  This would give the odds ratios for an increase of 10 (years, most likely).

Is this what you are asking?  I hope it was helpful, but I am kind of afraid I am making a Type III error of answering the wrong question.

SteveDenham

[KMun]

That answers it. Thank you very much.