07-25-2014 03:58 PM
I'm running glimmix for a group of categorical variables. I wanted to evaluate models using Akaike, therefore I put laplace as the METHOD, which allow be to obtain AIC values.
seBUT, my question is: Am I doing something wrong? Because results are very different if using the default method RSPL or the laplace, so I'm afraid I might me skipping something important....
I red that the laplace method wont let you model the R-effect, but I'm not sure if I understand what it means...
So.. please.... Help!!!
07-28-2014 11:24 AM
Yes, the results will often differ substantially between the pseudo-likelihood methods and the quasi-likelihood methods. Unfortunately, there doesn't seem to be a good way to compare competing models under the pseudo-likelihood methods, as the pseudo-data will not be constant.
So, model the repeated nature as a G side parameterization using method=laplace, and trust in those results, as they have been shown to be relatively less biased than the marginal estimates using an R side parameterization.
07-28-2014 02:50 PM
But I'm still worried. In the SAS documentation it says that:
'R-side effects in the RANDOM statement do not generate model matrices; they serve only to index observations within subjects"
The problem is my data. I'm using more that one observation per individual, therefore I'm using GLIMMIX to consider each individual as a "cluster" item. I'm studing dogs within a comunity, and each person has one or more dogs, therefore I'm using the "owner" as an item that should be considered by the procedure. If I understood correctly what SAS said, I would need R-side effects, because they would be serving me to index observations within subjects.
Am I understanding it correctly?
What do you think?
Could I still use laplace for my data?
07-28-2014 03:19 PM
In this case, R-side would index each dog within an owner, as if the owner were measured on successive time points or at geographically defined points. G-side would add columns to the Z matrix so that each owner is viewed as a cluster of dogs, and a variance component (or components) is estimated. For what you are doing, I don't see any need for R side approach.
Now a key here is the distribution associated with the response variable. Be sure that the link associated is appropriate--cumulative logit for ordinal categories and generalized logit for nominal categories.