Comparing covariance parameters from similar models in GLIMMIX

[skr01]

Hello all-

I am running the same model with two different response variables - pres/abs of two species. The models have binary responses with logit link, and include fixed effects, g-side random effects, and repeated measures (R-side). I am appending a simplified version that includes all the general categories of things in the model at the end of this.

From looking at the datat, I suspect that the variability in the probability of finding a particular species depends differently on the random predictors in the two species. I do not want to try to put the two into one model to test this, I just want to compare the proportion of the total variance explained by particular covariance parameters in these parallel models describing whether or not we expect to find  species A vs species B. I think therefore, that I want an intraclass correlation coefficient (ICC), so that I can say "Ah, "house" explains lots of variance in recovery of species A but not species B" (perhaps). I have found lots of information on how to calculate them in simple models, some discussions of how to calculate them in simple GLIMMIX models, and some in which people have said that they are not useful in GLIMMIX (though with little explanation of why not). I am looking for advice regarding whether computing ICCs for my two models and comparing them will adequately address my question, and if so, how to calculate them. Any input would be most appreciated.

proc glimmix data=work.mydata;

class House_Niche_Rep  house   seasonyrnum environment;

model PU(event='1') =  samplingperiod  environment/ dist=binary link=logit ddfm=residual;

random int samplingperiod environment/ subject=house ;

random samplingperiod / subject=House_Niche_Rep*house type=AR(1) residual;

covtest / wald;

nloptions tech=nrridg maxiter=250;

run;

Thanks,

susi

[SteveDenham]

This would be one of the cases where an ICC may be nearly impossible to calculate, due to the repeated nature and complex design of the approach.  Given that you do not want to fit a model with species included, which would enable a test of homogeneity if done well, I am unsure what to offer.  Are both species A and species B sampled for in all subjects?  This would help, as the data would be balanced.  If everything is balanced, then the absolute magnitude of the variance components will tell you something.  Keep in mind that the variance components are on the logit scale, and so may seem outlandish.

Steve Denham