Skip to content
Prev 6230 / 20628 Next

compound symmetry correlation structure in nlme

Rian--

I believe you would want to compare an lme() fit with random intercept 
to a gls() fit with compound symmetry (and hence, no random terms at all).

Sort of think Pinheiro and Bates make this comparison in their book, but 
I don't have it handy.

Hope that helps.

cheers, Dave


hello all,

I am analyzing some data with repeated measures on individuals, and want 
to test different correlation structures.
In Zuur et al (2007) they state "the inclusion of a random intercept in 
a GLM is imposing the compound symmetrical correlation structure, just 
as it did in the linear mixed model" (Chapter 13, p.323), which I 
understand as meaning that if a random intercept is included in a model, 
then the compound symmetry correlation structure is implied.

When I run the following two models (which are identical, except that in 
the second one I explicitly specify the correlation structure), I get 
the same estimates for the random and fixed effects, but the AIC and BIC 
values differ, because the second model has one more parameter (for the 
correlation structure).

global.ri.lme <- lme(Log.minH ~ site + f.year + cohort + emerg + emergSQ 
+ pri + priSQ
   + start.t + startSQ + sea + tidem + tided, data = SUSCforage, random 
= ~1|scoterID,
   method = "REML")

global.ri.cs1.lme <- lme(Log.minH ~ site + f.year + cohort + emerg + 
emergSQ + pri + priSQ
   + start.t + startSQ + sea + tidem + tided, data = SUSCforage, random 
= ~1|scoterID,
   correlation = corCompSymm(form = ~1|scoterID), method = "REML")

So, if I want to obtain an accurate AIC value, which should I use?

thank you,

Rian