[R] Problem with Variance Components (and general glmm confusion)

From: Toby Gardner <t.gardner_at_uea.ac.uk>
Date: Mon 04 Sep 2006 - 22:57:40 GMT


Dear list,

I am having some problems with extracting Variance Components from a random-effects model:

I am running a simple random-effects model using lme:

model<-lme(y~1,random=~1|groupA/groupB)

which returns the output for the StdDev of the Random effects, and model AIC etc as expected.

Until yesterday I was using R v. 2.0, and had no problem in calling the variance components of the above model using VarCorr(model), together with their 95% confidence intervals using intervals() - although for some response variables a call to intervals() returns the error: Cannot get confidence intervals on var-cov components: Non-positive definite approximate variance-covariance.

I have now installed R v. 2.3.1 and am now experiencing odd behaviour with VarCorr(lme.object), with an error message typically being returned:

Error in VarCorr(model) : no direct or inherited method for function 'VarCorr' for this call

Is this known to happen? For instance could it be due to the subsequent loading of new packages? (lme4 for instance?).

To get around this problem I have tried running the same model using lmer:

model2<-lmer(y~1 + (1|groupA) + (1|groupB))

Should this not produce the same model? The variance components are very similar but not identical, making me think that I am doing something wrong. I am also correct in thinking that intervals() does not work with lmer? I get: Error in intervals(model2) : no applicable method for "intervals"

GLMM I have a general application question - please excuse my ignorance, I am relatively new to this and trying to find a way through the maze. In short I need to compile generalized linear mixed models both for (a) Poisson data and (b) binonial data incorporating a two nested random factors, and I need to be able to extract AIC values as I am taking an information-theoretic approach to model selection. Prior to sending an email to the list I have spent quite a few days reading the background on a number of functions, all of which offer potential for this; glmmML, glmmPQL, lmer, and glmmADMB. I can understand that glmmPQL is unsuitable because there is no way of knowing the maximised likelihood, but is there much difference between the remaining three options? I have seen simulation comparisons published on this list between glmmADMB and glmmPQL and lmer, but it seems these are before the latest release of lmer, and also they do not evaluate glmmML. To a newcomer this myriad !  of options is bewildering, can anyone offer advice as to the most robust approach?

Many thanks for your time and patience,

Toby Gardner

School of Environmental Sciences
University of East Anglia
Norwich, NR4 7TJ
United Kingdom
Email: t.gardner@uea.ac.uk
Website: www.uea.ac.uk/~e387495

        [[alternative HTML version deleted]]



R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code. Received on Tue Sep 05 10:12:58 2006

Archive maintained by Robert King, hosted by the discipline of statistics at the University of Newcastle, Australia.
Archive generated by hypermail 2.1.8, at Thu 07 Sep 2006 - 07:51:18 GMT.

Mailing list information is available at https://stat.ethz.ch/mailman/listinfo/r-help. Please read the posting guide before posting to the list.