heteroscedastic model in lme4

Dear Vito,

I aggree that the glmer() and nlme() examples assume different distribution. That's why I called the nlme() version an apprioximation. If the steps described in P&B are only valid for linear models but not in the generalised models, then I have a dilemma. With glmer() I can use the appropriate distribution but a wrong correlation structure. A structure of which I'm certain that it is there (spatially clustered points). nlme() allows me to model the correlation structure but only unther the gaussion distribution. The latter is in my opinion a better alternative given that with enough data the residuals will behave approximately gaussian. Please do correct me if that is an incorrect statement.

Why nlme() and not lme() with a log-transformation? Well: the zero's in counts. Using a log(x + 1) transformation complicates the interpretation of the model. And what transformation would you suggest with binomial data? nlme() handles zero's with a log-link as well as true-false data with a logit-link.

Thierry

----------------------------------------------------------------------------
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature and Forest
Cel biometrie, methodologie en kwaliteitszorg / Section biometrics, methodology and quality assurance
Gaverstraat 4
9500 Geraardsbergen
Belgium 
tel. + 32 54/436 185
Thierry.Onkelinx at inbo.be 
www.inbo.be 

To call in the statistician after the experiment is done may be no more than asking him to perform a post-mortem examination: he may be able to say what the experiment died of.
~ Sir Ronald Aylmer Fisher

The plural of anecdote is not data.
~ Roger Brinner

The combination of some data and an aching desire for an answer does not ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey

-----Oorspronkelijk bericht-----
Van: vito muggeo [mailto:vmuggeo at dssm.unipa.it] 
Verzonden: donderdag 15 januari 2009 13:21
Aan: ONKELINX, Thierry
CC: Doran, Harold; Alan Cobo-Lewis; r-sig-mixed-models at r-project.org
Onderwerp: Re: [R-sig-ME] heteroscedastic model in lme4

dear Thierry,
I am adding a simple comment only on your second point.

If I am not wrong, I think that the two alternatives underlie different 
models

1)glmer(.., family = poisson) assumes a real Poisson distribution for 
your response y (conditioned to random effects), i.e. y=rpois(n,exp(mu)).

2) nlme(..) assumes a gaussian distribution for your response with a 
nonlinear mean model, i.e. y=rnorm(n,exp(mu))

Another (different) approach would be lmer() with log-transformed data, 
i.e. y=exp(rnorm(n,mu))

Probably, in a pure likelihood framework the first approach should be 
preferred if you have real count data..

Hope this helps,

vito

ONKELINX, Thierry ha scritto:

heteroscedastic model in lme4

Thread (7 messages)