[R-sig-ME] crossed effects with lmer but correlation structure with lme

[Denis Vile]

Hi all,

Was my problem not sufficiently well exposed or no one could help me ?

Sincerely,

Denis

Le 29/03/2012 10:09, Denis Vile a écrit :
> Dear R users,
>
> I'm trying to fit a crossed-effects mixed model that would include a 
> spatial correlation structure..
> The data come from four controlled experiments (control, treatment1, 
> treatment2, treatment1+treatment2) on plants grown in a growth 
> chamber.  Individual replicates of different genotypes were grown 
> together and response traits were measured. A covariate X is included 
> in the model with a quadratic form.
>
> We fitted the following model using lmer:
>
> fm1 <- lmer(Y ~ 1 + Trt1*Trt2*poly(X, degree=2, raw=T) +
> (1|idGenotype) + (1|Trt2:idGenotype) + (1|Trt1:idGenotype) +
>             (1|Trt1:Trt2:idGenotype), data=...)
>
> This model is very interesting because we can extract the BLUPs for 
> each genotype in each (crossed) environment.
>
> After discussion with colleagues, it appeared that we should try to 
> include the possible spatial heterogeneityof the 
> micro-environmentwithin the growth chamber. To this end, we tried to 
> fit a model with lme() because we cannot easily (if possible) include 
> a correlation structure using lmer(). The model is:
>
> fm2 <- lme(Y ~ (X+I(X^2))*idCondition,
>             random =~1|idGenotype/idCondition,
>             correlation=corGaus(c(15,0.95), 
> form=~x+y|idGenotype/idCondition, nugget=T),
>             data =...)
>
> where x and y are the coordinates of the plants within the growth 
> chamber.
>
> Since I was unable to fit the crossed effects Trt1 x Trt2 in lme() I 
> coded a new variable idCondition which is the combination of Trt1 and 
> Trt2, and treated genotypes within idCondition. This is not entirely 
> satisfying because it is impossible to extract all BLUPs as in fm1.
>
> Could you please tell me if I missed something and ifthere is a trick 
> to specify crossed effects using lme()?
> I assume that this should use pdClasses but I'm not at all at ease 
> with the matrix specification of mixed models.
> Alternatively, include a correlation structure in lmer seems to be 
> unfeasible, am I wrong?
>
> Thank you very much for your help,
>
> Denis
>
>

-- 

*Denis VILE*
Chargé de Recherche
Laboratoire d'Ecophysiologie des Plantes sous Stress Environnementaux 
(*LEPSE*)
UMR 759 *INRA*-SUPAGRO // Institut de Biologie Intégrative des Plantes 
(IBIP, bât 7)
2 place Pierre Viala
34060 Montpellier Cedex 2
Tel +33 (0)4 99 61 31 87
Fax +33 (0)4 67 52 21 16
http://www1.montpellier.inra.fr/ibip/lepse/