[R-sig-ME] significance of random effects and prediction

[Tahira Jamil]

Hi
I am fitting model in nlme/lme4) and interested testing the variance component and how the significance of variance component play a role in improving the prediction. First I fitted a model in nlme like this 

 v0 <- 0 * train.data$y + 1;

 fm.train<- lme(groupedData(y0 ~ X|v0), 
           random =list( v0=pdBlocked(list(pdIdent(~sp-1),pdIdent(~site-1),
                       pdIdent(~ X:Z0-1 ))),sp=pdIdent(~X-1)))  
  

and then only that variance component model which variance copmnent I wanted to test 

  fm.train1<- lme(groupedData(y0 ~ X|v0), 
         random =list( v0=pdIdent(~ X:Z0-1 )))


both the model fits well and  when I compare the above two models with exactRLRT test,
 exactRLRT( fm.train1, mA = fm.train$nlme)
 the variance component turn out to be siginifcant. 

then I fitted a reduce model 
 fm.train0<- lme(groupedData(y0 ~ X|v0), 
          random =list( v0=pdBlocked(list(pdIdent(~sp-1),pdIdent(~site-1)))
                       ,sp=pdIdent(~X-1))) 

But I do some cross validation and look for increse in R2 value for the left out data for both the models fm.train and fm.train0 give the approximately same value. 
So my question is if the variance component of a model is significant, does it improve or have some role for the prediction or how we can interpret this situation.
Hope for some good remarks 
Cheers

Tahira
Ph.D Student Biometris
Wageningen University