R help - Jul 2009 - multiple comparisons and generalized least squares

Dear R users,

 

I 'm working on a dataset consisting of 4 different dataframes with
tree, leaf, fruit and seed measurements made on 300 trees, coming from
10 provenances (30 trees per provenance, 10 leaves/fruits/seeds per
tree). Provenances are fixed effects (they were not randomly chosen),
but trees within provenances and leaves/fruits/seeds within trees were
randomly assigned. I wanted to check whether there were any
between-provenance differences concerning the measured characteristics.
For the leaves, fruits and seeds datasets, I tried some lme models of
the form:

model<- lme(characteristic~1, random=~1|tree, data=leaves) 

model<- lme(characteristic~provenance, random=~1|tree, data=leaves) 

or, in case of heteroscedasticity on the provenance level: 

model<-lme(characteristic~provenance, random=~1|tree, data=leaves,
weights=varIdent(form=~1|provenance)). 

Multiple comparisons were then performed on the best fitting model by
using the glht function of the package multcomp, with
linfct=mcp(provenance="Tukey").

 

For the tree characteristics, as far as I understand (but I am quite new
to R and mixed effect modeling), no random effects are needed for model
specification. But since most tree characteristics (height, stem
diameter, crown diameter etc.) have a very different variance for each
provenance, I used the function gls with a variance structure:

model<-gls(characteristic~provenance, data=trees,
weights=varIdent(form=~1|provenance))

Then, using the function glht as before, I get an error message: "no
'model.matrix' method for 'model' found!".

 

I tried to find an explanation in the mailing archives and the glht and
gls help files, but I couldn't, so therefore these questions:

1) Is the function glht just not designed for use with gls models, or is
there a problem with my model specification?

2) Is this approach (first mixed effect modeling, then multiple
comparisons of means) correct to determine the differences in
characteristics between provenances? (I cannot use anova because of the
variance structure and the hierarchical structure of the measurements)

3) If glht cannot work with gls models, is there another way to include
a variance structure in the analysis, when making multiple comparisons?

 

Many thanks in advance and sorry if my questions are not very clear or
if I missed some essential help file or mail. 

 

Katrijn

 


	[[alternative HTML version deleted]]