Pergunta
What is wrong with the following model:
# simulated data yr = 2; vg = 4, fm = 5, gen = 5
mbb <- data.frame( trait1 = rnorm(200, 15, 4),yr = c(rep (1:2, each = 100)),
vg = c(rep(rep(1:4, each =25), 2)), fm = rep(rep(1:5, each = 5), 8),
gen = sample(c(1:5), 200, replace = T))
require(lme4)
lmer(trait1 ~ (yr + vg + gen)^3 + (yr + vg + gen|fm:vg), data= mbb)# full model
I am getting following error:
Error in validObject(.Object) :
invalid class "mer" object: Slot Zt must by dims['q'] by dims['n']*dims['s']
In addition: Warning messages:
1: In fm:vg : numerical expression has 200 elements: only the first used
2: In fm:vg : numerical expression has 200 elements: only the first used
Solução
The problem is precisely that fm
and vg
are stored as numeric, not as factors, and so lmer
tries to interpret fm:vg
as a sequence operator (see ?seq
) rather than an interaction operator (see ?interaction
). You can:
convert
fm
andvg
to factors within the data frame (mbb <- transform(mbb,vg=factor(vg),fm=factor(fm))
) [it's not clear from your setup whether you wantvg
andfm
to be factors or continuous predictors ... that distinction would be very important, of course ... if you want them as continuous predictors, then it's a bit weird to treat them as factors for the purposes of grouping ... ]write the interaction explicitly as
interaction(fm,vg)
on the flyconvert to factors on the fly (
(yr+vg+gen|factor(fm):factor(vg))
)use Jim M.'s solution
I think these will all work, although I have to admit that I haven't tested them.
Outras dicas
One possible solution to model the interaction as a random effect is to add the interaction term as an additional column in the mbb data frame.
mbb$fmvg <- with(mbb, interaction(fm,vg, sep=":"))
The model then becomes
lmer(trait1 ~ (yr + vg + gen)^3 + (yr + vg + gen|fmvg), data= mbb)