[R] Difficulty with lme

[Ben Bolker]

Kevin Wright <kw.stat <at> gmail.com> writes:

> 
> Generally, the only way to estimate f1:f2 is if you have all combinations of
> data present for these two factors.

  Well, he said it was unbalanced, he didn't say how unbalanced --
i.e. it's not clear (to me) whether there are any completely missing cells
or not  ...

> On Wed, Oct 5, 2011 at 2:00 PM, Brad Davis <bhdavis1978 <at> 
> gmail.com> wrote:
> >
> > I'm having some difficulty with lme.  I am currently trying to run the
> > following simple model
> >
> > anova(lme(x ~ f1 + f2 + f1:f2, data=m, random=~1|r1))

  [which you could also specify as ~f1*f2]

> > Which is currently producing the error
> >
> > Error in MEEM(object, conLin, control$niterEM) :
> >  Singularity in backsolve at level 0, block 1
> >
> > x is a numeric vector containing 194 observations. f1 is a factor vector
> > containing two levels, and f2 is a different factor vector containing 5
> > different levels.   R1 is a another factor vector containing 13 different
> > levels, and it is again, unbalanaced.  f1, f2 and r1 are unbalanced, but I
> > can't do anything about it.  The data comes from wild-caught samples and
> > not
> > from a nice, neat experiment.  If I change the model specification slightly
> > removing the interaction term (e.g. anova(lme(x ~ f1 + f2, data=m,
> > random=~1|r1)) ), then lme proceeds without producing any errors.

  I have a couple of suggestions:
(1) try lmer (it will at least work differently, and might work better)
(2) try expanding your model out to a one-way design --
 lme(x~interaction(f1,f2),data=m,random=~1|r1)

   Follow-ups should probably be sent to r-sig-mixed-models at r-project.org