[R] anova on binomial LMER objects

[Douglas Bates]

On 9/25/05, Horacio Montenegro <nepossiver at yahoo.com> wrote:
>
>     Hi Spencer and Robert,
>
>     I have found the same behaviour, but only for lme4
> and Matrix after the 0.96 release. lme4 0.95-10 and
> Matrix 0.95-13 releases gave "sensible" results. This
> could be an introduced bug, or a solved bug - you
> should ask Prof. Bates.
>
>     hope this helps, cheers,
>
>     Horacio Montenegro

I have run into a couple of other things that the "improvements" from
the 0.95 series to the 0.96 series has made worse.  This may take a
while to sort out.  Thanks to Robert Bagchi for the very thorough
error report.


>
> --- Spencer Graves <spencer.graves at pdf.com> wrote:
> >         I agree:  Something looks strange to me in this
> > example also;  I have
> > therefore copied Douglas Bates and  Deepayan Sarkar.
> >  You've provided a
> > nice simulation.  If either of them have time to
> > look at this, I think
> > they could tell us what is happening here.
> >
> >         If you need an answer to your particular problem,
> > you could run that
> > simulation 1000 or 1,000 times.  That would tell you
> > whether to believe
> > the summary or the anova, or neither.  If you want
> > to understand the
> > algorithm, you could walk through the code.
> > However, "lmer" is a
> > generic, and I don't have time now to figure out how
> > to find the source.
> >   A response from Brian Ripley to a question from me
> > a couple of days
> > ago provides a nice summary of how to do that, but I
> > don't have time to
> > check that now.
> >
> >         Sorry I couldn't help more.
> >         spencer graves
> >
> > Robert Bagchi wrote:
> >
> > > Dear R users,
> > >
> > > I have been having problems getting believable
> > estimates from anova on a
> > > model fit from lmer. I get the impression that F
> > is being greatly
> > > underestimated, as can be seen by running the
> > example I have given below.
> > >
> > > First an explanation of what I'm trying to do. I
> > am trying to fit a glmm
> > > with binomial errors to some data. The experiment
> > involves 10
> > > shadehouses, divided between 2 light treatments
> > (high, low). Within each
> > > shadehouse there are 12 seedlings of each of 2
> > species (hn & sl). 3
> > > damage treatments (0, 0.1, 0.25 leaf area removal)
> > were applied to the
> > > seedlings (at random) so that there are 4
> > seedlings of each
> > > species*damage treatment in each shadehouse.
> > There maybe a shadehouse
> > > effect, so I need to include it as a random
> > effect. Light is applied to
> > > a shadehouse, so it is outer to shadehouse. The
> > other 2 factors are
> > > inner to shadehouse.
> > >
> > > We want to assess if light, damage and species
> > affect survival of
> > > seedlings. To test this I fitted a binomial mixed
> > effects model with
> > > lmer (actually with quasibinomial errors). THe
> > summary function suggests
> > > a large effect of both light and species (which
> > agrees with graphical
> > > analysis). However, anova produces F values close
> > to 0 and p values
> > > close to 1 (see example below).
> > >
> > > Is this a bug, or am I doing something
> > fundamentally wrong? If anova
> > > doesn't work with lmer is there a way to perform
> > hypothesis tests on
> > > fixed effects in an lmer model? I was going to
> > just delete terms and
> > > then do liklihood ratio tests, but according to
> > Pinheiro & Bates (p. 87)
> > > that's very untrustworthy. Any suggestions?
> > >
> > > I'm using R 2.1.1 on windows XP and lme4 0.98-1
> > >
> > > Any help will be much appreciated.
> > >
> > > many thanks
> > > Robert
> > >
> > >
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
>