[R-sig-ME] Mixed model (with interaction) for gene expression and iteration
Rolf Turner
r.turner at auckland.ac.nz
Thu Jun 4 22:35:31 CEST 2009
Dear Doug,
Two issues have arisen in respect of this thread that I would like to
comment upon.
(1) In my previous posting I included the link to Doug's explanation
of the
lmer() syntax:
http://www.nabble.com/lme-nesting-interaction-advice-
td17131600i20.html#a17213604
I found this explanation incredibly useful, and apparently others
have as well.
(I was thanked off-list for posting this link.)
Could, perhaps, the content of this link, in appropriately edited
form, be
included in the documentation for lmer(), or perhaps in a vignette
(to which
the documentation might direct the reader)? To save on frustrating
searching
for those who are coming in to this area fresh.
(2) In one of his posts Juan Pedro Steibel wrote:
> Note that m1c is actually a fixed effects model, and can not be fit
> with
> lmer, so you'll need to reconstruct the -2LL from another source.
He goes on to say:
> But with a normal model that should not be a problem.
Hah! Not a problem for some, maybe. But for those of us whose
initially
few brain cells have been decimated by alcohol ( :-) ) it is a problem.
Might I make a ``feature request'' that it be made possible to fit
models
with fixed effects only in lmer()? (Basically for convenience of
testing
whether random effects are ``useful''.) I personally would also be
comforted
by being able to do things like
fit1 <- lmer(y ~ x, REML=TRUE)
fit2 <- lmer(y ~ x, REML=FALSE)
and compare the resulting estimates of sigma^2 (which presumably
should differ
by a factor of n/(n-2) if I'm understanding things correctly). Being
able to
do this would give me the comforting illusion ( :-) ) that I ***am***
understanding
things correctly. Or perhaps disabuse me of this illusion.
Would it be a Herculean task to adapt lmer() to have this capacity?
Naively I would
have thought that if a model with fixed and random effects can be
fitted, then
surely a (simpler) model with only fixed effects can be fitted. OTOH
the fixed effects
only scenario is in some sense ``on the boundary'', and boundary
cases can be vexatious.
Not knowing how lmer() works, I have no idea how hard or easy making
such a change
would be.
I would appreciate hearing your thoughts on this.
cheers,
Rolf
######################################################################
Attention:\ This e-mail message is privileged and confid...{{dropped:9}}
More information about the R-sig-mixed-models
mailing list