lme vs. lmer

Ben Bolker wrote:

Douglas Bates wrote:

On Tue, Sep 29, 2009 at 1:02 PM, Ben Bolker <bolker at ufl.edu> wrote:

Christopher David Desjardins wrote:

I've started working through Pinheiro & Bates, 2000 and noticed the use
of lme from the nlme package. I am curious if lmer from lme4 has
superseded lme or if lme still holds its own? The reason I ask is that
I
have taken a few classes where we've solely used lmer and just read
about lme today. If both functions are on equal footing, can the
p-values from lme be trusted?
Thanks!
Chris

?You should read the extended discussion of p-values, degrees of
freedom, etc. that is on the R wiki (I think) and referenced from the R
FAQ. ?At least in my opinion, (n)lme is still fine (and indeed necessary
at this stage for fitting heteroscedastic and correlated models). ?The
df/p-value estimates, however, are "use at your own risk" -- you'll have
to read the literature and decide for yourself.
?I still think there's room for someone to implement (at least)
Satterthwaite and (possibly) Kenward-Roger corrections, at least for the
sake of comparison, but I'm not volunteering.

You may need to define them first. ?Many of the formulas in the mixed
models literature assume a hierarchical structure in the random
effects - certainly we used such a formula for calculating the
denominator degrees of freedom in the nlme package. But lme4 allows
for fully or partially crossed random effects so you can't think in
terms of "levels" of random effects.

Referring to the "Satterthwaite and Kenward-Roger corrections" gives
the impression that these are well-known formulas and implementing
them would be a simple matter of writing a few lines of code. ?I don't
think it is. ?I would be very pleased to incorporate such code if it
could be written but, as I said, I don't even know if such things are
defined in the general case, let alone easy to calculate.

I am not trying to be argumentative (although of late I seem to have
succeeded in being that). ?I'm just saying that I don't think this is
trivial. (It I wanted to be argumentative I would say that it is
difficult and, for the most part, irrelevant. :-)

?Fair enough. Actually, I'm not sure I meant implementing them in lmer
-- even implementing them in nlme would be useful (and perhaps more
straightforward, if not trivial). I also wouldn't impose the requirement
that they have to be feasible for huge data sets -- I'm just curious if
they can be implemented within lme in a relatively straightforward/
boneheaded way.
? But again, this is very far down my to-do list (and at the edge
of my abilities) and completely off yours, so unless someone else bites
it won't happen.

I don't think (n)lme is all that easy either; you still need to sort out the
connection between its multilevel formulation and the projection matrices in
the K-R paper. In both nlme and lme4, an implementation is almost certainly
possible, although probably complicated and perhaps at the expense of all
computational efficiency.

One main problem is that even when they can be calculated, the corrections
rely on a normal distribution assumption which is more than likely wrong in
practice. This isn't any different from ordinary t-tests: once you get into
the single-digit df regime, you really don't know what you are doing, and if
there are more than 30 df, the normal approximation works well enough
without the correction.

Accordingly, I tend to see low df more as a warning flag than as something
that gives accurate p values, and I sometimes wonder whether there is a way
to raise such a flag more expediently.