Skip to content

general GLMM questions

6 messages · Ben Bolker, Dimitris Rizopoulos, vito muggeo +2 more

#
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Dear sig-mixed readers,

~  Some of my students and I are foolishly attempting to write a review
of GLMMs for an ecology/evolution audience.  Realizing that this is a
huge, gnarly, and not-completely-understood subject (even for the
experts -- see fortune("mixed")), we're trying to provide as much
non-technical background and guidance as can be squeezed into a
reasonable sized journal article ...

~  We've run into quite a few questions we have been unable to answer
... perhaps because no-one knows the answers, or because we're looking
in the wrong places.  We thought we might impose on the generosity of
the list: if this feels ridiculous, please just ignore this message.
Feedback ranging from "this is true, but I don't know of a published
source" to "this just isn't true" would be useful.  We are aware of the
deeper problems of focusing on p-values and degrees of freedom -- we
will encourage readers to focus on estimating effect sizes and
confidence limits -- but we would also like to answer some of these
questions for them, if we can.

~  Ben Bolker

1. Is there a published justification somewhere for Lynn Eberly's
( http://www.biostat.umn.edu/~lynn/ph7430/class.html ) statement that df
adjustments are largely irrelevant if number of blocks>25 ?

2. What determines the asymptotic performance of the LRT (likelihood
ratio test) for comparison of fixed effects, which is known to be poor
for "small" sample sizes?  Is it the number of random-effects levels
(as stated by Agresti 2002, p. 520), or is it the number of levels of
the fixed effect relative to the total number of data points (as
stated by Pinheiro and Bates 2000, pp. 87-89)?  (The example given by
PB2000 from Littell et al. 1996 is a test of a treatment factor with
15 levels, in a design with 60 observations and 15 blocks.  Agresti's
statement would imply that one would still be in trouble if the total
number of observations increased to 600 [because # blocks is still
small], where PB2000 would imply that the LRT would be OK in this
limit.  (A small experiment with the simulate.lme() example given on
PB2000 p. 89 suggests that increasing the sample size 10-fold with the
same number of blocks DOES make the LRT OK ... but I would need to do
this a bit more carefully to be sure.)  (Or is this a difference
between the linear and generalized linear case?)

3. For multi-level models (nested, certainly crossed), how would one
count the "number of random-effects levels" to quantify the 'sample
size' above?  With a single random effect, we can just count the
number of levels (blocks).  What would one do with e.g. a nested or
crossed design?  (Perhaps the answer is "don't use a likelihood ratio
test to evaluate the significance of fixed effects".)

4. Does anyone know of any evidence (in either direction) that the
"boundary" problems that apply to the likelihood ratio test (e.g. Self
and Liang 1987) also apply to information criteria comparisons of
models with and without random effects?  I would expect so, since the
derivations of the AIC involve Taylor expansions around the
null-hypothesis parameters ...

5. It's common sense that estimating the variance of a random effect
from a small number of levels (e.g. less than 5) should be dicey, and
that one might in this case want to treat the parameter as a fixed
effect (regardless of its philosophical/experimental design status).
For small numbers of levels I would expect (?) that the answers MIGHT
be similar -- among other things the difference between df=1 and
df=(n-1) would be small.  But ... is there a good discussion of this
in print somewhere?  (Crawley mentions this on p. 670 of "Statistical
Computing", but without justification.)

lme4-specific questions:

6. Behavior of glmer: Does glmer really use AGQ, or just Laplace?
Both?  pp. 28-32 of the "Implementation" vignette in lme4 suggest that
a Laplace approximation is used, but I can't figure out whether this
is an additional approximation on top of the AGQ/Laplace approximation
of the integral over the random effects used in "ordinary" LMM.  When
I fit a GLMM with the different methods, the fitted objects are not
identical but all the coefficients seem to be.  (I have poked at the
code a bit but been unable to answer this question for myself
... sorry ...)

(The glmmML package claims to fit via Laplace or Gauss-Hermite
quadrature (with non-adaptive, but adjustable, number of quad points
- -- so it's at least theoretically possible?)

library(lme4)
set.seed(1001)
f = factor(rep(1:10,each=10))
zb = rnorm(1:10,sd=2) ## block effects
x = runif(100)
eta = 2*x+zb[f]+rnorm(100)
y = rpois(100,exp(eta))

g1 = glmer(y~x+(1|f),family="poisson",method="Laplace")
g2 = glmer(y~x+(1|f),family="poisson",method="AGQ")
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFIIc1Lc5UpGjwzenMRAr4uAJ90myt79pJZCa1a801FkxHRnAHYdgCfUYy+
P0ljXHs4lt8aTwpWKncRkBg=
=nd22
-----END PGP SIGNATURE-----
#
----- Original Message ----- 
From: "Ben Bolker" <bolker at ufl.edu>
To: "R Mixed Models" <r-sig-mixed-models at r-project.org>
Sent: Wednesday, May 07, 2008 5:39 PM
Subject: [R-sig-ME] general GLMM questions
Intuitively, I think this will depend on the correlation between the
measurements within each block. At the one extreme, assume that all
the measurements within each block are the same, then actual sample
size would be the number of block. At the other extreme, assume that
all the
measurements within each block are random, then the sample size would
be the total number of observations.

I know some people from Hasselt University in Belgium that have worked
on the "Effective Sample Size" for mixed models; you can check at the
following presentation given in ISCB last summer
(http://www.iscb2007.gr/ppt/Wednesday/Orfeas/16.54-17.12/slides_ISCB2007.pdf)
I think this is an interesting question. Let

# AIC under the null
AIC.0 = -2*logLik.0 + 2npar

# AIC under the alternative
AIC.1 = -2*logLik.1 + 2(npar + 1)

then you reject when

AIC.1 < AIC.0 =>

-2*logLik.1 + 2(npar + 1) + -2*logLik.0 + 2npar < 0 =>

LRT > 2.

Now for boundary problems and according to Stram and Lee (1994,
Biometrics), LRT ~ 0.5 * chisq(0) + 0.5 * chisq(1), for which the
critical value is 1.923. Thus, it seems to work more or less ok in
this case. However, if you wanted to test wether the variance is 10,
then LRT ~ chisq(1), for which the critical value is 3.841!
Well, in Linear Mixed Models the integral over the random effects can
be analytically evaluated and thus no approximation (i.e., AGQ or
Laplace) is required. In GLMMs this is not the case and thus the
log-likelihood needs to be calculated approximately. One method for
approximating the integral is the AGQ, and in fact Laplace is AGQ with
one quadrature point.

AFAIK (but Doug can correct if I'm wrong), glmer() uses Laplace since
AGQ is not yet implemented.
Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm
#
Dear Ben,
I am going to reply just to your question #4.

Yes, the AIC suffers from the same drawbacks of the log-Likelihood; 
therefore if the LRT does not work (for testing for variance components 
in LMM and also for any non-regular models, say), the AIC is not 
expected to work too. (I don't remember the references where I read 
this,..sorry). For instance, in problems related to breakpoint 
estimation the logLik is  just piecewise differentiable and if one is 
interested in testing for the existence of the breakpoint, the LRT and 
the AIC do not work. However simulation studies have shown that the BIC 
works (this makes sense because the BIC has a Bayesian justification and 
nondifferentiable logLik typically does not matter..)

best,
vito

Ben Bolker ha scritto:

  
    
#
On 08/05/2008, at 1:39 AM, Ben Bolker wrote:

            
Actually denominator df > 25. This seems to derive from t  
distributions with df greater than 25 all being much the same, in fact  
close to a normal distribution. In reality variations in the data from  
normality are more important.
This is probably dependent on whether the comparisons are within- or  
between- block. The PBIB has lots of within- block comparisons so  
increasing block size will tend to make things asymptotic. Try blocks  
where all within a block receive the same treatment and see how much  
increasing block size helps.
This is a good question. For choosing the number of classes for  
mixture models it has been shown that BIC fails theoretically but  
works well in practice (proven with simulations) and compares well to  
results from parametric bootstrapping of the LRT. Some simulations for  
random effects would be interesting.

Ken
1 day later
#
On Wed, May 7, 2008 at 10:39 AM, Ben Bolker <bolker at ufl.edu> wrote:
To answer this question I must again, I regret, distinguish between
the CRAN version of the lme4 package and the R-forge development
version of lme4.

In the R-forge version the only method for generalized linear mixed
models and for nonlinear mixed models is direct optimization of the
Laplace approximation to the deviance.  One of the Summer of Code
projects that Google has funded for the R Foundation (see
http://code.google.com/soc/2008/rf/about.html) is implementation of
the Adaptive Gauss-Hermite Quadrature approximation to the deviance.
That will be implemented in the development version (i.e. the R-forge
version) of the lme4 package.  AGQ will only be offered for models
with a single grouping factor for the random effects.

I realize that it is somewhat irritating and confusing to readers of
this list to have descriptions of the R-forge version of the package
contrasted with the CRAN version.  It is natural to expect that the
R-forge version should be the version on CRAN.  The reason that I have
not yet released the R-forge to CRAN is because of problems with the
mcmcsamp function in the R-forge version.  If I moved the R-forge
version to CRAN then code from Harald Baayen's book and probably code
from Gelman and Hill's book would no longer work.  Software versions
can be changed much more readily than can editions of a book.  I think
there is a way around the problem with mcmcsamp but I won't be able to
say for sure until it is coded and tested, which will take time.  I
don't want to predict exactly how much time - I always manage to
underestimate drastically.

I do not plan to provide an implementation of penalized
quasi-likelihood (PQL) in what is currently the development version
and what will become lme4_1.0.  PQL for GLMMs and the
"Lindstrom-Bates" algorithm for nonlinear mixed models (the approaches
are related) are examples of alternating conditional optimization
(think of the Gibbs sampler approach with optimization instead of
sampling).  It is a dangerous practice in that it can result in
oscillation between conditional optima.  I now prefer direct
optimization techniques where the fixed-effects parameters and the
variance components are optimized simultaneously.  There may be
circumstances where PQL is advantageous but usually those are because
of over-parameterized models.
Yes.  I think the adaptive part is important (in fact, I know it is
important) and probably more important than the distinction between
the Laplace approximation and Gauss-Hermite quadrature.  That is, you
gain more from using the Laplace approximation at the conditional
modes of the random effects (which is the "adaptive" part) than
increasing the number of Gauss-Hermite quadrature points at the
marginal modes.  The tricky bit of AGQ or Laplace is determining the
conditional modes and that part is already done.
3 days later
#
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

| lme4-specific questions:
|
| 6. Behavior of glmer: Does glmer really use AGQ, or just Laplace?
| Both?  pp. 28-32 of the "Implementation" vignette in lme4 suggest that
| a Laplace approximation is used, but I can't figure out whether this
| is an additional approximation on top of the AGQ/Laplace approximation
| of the integral over the random effects used in "ordinary" LMM.  When
| I fit a GLMM with the different methods, the fitted objects are not
| identical but all the coefficients seem to be.  (I have poked at the
| code a bit but been unable to answer this question for myself
| ... sorry ...)
|
|> To answer this question I must again, I regret, distinguish between
|> the CRAN version of the lme4 package and the R-forge development
|> version of lme4.
|
|> In the R-forge version the only method for generalized linear mixed
|> models and for nonlinear mixed models is direct optimization of the
|> Laplace approximation to the deviance.  One of the Summer of Code
|> projects that Google has funded for the R Foundation (see
|> http://code.google.com/soc/2008/rf/about.html) is implementation of
|> the Adaptive Gauss-Hermite Quadrature approximation to the deviance.
|> That will be implemented in the development version (i.e. the R-forge
|> version) of the lme4 package.  [snip]

~   Great!

~  A minor feature request: does it make sense to update the
documentation and code of glmer to make it clear that it does
*not* do AGQ at the moment?  I guess that depends how soon you
would expect the GSoC code to come online ...

|> I realize that it is somewhat irritating and confusing to readers of
|> this list to have descriptions of the R-forge version of the package
|> contrasted with the CRAN version.  It is natural to expect that the
|> R-forge version should be the version on CRAN.  The reason that I have
|> not yet released the R-forge to CRAN is because of problems with the
|> mcmcsamp function in the R-forge version.  If I moved the R-forge
|> version to CRAN then code from Harald Baayen's book and probably code
|> from Gelman and Hill's book would no longer work.  Software versions
|> can be changed much more readily than can editions of a book.  I think
|> there is a way around the problem with mcmcsamp but I won't be able to
|> say for sure until it is coded and tested, which will take time.  I
|> don't want to predict exactly how much time - I always manage to
|> underestimate drastically.

~   If (just hypothetically speaking) I were writing a review that
would be published in 6 months or so, do you have a recommendation?
(Would you still trust GLMM/mcmcsamp results from the CRAN version?)

| (The glmmML package claims to fit via Laplace or Gauss-Hermite
| quadrature (with non-adaptive, but adjustable, number of quad points
| -- so it's at least theoretically possible?)
|
|> Yes.  I think the adaptive part is important (in fact, I know it is
|> important) and probably more important than the distinction between
|> the Laplace approximation and Gauss-Hermite quadrature.  That is, you
|> gain more from using the Laplace approximation at the conditional
|> modes of the random effects (which is the "adaptive" part) than
|> increasing the number of Gauss-Hermite quadrature points at the
|> marginal modes.  The tricky bit of AGQ or Laplace is determining the
|> conditional modes and that part is already done.

~  OK, I was confused about the distinction in the meaning of
"adaptive" (as you pointed out previously on r-help ...)  I think
I have it now.

https://stat.ethz.ch/pipermail/r-help/2007-March/128043.html

~  thanks for your help!

~  Ben Bolker
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFIKGYLc5UpGjwzenMRAoBnAJ0R7w6FHL8uTaz3OKmMvJWHByS0tACdEy4x
ABU30kHrB/eNoXcHMvJJ4LE=
=whoY
-----END PGP SIGNATURE-----