R help - Nov 2007 - lmer and method call

Hello all,

I'm attempting to fit a generalized linear mixed-effects model using lmer
(R v 2.6.0, lmer 0.99875-9, Mac OS X 10.4.10) using the call:

vidusLMER1 <- lmer(jail ~ visit + gender + house + cokefreq + cracfreq +
herofreq + borcur + comc + (1 | code), data = vidusGD, family = binomial,
correlation = corCompSymm(form = 1 | ID), method = "ML")

Although the model fits, the summary indicates the model is a "Generalized
linear mixed model fit using Laplace". I've tried any number of
permutations; is only Laplace supported in lmer, despite the text of the
help file?

Thanks for any guidance you might give.


M-J


--

Research assistant
Vancouver Injection Drug User Survey
BC Centre for Excellence in HIV/AIDS
St. Paul's Hospital, Vancouver

MSc candidate
Department of Healthcare and Epidemiology
Faculty of Medicine
University of British Columbia

On Nov 29, 2007 8:09 PM, M-J Milloy <mjmilloy at cfenet.ubc.ca>
wrote:
>
> Hello all,
>
> I'm attempting to fit a generalized linear mixed-effects model using
lmer
> (R v 2.6.0, lmer 0.99875-9, Mac OS X 10.4.10) using the call:
>
> vidusLMER1 <- lmer(jail ~ visit + gender + house + cokefreq + cracfreq +
> herofreq + borcur + comc + (1 | code), data = vidusGD, family = binomial,
> correlation = corCompSymm(form = 1 | ID), method = "ML")
>
> Although the model fits, the summary indicates the model is a
"Generalized
> linear mixed model fit using Laplace". I've tried any number of
> permutations; is only Laplace supported in lmer, despite the text of the
> help file?
The help file does say that for a generalized linear mixed model
(GLMM), which is what family = binomial implies, the estimation
criterion is always "ML" (maximum likelihood) as opposed to
"REML"
(restricted, or residual, maximum likelihood).  So stating method "ML"
is redundant.

For a GLMM, however, the log-likelihood cannot not be evaluated
directly and must be approximated.  Here the help file is misleading
because it implies that there are three possible approximations, "PQL"
(penalized quasi-likelihood), "Laplace" and "AGQ" (adaptive
Gaussian
quadrature).  AGQ has not yet been implemented so the only effective
choices are PQL and Laplace.  The default is PQL, to refine the
starting estimates, followed by optimization of the Laplace
approximation.  In some cases it is an advantage to suppress the PQL
iterations which can be done with one of the settings for the control
argument.

Douglas Bates <bates <at> stat.wisc.edu> writes:

(lmer)
> The default is PQL, to refine the
> starting estimates, followed by optimization of the Laplace
> approximation.  In some cases it is an advantage to suppress the PQL
> iterations which can be done with one of the settings for the control
> argument.
I had found out the hard way that it is often better to let PQL 
play the game rather loosely.  Yet I never dared to tell someone, for fear
the approximation could end up in the wrong slot,

Any rules (beside trying variants) if I can trust such a result?

Dieter