<html>
<head>
<meta http-equiv="content-type" content="text/html; charset=ISO-8859-1">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<blockquote type="cite">
<pre>On 13-03-15 11:04 AM, dave fournier wrote:
><i>
</i>><i> A while back it became clear that the Laplace approximation is not
</i>><i> good enough for some simple RE models. With no crossed effects one
</i>><i> can use Gauss-Hermite quadrature (This can be done for more than one
</i>><i> RE per group in ADMB although I have seen reports on the notoriously
</i>><i> inaccurate R list claiming that ADMB can only do GH for one RE per
</i>><i> block).
</i>
Dave: those reports were from me, following the published documentation.
When I go to
<a href="http://admb-project.org/documentation/manuals/admb-user-manuals">http://admb-project.org/documentation/manuals/admb-user-manuals</a>
and open
<a href="http://admb-project.googlecode.com/files/admbre-10.0-rev1.pdf">http://admb-project.googlecode.com/files/admbre-10.0-rev1.pdf</a>
and search for "Gauss-Hermite quadrature", I find this:
><i> In the situation where the model is separable of type “Block diagonal Hessian,” with only a
</i>><i> single random effect in each block (see Section 4), Gauss-Hermite quadrature is available as
</i>><i> an option to the Laplace approximation and to the -is (importance sampling) option. It is
</i>><i> invoked with command line option -gh N, where N is the number of quadrature points.
</i>
I also checked in the LaTeX code of the latest version from the SVN
repository -- this text is current.
Perhaps the documentation can be corrected? If you tell me what it
should say, I can submit a documentation patch ...
Ben
><i>
</i>><i> However for models with crossed effects in general GH would be
</i>><i> impractical. The alternative is importance sampling. We should
</i>><i> encourage users to routinely do importance sampling to test the
</i>><i> stability of the estimates. We could make this simple with a command
</i>><i> like
</i>><i>
</i>><i> -is 25 121 50
</i>><i>
</i>><i> this would do 50 fits with 25 random points and starting with the
</i>><i> random number seed 121. The seed would get changed for each fit. A
</i>><i> report would get generated showing the variability of the LL and
</i>><i> parameter estimates.
</i>><i>
</i>><i> One problem is that since lme? in R can not do this the idea will
</i>><i> probably get surpressed on the R list. Why do I care? Well users of
</i>><i> glmmadmb should be encouraged to take advantage of importance
</i>><i> sampling. I worry that the kind of sloppy practice encouraged in the
</i>><i> R community extends to glmmadmb which uses my software.
</i>
You'll also notice that I *suggested* using ADMB with importance
sampling on the R mailing list in the first place (in the same message
that you're referring to obliquely above) ...
</pre>
</blockquote>
<br>
I extended the software to have multiple random effects per group<br>
some time back. People seem to be aware of this as I have been sent<br>
code to examine with exactly this setup.<br>
<br>
As for the importance sampling, we all became aware that for a
simple RE<br>
model the Laplace approximation was not good enough to get reliable<br>
estimates from either lme? or ADMB. At that time I discovered<br>
that the GH quadrature in lme? was broken and you decided to fix it<br>
quietly without having the users worry their little heads about it.<br>
Well fine. However GH can not be used for crossed effects due to<br>
computational limitations. One has every reason to expect therefore<br>
that neither ADMB or lme? can fit these models reliably based on the<br>
LA. However ADMB can do importance sampling for these models,<br>
and IS proved to be adequate for the simple model referred to above.
<br>
<br>
Now in many cases I have found by using IS that for negative
binomial<br>
models the LA appears to be good enough. The kind of problem<br>
for which the LA did not perform well was a binary response model.<br>
<br>
So I conclude that lme? should not be relied on for binary response
models<br>
with crossed effects. However I doubt that anyone will deal with
this issue on<br>
the R lists.<br>
<br>
In any event it should be simple with ADMB to use IS to investigate
the<br>
adequacy of the LA for any RE model. This should be a routine part
of<br>
any serious analysis.<br>
<br>
<br>
<br>
<br>
<br>
<br>
<blockquote type="cite">
<pre>
Ben</pre>
</blockquote>
</body>
</html>