[ADMB Users] ADMB versus R

[Seth W. Bigelow]

Dear Dave & admb users:

When I recently began using the glmmADMB package I picked up from R forge, I
noticed right away that 1) the thing was awesome and let me cleanly do an
analysis that other R packages didn't seem to be able to and 2) that there
was some missing functionality, e.g., some of the usual R 'methods' didn't
work. After I asked the R mixed models listserv about the lack of change in
likelihood when I changed the variance structure, I got suspicious that
there was something wrong with the analysis (because of the huge, many
orders of magnitude differences in the error variances) but I didn't pursue
it further on the R listserv because the basic analysis seemed fine. I just
assumed that the translation of more advanced variance structures from R to
ADMB had not been done yet. 

I appreciated Dave's analysis in ADMB, using the more complicated variance
structure; it confirmed my intuition that the complicated variance structure
was not necessary, and gave me a little more insight into ADMB. I'm not sure
I'm going to skip the R interface and learn to program ADMB directly, but
for the moment I'm continuing to use glmmADMB--cautiously--and it has been a
godsend.

So, many thanks to Dave and other ADMB (& glmmADMB) developers!

Seth W. Bigelow 


-----Original Message-----
From: dave fournier [mailto:davef at otter-rsch.com] 
Sent: Thursday, November 08, 2012 12:00 PM
To: users at admb-project.org
Cc: Seth W. Bigelow
Subject: Re: [ADMB Users] ADMB versus R

I think you a re missing the point.

This all began because a guy asked a question about a seeming 
contradiction in
his model fits on the R list. He got the usual runaround about his data 
not being good
enough or whatever.

Now after years of watching this stuff my general opinion is that people 
who trust
the R gurus for information more or less deserve what they get.

However in this case glmmadmb is something that I am partly responsible 
for so
I react when I think someone is getting bad advice.  I took a fair 
amount of time
and R pain to recreate his analysis and discovered that glmmadmb is not 
doing what one
might assume in this case.

I quickly hacked together what I assume is the model the guy wanted and 
it did
produce a better fit as one would expect. However the improvement was 
not significant.

Then Ben modified the R code to produce what he assumed was the correct 
model
and posted the results, noting that they were different from mine with the
implication that I must be doing something wrong.

However he never checked the log-likelihood for his new model. It is a 
lot worse
than the original user's LL so the model is not what he thinks it is.

I ran my model on the glmmadmb.dat file that Ben's model produced and 
got the
same LL as he got.  I agree with him that the design matrix Z appears to be
correct, but there must be something wrong with the dat file.  I wonder
if it is the II's.

Anyway this is neither convenient or quick and I have more interesting 
things to work on.