[ADMB Users] New error messages over night

[Ben Bolker]

On 12-04-28 01:19 PM, dave fournier wrote:
> Hi everybody,
> 
> As some have posted earlier, I experience error messages on a model
> which worked yesterday (couldn't find STD file). Downgrading to v 0.7
> didn't help, and including the admbControl modification suggested by
> Ben Bolker in earlier posts didn't work, either. Changing the model
> structure, I get another error message which I didn't get before,
> either:
> 
> Error in II[, ii] = II[, ii] + REmat$codes[[i]] : number of items to
> replace is not a multiple of replacement length
> 
> I wonder why I get these errors now and not yesterday?

  I suspect (although obviously can't guarantee) that you changed
something without realizing.  (I have done this myself many times, and
most times until I actually find the problem I am (wrongly) confident
that I couldn't possibly have changed anything.)  What did you change
between yesterday and today?  Was it *only* your glmmADMB version?

  While I can't diagnose the previous issue where the ADMB model fails,
searching for the error message you are getting now ("Remat glmmadmb not
a multiple" on Google) finds

https://stat.ethz.ch/pipermail/r-sig-mixed-models/2012q1/017533.html

  in which I said this:

* the basic problem is that you have NA values in your data: these get
removed automatically in one place in the code and not in the other,
hence the length mismatch.  This is not absolutely trivial to fix
automatically -- the fixed and random effect predictors are handled
separately, and there may be extra variables in the data that should be
disregarded completely -- but in the meantime I have at least put in an
informative warning message in this case (for the next release).  The
simplest solution is to use na.omit() to get rid of these values (which
can't be used in the fit anyway).

  By downgrading, I think you have gone back to a version that *didn't*
have the informative error message.

  So:

  (1) take out NA values in your data
  (2) post your data, or a reproducible example, in some form we can
work with.  That could be the data and the code (please make your code
short and self-contained, and see <http://tinyurl.com/reproducible-000>
for help on posting reproducible data), or you can run glmmADMB with
'save.dir' set to a temporary directory; then you can copy the
glmmadmb.par and glmmadmb.dat files from that directory, for diagnosis
by ADMB gurus (who probably don't want to mess with your R code).



> 
> I installed the glmmADMB package from CRAN, and downgraded later with
> the file from Ben's webpage. When I tried to upgrade again, I saw
> that the build on CRAN disappeared. Does this mean somebody is
> working on it at the moment?

  Well, you couldn't have installed from CRAN, because it doesn't live
on CRAN (and never did -- packages that rely on binary 'blobs' , as
glmmADMB does, are not allowed on CRAN) -- it lives on R-forge.  Based
on this <http://r-forge.r-project.org/R/?group_id=847>, it *appears*
that the R-forge build is currently working.
> 
> I'm really thankful that you guys are working on this, since in my
> very limited understanding of stats (plus lack of time)
> 
> 
> this is the only useful method I found so far to somehow get results
> for my PhD thesis (ecology, blocked experiment, counts and
> proportions, comparison of predictors). However, with some things not
> functioning, I don't dare relying on this package too much, which
> leaves me without a tool to analyze my data. I'm a bit desperate now
> and wondering if you think if you will get a well functioning version
> out soon, i.e. if it's worth waiting a while, or if I should try to
> find other solutions (which I have no clue what else to try)?

  We need a reproducible example before we can help you ... see for
example <http://tinyurl.com/reproducible-000> ...