[R] ADMB error- function maximizer failed (couldnt find STD file)

[Ben Bolker]

gNathan Svoboda <nsvoboda <at> CFR.MsState.Edu> writes:

> Greetings glmmADMB function users, I am trying to run a series of
> models using the glmmADMB function with several different
> distribution families (e.g., poisson, negbinom). I am using a
> Optiplex 790 PC with Windows 7, 16.0 GB of RAM and a 64-bit
> operating system. I am running R version 2.15.0 and started out
> using the most recent version of glmmADMB (I believe version
> 7.2.15).  My data is zero inflated count data that is
> overdispersed. When I ran the following code for either a zero
> inflated poisson (family="poisson") or a neg binomial (i.e.,
> family="nbinom"):
 
> fit_zipoiss1 <- glmmadmb(LOCS~D_ROADS + (1|YEAR) , data=FAWNS, 
> zeroInflation=TRUE,
> family="poisson", mcmc=TRUE)
> 
> I get the error: 
> Error in glmmadmb(LOCS ~ D_ROADS + (1 | YEAR), data = FAWNS, 
> zeroInflation = TRUE,  :
> The function maximizer failed (couldn't find STD file)
> 
> I tried setting the number of mcmc iterations by running the following code: 
> fit_zipoiss <- glmmadmb(LOCS~D_ROADS + (1|YEAR) , data=FAWNS, 
> zeroInflation=TRUE, 
> family="poisson", mcmc=TRUE, mcmc.opts=mcmcControl(mcmc=50000))
> but again received the same "Function maximizer...." error. 
 
# As suggested by previous users (specifically Ben Bolker) I
# downloaded an earlier version of glmmADMB ("glmmADMB_0.7.tar.gz")
# and re-ran the same models (e.g., poisson and neg. binomial) and
# received the same error. As suggested by Mr. Bolker and others I
# also tried running both a poisson and neg. binomial model with
# "admb.opts=admbControl(shess=FALSE,noinit=FALSE) " and received the
# same error.  I am not sure if an even older version of glmmADMB
# (i.e., glmmADMB v. 0.6.5) would work as it seems to have worked for
# others but I am unsure of where I can find this version?

  A couple of points:

* http://www.math.mcmaster.ca/~bolker/R/src/contrib/ has earlier versions of
the glmmADMB package, although not version 0.6.5 (nothing between 0.5-2 and
0.7) ...

* how many years do you have?  Especially if you have a giant data set, you
might try exploring the effects of treating YEAR as fixed.  This will have
a couple of advantages -- may get you around the problem, and
will also allow you to fit your model using the pscl package.  Treating YEAR
as fixed will work well if you (1) have lots of data per year and/or (2) have
a small number of years.

* have you done exploratory plots of LOCS vs D_ROADS by year?  Are your
data very unbalanced, either in terms of observations or in terms of numbers
of zero counts? (i.e. do you have a very small number of observations, or
a very small number of non-zero observations, for some combinations of
D_ROADS and year?  in principle random-effects models are supposed to handle
these issues, but they can't always ...)

* it's probably not worth bothering with mcmc at all while you are
still in the debugging stage -- it is an add-on, not an integral part of
the model fitting process.  (Just leave out the mcmc argument entirely:
the default value is FALSE)

* I think the current convention is that glmmADMB questions belong on the
r-sig-mixed-models at r-project.org list (not here in the general R help, and
not on the ADMB mailing list).

  stuff
    to 
      make
         gmane 
           happy ... it doesn't like too much quoted