[R-sig-ME] previous posts about error message in glmmADMB

[Ben Bolker]

On 13-05-20 06:19 AM, Julia Sommerfeld wrote:
> Dear list,
> 
> I'm trying to fit a glmmadmb, but I'm always getting the following error
> message:
> 
> *mod <- glmmadmb(anzahl_round~wea1*seg + offset(log(flaeche)) +
> (1|day_ID) + (1|seg), zeroInflation=TRUE, family="poisson", data=dat1)  *
> 
> /Parameters were estimated, but not standard errors were not: the most
> likely problem is that the curvature at MLE was zero or negative
> Error in glmmadmb(anzahl_round ~ wea1 * seg + offset(log(flaeche)) + (1
> |  :
>   The function maximizer failed (couldn't find STD file) Troubleshooting
> steps include (1) run with 'save.dir' set and inspect output files; (2)
> change run parameters: see '?admbControl'
> In addition: Warning message:
> running command 'C:\Windows\system32\cmd.exe /c
> "C:/Users/Hawksbill/Documents/R/win-library/3.0/glmmADMB/bin/windows32/glmmadmb.exe"
> -maxfn 500 -maxph 5 -noinit -shess' had status 1 /
> 

  Update:

I can replicate this problem, and we are working on a tweak that will
allow glmmADMB to fit this model.  However, in the meantime I have some
advice ...

 This case is similar to the previous one posted here.  The essential
problem is that the data are highly overdispersed (with or without
zero-inflation), so that estimates of the predicted Poisson mean
underflow to zero, which throws off the estimation.  In the absence of
being able to fit the ZIP, you can confirm pretty well that the ZINB is
a better model by looking at the summary of the ZINB and seeing that the
estimated alpha (overdispersion parameter) is small, with a small
standard error -- therefore NB will be better than Poisson.

  A couple of other points to note, though:

* you should not be trying to fit seg as both a fixed and a random
factor. By putting wea1*seg in the formula, you are implicitly including
the main effects of wea1 and seg and their interaction (the interaction
alone would be written wea1:seg.  You might want ~wea1+wea1:seg ?

* the results of fitting ZINB and NB are a little weird, possibly for
this reason (the NB fits better [has a lower negative log-likelihood]
than the ZINB, which should be impossible)

* you might also want to check the mean-variance scaling/consider
family="nbinom1"

* in all the models I fitted (ZIP, P, ZINB, NB) the 1|seg variance comes
out as zero, again probably because of the redundancy described above.
Even when you don't have redundancy in your model this sort of thing
happens sometimes, indicating that your random-effects model is slightly
overfitted.